Professional Documents
Culture Documents
Supply Chain Management in Manufacturing and Service Systems
Supply Chain Management in Manufacturing and Service Systems
Sharan Srinivas
Suchithra Rajendran
Hans Ziegler Editors
Supply Chain
Management in
Manufacturing and
Service Systems
Advanced Analytics for Smarter
Decisions
International Series in Operations Research
& Management Science
Founding Editor
Frederick S. Hillier
Stanford University, Stanford, CA, USA
Volume 304
Series Editor
Camille C. Price
Department of Computer Science, Stephen F. Austin State University,
Nacogdoches, TX, USA
Associate Editor
Joe Zhu
Foisie Business School, Worcester Polytechnic Institute, Worcester, MA, USA
More information about this series at http://www.springer.com/series/6161
Sharan Srinivas • Suchithra Rajendran •
Hans Ziegler
Editors
Hans Ziegler
School of Business, Economics and
Information Systems
University of Passau
Passau, Germany
This Springer imprint is published by the registered company Springer Nature Switzerland AG.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
To the memory of my father, B. Srinivasan,
who always believed in me and supported all
my choices in life.
—Sharan Srinivas
To my husband, parents and my late
grandmother (Dr. A.B Vasanthalakshmi)
—Suchithra Rajendran
To Sarah Madeleine and Ulrike
—Hans Ziegler
Preface
Management of supply chains has evolved rapidly over the last few years owing
to the inception of Industry 4.0, where digitization using Internet of things (IoT),
advanced robotics, and sensors have taken precedence. The traditional supply chain
systems are no longer efficient as these transformations have led to dynamic and
interconnected systems that require robust management capabilities. Moreover,
today’s customers are being exposed to numerous products and services through
the Internet, leading to high expectations requiring superior supply chain systems
with cost-effective, responsive, and sustainable capabilities. Nevertheless, digital
transformation, both in manufacturing and service industries, has led to large
volumes of real-time data which can be leveraged to obtain actionable insights and
drive profitable supply chain decisions. This book focuses on providing an overview
of current trends in supply chain as well as publishing state-of-the-art original
research work dealing with advanced analytical models (predictive and prescriptive
analytic models) for the design, planning, and operation of supply chains in the era
of digitization and Industry 4.0. It intends to empower supply chains with smarter
decisions at all levels and stages.
The key characteristics of this book are as follows:
1. Covers recent trends, developments, and applications in supply chain manage-
ment
2. Includes a wide collection of analytical methodologies for optimizing key supply
chain decisions
3. Bridges the theory–practice gap in supply chain management
4. Designed, organized, and edited considering non-experts
5. Holistic with contributions from leading academicians and industry practitioners
6. Unified single-source guide
7. Versatile reference book for students, researchers, educators, and practitioners,
alike.
vii
viii Preface
The vital areas of supply chain management empowered with smart decisions
include:
1. Supply chain network design
2. Logistics and distribution
3. Process optimization
4. Product life cycle management
5. Visibility
6. Risk management
7. Sustainability
8. Customer feedback
Book Overview
tactical decisions are made every month or quarter and are related to supply chain
planning activities like purchasing, production and planning, transportation, inven-
tory management, and distribution. The chapters entitled “An Integrated Problem
of Production Scheduling and Transportation in a Two-Stage Supply Chain with
Carbon Emission Consideration” and “A Simulation-Based Evaluation of Drone
Integrated Delivery Strategies for Improving Pharmaceutical Service” prescribe the
best course of action for tactical decision-making. Finally, operational decisions are
related to determining the day-to-day tasks needed to satisfy individual customer
orders. They involve activities like allocating inventory, assigning logistics, updating
delivery dates, and placing replenishment orders in response to customer orders.
The chapters entitled “Pro-active Strategies in Online Routing” and “Prescriptive
Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces” provide
insights to make effective real-time operational decisions.
While this book is organized based on the decision level each chapter addresses,
the chapters can also be understood by the methods they employ. The chapters
entitled “Product Life Cycle Optimization Model for Closed Loop Supply Chain
Network Design,” “An Integrated Problem of Production Scheduling and Trans-
portation in a two-Stage Supply Chain with Carbon Emission Consideration,”
Pro-active Strategies in Online Routing,” and “Prescriptive Analytics for Dynamic
Real Time Scheduling of Diffusion Furnaces” leverage mixed-integer linear pro-
gramming (MILP) techniques to obtain the optimal values for the decision variables
in a supply chain, given practical constraints. Further, the chapters entitled “Pro-
active Strategies in Online Routing” and “Prescriptive Analytics for Dynamic
Real Time Scheduling of Diffusion Furnaces” integrate heuristics into their MILP
formulations to quality solutions more quickly. On the other hand, the chapter
entitled “A Simulation-Based Evaluation of Drone Integrated Delivery Strategies
for Improving Pharmaceutical Service” evaluates the effectiveness of integrating
new strategies to improve last-mile delivery in supply chain using discrete event
simulation modeling. The chapter entitled “Supply Chain Risk Management in
Indian Manufacturing Industries: An Empirical Study and a Fuzzy Approach”
employs fuzzy cognitive map (FCM) to foresee and manage SC risks. Finally,
the chapter entitled “Improving Service Supply Chain of Internet Services by
Analyzing Online Customer Reviews” uses text analytics techniques like bigram
and trigram analysis on online customer reviews to gain insights to make better
strategic decisions.
Chapter Summaries
drivers, and varied goals associated with supply chains. Moreover, this chapter also
familiarizes the readers with analytics, setting the stage for the book.
The chapter entitled “Intelligent Digital Supply Chains” reports on how modern
supply chains are leveraging Industry 4.0 technologies to gain end-to-end visibility
and intelligence in their supply chains. Further, it discusses planning, execution,
evaluation, monitoring, risk management, and opportunities from the perspective
of digitized supply chains. This chapter is vital for both academic professionals
who want to understand the tools employed by industry practitioners and industry
personnel who need to catch up with the latest trends.
The chapter entitled “Product Life Cycle Optimization Model for Closed Loop
Supply Chain Network Design” proposes a product life cycle optimization model
for a closed-loop supply chain network design. In environmentally concerning
times, a closed-loop supply chain capable of reclaiming and reusing post-consumer
materials to reduce wastes and dependence on raw materials is crucial. Further,
remanufacturing is efficient and profitable. This chapter proposes an integrated
multi-period optimization model to design the closed-loop supply chain for OEMs.
The authors explicitly model the optimal collection of remanufactured products
through suppliers in a dynamic manner across multiple time periods over the product
life, accounting for demand, quality, and remanufacturability of returned products.
The model is applied to a realistic case study of Apple’s iPhone 7 for a product life
cycle of 8 years.
The chapter entitled “Supply Chain Risk Management in Indian Manufacturing
Industries: An Empirical Study and a Fuzzy Approach” seeks to predict supply chain
risks via early warning signals and implement appropriate mitigation strategies.
With many organizations solely focusing on efficiency to win today’s fierce
competition, supply chains have become more vulnerable to risks. The authors strive
to address this issue using a twofold method. In the first part, the critical factors
of supply chain risk management are understood by developing a framework, and
the views of the practicing managers about risks perceived in their organization
are captured using an empirical study. Second, this information is used to develop a
fuzzy cognitive map to identify all plausible risks in the future, given a risk observed
from a point in time, and suggest proactive mitigation strategies for practicing
managers.
The chapter entitled “Improving Service Supply Chain of Internet Services by
Analyzing Online Customer Reviews” proposes a methodology to leverage online
customer reviews using text analytics to improve service quality and customer
satisfaction. It specifically focuses on Internet service providers (ISPs), as the
service sector is gaining more attention, and the wireless communications industry
has become highly competitive. The chapter seeks to extract the current strengths,
weaknesses, opportunities, and threats (SWOT), along with their corresponding
root causes for several leading ISPs by exploring consumer reviews using text
analytics. The proposed approach consists of four different stages, bigram and
trigram analyses, topic identification, SWOT analysis, and root cause analysis
(RCA), and provides compelling managerial insights. This chapter is critical, as
all businesses can benefit from learning to leverage their online customer reviews.
Preface xi
First and foremost, we would like to thank the authors as this book would not have
been possible without their timely and novel contributions. We also acknowledge
their efforts in preparing concise chapters for addressing recent and practically
relevant supply chain issues using advanced analytical methodologies. Besides, we
would also like to thank the authors for promptly revising the chapters based on the
reviewer’s comments.
We want to express our sincere appreciation to the editorial assistant, Surya
Ramachandiran, for his outstanding assistance in several preparatory aspects of this
edited book, especially for typing the first chapter, reviewing the language style,
and suggesting insightful edits to improve the overall quality of the chapters. We
want to thank Christian Rauscher, senior editor of Business, Operations Research
& Information Systems at Springer, for his guidance from the conception to
the completion of this edited book. We also acknowledge the timely help and
information provided by Sayani Dey, production editor at Springer, during different
book preparation stages. Finally, we would like to thank our families for their
encouragement and support in our research endeavors.
xiii
Contents
xv
Contributors
xvii
xviii Contributors
While managing a supply chain, numerous decisions need to be taken regarding the
flow of materials, information, and funds. The various decisions taken in a supply
chain fall under one of the three levels, namely, strategic, tactical, and operational,
based on how frequently that decision is taken and duration over which its impact is
experienced.
4 S. Srinivas et al.
2.1 Strategic
Strategic decisions primarily determine how a supply chain is designed and who will
be the partners for the upcoming years. Unlike other decisions, strategic decisions
have a significant long-term impact, and making sudden changes is generally both
expensive and not feasible. However, both the customer demands and market are
ever-changing, making it crucial for the companies to carefully account for the
uncertainties tied with the future before making any strategic decisions. These
decisions majorly decide the structure of the supply chain, method for material
procurement, strategy for allocating resources, and processes undertaken at each
stage. Broadly all the strategic decisions taken in a supply chain fall under one
of the three categories—network design, production and sourcing, and information
technology.
First, in network design, decisions regarding the number and type of facilities
needed, their geographic locations, and their production and storage capacities are
made. Other strategic decisions regarding the mode of transport between these
facilities also fall into this category. Second, in production and sourcing, decisions
on making or buying (whether to outsource or conduct the activity in-house) at each
stage of a supply chain are taken. Moreover, decisions on the selection of vendors,
sub-contractors, and other alliances also belong to this category. Finally, strategic
decisions can also be related to managing information technology infrastructure.
Decisions like what type of information systems are needed, and whether to develop
them internally, buy the related commercially available version or employ the freely
available open-source alternatives are taken here.
An American multinational food and beverage corporation, PepsiCo, is an apt
example of how effective strategic decisions can better satisfy customers. While
PepsiCo’s best-known products were carbonated drinks packaged in metal cans or
plastic bottles, the change in customer’s preferences towards nutritious food and
environmentally friendly products created a challenge. To satisfy this new set of
customer preferences, PepsiCo took the strategic decision to empower alternatives
like Naked Juice and O.N.E Coconut Water (Forbes 2016). The company sourced
ingredients from across the world, including certified non-GMO, fresh, and organic
alternatives. Further, to enable the production of these products, global supply
chains were set-up with refrigeration capabilities throughout the chain. Finally, to
satisfy environmentally conscious customers, Naked Juice products were packaged
in recyclable bottles. Furthermore, these bottles were explicitly designed in cuboidal
shapes to improve packing efficiency during transportation, and the modes of
transportation employed were also changed to reduce the overall carbon footprint.
Chapters “Product Life Cycle Optimization Model for Closed Loop Supply
Chain Network Design”, “Supply Chain Risk Management in Indian Manufacturing
Industries: An Empirical Study and a Fuzzy Approach”, “Improving Service Supply
Chain of Internet Services by Analyzing Online Customer Reviews”, and “An
Integrated Problem of Production Scheduling and Transportation in a Two-Stage
An Overview of Decisions, Performance and Analytics in Supply Chain Management 5
Supply Chain with Carbon Emission Consideration” address some of the key
strategic issues in a supply chain.
2.2 Tactical
2.3 Operational
Enablers can be understood as things that are needed to achieve a goal. Marien
(2000) reports four enablers that are needed for the effective functioning of a supply
chain, namely
• Organizational infrastructure: It is crucial for an effective supply chain as it
determines how the different stages coordinate together to accomplish its goals.
The key concern here is organizing supply chain activities, both within the firm
and across firms, for vertical orientation or more decentralization.
• Technology: Two types of technologies, manufacturing technology, and informa-
tion technology, are necessary to enable superior supply chains.
• Strategic alliances: Establishing long-term partners is a key enabler for a supply
chain. Alliances specifically play a vital role in a decentralized supply chain
where great power and responsibility are present with the suppliers.
• Human resources: Includes technical and managerial employees with a holistic
understanding of supply chain management concepts and tools. These employees
are needed to design and operate an effective supply chain.
While enablers support the smooth functioning of a supply chain, the drivers
are areas of crucial decision making. The four major drivers of a supply chain, as
described by Ravindran and Warsing (2016, pp. 7–9) and Chopra and Meindl (2013,
pp. 41–42), include inventory, transportation, facilities (plants and distribution
centers) and suppliers. These drivers do not function independently but interact
with each other to establish the overall performance of a supply chain. For example,
setting up facilities in remote places away from major cities may reduce rental costs,
but will increase transportation costs and affect delivery time. Similarly, procuring
and storing supplies in large quantities can reduce cost in terms of raw material and
transportation and even improve customer satisfaction levels, but will drastically
increase inventory storage costs. Hence understanding how these drivers interact
and making efficient trade-offs between them is crucial to achieving superior supply
chain performance.
Each company has a competitive strategy that involves satisfying the needs of
the customer belonging to a particular segment. The overall supply chain and its
individual stages need to align with this strategy. For example, customers of a
supermarket prioritize availability and variety over the price of the products. These
customers are willing to pay higher prices, provided they can buy everything from
vegetables to pastries at the same place. Hence, these stores have robust supply
8 S. Srinivas et al.
chains that provide a vast range of 15,000 to 60,000 SKUs. While on the other
hand, customers of limited-assortment stores want lower prices and are ready to
compromise on variety. These stores tend to have cost-effective supply chains that
offer a limited range of items (fewer than 2000), with many stores not offering
any perishables. The requirements of the customers prioritized by the competitive
strategy must match with the capabilities of the supply chain, and this consistency
is referred to as strategic fit by Chopra and Meindl (2013, p. 21). In case of a lack of
strategic fit, the company must alter the supply chain to meet its competitive strategy
or modify its competitive strategy based on what its supply chain is designed to do
well. Depending on these supply chain capabilities, supply chains can be understood
as one or a mix of the following types:
• Responsive
• Efficient
• Resilient
• Humanitarian
• Green
• Sustainable
Figure 2 illustrates the key capabilities of the different supply chain types
with respect to seven different criteria, namely, profitability, cost reduction, speed,
flexibility, social responsibility, environmental concern, and ethical practice.
Fig. 2 Capabilities based on supply chain type with respect to different criteria
Fig. 3
Responsiveness-efficiency
trade-off frontier
An Overview of Decisions, Performance and Analytics in Supply Chain Management 11
Barve and Yadav (2014) describe the humanitarian supply chain as the “flow of
relief aid and the related information between the beneficiaries affected by disaster
and the donors so as to minimize human suffering and death”. In a humanitarian
supply chain, the customers include the affected people and the intermediate storage
facilities, while the supplies include relief aids like materials, logistics, and even
volunteers. Since humanitarian supply chains face a lot of unknowns, uncertainties
and need coordination among numerous stakeholders (like donors, volunteers,
government, NGOs, and military), along with relief aids, the effective flow of
information is vital. This type of supply chain is complex as they tend to have
limited infrastructure and other resources, making them dependent on donors and
volunteers.
Humanitarian supply chains need to be highly responsive to the disaster type and
its changing phases. While they may be newly established during a particular crisis,
humanitarian works during the COVID-19 pandemic showed the importance of
12 S. Srinivas et al.
With increasing customers showing concern for the environment, companies seek
to identify and incorporate environmentally friendly practices in their supply chain
to gain competitive advantage. Further, the government imposed environmental
regulations that have made it imperative for companies to build greener supply
chains, especially in countries like France, Spain, Morocco, and Kenya. Building
a green supply chain requires a unified effort from all the stakeholders and stages.
Manufacturers must work alongside both the suppliers and customers to enable
their environmental goals. These environmental goals usually include reducing solid
waste, effluent waste, air emissions, and usage of toxic materials.
While employing a green supply chain can give a competitive advantage, there
still are concerns about whether they will translate into substantial improvement
in profits or market share. However, employing environmentally harmful practices
have impacted businesses negatively. For example, Nestle, a global food processing
company, had to face a myriad of issues when Greenpeace International held the
production of Nestle’s confectionery product, KitKat, responsible for deforestation
(Purkayastha and Chaudhari 2012). The company was accused of destroying
precious rainforests to increase palm plantations and palm oil produce needed to
manufacture their confectionary. Further, by facilitating high impact social media
campaigns, Green peace could pressurize Nestle to stop sourcing palm oil from
Sinar Mars, an organization accused of illegally clearing rainforests, and source its
palm oil responsibly. Similarly, Unilever too had to stop purchasing palm oil from
controversial vendors.
With dwindling resources, today’s supply chains need to focus on social and envi-
ronmental facets along with common economic goals to achieve sustained growth.
These three dimensions of sustainability are together referred to as the “triple bottom
line (3BL)”. The goals of a sustainable supply chain can be understood as the
following:
An Overview of Decisions, Performance and Analytics in Supply Chain Management 13
The current industrial revolution, Industry 4.0, is about integrating physical and dig-
ital systems to enable effective decisions that require minimal human supervision. It
focuses on inter-connectivity using technologies such as the internet of things (IoT),
cloud computing, artificial intelligence (AI), and advanced robotics. While these
intelligent technologies have transformed numerous areas, they specifically have
a profound impact on supply chains. Industry 4.0 has led to major improvements
in different supply chain stages, including procurement, inventory management,
logistics, production, and retailing, by facilitating process integration, automation,
digitization, and analytical power. The ability to exchange data and make decisions
is particularly useful in supply chain as they have a dynamic network consisting of
multiple stakeholders and stages needing collaborative decisions at every level.
Despite numerous benefits, a majority of companies are yet to adopt Industry
4.0 technologies into their supply chain due to doubts on return on investments,
struggle to find qualified talent to implement and maintain these systems, shortage
of financial resources, concerns over data security, lack of information technology
infrastructure or even due to the sheer lack of knowledge about its benefits (Horváth
and Szabó 2019). However, there have indeed been numerous success stories. One of
the best examples is the widespread adoption of intelligent tools that can accurately
predict consumer behavior and guide demand planners. Demand planning is a
manually intensive week-long task repeated at the beginning of every month (SAS
2018). Demand planners spend 40% of their time cleaning and managing inventory
and sales data, an additional 30-40% of time reviewing forecasting models and fine-
14 S. Srinivas et al.
tuning them, and 10% of time reporting their findings (SAS 2018). This further
becomes cumbersome when a supply chain deals with numerous products across
different categories. Nevertheless, with developments in analytical technologies,
supply chains leverage AI tools to effectively handle most of these manual and
repetitive tasks, leaving time for demand planners to focus on other value-adding
work.
Apart from demand planning and then sourcing from reputed suppliers, ensuring
the quality of a product/service is a vital aspect of supply chain management
(Romano and Vinelli 2001). With improvements in capturing and analyzing sensory
data, companies continuously monitor their product/service to ensure consistent
quality. For example, manufacturing industries capitalize on image processing
and machine learning developments to automate visual inspections, which would
otherwise be time-consuming, labor-intensive, and prone to human error. Like few
modern call centers, other service providers also leverage voice recognition and
natural language processing tools to ensure their agents’ quality and drive customer
satisfaction. Further, with developments in IoT technologies, not only quality is
monitored but also controlled. For example, many plastic goods manufacturing
companies continuously monitor their manufacturing process at all stages. When
any temperature deviation or product abnormality is identified, the production
process enabled using IoT is automatically modified to produce ideal results. Thus,
Industry 4.0 also supports supply chains by enabling consistent quality, cutting-
down costs, improving resource utilization, and driving process efficiency.
Due to the developments in Industry 4.0 technologies, specifically IoT and
sensor technology, both manufacturing and service organizations generate massive
amounts of industrial data. To a large extent, the success of a supply chain lies in its
ability to effectively capitalize on this data by leveraging advanced analytical tools
to gain intelligence. Supply chain analytics deals with the use of quantitative models
for data-driven management of all the decision levels—from helping in supply
network design and vendor selection at the strategic level to managing procurement,
inventory, demand planning, and logistics at tactical and operational levels.
The traditional sources of data for supply chain analytics include radio frequency
identification (RFID) systems that automatically track items attached with RFID
tags, global positioning systems (GPS) that provide the locations of shipments in
transit, and barcode enabled systems that capture transactions. Further, supply chain
analytics also heavily depends on data visualization techniques to report its findings
in a human interpretable manner. Depending on the complexity and value addition,
supply chain analytics can be classified into three types—descriptive, predictive,
and prescriptive.
• Descriptive analytics uses aggregation, visualization, and mining of historical
data to provide insights on prior trends and patterns. It helps supply chain
practitioners to learn from the past and also uncover relationships between
variables, thus empowering them to make better decisions in the future. Apart
from past data, descriptive analytics also uses current data to provide valuable
real-time information and visibility needed to effectively manage the supply
An Overview of Decisions, Performance and Analytics in Supply Chain Management 15
References
Barve, A., & Yadav, D. K. (2014). Developing a framework to study the various issues of
humanitarian supply chains. In Proceedings of 10th Asian Business Research Conference–
Bangkok, Thailand.
Blackstone, J. H. (2010). APICS dictionary, APICS the Association for Operations Management
(13th ed. revised, p. 148).
Chopra, S., & Meindl, P. (2013). Supply chain management: Strategy, planning, and operation (5th
ed.). London: Pearson Education.
Deloitte. (2020). COVID-19 - Managing supply chain risk and disruption.
Fisher, M. L. (1997). What is the right supply chain for your product? Harvard Business Review,
75, 105–117.
Forbes. (2016). PepsiCo’s practical application of supply chain resilience strategies. Retrieved
from https://www.forbes.com/sites/stevebanker/2016/10/01/pepsicos-practical-application-of-
supply-chain-resilience-strategies/#1d81bfea6293
Fortune. (2015). Why target failed in Canada. Retrieved from https://fortune.com/2015/01/15/
target-canada-fail/
Horváth, D., & Szabó, R. Z. (2019). Driving forces and barriers of Industry 4.0: Do multinational
and small and medium-sized companies have equal opportunities? Technological Forecasting
and Social Change, 146, 119–132.
Lepenioti, K., Bousdekis, A., Apostolou, D., & Mentzas, G. (2020). Prescriptive analytics:
Literature review and research challenges. International Journal of Information Management,
50, 57–70.
Marien, E. J. (2000). The four supply chain enablers. Supply Chain Management Review, 4(1),
60–68.
An Overview of Decisions, Performance and Analytics in Supply Chain Management 17
Performance Magazine. (2020). Ben & Jerry’s – An example of how to integrate sustainability
in business. Retrieved from https://www.performancemagazine.org/ben-jerrys-sustainability-
business/
Purkayastha, D., & Chaudhari, A. (2012). Greenpeace, nestle and palm oil controversy: Social
media driving change? IBS Center for Management Research.
Ravindran, A. R., & Warsing, D., Jr. (2016). Supply chain engineering: Models and applications.
Boca Raton, FL: CRC Press.
Romano, P., & Vinelli, A. (2001). Quality management in a supply chain perspective. International
Journal of Operations & Production Management, 21(4), 446–460.
SAS Institute. (2018). Assisted demand planning using machine learning for CPG and retail.
TradeGecko. (2018). IKEA supply chain: How does IKEA manage its inventory? Retrieved from
https://www.tradegecko.com/blog/supply-chain-management/ikeas-inventory-management-
strategy-ikea
Vogue Business. (2020). Luxury’s war effort against coronavirus. Retrieved from https://
www.voguebusiness.com/companies/european-luxury-fashions-war-effort-against-coronavirus
Intelligent Digital Supply Chains
Raghav Jandhyala
1 Introduction
Supply chains are growing increasingly complex with global, multi-enterprise and
multi-channel operations to meet the ever-growing customer expectations and need
for individualized products. In this intelligent era, supply chains need to be digitized
in order to be responsive, sustainable, profitable and resilient to disruptions. The
flow of information across all aspects of the supply chain from strategy to execution
and the collaboration across inter- enterprises and cross functional units are a
necessity of modern supply chains. An end-to-end visibility of connected supply
chain data is key to analyze, sense and respond to changing market situations. In
this chapter, we will cover the digitization of supply chains and intelligent visibility
with a focus on the following topics:
• Digital supply chains.
• Intelligent visibility connecting strategic, tactical and operational processes.
• Supply chain control tower—global and local.
• Next-gen supply chain analytics.
• Supply chain alerts and exception management.
• Insight-to-action on supply chain issues.
• Supply chain performance metrics—KPIs.
• Cognitive supply chains—role of technology in Industry 4.0.
• Resilient supply chains.
• Evaluation of supply chain assumptions, risks and opportunities.
• Collaborative enterprise planning integrating finance with operations.
R. Jandhyala ()
SAP Labs, Tempe, AZ, USA
e-mail: raghavjv@gmail.com
Digital supply chains are the next generation real time supply chains with seamless
visibility of data and well-orchestrated planning, execution, cross-collaboration,
analytics and intelligence and automation across all operations of a supply
chain.
Digital supply chains enable supply chain business processes such as Design-
to-Operate. Before we get into the digital supply chain, let us first understand the
challenges and opportunities in a supply chain network.
With the changes in supply chain processes, thin margin and availability of
technology, business processes have changed over time, moving from a traditional
waterfall planning and execution to a more continuous planning and execution.
A closed loop and continuous alignment of planning and execution take the
business to a new level, which was not possible before, connecting the strategic
and tactical processes to the operations/execution and responding to short term
disruptions along with incorporating better customer and product experiences.
Integrated Processes Supply chains have evolved from being fragmented with
departmental planning and execution to a collaborative one with integrated planning
across demand, supply and finance.
Faster Planning Cycles By bringing data to one system of planning and execution,
customers now have a greater visibility of data and can plan faster and frequently
compared to the traditional monthly or quarterly planning processes.
Accuracy of Supply Chain Models With the advances in forecasting and planning
algorithms across demand, supply and inventory, enterprises have shifted from being
supply driven to demand driven by able to sense customer demand and respond to
changes more efficiently.
Improved Experiences With better visibility into customer and product expe-
rience, enterprises can support mass manufacturing of products and bring more
customization choices to their end customers.
Figure 1 shows an evolution of supply chain planning from efficiency to
experience. This figure is adopted from SAP positioning of digital supply chain.
Though the evolution of supply chains over the last few decades has dramatically
increased, the next generation, state of the art supply chains are moving towards
bringing:
Autonomous Self-Regulated Capabilities Supply chains that consider historical
trends, internal and external events, and signals to provide recommendations, based
on corporate goals assist the planners with informed decisions and supply chain
executives with strategies to drive the business forward.
Real time Synchronized Planning The traditional planning with monthly and
weekly cadence will evolve towards continuous and real time planning. This will
help to think beyond the traditional demand and supply streams to a new value-
chain oriented supply chain view across the end to end supply chain.
Network Collaboration The business processes collaboration and visibility are
increasingly getting better within an enterprise. However, as seen in the recent
COVID-19 pandemic, supply chains are global, and business processes are inter-
connected with enterprises beyond their own chains. A disruption in one of the
22 R. Jandhyala
enterprises has an impact on the end to end supply chain. Therefore, a network of
inter-enterprises with collaboration and visibility into business processes, enabling
easy onboarding of business partners into a truly scalable network, is required.
By digitizing the supply chain, companies can expand beyond their traditional
boundaries and enter new markets with new business models. With the increasing
competition and continuously changing customer behavior, companies need to
evolve to be innovative and survive the market disruptions. Figure 2 shows the
new business models enabled by the digital supply chain, as adopted from SAP’s
positioning of Digital Supply Chain of One.
Consumer Buying Patterns Consumers in the current digital world have multiple
avenues to shop for the best product and get the best competitive price. Customers
have different buying patterns. They can even search online and buy at store, or
physically view that product in a store and buy online. Customers’ buying decisions
are influenced based on product reviews and trends on social media. The physical
and virtual boundaries are blurring for consumers, and the companies need to be
innovative to meet or exceed consumers’ expectations.
Examples of digital disruption by innovative companies: Uber is the largest
taxi company without owning any vehicle. Netflix is the largest streaming company
without owning movies; however, they later changed their business model to launch
Intelligent Digital Supply Chains 23
Netflix movies. Airbnb is the largest accommodation provider without owning any
property.
Wholesale to Consumer Model
Consumer product companies are expanding their business boundaries to open
online channels to reach consumers directly; they now ship their products directly
to the consumers in addition to supplying them to retailers. This also helps them
to sense and respond directly to the changing market or customer buying patterns
and serve better products to the customer. These companies now have a better pulse
on the consumers as they do not solely rely on the retailers to provide the demand
signals.
Retail Supply Chains
Retailers are entering the consumer product market by providing their own brands
and products at a competitive price. Retailers have a better sense of the customer
demand and buying patterns by analyzing the point-of-sales transactions (POS) and
are expanding their supply chain footprints with their own distribution centers and
manufacturing plants for products that can bring additional revenue. For example,
Safeway provides signature branded products across different departmental units.
Walmart offers “Great Value” brand for various products manufactured by Walmart.
Retailers to Logistic Providers
With customer demand ever-increasing to get the product physically within the same
day or the same hour, we see retail companies entering the logistic business. Online
retailers like Amazon are entering the logistic space with their own fleet of logistics
or a mix of inhouse or external logistics partners. Companies like Albertsons own
and partner with logistic providers for same-day delivery of items purchased online.
Best Buy and Home Depot offer in-store pickup and delivery services. Companies
that offer a high level of logistics flexibility see increasing loyalty with customers
and now have firsthand information on customer demand and better visibility of
their inventory position.
24 R. Jandhyala
Logistics to Manufacturers
Logistic companies, to be competitive and provide the best customer experience,
are investing in 3D manufacturing to reduce the costs and carbon impact of
transportation wherever it is feasible. For example, auto logistic companies for spare
parts are partnering or establishing 3D manufacturing sites for spare parts in the
hubs where there is more customer demand, to reduce interstate transportation and
optimize intercity delivery.
Manufacturers to Service Delivery
With the digital supply chain and online digital twin representation of a physical
asset, manufacturers can get information on the asset usage and proactively provide
services for maintenance to the consumer. This integrated manufacturing and
servicing model generates additional revenue streams, better customer service, and
reduces costs.
Organizations need to have a unified view of supply chain plans and the execution
of such plans to enable a real time supply chain that profitably balances demand and
supply. Organizations need to react faster to changes and break departmental silos
of information by enabling a connected supply chain that can reduce planning cycle
26 R. Jandhyala
time and run scenarios for decision making. An intelligent supply chain enables
organizations to profitably analyze, sense and react to changes in real time to meet
the supply chain goals of
• Increasing supply chain agility
• Increasing end-to-end visibility
• Reducing cycle time for planning
• Increasing on-time delivery and customer service levels
• Reducing supply chain cost
In order to realize the end to end digital supply chain, it is important to have an
end to end visibility of all the processes and activities across several stakeholders
and interconnected networks. These include:
• Visibility of planning processes for strategic, tactical and operational planning
• Close loop alignment of planning and execution
• KPIs to measure supply chain performance and health
• Real time visibility of events that happen during execution.
Figure 4 shows the systems, people, processes and tools across a supply
chain. These include planning types ranging from strategic to execution, business
processes covering sales and operations planning to manufacturing and sales &
distribution, time horizons ranging from years to hours, dimensionality ranging from
aggregated business units to detailed order level, systems ranging from planning
to logistics systems, and supply chain roles ranging from sales and marketing to
operations.
The subsequent chapters will cover some details of each of these areas that make
an end to end supply chain visibility challenging and the benefits of having a truly
integrated supply chain system.
• Multiple systems are involved in a supply chain process with disconnected data
and processes
• Data warehouses with a lack of visibility across all involved parties. Misaligned
data and rigid models.
• Weekly/nightly processing of data to warehouses that delays visibility into
problems
• No proper visibility related to tracking of adherence to the plan.
• Disconnected functional units and processes—siloed and stale data.
• No visibility outside of the four walls of the supply chain network. e.g., lack of
knowledge on whether the supplier has enough inventory, and if they can commit
to the forecasts to keep up customer service levels.
• Lack of visibility to demand, supply, inventory along with insights into strategic
plans, execution, and financial targets.
A supply chain control tower is a system of record for supply chain visibility
that connects the data across multiple sources to provide an end-to-end visibility,
performance management, exception management, case resolution, and simulation
capabilities. It provides insights into the current problems in the supply chain and
events that result in plan deviations, and it projects the future outcome by mitigating
the risks.
There is an increasing demand in the market for supply chain control towers
that provide real time insights into the internal and external events which affect the
normal operations of the supply chain and reduce the time spent by users to identify,
analyze and resolve issues in a timely manner. With the advances in technologies
and artificial intelligence, there is a need for autonomous supply chain control
28 R. Jandhyala
towers that can self-identify and heal common and repeating pattern issues. Supply
chain control towers are not the systems for planning or execution, but they provide
visibility across the end to end supply chain with a single source of truth for decision
making. For example, a supply chain control tower can provide visibility that a sales
order will not be delivered on time and will affect the order fulfillment KPIs with
the strategic customer. The resolution for this issue is typically performed in the
planning or execution system. Supply chain control towers can be classified into the
global control tower and local control tower.
Global control tower brings focus on key aspects that affect the movement
of materials across the supply chain with visibility to inbounds, outbounds and
in-transits across the internal and external networks. A global control tower may
provide visibility to critical events, and the details of such events are managed in
a local control tower. For example, a local control tower for manufacturing will
have visibility over the production plans, deployments and detailed scheduling,
whereas the control tower for logistics will have information about freight orders,
transportation schedules and other execution data. Once the data is sourced from
multiple systems, it is important to identify which data is relevant for visibility and
only bring that data to a global control tower and manage details in a local control
tower.
A global control tower would primarily have insights into the overall demand,
supply and inventory positions along with insights into execution data related to the
movement of goods. Further, any events from the local control tower, for example,
a truck delay should be visible in the global control tower for visibility into the
issues and decision making. Figure 5 shows local control towers of sales, assets,
manufacturing and logistics providing insights to the global control tower for end-
to-end visibility, scenario planning and collaboration. This figure is adapted from
SAP’s positioning of Integrated Business Planning (IBP) Supply Chain Control
Tower.
Intelligent Digital Supply Chains 29
Supply chain control towers also connect digitally with the extended supply chain
networks including suppliers, contract manufacturers and logistic providers. For
example, a forecast commit scenario between manufacturer and supplier digitally
connects the requirements of raw materials for manufacturing of finished goods
by the buyer with the forecast commit of raw materials from the supplier. Further
visibility of supplier’s inventory allows the buyer to make better decisions on
which customer demands to fulfill. This reduces uncertainty in planning and better
transparency across all stakeholders of the extended supply chain.
Supply Chain Systems of Record
Most enterprises have multiple supply chain systems to manage the end to end
supply chain, each with different purpose. Thus, for an organization, a supply chain
is an interconnected system of systems. Some of the common systems in the supply
chain are listed below:
• One or more enterprise resource planning (ERP) systems as execution system
of records based on different geographic regions or different business units.
Examples include SAP ECC, SAP S4HANA and JDA.
• One or more planning systems as planning system of record based on different
geographic regions or different business units. Examples include SAP Integrated
Business Planning, JDA, Kinaxis, and E2OPEN.
• External systems for collaboration with suppliers and customers. Examples
include SAP ARIBA and Oracle.
• External pricing systems.
• Warehouse management systems.
• Transportation management systems.
• Logistic business network providers.
• Systems from mergers and acquisitions.
• Experience management systems for unstructured data/customer sentiment anal-
ysis.
• Data warehouses for standard reporting and KPIs.
• 3PL third party systems.
Although it would be desirable to have a single system which can manage the end
to end supply chain, in reality, the enterprises have multiple supply chain systems.
Visibility to data across all these systems gives insights into supply chain health to
identify potential issues and take corrective measures.
A supply chain control tower connects the data across these multiple sources to
provide an end-to-end visibility and single source of truth across all stakeholders. It
provides instant insights into the demand, inventory and supply across the entire
supply chain with visualizations in the form of supply chain network diagram,
geographic views, heatmaps and charts.
However, it should be carefully managed as to what data should be included
for visibility because bringing data across all these systems into a single system
for visibility would be a massive amount of records with high latency, multiple
points of failure, storage costs, and noise which defeats the purpose of gaining useful
30 R. Jandhyala
Supply Chain Control Tower integrates performance data from multiple systems
ERP
SNC
ERP
S/4 HANA S/4 HANA
3rd Party
insights from data. Therefore, it is important for end-to-end visibility to bring in the
key information from these different systems with aligned master data and business
semantics and present to the end users at the right level of granularity for analysis
and decision making. Figure 6 is adapted from SAP’s positioning of IBP Supply
Chain Control Tower and shows an example of several supply chain systems across
different regions and business units. These include systems like SAP ERP or SAP
S/4 HANA for supply chain execution, Advanced Planning and Optimization (APO)
for supply chain planning, SAP Integrated Business Planning (IBP) for next-gen
supply chain planning, SAP SNC and SAP ARIBA for collaboration with external
suppliers and several other third party systems that are integrated to SAP IBP Supply
Chain Control Tower.
Analytics on real-time supply chain data is very much needed to understand the
health of the supply chain and make data-driven decisions. Analytics has been a
very important and integral part of supply chain management, especially needed
to visualize and make a meaningful analysis of the vast amount of supply chain
data. Analytics provides a graphical way to understand the supply chain data,
trends, seasonal and buying patterns, comparisons (cost-to-serve by region), and
view anomalies and outliers (e.g., promotions in sales history) in the data. Analytics
provides end to end visibility of inbound and outbound activities that happen in a
supply chain by integrating important information across multiple source systems. It
Intelligent Digital Supply Chains 31
is the basis for managing exceptions and driving actionable results from the derived
insights in supply chain networks.
Supply Chain Data Is a Big Data Problem An enterprise with 1000 customers,
100 products, 10 distribution centers, and a few factories can plan millions of
combinations of products, customers and locations. Add to that the time series data
of past 2 years and predictions on next 3 years in weekly granularity and multiple
plans across sales, marketing, supply chain and finance. This can easily result in
billions of planning data points. As supply chain data grows and changes rapidly
with accumulating sales, production, procurement, transportation and events data,
it is necessary to view the complete picture of a supply chain rather than a narrow
focus on a particular business process or business unit or region.
Even with technological advancements, there are several challenges that organi-
zations face to get decent analytics
1. No real time visibility of data because of long ETL (Extract, Transform and
Load) processes that bring data to an enterprise data warehouse from transac-
tional systems. This can take sometimes days or weeks for users to see the latest
and accurate data in the enterprise data warehouse.
2. Multiple supply chain systems of data (ERP, planning, logistics, warehouse, etc.)
with a lack of integration and semantics matching across systems.
3. IT overhead to create and maintain analytic charts and reports.
4. Fixed analytics with a lack of flexibility to create different chart types or perform
drill-downs or filter data.
5. Offline spreadsheets and charts to multiple copies of data that get stale over time.
6. No single version of truth resulting in poor management review meetings where
the discussions and firefighting are on who has the right data rather than making
decisions based on one version of the truth.
Beyond Visualization The technology advances in storing and analyzing data have
opened new avenues for supply chain executives to view one version of the truth and
make data driven decisions based on real time data. Such an integrated analytics
system allows supply chain analysts to go beyond simple visualizations of charts to
a comprehensive analysis and decision-making process, which includes:
1. Aligned supply chain KPIs and metrics.
2. Balances scorecard of key supply chain functions.
3. Transparency of information and collaboration across various functional units.
4. Exception management with insights: Automated alerts along with insights into
the root cause and solution recommendations.
5. Predictions of events that disrupt to supply chain. For example, container
shipping late.
6. Real time ‘what if’ analysis and scenario planning to evaluate several solutions
along with the financial impact of the decisions.
7. Connected supply chain sources with an end to end visibility of the network, both
internal and external.
32 R. Jandhyala
8. Navigation to the source system of records e.g., sales orders in ERP to make
changes to data.
Figure 7 shows an example of a supply chain dashboard with the following chart
types
• Geo charts showing manufacturing locations
• Heatmap presenting capacity utilization
• Horizontal bar chart indicating resource overload
• Pie chart showing sales by country
• Combination line and bar chart comparing financial plan and forecast projections
• Vertical bar chart reporting actual revenue by products segmented to A, B and C
categories.
Analytical charts can be classified for different functional areas of supply chain e.g.,
demand, supply, inventory, financial reconciliation, or across process types from
strategy, tactical to operational plan across different time horizons, organizational
structures and roles (Kepczynski et al. 2019a, 2019b; Kusters et al. 2018; Chopra et
al. 2013). Figures 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, and 21 are some
of the common charts in a supply chain as adapted from SAP Integrated Business
Planning.
Intelligent Digital Supply Chains 33
1. Forecast analysis: Line chart of forecast vs. actuals comparing statistical forecast
generated from sales history, sales quantity and actuals quantity
Intelligent Digital Supply Chains 35
2. Customer service levels: service levels by product family, customer segment and
location with drill down options to view customer fill rates.
3. Time series analysis: Distribution of time series property across the demand
inputs. This shows a pie chart of data distributed by continuous, seasonal and
trend data.
4. Forecast error and bias: A vertical bar chart by time series comparing the forecast
error and bias in the planned forecast vs. actuals for a 3 month lag.
5. Adherence to financial targets: compare consensus plans to the annual operating
plans and actuals quantity projection from the prior year across all product
families.
6. Inventory types valuation: inventory valuation for the next 12 weeks distributed
across safety stock, cycle stock, merchandise stock, pipeline, and in-transit
inventory.
36 R. Jandhyala
15M
10M
5M
0
Finished Goods Semi-Finished Raw-Materials Packaging
7. Global inventory position: geo chart of stocking nodes in the supply chain
showing the on-hand position of inventory and locations with excess inventory
and low inventory.
8. Inventory projections by material types: view inventory position by finished
goods, raw materials, semi-finished goods and packing materials comparing
against inventory targets. Provides visibility into projected inventory shortages
below safety stock.
9. Production and distribution network: gives visibility to material flows, in-transit
quantities across the network, late shipments, etc. Figure 16 shows a supply chain
network visualization of all the downstream and upstream nodes of the network
from customers to distribution centers to manufacturing plants to suppliers along
with the production process from raw materials to finished goods.
Intelligent Digital Supply Chains 37
250
200
150
100
50
0
Sim - Cap Usage PR
10. Resource capacity views: shows the available capacity vs. the capacity usage of
the products. It provides visibility into resources that are over or under-utilized
with drill down capabilities to see which products are consuming more capacity.
38 R. Jandhyala
800K
600K
400K
200K
W1 W2 W3 W4 W5 W6 W7 W8
14. Complete demand, supply and inventory: include charts showing both volumes
and financial valuation of quantities allowing drill downs into demand (forecast,
sales orders, dependent demand), inventory (current and projected inventory by
finished goods, raw, pack and work-in-progress) and supply elements (distribu-
tion receipts, in-transit inventory, production planned, production confirmed)
Modern supply chain systems should provide business user friendly analytics
capabilities that help the user to quickly get to the right information without heavily
relying on the IT team to provide the necessary reports, data and models. Usability
of analytics and good visualizations are critical for the adoption of analytics by the
end user community. Below are some of the characters of analytical tools which
help users unleash the business value of data.
• Pre-delivered content with the standard set of dashboards, charts and reports for
most common areas in supply chain
• Self-service creation of charts specific to the user’s needs with very minimal
training.
• Multiple visualizations to represent the data in the right format which could be
both table and chart format.
• Various charting options like network charts, geographic chart, pie charts, bar
chart, dual axis charts and waterfall charts.
• Define and visualize thresholds for the key figure measures on the charts
• Easy filtering options on the data represented in the charts by time ranges or
supply chain dimensions like products, customers, locations, resources, etc.
• Drill down into the details of a specific section of the chart allowing users to
narrow down on an issue.
• Flexible slice and dice of data across different hierarchy levels.
• Easily define and view top N and bottom N values.
• Smart insights into the information beyond what is displayed on the chart
• Annotations on different sections of the chart
• Faster performance on charts with faster refreshes of data.
• Collaboration of analytics with other stakeholders by sharing charts.
Alerts or exceptions to the planned results occur very frequently in supply chains.
In order to focus attention on the deviations in the supply chain, users require
robust alert management functions that work on the vast amount of data and provide
manageable alerts at the level which can be understood acted upon.
40 R. Jandhyala
Time horizon: the level and period range for the alert to be calculated, e.g., 2020
calendar week 20 to 2021 calendar week 19.
Severity: the priority of alerts: high, medium, or low.
Alert rules: flexible conditions that can be defined by the user. The conditions can
be defined as absolute conditions, e.g., capacity utilization over 100%, or relative
conditions, e.g., projected stock >110% max inventory. Alert conditions can be
further defined in rules groups with or/and on alert conditions
Alert impact key figures: the measure of the impact of the alert: e.g., cost of
overstock.
Alert charts: the additional analytical charts attached as part of the alert definition
to understand the alert situation better.
Figure 23 shows the results of alerts generated by running an alert definition on
the data. The results show the combinations of resource, location and time for the
alert and its impact value. e.g. resource 1011 for location 1010 in week 2020 CW44
has a capacity utilization of 102%.
Alerts get notified to the user through emails or are shown in the user interface
based on the user’s preference. Users can take further actions from the generated
alerts. For example, snoozing the alert or defining a case to manage the mitigation
of alert with other stakeholders.
Machine Learning for Alert Recommendations
Alerts conditions in the alert framework are typically defined as static rules. It works
if data is not changing and the thresholds can be well defined. However, many
planners find it hard to define the right thresholds and rules for alerts, leading to
too little or too many alerts if not defined correctly.
Machine learning can perform pattern recognition on the data to detect anomalies
and report as alerts. One common outlier technique is to use DBSCAN unsupervised
learning clustering method on the data. DBSCAN detects outliers from the data
points that do not fit to the clusters. Typically, Planners have to manually provide
the lower and upper threshold to detect outliers. However, with machine learning,
the system can identify outliers based on the similarity of the key figure value in
other time buckets.
7 Insight to Action
True control tower systems provide insight to action, while allowing supply chain
analysts to visualize information, perform analysis, view exceptions and their
impacts and finally take actions to resolve the issue (Fig. 24).
Scenario: Depleting Inventory
For example, the projected inventory in a global inventory position chart shows that
in the next quarter, the inventory will be depleting and will go below the safety stock
levels. The supply chain analyst needs to analyze what is causing this issue. A drill
down into the specific region and time shows that the available supply is less than
demand for that quarter.
On further analysis, the reason could be that there is a planned factory shut down
for maintenance in that quarter and no additional production. Therefore, the demand
needs to be fulfilled from the existing inventory.
The analysis can further drill down to view which products at which locations
are affected and take corrective measures like rebalancing inventory from other
locations or using a contract manufacturer.
Scenario: Delayed Customer Orders
In this scenario, the control tower provides visibility into the sales orders execution
data and alerts on the orders that are not confirmed on time or confirmed below the
requested quantities. The supply chain planner can then look into the gating factors
or root cause analysis related to the orders with the late delivery and then navigate
to the source system to change the order dates or quantities to reduce the impact on
the customer service levels.
An intelligent visibility application brings together all the steps from insight to
action in a single user interface, without the planner having to login to different
Intelligent Digital Supply Chains 43
takes days to identify and resolve, is now reduced to hours, increasing the user’s
productivity so they can focus on other important activities.
transparent across all stakeholders. KPIs can be defined across areas like customer
service, demand, supply, inventory, production, logistics and procurement.
Supply chain data is vast, with several billions of planning combinations and data
points. In order to make a meaningful business representation of the data, KPIs are
necessary both to keep track of the progress and measure against the set objectives
of the business.
The KPIs are reviewed by all roles in the organization from the boardroom to
the shop-floor. C-suite executives have well defined KPIs that measure the business
growth, e.g., revenue margin and customer service levels. Production planners have
KPIs to measure the utilization rate of resource capacity.
KPIs should be measurable, transparent, mutually agreed, value oriented and at
the right business level to facilitate collaboration, communication and adherence to
common goals for an organization. KPIs should be adequate with a proper RACI
matrix across the stakeholders, and defining too many or too few KPIs without
owners of such KPIs will not add much value.
KPIs are visually represented for reporting in dashboards and scorecards.
They are usually easy to understand single metrics with visualizations such as
speedometer, percentage meter, waterfall chart, traffic lights or single values. They
are measured against target values or thresholds.
KPI are driven by organizational objectives, as illustrated below:
• Sales and marketing: improve promotion effectiveness by 20% by aligning
activities across all functional units from commercial, supply chain to finance.
• Finance: topline revenue increase by 10% and margin growth by 6% by
identifying growth opportunities.
• Inventory: reduce out of stocks and balance inventory to reach higher customer
service levels, reduce inventory write-offs by 5mi USD by better planning and
visibility of products with expiring shelf life.
• Demand planning: improve forecast accuracy to 86% from the current 78%,
launch new products faster to the market by improving the NPI (New product
introduction) process.
• Supply chain: reduce the lead time of products from 45 days to 30 days, improve
resource utilization from 82% to 95% by streamlining production activities.
• Process Improvements: improve planners’ productivity by reducing planning
cycle preparation from 8 days to 3 days, increase process automation and improve
process efficiency to better utilize time and critical resources.
• Customer service: increase customer service levels from the current 92% to a
target 96% with improved case fill rates.
Common Supply Chain KPIs
Some of the KPIs commonly used in supply chain across demand, inventory, supply
and production are listed below. These KPIs also bring together the data from
planning and execution to measure adherence to the plan.
Demand KPIs
• Forecast accuracy: it is measured as forecast error (MAPE), where MAPE (mean
absolute percentage error) is the average of absolute percentage error between
46 R. Jandhyala
where A(t) is actual quantity at time t, F(t) is forecasted quantity at time t, and n
is the total number of periods
• Forecast bias: it measures the average percentage error, which is the deviation of
forecasts from actuals.
1 At − Ft
n
MPE =
n At
t =1
where A(t) is actual quantity at time t, F(t) is forecasted quantity at time t, and n
is the total number of periods
Customer Service KPIs
• Projected fill rate: ratio of constrained demand (supply) vs customer demand in
current and future periods.
• Historical case fill rate: measures the ratio of quantities ordered vs. quantities
fulfilled across items ordered.
• Historical line fill rate: a measure of what percentage of items in the order are
fulfilled on time with the right quantity.
Inventory KPIs
• Days of supply: a measure of the number of days the projected inventory will
cover the demand of the subsequent periods
Supply KPIs
• Production adherence: a measure of the ratio of planned production (scheduled
production) to the actual production from the execution system
• Supply shortage: a measure of the missing supply quantity compared to demand
forecast across the combinations of products and locations for a given time range.
• Capacity utilization: the ratio of capacity usage of the resource to its available
capacity.
Financial KPIs
• Total inventory valuation: total valuated cost of all inventory types including
finished goods inventory, raw materials and work-in-progress inventory
• Cost of goods sold: cost to serve the goods sold to customers, which includes
direct and indirect costs related to materials, labor, production, procurement and
transportation.
• Inventory turnover: the number of times the inventory should be cycled during
a year. It is based on valuated annual inventory and annual COGS i.e. ratio of
annual COGS to annual inventory
SCOR KPIs
SCOR KPIs (APICS SCC 2015) are industry standard KPIs provided by APICS that
spans all supply chain business process with a standard set of hierarchical KPIs of
around the six primary management processes: plan, source, make, deliver, return
and enable. Below is an example of perfect order fulfillment KPI.
Perfect order fulfillment: a measure of how the customer orders are serviced
based on other detailed KPIs like
• Orders delivered in full: a measure of whether an order is delivered to the
customer in the right quantities.
• On-time performance: a measure of whether orders are delivered within the
committed dates and to the correct customer location.
• Orders delivered in condition: a measure of whether the orders are delivered with
current documentation, e.g., billing, shipping and other documents.
KPI Templates
Organizations should follow a good template for defining and managing the KPIs.
Each KPI should be structured around one focus area, for example, supplier
performance with well-defined purpose of the KPI. Below is a guideline for defining
KPIs for managing supply chain
• Name: choose a name that provides a clear short definition of the KPI without
ambiguity.
• Definition: provide the business context of the KPI with clarity on the measure
(e.g., ratio, percentage, quantity, etc.) and how the results can be interpreted.
• Category: tag with the primary focus areas to which this KPI belongs, for
example, customer service.
• Calculation: provide a precise calculation using mathematical formulas to clearly
define the calculation.
48 R. Jandhyala
• Roles: define the users who are responsible for the KPI along with other roles
players who use this KPI. A RACI matrix of the KPI will help to define the clear
accountability of the KPI.
• Data source: define the sources of data and transformations required to retrieve
and load to the target system where the KPIs are stored or calculated. E.g., sales
orders and deliveries with details from order to line items to schedule lines from
SAP S/4 HANA, JDA, etc.
• Related KPIs: Define how this KPI is related to other KPIs in the same or
different categories. For example, KPI of safety stock projection may depend
on the KPI forecast accuracy.
Interested readers can find more information on supply chain KPI from
references (APICS SCC 2015; Ravindran and Warsing 2016).
Advances in technologies have played a key role and continue to transform supply
chains from traditional supply chains to next generation cognitive supply chains. A
cognitive supply chain is proactive to sense supply chain disruptions and is capable
self-correcting, triggering re-planning and providing intelligent recommendations
through advances in artificial intelligence like reinforcement learning where mul-
tiple agents work towards the common goal. With technological advancements to
store and process the complex and vast amount of growing big data along with
flexibility with cloud deployments, supply chains are smarter than ever before to
quickly respond in real-time to changing environments.
Figure 28 shows the technologies currently available and being used in enterprise
business applications to transform the business processes, business models and
workspaces. It is adapted from SAP’s positioning of intelligent technologies for
the intelligent enterprise.
We will look at some of the important technology areas that play a major role in
supply chains.
• Big data and in-memory computing
Advances in big data have made it possible to store a vast amount of supply
chain data from multiple sources into data lakes for real time visibility into the
supply chain and real time simulations. Big data provides the much needed 4Vs
for an enterprise:
– Velocity: Frequency of data which can be real time, batch or streaming data.
– Volume: store and process petabytes of supply chain data and aggregate results
on-the-fly.
– Variety: bring together structured and unstructured data.
– Value: enable smarter supply chains by deriving value from the data.
Intelligent Digital Supply Chains 49
In-Memory computing,
Internet of things
Big data
Machine learning,
Artificial intelligence
Cloud scalability,
elasticity, availability Cloud Platform
Blockchain &
smart agents
Mobile &
conversational 3D Printing
and significance in the supply chain, especially in the execution areas like
transportation management and digital manufacturing.
In the digital manufacturing area, hundreds of IoT sensors are part of the asset,
e.g., production resource providing asset intelligence. These sensors monitor
the temperature, vibrations, and other performance metrics and can intelligently
sense if the asset needs maintenance. In transportation, IoT sensors monitor
the state and conditions of goods movement and automate material handling
with RFID technology across several logistic hand-over points without human
intervention. For example, in medical devices transportation, IoT sensors can
monitor the temperature fluctuations and severe vibrations that can damage the
equipment. IoT is an essential technology in the Industry 4.0 revolution of supply
chains.
• Blockchain
Blockchain provides the transparency, traceability, and auditability of all
documents and events that happen in the supply chain for material movements.
There have been good use cases for blockchain in supply chains in the areas of
agriculture, logistics and life sciences.
For example, in agriculture, blockchain helps track every stage of the process
for the fresh food items to reach consumer, and in an event of outbreak, trace back
to the origin. In logistics, blockchain helps track every movement of material
across logistic providers, shipping ports and inter-continental trading hubs to
instantly track and respond to delays.
• Machine learning, artificial intelligence and robotic process automation
Intelligence and automation elevate supply chains to focus on high value
activities. The digital transformation of supply chains, together with advances in
big data, algorithms and hardware, have enabled the new AI era of the cognitive
supply chain where systems learn from the data to augment human decision with
intelligent recommendations.
Supply chain systems have traditionally relied on statistical and deterministic
models for forecasting, supply network optimization and several other planning
and execution processes. With machine learning and AI techniques for super-
vised learning, un-supervised learning, reinforcement learning and deep learning,
systems can analyze patterns in the data, correlate with external data and events
and augment or replace the traditional planning with cognitive planning. For
example, demand sensing augments the statistical demand planning process by
sensing short term demand based on pattern recognition of sales orders, sales
history, demand history and external data like weather and point of sales.
Using robotic process automations, supply chain users can deploy bots that
can carry out mundane tasks of users, thereby freeing up their time for other
productive activities. The bots can carry out tasks like reading email for new
procurement requests, logging into multiple systems for data entry, solving
supply chain alerts, and trigger workflows.
A cognitive supply chain self-learns from the patterns in the data and provides
predictive and proactive insights to the users to make timely decisions before
disruptions occur. It provides:
Intelligent Digital Supply Chains 51
Descriptive Analytics
Many organizations see the need to view and report on the data and have invested
over years to build analytics using Data Warehouse platforms and powerful analyti-
cal and visualization tools. These are descriptive analytics with fixed representation
and visualization of data. Figure 29 shows a dashboard which has both structured
and unstructured data, as adapted from SAP for illustration. By integrating customer
sentiment data, the User is able to see how the negative perception of the product
is affecting the customer satisfaction. Such analytics help users to instantly see the
related data together to make some inferences.
Predictive Analytics
Predictive analytics provide smart data discovery tools for predictive insights into
the data and predicting future outcomes using Machine Learning. This helps with for
proactive analysis rather than being reactive. For example Figure 30 is based on the
customer sentiments analysis, it is predicted what the future customer satisfaction
will be and how it affects the revenue and profitability. This figure is adapted
year. Promotions can be thought of as a variant of an event with a short time span,
e.g., 2 with 1 offer and back-to-school promotions.
A well-orchestrated process for managing business drivers has the following
benefits:
• Captures all qualitative drivers like internal or external and short term or long
term across all functional units
• Reacts to vulnerabilities or new opportunities much quicker.
• Increases transparency and trust across different functional units and stakehold-
ers.
• Consolidates drivers across top-down drivers from strategy and finance, with
operational drivers entered by local and regional planners and cross-functional
across different hierarchies.
• Increases productivity by capturing all influencing drivers that impact the supply
chain in one place.
• Collaborates on decisions making on which drivers to be included in planning
cycles
• Improves business review meetings by evaluating multiple scenarios of which
drivers to include in the plan, which is otherwise done offline with a toll on time
and resources.
Process for Managing Business Drivers
A sample business process of managing business drivers, including assumptions,
risks and opportunities is shown in Fig. 32. This can be different for each orga-
nization based on its business purpose. Assumptions are typically defined once a
year with collaboration across all the functional units, including commercial, supply
chain, finance and strategy. Assumptions are defined at an aggregate hierarchical
level, example product category along with the influencing drivers, example growth,
market share, competition etc. for each month or each quarter.
During the monthly review meeting, the assumptions are reviewed and updated
with any changes based on the market conditions and actual performance of the
business. Risks and opportunities are defined for each planning phase i.e., demand,
supply, financial reconciliation, etc. Next, the quantitative plans are adjusted to
include the agreed risks and opportunities. Further, any changes to the forecasts are
tagged with the change in assumption. This provides a context for the quantitative
changes by associating them with the assumption changes. During the review
meetings, an assumption changes report is generated, which brings together the
qualitative influencing factor and the quantitative change, e.g., forecast change
associated with it.
SAP integrated business planning product provides the required tools and
functionality to easily manage the business drivers. Figure 33 is an example of
managing risks and opportunities in a driver-based planning application of SAP
IBP that provides the required user interface, sample data model and best practice
content to easily manage risks and opportunities and collaboratively include them
in the plan.
Intelligent Digital Supply Chains 55
The planning views provide the flexibility to define a business driver e.g., risks
and opportunities across multiple planning hierarchy levels, time ranges and driver
impact key figures for the functional area. For example, there can be demand risks
and opportunities defined for the North America region for current and future
quarters; supply assumptions defined for all regions for the next 2 years.
From the planning view, which defines the structure of the business drivers,
planners can navigate to view and define the individual drivers. For example, below
is a summary view of all risks and opportunities defined at the hierarchical level
product family and customer region. Each driver has the following properties:
1. Qualitative information about the driver: this includes name, description, plan-
ning cycle and type.
2. Hierarchical level: a driver can be defined for one or more hierarchical levels,
e.g., risk 1 is for product family 1 and product family 2 for all customer regions,
and risk 2 is for product family 3 for EMEA region.
56 R. Jandhyala
3. Time range: each driver can be defined for different time ranges at a defined time
granularity, e.g., Jan–May 2020.
4. Impact key figures: these are the qualitative measures representing the impact of
the key figure. For example, opportunity 1 has an impact of 5mi USD in revenue.
5. Qualifying attributes: these include driver attributes like budgeted, status, proba-
bility and include-in-plan.
The drivers can be defined in one go at multiple planning hierarchical levels and
time ranges, with a qualitative value for the driver impact key figure. The key figure
value can also be adjusted for different time periods or planning combinations. For
example, opp1 has higher weightage of revenue in the later quarter compared to the
first 3 quarters.
The summary view lists all the risks and opportunities along with analytical
charts to view the impact of the drivers on planning data like top 5 risks, top 5
opportunities, alerts on risks and opportunities above thresholds (Fig. 34). Planners
can review all the risk and opportunities in one go and must maintain drivers by
changing attributes like probability and status or carry forward drivers to the next
planning cycle. Qualified drivers with the right granularity of planning attributes,
which are budgeted and financially viable, and with higher probability, can be
included in the plan by changing the attribute value include-in-plan to 1, and its
effects can immediately be seen in the planning data.
The contribution of the risk or opportunity can immediately be seen in the supply
plan. For example, risk1, risk 2 and opp1 are included in the plan. These drivers are
defined at an aggregated level of the product family and customer region than the
final forecast key figure, which is at the product/customer level. The contribution of
these three drivers is disaggregated from aggregate level to product customer level
of the final forecast key figures and added or negated from the planned forecast
to arrive at the final forecast value that represents the plan considering risks and
opportunities, as illustrated in Fig. 35.
Fig. 37 Planning types and horizons for finance and supply planning
(internal and external) to follow, which positively impacts the balance sheet and
profit and loss statements of the business.
Most companies follow these planning cycles shown below for strategic plan-
ning, which is 3–5 year long term plan followed by an annual operating plan or
budget plan followed by sales and operations planning for the Rolling 18 months.
An example of the planning processes is shown in Fig. 37.
Financial planning systems need the volume-based supply chain plans for
running processes like budget planning and product cost planning, and the supply
chain planning systems need the valuation based financial plans for consensus
demand planning and financial reconciliation.
Supply chains have financial costs associated with each step to fulfill customer
demand, as shown in Fig. 38. These include raw material costs, production costs,
productive costs, direct and indirect costs, transportation costs and storage costs.
Supply chains employ profit optimization techniques to arrive at a profitable cost-
effective strategy to fulfill customer demands.
Figure 39 shows a consensus planning view in SAP Integrated Business Planning
consolidating inputs from all functional units -sales, marketing, demand planning,
finance to arrive at a consensus demand plan. AOP represents the annual operating
62 R. Jandhyala
plan. Both volumes and values of different plans are recorded or calculated, and the
gap between AOP and consensus plans is visible month to month. Planners have
a consolidated view of the complete demand picture and can make adjustments to
close the gap to the financial targets. Revenue and profit are calculated automatically
based on adjustments to the volumes.
The scope of financials in supply chains is to provide directional finance and
valuation for supply planning decisions. These include:
• Financial projections (valuation of the plans) that convert volumes to values and
report at different aggregation levels
• Currency conversions including simulations on price and exchange rates changes
• Rolling revenue and gross margin projections
• Planned price and costs as input at different aggregate levels than supply chain
plans.
• Budget plans or annual operating plans as financial targets to drive operation
plans
• Calculations of price based on revenues and volumes
• Weighted average of price and costs calculations
• Basic cost rollups for cost-to-serve calculations
The overall process for finance and supply chain integration is shown in Fig. 40
across tactical annual budget planning and operational rolling monthly revisions of
the plan. This figure is adapted from SAP’s positioning of collaborative enterprise
planning between finance and supply chain solutions.
Annual Financial Planning
1. Tactical annual financial planning is performed in financial planning systems for
the next fiscal year combining inputs from strategic planning, prior year plans
and operations plans from a supply chain that consolidates demand and supply
plans for the next 1–3 years.
Intelligent Digital Supply Chains 63
2. Supply chain systems send the operational plan to financial planning systems as a
starting point for granular financial planning (e.g., detailed annual budget—sales
units by region, cost assumptions and investments in capacity) that integrates the
operational plan.
3. The resultant annual operating plan is sent to supply chain systems as a financial
target for consensus and constrained demand planning to evaluate scenarios that
bridge the gap to meet financial targets.
Operational Rolling Financial Planning
1. The annual operating plan is sent from financial planning to supply chain
planning systems once a year. Iterations occur monthly in the supply chain
until a consensus is reached on a single plan considering the financial budget
and constrained supply plans that allocate the critical resources, inventory and
materials most efficiently to satisfy customer demands while maintaining high
profit margins and working capital.
2. The process in supply chain starts with arriving at a consensus demand plan
in S&OP based on sales, marketing, demand and financial plans followed by
constraining the plan based on profit optimization considering supply chain
constraints and costs, along with business assumptions, risks and opportunities
(Kusters et al. 2018).
64 R. Jandhyala
3. Every stage of the process in the supply chain also includes financial projections
like revenue, margin, costs for the demand and supply plans to see both the
volume and value for the plans.
4. Finance and operations collaborate monthly or quarterly on variance analysis
reporting (plan vs. actuals) to reforecast and model changes in the business,
industry or economy.
5. The revised demand, supply and inventory quantities are sent to the financial
planning system for monthly/quarterly revision of the plan. This includes running
processes like sales revenue planning, cost and activity planning, product cost
planning, profitability planning, and P&L planning. The net result is an aligned
profit and loss, balance sheet, and cash flow statements with the supply chain
plan.
6. The results are revised financial targets that are sent to supply chain for visibility
into the latest P&L. The revised financial plans, i.e., Q2 revision, Q3 revision,
Q4 revision are the new financial targets for the supply chain.
References
APICS SCC. (2015). The supply chain operations reference model (SCOR® ).
Chopra, S., Meindl, P., & Kalra, D. V. (2013). Supply chain management: Strategy, planning, and
operation (Vol. 232). Boston, MA: Pearson.
Kepczynski, R., Ghita, A., Jandhyala, R., Sankaran, G., & Boyle, A. (2019a). Enable IBP with
SAP integrated business planning. In Implementing integrated business planning (pp. 23–110).
Cham: Springer.
Kepczynski, R., Dimofte, A., Jandhyala, R., Sankaran, G., & Boyle, A. (2019b). Implementing
integrated business planning (Management for Professionals). Cham: Springer.
Kusters, J., Jandhyala, R., Mane, P., & Sinha, A. (2018). Sales and Operations Planning (S&OP)
with SAP IBP (SAP PRESS).
Ravindran, A. R., & Warsing, D., Jr. (2016). Supply chain engineering: Models and applications.
Boca Raton, FL: CRC Press.
Product Life Cycle Optimization Model
for Closed Loop Supply Chain Network
Design
1 Introduction
Closed Loop Supply Chain (CLSC) management is defined as ‘the design, control
and operation of a system to maximize value creation over the entire life cycle
of a product with dynamic recovery of value from different types and volumes of
returns over time’ (Guide and Li 2010). According to the authors, a business process
perspective on CLSC consists of three subprocesses, namely:
(i) Front-end that deals with product return management, ensuring the availability
of products at required quality and quantity.
(ii) Engine that deals with operational issues in remanufacturing, ensuring profit
from remanufacturing.
(iii) Back-end that deals with market development of remanufactured products,
ensuring demand for recovered products.
OEM has to make two major strategic decisions at the Front-end of CLSC,
namely, (i) identify the most profitable channel of product recovery and (ii) identify
the profit maximizing remanufacturing market that OEM should participate in.
We explicitly model the optimal collection of remanufactured products through
suppliers in a dynamic manner across multiple time periods over the product
life, accounting for demand, quality and remanufacturability of returned prod-
ucts.
A. Dhamodharan ()
Tesla Motors, San Carlos, CA, USA
e-mail: adhamodharan@tesla.com
A. Ravi Ravindran
Pennsylvania State University, State College, PA, USA
e-mail: axr32@psu.edu
At the back-end of the CLSC, the OEM has to (i) make pricing decisions for
new and remanufactured products based on the consumer willingness to pay (WTP)
and (ii) estimate the demand for new and remanufactured products throughout
the product lifecycle. We use the extended Bass model (Debo et al. 2006) to
model total demand for the product, and cannibalization of new product demand
is reflected in pricing decisions of new and remanufactured products. We use
pricing decisions for new and remanufactured products based on Abbey et al.
(2015a, b).
We develop a multi-period CLSC network design planning model, incorporating
product life cycle for a supply chain consisting of three stages: (i) OEM manufactur-
ing facility, (ii) distribution centers, hybrid facilities and recovery centers and (iii)
retailers. At the strategic level, the OEM has fixed locations for manufacturing plants
and retailers, and potential locations for distribution centers, hybrid facilities and
recovery centers. In the forward flow, OEM manufactures new and remanufactured
products and distributes them through warehouses and hybrid facilities to satisfy
demand at the retailers in fixed locations. In the reverse flow, retailers collect
commercial returns, end of use returns and end of life returns from consumers for a
collection cost, provided the return has a certain level of functionality. These returns
are then sent to hybrid facilities and recovery centers, where they are disassembled,
inspected and stored. The OEM’s manufacturing facilities pull these returns from
the hybrid facilities based on the demand for remanufactured products. At the
tactical level, we consider an uncapacitated model, where demand for new and
remanufactured products follow an extended Bass model (Debo et al. 2006). We
consider a segmented market with quality-sensitive and price-sensitive consumers
based on the empirical evidence from Abbey et al. (2015a). We assume that the
WTP functions for both quality-sensitive and price-sensitive consumers overlap.
We use monopolist pricing decisions for new products in all periods (Abbey et
al. 2015b). When making a pricing decision on remanufactured products, we use
monopolist pricing expression (Abbey et al. 2015b) only when there are enough
available returns to satisfy the demand. Under constrained availability of returns,
we use a myopic pricing decision, where the OEM maximizes profit from selling
remanufactured products for that period only. We also penalize any lost demand due
to shortage of returns.
In the literature review, we briefly discuss contributions to dynamics of product
lifecycle; consumer perception of new and remanufactured products, the demand
cannibalization for new products from remanufactured product sales and finally,
pricing decisions under remanufacturing. We introduce the product life cycle
optimization model and use a case-study to show that the OEM’s approach to supply
chain network design can significantly impact the profitability in the reverse supply
chain over the product lifecycle. Finally, we present the conclusions from sensitivity
analysis on various parameters in the model.
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 67
2 Literature Review
We use diffusion based models to study product lifecycle dynamics. Bass (2004)
has developed a growth model to forecast demand for consumer products, called
the Bass model. The Bass model is based on the behavioral assumption that the
probability of a new purchase of a product at any time depends on the number
of previous purchases of the product. The model is empirically tested and found
to accurately predict the highest number of sales in a period and its timing, for
various consumer durables, predominantly between 1950 and 1960. Debo et al.
(2006) are the first to consider product life cycle dynamics in a closed-loop setting.
The authors extend the Bass diffusion model to: (i) study life cycle demand of
new and remanufactured products and (ii) analyze investments in flexible capacity.
The authors find that slow diffusing products are best suited for remanufacturing,
while fast diffusing products and products with repeat sales, make investment in
flexible capacity valuable. They also comment that the right time for the OEM to
introduce remanufactured products is after the sales for the new product has peaked.
In addition to the closed-loop setting and the factors studied in this work, our model
additionally considers the evolution of product pricing over the product lifecycle,
since profit is a function of pricing and demand.
There are very few studies in the literature that explicitly consider supply chain
decisions, including pricing, across the life cycle of a product, in a closed loop
setting. To the best of our knowledge, Chen and Chang (2013) is the only study that
analyzes pricing decisions across the life cycle of a product, but does not model the
product demand. Papers that analyze specific CLSC problems, such as, Savaskan et
al. (2004), Ferrer and Swaminathan (2010), Toktay and Wei (2011), consider one
or two-period product life cycle to keep their models tractable. On the other hand,
papers analyzing broader CLSC problems, such as, Jayaraman (2006), Min et al.
(2006), Pazhani and Ravindran (2014), do not model demand from the product life
cycle view.
Ovchinnikov (2011) shows empirical evidence that, because of the lower price of the
remanufactured product, a new market segment is ready to purchase remanufactured
products. Ovchinnikov (2011) also provides empirical evidence that, as the price
of remanufactured product increases, shifting of consumers from new product
purchase to remanufactured product purchase follows an inverted U shape. Pang
et al. (2015) analyze data on purchases of electronic products available in eBay,
UK. The authors determine seller identity and reputation, length of warranty period,
supply of remanufactured products, proxy demand, duration of sale and end day
68 A. Dhamodharan and A. Ravi Ravindran
of sale as significant factors that determine the price differential between new
and remanufactured products. While the studies above characterize customer WTP,
they lack extensive empirical evidence. In our work, we use the market structure
proposed by Abbey et al. (2015a, b), who conduct multiple studies to understand
consumer perceptions of remanufactured products, including technological, house-
hold and personal products. The authors find that product quality is significantly
more important than the discount offered in all the studies. They also find that
brand equity has insignificant impact and educating consumers on remanufacturing
process does not help in alleviating their negative perceptions. Based on their
extensive experimental results, the authors conclude that consumers fall into one
of the two categories:
(i) A new product only segment, where the consumer always has a significantly
higher preference for a new product. Even under extreme discounts for reman-
ufactured products, the consumer prefers only new products. For all future uses,
we refer to this consumer segment as Quality-sensitive consumers.
(ii) An indifferent segment, where the consumers are indifferent between new and
remanufactured products. When both new and remanufactured products are
presented at the same discount level, the consumer segment only shows a slight
preference for new products. For all future uses, we refer to this consumer
segment as price-sensitive consumers.
We follow the consumer behavior model of Abbey et al. (2015b), overlaying it
on the demand model, and explicitly model the OEM pricing decisions over the
product lifecycle.
Majumder and Groenevelt (2001) initiated a stream of papers on the pricing strategy
for remanufacturing. The majority of the articles begin with the assumption of
a theoretical model for consumer behavior (Abbey et al. 2015b). Hammond and
Beullens (2007) are the first to extend the pricing model to CLSC. The authors
consider oligopolistic CLSC under the Waste Electrical and Electronic Equipment
(WEEE) directive. They derive the Nash equilibrium for the supply chain and find
that the legislation encourages reverse supply chain activities. Abbey et al. (2015b)
compute the optimal pricing strategy based on an empirically observed market
structure. The authors initially derive optimal pricing expressions of a two-period
model, where the OEM sells only new products in the first period and sells both
new and remanufactured products in the second period. The main assumption in the
initial model is no constraint on the availability of product returns. The authors then
impose this constraint and propose an algorithm to compute the optimal pricing
strategy. We use both the market structure and the optimal pricing strategy given
by Abbey et al. (2015b). We observe that OEM’s pricing strategy influences the
sales of new and remanufactured products. Also, the quantity of new products sold
impacts the quantity of returns. Hence, we analyze the OEM’s pricing strategy at
the network design level.
3 Methodology
Distribution
Centers
(DC)
Recovery
Centers
(RC)
Fig. 1 Structure of CLSC network examined in PLCOM. Solid arrows represent the forward flow
of new and remanufactured products from the manufacturer to consumers. Dotted arrows represent
the backward flow of returns from consumers to the manufacturer
model (Appendix 2) that reflects product life cycle parameters. The optimization
model decides the optimal demand fulfillment and the optimal selling price based
on the pricing model (Appendix 1). The output of the optimization model gives the
optimal network design and the optimal distribution plan that maximizes the profit
over the product life cycle.
Sets:
N: planning horizon of the optimization model, t = 1,2,..,N
MP: set of all manufacturing plants in the network
DC: set of all potential distribution centers
HF: set of all potential hybrid facilities
RC: set of all potential recovery centers
R: set of all retailers in the network
Parameters used from integrated pricing and demand model (Appendix 3):
market size: the total size of the market considered. marketsize is normalized to 1
innov: innovation coefficient of the product that creates new demand. The coefficient
is expressed as a fraction of the total market size
imitation: imitation coefficient that creates new demand through word-of-mouth
effect from existing customers. This coefficient is also expressed as a fraction
of the total market size
marketexpt : market size expanded in period t, expressed as a fraction of total market
size
markett : market size at the end of period t, expressed as a fraction of total market
size
ϕ: the proportion of returns that can be remanufactured (quality of consumer returns)
residencej : the fraction of current product owners, markett , whose new product will
fail at period j
repeatt : demand created from repeat purchase by current product owners in period
t. This unit is a fraction of markett
L: product lifetime
newprdownt : number of consumers who own the new product in period t
returnst : number of consumer returns of new product in period t. Note that we do
not consider returns of remanufactured products. The consumer return may be a
commercial return or end of life return
Qnewt : optimal production quantity for the new product in period t
Qremt : optimal production quantity for the remanufactured product in period t
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 73
λ: the fraction of consumers that only buy a new product, referred to as Quality-
sensitive consumers
1-λ: the fraction of consumers that only buy a cheaper product, referred to as Price-
sensitive consumers
α 1 + β 1 θ 1 : willingness to pay function for quality-sensitive consumers
α 1 : intercept of willingness to pay function of quality-sensitive consumer
β 1 : the slope of willingness to pay function of quality-sensitive consumer
θ 1 : variable between 0 and 1 to give cumulative demand for new product
α 2 + β 2 θ 2 : willingness to pay function for price-sensitive consumers
α 2 : intercept of willingness to pay function of price-sensitive consumer segment
β 2 : slope of willingness to pay function of price-sensitive consumer segment
θ 2 : variable that varies between 0 and 1 to give cumulative demand for new product
innov: innovation coefficient of the product
imitation: imitation coefficient of the product
ϕ: the fraction of consumers that repeat purchase of a new product
remtime: time to remanufacture a product return
repcust: the fraction of current product owners making a repeat purchase (treated as
a constant)
cn : manufacturing cost for new product
intro: time period at which remanufactured product is introduced
Cost components:
cn : manufacturing cost for new product
pn : selling price for new product
pr : selling price for remanufactured product
cn : manufacturing cost of a new product
cr : remanufacturing cost
dispcost: unit cost of disposing a product
hc : unit cost of holding a returned product in inventory per period
hr : unit cost of holding a remanufactured product in inventory per period
sc: cost of not fulfilling a demand per unit
fw : fixed cost of opening a DC
fh : fixed cost of opening a HF
fr : fixed cost of opening a RC
tf 1mp,dc,s: unit transportation cost from a manufacturing plant mp, to retailer s,
through a distribution center dc
tf 2mp,hf,s: unit transportation cost from a manufacturing plant mp, to retailer s,
through a hybrid facility hf
tb1s,rc,mp: unit transportation cost from a retailer s, to a manufacturing plant mp,
through a recovery center rc
tb2s,hf,mp: unit transportation cost from a retailer s, to a manufacturing plant mp,
through a hybrid facility hf
74 A. Dhamodharan and A. Ravi Ravindran
Capacity components:
cap1mp: maximum capacity of manufacturing plant mp
cap2dc: maximum capacity of distribution center dc
cap3hf : maximum capacity of hybrid facility hf
cap4rc: maximum capacity of recovery center rc
Demand at retailers:
ωs : fraction of the total demand at retailers
γ s : fraction of consumer returns at retailer s
Qnews,t : demand for the new product at retailer s during period t, and is given by:
Qnews,t = (Qnewt ) ωs ;
where
∗
For t < intro, Qnewt = qn1 marketdemandt
∗
For t ≥ intro, Qnewt = qn2 marketdemandt
Qrems,t : demand for the remanufactured product at retailer s and is given by:
Qrems,t = Qremt ωs
where Qremt = qr∗ marketdemandt
Decision variables:
Iremt : inventory of remanufactured product at the beginning of a period
Irett : inventory of returned product at the beginning of period t
remcollecs,t : maximum number of product returns available at retailer s in period t
zt : binary variable that takes a value of 1 if there is a shortage of remanufactured
product in period t
shortaget : quantity of shortage of remanufactured product in period t
OpenDCdc : binary variable that takes a value of 1 if DC at d is open and 0 otherwise
OpenHFhf : binary variable that takes a value of 1 if HF at hf is open and 0 otherwise
OpenRCrc: binary variable that takes a value of 1 if RC at rc is open and 0 otherwise
Qnewtr1mp,dc,s,t: quantity of new products, transported from manufacturing plant
mp, through distribution center dc, to retailer s during period t
Qnewtr2mp,hf,s,t: quantity of new products, transported from manufacturing plant
mp, through hybrid facility hf, to retailer s during period t
Qremtr1mp,dc,s,t: quantity of remanufactured products, transported from manufac-
turing plant mp, through distribution center dc, to retailer s during period t
Qremtr2mp,hf,s,t: quantity of remanufactured products, transported from manufactur-
ing plant mp, through hybrid facility hf, to retailer s during period t
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 75
N
(pn − cn ) Qnews,t (1)
t =1 sεR
(i) For t < intro, we use Eq. (2) for pn and Eq. (3) for Qnewt , given below.
∗ cn 1 λβ2 (α1 + β1 ) + (1 − λ) β1 (α2 + β2 )
p =
n
pn1 = + (2)
2 2 λβ2 + (1 − λ) β1
∗
Qnewt = qn1 marketdemandt (3)
α1 + β1 − pn α2 + β2 − pn
Qnewt = λ + (1 − λ) marketdemandt
β1 β2
(4)
(ii) For t ≥ intro, we use Eq. (6) for pn and Eq. (7) for Qnewt , as shown below.
∗ cn α1 + β1
pn = pn2 = max α1 , + (6)
2 2
∗ (marketdemand ) λ =
Qnew t= qn2 t
α1 +β1 −cn (7)
λ2 min 2β1 ,1 marketdemand t
N
(pr − cr ) Qrems,t − shortaget (8)
t =1 sεR
76 A. Dhamodharan and A. Ravi Ravindran
where:
cr α1 + β1
pr = max α2 , + (9)
2 2
α2 + β2 − cr
Qremt = λ (1 − λ) min , 1 marketdemandt (10)
2β2
N
hr I remt + hc I rett (12)
t =1
N
sc shortaget (13)
t =1
N
(1 − ϕ) returnst dispcost (14)
t =1
• Transportation cost (TRANS):
N
Qnewtr1mp,dc,s,t tf mp,dc,s +
t =1 mp ε MP dc ε DC s ε R
N
Qnewtr1mp,hf,s,t tf mp,hf,s +
t =1 mp ε MP hf ε H F s ε R
N
Qnewtr1mp,dc,s,t tf mp,dc,s +
t =1 mp ε MP dc ε DC s ε R
N (15)
Qnewtr1mp,hf,s,t tf mp,hf,s +
t =1 mp ε MP hf ε H F s ε R
N
Qnewtr1s,rc,mp,t tf s,rc,mp +
t =1 s ε R rc ε RC mp ε MP
N
Qnewtr1s,hf,mp,t tf s,hf,mp
t =1 s ε R hf ε H F mp ε MP
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 77
4. Capacity constraint for each DC where the sum of all products transported to a
DC in every period should not exceed its capacity, if the DC is built.
Qnewtr1mp,dc,s,t + Qremtr1mp,dc,s,t
mp ε MP dc ε DC (20)
≤ capacitydc OpenDCdc ∀dc ε DC, t ε T
78 A. Dhamodharan and A. Ravi Ravindran
5. Capacity constraint for each HF where the sum of all products transported to a
HF in every period should not exceed its capacity, if the HF is built.
Qnewtr2mp,hf,s,t + Qremtr2mp,hf,s,t + Qrettr2s,hf,mp,t
mp ε MP s ε R
≤ capacityhf OpenH F hf ∀hf ε H F, t ε T
(21)
6. Capacity constraint for each RC, where the sum of all consumer returns
transported to a RC in every period should not exceed its capacity, if the RC
is built.
Qrettr1mp,rc,s,t + ≤ capacityrc OpenRC rc ∀rc ε RC, t ε T
mp ε MP s ε R
(22)
7. Constraint to update the inventory of consumer returns. Until the remanufactured
product is introduced in the market, there is no production of remanufactured
product. Hence, the inventory of consumer returns keeps adding up with the
collection of returns.
I rett = I rett −1 + remcollecs,t −1 ∀t < intro&t ε T (23)
sεR
I remt = I remt −1 +
⎛ Qremt −1 − shortaget −1 ⎞
− ⎝ Qremtr1mp,dc,s,t −1 + Qremtr2mp,hf,s,t −1 ⎠ ∀t
mp ε MP s ε R dc ε DC hf ε H F
(26)
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 79
shortaget ≤ M zt ∀t (27)
OpenDCdc ε {0, 1} ∀dc, Open H Fhf ε {0, 1} ∀hf, OpenRCrc ε {0, 1} ∀rc
(32)
zt ε {0, 1} ∀t (33)
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
Time Periods
Residence time distribution
function for residence time, as shown in Fig. 3. We also assume that remanufactured
products are introduced into the market in period 25.
Based on the lifecycle assumptions above and considering iPhones to be a
fast-moving product, we arrive at the innovative coefficient, innov = 0.0005 and
imitation coefficient, imitation = 0.000025, so that the sales in the initial period of
the lifecycle are significantly higher than the sales at the latter stage and the market
consumes the product at least once within the product lifetime. We assume a realistic
market structure with two groups of consumers: (i) quality-sensitive consumers with
WTP function 550 + 250θ and (ii) price sensitive consumers with WTP function
450 + 200θ . We obtain the WTP function based on the iPhone market prices at
different retailers. For quality-sensitive customers, we model the WTP function
based on the iPhone7 (32 GB)‘s market price of $699, from the retailers like
Amazon and Walmart. For price-sensitive customers, we model the WTP function
based on the pre-owned iPhone6 Plus (16 GB) market price of $540, from the
retailer Bestbuy. We assume the new product manufacturing cost to be $400 and the
remanufacturing cost to be $100 (total price of all the components inside an iPhone
is roughly estimated at $220). We assume 60% of the market is of quality-sensitive
consumers (λ = 0.6) and 80% of the customers make a repeat purchase. We also
assume all the collected returns are of good quality and can be remanufactured. We
consider a CLSC network with 3 manufacturing plants, 3 retailers and 10 potential
locations for distribution centers (DCs), recovery centers (RCs) and hybrid facilities
(HFs), each with a capacity of processing 100 items. Recall that DCs handle only
forward flows, RCs handle only reverse flows and HFs handle both forward and
reverse flows. We assume the fixed cost of opening a DC and a RC to be $500,000,
and the fixed cost of opening a HF to be $550,000. We assume per period inventory
holding cost for a unit of new product, returned product and remanufactured product
to be $5. Table 1 shows the proportion of market demand and product returns at each
retailer. Appendix 4 lists unit transportation cost within the CLSC network.
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 81
Problem Size
The maximum profit obtained by the CPLEX solver was $1,277,540. The optimal
network structure uses 3 DCs (locations 1, 7 and 10) and 1 HF (location 10). No
RC is built. Both a DC and a HF are built at location 10 since transportation cost to
retailer 2 from location 10 is minimum. Retailer 2 has the maximum proportion of
demand and the maximum proportion of returns (see Table 1). The optimal network
flow for new products, remanufactured products and consumer returns are shown in
Figs. 4, 5 and 6.
In Fig. 4, we observe that the DC at location 1 serves new products to retailer
1, the DC at location 10 serves new products to retailer 2 and the DC at location
7 serves new products to retailer 3. In Fig. 5 we observe that DC at location 1
serves remanufactured product to retailer 1 and retailer 3, the DC and the HF at
Since we limit the number of times a customer can repeat the purchase of a
product to one, we see that new product sales has a bimodal shape. The first peak
in the new product sales curve is caused by the first-time customers, both quality-
sensitive and price-sensitive consumers. The second peak is caused by the quality-
sensitive repeat consumers. Hence, the cumulative new product sales volume before
the new product sales peak is lower than the cumulative new product sales volume
after the peak. If the spread of the residence time is narrower, we can see a more
pronounced second peak due to repeat purchase, as discussed in Debo et al. (2006).
The remanufactured product sales have a bell-shaped curve with a sharp increase to
the peak, followed by a smooth decrease. The sharp increase to the peak is caused
by the price-sensitive first-time consumers. The smooth decrease is due to the price-
sensitive repeat consumers who replaced their first-time purchase of a new product
with a remanufactured product.
Figure 8 shows the evolution of product sales over the product life cycle with
respect to first-time sales and repeat sales. We see the decrease in first-time sales at
period 25, which can be attributed to the introduction of the remanufactured product
into the market. We also observe that the peak of first-time sales occurs in period
43, and the peak of repeat sales occurs in period 68. The gap of 25 time periods
between the peaks is explained by the unimodal shape of the residence function,
with its maximum value at 25 periods after the purchase (see Fig. 3). The repeat
sales is more skewed than the first time sales due to the delayed introduction of the
remanufactured product.
Figure 9 shows the availability of returns to remanufacture and shortages in
remanufactured products. In period 25, we see a sharp increase in shortages as
84 A. Dhamodharan and A. Ravi Ravindran
Fig. 8 Sales across product life cycle: first-time sales and repeat sales
160
140
120
100
sufficient returns are not available in the market to satisfy the demand for the
remanufactured product. However, from period 57 onwards, we see no shortages
of remanufactured products.
Table 2 Comparison of optimal network designs and profit between integrated and sequential
approaches
Factors Integrated CLSC Sequential CLSC
Optimal network design DC:1,7,10; HF:10 DC:1,7,9,10; RC:7
Profit during product lifecycle $1,277,540 $923,222
In this subsection, we conduct a sensitivity analysis on the parameter intro, the time
period when the remanufactured product is introduced into the market, using the
illustrative example. The value of the parameter intro is varied from 25 to 75, and
the PLCOM is solved in an integrated manner to study its impact on the following:
• network design,
• profit,
• shortage,
• new product sales,
• remanufactured product sales,
• first time sales and
• repeat sales.
Impact on Profits
Figure 10 plots optimal profit and shortage of remanufactured products, over the
product lifecycle, against the intro values from 25 to 75. The optimal profit increases
as the intro value increases to 45, then decreases smoothly until the intro value of
55. From the intro values of 60 onwards, we see a sharp decrease in the optimal
profit. We conclude that 40 is the most profitable time period to introduce the
remanufactured product into the market. Time period 60 and beyond is too late for
the OEM to realize maximum profits from the remanufacturing market.
Impact on Shortages
Also note that in Fig. 10, the total shortage over the product life cycle is 3500 for
the intro value of 25. As the intro value increases, that is, when the remanufactured
products enter the market later, there is a decrease in shortages because there are
more product returns available for remanufacturing. Interestingly, the minimum
shortage value of 0 occurs for intro values of 55 or more. The maximum profit
scenario with intro value of 40 has a shortage of 744 remanufactured units.
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 87
Figure 11 plots new product sales, remanufactured product sales and total sales
over the product lifecycle against the intro values from 25 to 75. We observe that
both remanufactured product sales and the total sales decrease in a linear fashion
with higher values of intro. This is explained by the fact that as the value of
Fig. 11 New and remanufactured product sales over product life cycle
88 A. Dhamodharan and A. Ravi Ravindran
Fig. 12 First time and repeat sales over product life cycle
intro increases, the time periods left for the price-sensitive consumers to purchase
the remanufactured products reduce. On the other hand, we see an increase in
new product sales for higher values of intro because the new product demand is
cannibalized after intro.
Figure 12 plots first time sales repeat sales and total sales, over the product lifecycle
against the intro values from 25 to 75. We observe that first-time sales decreases
with an increase in intro value, while the repeat purchase remains constant. First-
time sales made by the price-sensitive and the quality-sensitive consumers decreases
with an increase in intro value since price-sensitive consumers do not have enough
time to make first-time purchases of remanufactured products. There are only a
fixed number of quality-sensitive consumers in the market. The model restricts the
number of repeat purchases to one. Hence, repeat purchases are flat and independent
of intro value.
Table 3 shows the optimal CLSC network design for different values of intro. It is
interesting to note that the DCs and the HFs chosen are the same irrespective of the
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 89
intro value. As intro value is increased from 25 to 30, 35, 40 and 45, we find that it is
profitable for the CLSC network to operate an additional RC in order to reduce the
shortage of the remanufactured products. For intro value of 50, in spite of a shortage
of 534 units of the remanufactured products (see Fig. 11), it is not profitable to
operate the additional RC as the fixed cost of the facility exceeds the profit made
from the sales of the 534 units. For intro value of 55 onwards, it is not profitable to
operate the additional RC due to the reduced volume of the remanufactured sales.
In summary, for the case study, the OEM can achieve a maximum profit of
$1,514,000 by introducing the remanufactured product at time period 40. The
optimal CLSC network will have:
• DCs at locations 1, 7 and 10
• an RC at location 10
• a HF at location 10
In other words, just at location 10, the OEM will have a DC, RC and HF.
The reason behind choosing location 10 is the minimum transportation cost to
retailer 2, who faces the maximum proportion of demand and collects the maximum
proportion of returns (see Table 1).
periods are considered slow diffusing products. We solve the illustrative example
using the PLCOM for lifecycle values of 100 periods, 80 periods, 60 periods and
40 periods. The products with lifecycle values of 40 and 60 periods are considered
as fast diffusing products, while the products with lifecycle values of 80 and 100
periods are considered as slow diffusing products. Note that in all the 4 cases the
total market size is maintained at 10,000.
Table 4 compares the profit and the optimal network designs for the above-
mentioned values of the product lifecycle. We find that the fast diffusing product
with a lifecycle value of 40 periods results in a loss to the OEM, as it requires
extensive investment to open DCs at 6 different locations and HF at location 10,
in order to handle the high volume of product flow in each period. Hence, when
designing a supply chain for fast diffusing products, we recommend the OEM to
plan for flexible capacity, so that the fixed cost of opening facilities is minimal, as
noted by Debo et al. (2006). In the case of a fast diffusing product with a lifecycle
value of 60 periods, the OEM is able to make a profit by opening DCs at 3 different
locations and HF at location 10. The profit is significantly lower than the profit made
from slow diffusing products due to a high shortage of consumer returns. In the case
of slow diffusing products, we notice that the profit is higher for a product with a
longer lifecycle.
Under a fixed capacity setting, as the product lifecycle grows longer, the demand
is more evenly distributed. Thus, we expect the number of facilities in the supply
chain network to remain about the same. From Table 4, we notice that the number of
facilities in the network for the lifecycle value of 60 periods is lower than the number
of facilities for the lifecycle value of 40 periods. Similarly, the number of facilities
in the network for the lifecycle value of 80 periods is lower than the number of
facilities for the lifecycle value of 40 periods. Finally, the number of facilities in the
network for the lifecycle value of 100 periods is lower than the number of facilities
for the lifecycle value of 40 periods and 80 periods and is the same as the number
of facilities in the network for the lifecycle value of 60 periods. Interestingly, we
observe that the number of facilities in the network for the lifecycle value of 80 has
an additional facility compared to the network for the lifecycle value of 60. The
additional facility is to satisfy the demand for the remanufactured products from
cannibalization, explained later in this section.
Table 4 Profit and optimal network design for slow and fast moving products
Category Lifecycle Profit DC HF RC
Fast diffusing 40 −$1,853,000 1,2,7,8,9,10 10 –
Fast diffusing 60 $390,900 1,7,10 10 –
Slow diffusing 80 $1,035,000 1,7,10 10 7
Slow diffusing 100 $1,277,540 1,7,10 10 –
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 91
Figure 13 plots new product sales and remanufactured product sales over the
product lifecycle for different lifecycle values. We notice that as the product
lifecycle shortens, the peaks from the first-time sales and the repeat sales are more
pronounced. In the case of a fast diffusing product with a lifecycle value of 40
periods, we see a big drop in the new product sales when the remanufactured product
is introduced. This indicates a big proportion of new product consumers switching
from the new product to buy the remanufactured product. The revenue loss from
the demand cannibalization also results in a net loss to the OEM, as shown in Table
4. Thus, in this case, it is clearly not profitable for the OEM to participate in the
remanufacturing market. In the case of fast diffusing products with a lifecycle value
of 60 periods, we notice that the demand cannibalization is significantly lower, and
the OEM makes a net profit. In order to analyze whether it is profitable for the
OEM to participate in the remanufactured market, we solve only for the forward
supply chain and find that the OEM would incur a net loss operating only in the
forward supply chain, due to fixed capacity. Hence, we conclude that in this case, it
is profitable for the OEM to participate in the remanufacturing market.
Fig. 13 New and remanufactured product sales for different values of product lifecycle
92 A. Dhamodharan and A. Ravi Ravindran
Figure 14 plots first-time sales, repeat sales and shortage over the product lifecycle,
for different lifecycle values. We notice that, as the product lifecycle grows shorter,
there is an increase in the amplitudes of the first-time sales, repeat sales and the
shortage functions. We also find that the total first-time sales increases and the total
repeat sales decreases, due to lack of time for the product to fail. Initially, the total
shortage increases due to the shortage of consumer returns. In the case of lifecycle
value of 60 periods, we notice a second peak in shortage, coinciding with the peak
of repeat sales. This second peak is due to increased demand for remanufactured
products from the delayed cannibalization effect. The delayedcannibalization effect
occurs when the price-sensitive consumers, who initially bought the new product,
before the introduction of the remanufactured product, switch to buy the remanu-
factured product when making their repeat purchase. In the illustrated example, the
delayed cannibalization effect is expected to occur from periods 38 to 61, as shown
in Fig. 14. As the product lifecycle grows longer, the number of consumers causing
the delayed cannibalization effect is smaller and hence more evenly distributed,
providing time to collect consumer returns and satisfy the increased demand for
the remanufactured product. Thus, in the case of the lifecycle value of 80 periods,
we observe the second shortage peak with a significantly lower amplitude, and in
the case of the lifecycle value of 100 periods, there is no second shortage peak. This
also explains why the optimal network design for the lifecycle value of 80 periods
has an additional RC at location 7 compared to the lifecycle value of 60 periods.
Fig. 14 First-time sales, repeat sales and shortage for different values of product lifecycle
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 93
The additional RC is used to collect more returns and minimize the shortage, as
illustrated in Fig. 14. However, the additional RC is not required in the case of the
lifecycle value of 100 periods, because of the very small increase in the demand for
remanufactured products from the delayed cannibalization.
In summary, we conclude the following from the sensitivity analysis of the
imitation coefficient:
• Slow diffusing products are more profitable for remanufacturing compared to fast
diffusing products because there is enough time to collect consumer returns
• OEMs dealing with fast diffusing products should plan for flexibility in capacity
in order to avoid the fixed cost of opening additional facilities
• Managers should understand the effect of delayed cannibalization when design-
ing the reverse channel and planning for the collection of consumer returns
Table 5 Profit and optimal network design for different values of the quality of consumer returns
Quality of consumer returns Profit DC HF RC
60% $960,000 7,10 1 7
70% $1,190,000 1,7,10 – 10
80% $1,277,540 1,7,10 10 –
90% $1,454,000 1,7,10 10 1
94 A. Dhamodharan and A. Ravi Ravindran
It is in the interest of the OEMs to collect consumer returns of high quality. Thus,
the OEMs can provide incentives to consumers to encourage high-quality returns.
Such a practice is followed in the smartphone industry, where the consumer gets
a rebate when buying a new smartphone for returning a good quality smartphone.
The difference in profits for the various quality values (Table 5) provides an upper
bound on the incentives that can be provided by the OEM, without incurring loss.
For example, if the OEM is collecting consumer returns of 60% quality, then the
OEM is justified to spend up to $229,700 (the difference between $1,1,90,000 and
$960,300) in consumer incentives, to ensure a 70% quality of consumer returns.
Figure 15 plots the new product sales and the remanufactured product sales over
the product lifecycle, for different values of the quality of consumer returns. We
notice that, as the quality of the consumer returns increases, the second peak of the
new product sales and the remanufactured product sales increase due to an increase
in repeat purchase. We also find that as the quality of the consumer return increases,
the total new product sales and the total remanufactured product sales increase in a
linear fashion.
Figure 16 plots the first-time sales, the repeat sales and the shortage over the
product lifecycle, for different values of quality of consumer returns. We notice that
as the quality of the consumer returns increases, the total first-time sales and the
total repeat sales increase. However, as the quality of consumer returns increases,
the total shortages do not monotonically increase. This can be explained by the
following example: When the quality of consumer return value increases from 60%
from 70%, the total shortage increases from 2743 units to 3079 units. The profit in
the case of 70% quality of consumer return is still higher than the profit in the case
Fig. 15 New and remanufactured product sales for different values of the quality of consumer
returns
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 95
Fig. 16 First-time sales, repeat sales and shortage for different values of the quality of consumer
returns
of 60% quality of consumer returns (see Table 5) due to the cost savings from the
network design. Note that the optimal network design for 60% quality of consumer
returns opened a HF at location
1, which costs $50,000 more than the DC at location 1 opened in the optimal
network design for 70% quality of consumer returns (see Table 5). Since a HF can
handle both products and returns but a DC can only handle products, we notice an
increase in shortage in the case of 70% quality of consumer returns.
In summary, we conclude the following from the sensitivity analysis on the
quality of the consumer returns:
• Higher quality of consumer returns provide higher profit for the OEMs
• The analysis provides OEMs insights on how much to spend on consumer
incentives in order to improve the quality of the consumer returns, without
incurring a net loss from the incentives
Total sales usually increase with an increase in the quality of consumer return
• Total shortages do not always decrease with an increase in the quality of
consumer returns. Sometimes it may be optimal to have a higher shortage due
to the high fixed cost of opening an additional facility.
96 A. Dhamodharan and A. Ravi Ravindran
4 Conclusions
In this chapter, we began by introducing the PLCOM and illustrated the evolution
of new and remanufactured product sales, first-time and repeat sales and shortage
of remanufactured products over the product lifecycle using a case study. The case
study also quantified the cannibalization of the new product demand, a phenomenon
that deters the OEMs from participating in the market for the remanufactured
products. We continued with the same example to illustrate that an integrated
approach to design a CLSC network is more profitable for the OEM to participate
in the reverse supply chain.
Sensitivity analysis on the time period to introduce the remanufactured product in
the market identified the optimal time period maximizing the profit and showed the
tradeoff between the shortage of remanufactured products and the new product sales.
Sensitivity analysis on imitation coefficient showed that slow diffusing products
are more profitable in the reverse channel compared to fast diffusing products. We
also identified the delayed cannibalization phenomenon in case of fast diffusing
products caused by the price-sensitive consumers switching to the remanufactured
product during their repeat purchase. Finally, the sensitivity analysis on the quality
of consumer returns provided insights into how much the OEM can spend on
incentives to improve the quality of consumer returns, without incurring a net loss
from the incentives.
Appendices
Appendix 1
Pricing Model
Quality-sensitive consumers will only buy new products and are not motivated
by price discounts. The willingness to pay (WTP) function for quality-sensitive
customer segment is modeled as a linear function given by:
f (θ ) = α1 + β1 θ (34)
where θ in (0,1) and (1- θ) is the fraction of the customer segment with
WTP = α1 + β1 θ.
Price-sensitive consumers are indifferent to whether the product is new or
remanufactured and are purely motivated by the price of the product. Hence, there
is no need to model new and remanufactured products separately. The willingness
to pay (WTP) function for this customer segment is modeled as a linear function
given by:
f (θ ) = α2 + β2 θ (35)
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 97
where θ in (0,1) and (1- θ) is the fraction of the customer segment with
WTP = α2 + β2 θ, α2 is the maximum selling price at which all the customers in
this price-sensitive category are willing to buy the product (new or remanufactured).
The proportion of price-sensitive consumers, who are willing to buy
new/remanufactured product for price pr , denoted by qr , is given by:
pr − α2
qr = (1 − λ) 1 − (36)
β2
∗ cn 1 λ β2 (α1 + β1 ) + (1 − λ) β1 (α2 + β2 )
pn1 = + (37)
2 2 λ β2 + (1 − λ) β1
∗ λ β2 (α1 + β1 ) + (1 − λ) β1 (α2 + β2 )
qn1 = (38)
λ β2 + (1 − λ) β1
where:
cn : manufacturing cost of new product
cr : remanufacturing cost
λ: the fraction of quality-sensitive consumer segment
1- λ: the fraction of the market falling under price-sensitive consumer segment
In order to make use of the above equations, we assume that α2 < α1 ≤ α2 + β2 ,
i.e., there is overlapping in WTP between the two consumer categories. Also, pn1 ∗ ≥
∗
α1 and α2 ≤ pn1 ≤ α2 + β2 .
When the OEM introduces the remanufactured product into the market, assuming
no constraint on the production of remanufactured products, the optimal pricing pn∗ ,
pr∗ and the optimal production quantity qn∗ , qr∗ , expressed as a fraction of the market
demand for new products and remanufactured products, respectively, are given by
the following equations (Abbey et al. 2015a).
∗ cn α1 + β1
pn2 = max α1 , + (39)
2 2
∗ α1 + β1 − cn
qn2 = λ min ,1 (40)
2β1
∗ cr α2 + β2
pr2 = max α2 , + (41)
2 2
∗ α2 + β2 − cr
qr2 = (1 − λ) min ,1 (42)
2β2
product until the introduction of the remanufactured product. During those periods,
we use Eqs. (37) and (38) to decide the pricing and quantity of the new product
produced. Let intro be the time period when the OEM introduces the remanufactured
product into the market. During the periods after the introduction of remanufactured
products, assuming no capacity restriction on producing new products, we use
Eqs. (39) and (40) as the new price and the new production quantity for the
new product. For the remanufactured products, however, the optimization model
constraints the production quantities based on the quantity of consumer returns and
the remanufacturability of the consumer returns. Hence, if the availability of the
consumer returns in a period exceeds the demand for remanufactured products, we
use the Eqs. (41) and (42) to make the pricing and the production quantity decisions
for the remanufactured products. If the availability of the consumer returns in a
period is lower than the remanufactured product demand, then we assume that the
OEM sells all the available inventory of remanufactured products at pr∗ , and the
shortages are lost sales.
In order to make use of the Eqs. (39) through (42), we need to make the following
assumptions:
cn > cr . The manufacturing cost of the new product is greater than the remanufac-
turing cost
α1 > α2 . The minimum WTP for the new product is greater than the minimum WTP
for the remanufactured product
α1 + β1 > α2 + β2 . The maximum WTP for the new product is greater than the
maximum WTP for the remanufactured product
All the three assumptions are very practical. Under such assumptions, pn∗
> pr∗ . Hence, when remanufactured product is available in the market, price-
sensitive customers will not buy the new product. Thus, new product demand might
decrease due to the introduction of the remanufactured product, resulting in demand
cannibalization.
Appendix 2
Demand Model
Debo et al. (2006) extended the Bass model to study life cycle demand of new
and remanufactured products and analyzed the cost-effectiveness of investments in
capacity for remanufacturing, as explained in the literature review section. Debo et
al. (2006) assumed a specific OEM selling price function for both the new and the
remanufactured products, without considering the market structure. We adopt their
model to generate demand for both new and remanufactured products throughout
the product life cycle and then integrate it with the OEM’s pricing model, which is
described in the previous section. The demand model is described below.
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 99
Demand Equations
In the extended Bass model, the market size for new and remanufactured products
grows each period due to the appeal of the product to the market, denoted by the
innovation coefficient, and word-of-mouth from the existing customers, denoted by
the imitation coefficient. Market size expanded during period t is a fraction of the
uncaptured market (1-markett-1) that is open to buying the product, as given by the
Eq. (43).
Market demand for new and remanufactured products in period t, given by the
Eq. (44), is the sum of market expanded during period t and repeat consumers.
Market demand is split into new and remanufactured product demand using
the parameter λ, defined as the fraction of quality-sensitive customers in the
market. Hence, the new product demand is given by λ (marketdemandt) and the
remanufactured product demand is given by (1 − λ) (marketdemandt).
The market size of first-time product owners, given by Eq. (45) using both new
and remanufactured products at the end of period t, is updated as the sum of the
market size at the end of the previous period and the market size growth during
period t.
In the extended Bass model, the term residence time distribution is used to refer
to the probability distribution of the number of time periods for which the product
(new or remanufactured) stayed with a customer. Hence, the number of new product
owners at the beginning of period t is obtained by subtracting the number of new
product consumers who stopped owning the product at the end of the period (t-1)
and adding the number of consumers who bought the new product at the beginning
of period t, as given by the Eq. (46).
L
newprodownt = newprodownt −1 +Qnewt − residencej newprodownt −j −1
j =1
(46)
min(t,L)
repeatt = repcust residencej − marketexpt −j + repeatt −j (47)
j =1
L
returnst = residencej newprodownn,t −j (48)
j =1
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 101
Appendix 3
Integration of Pricing and Demand Models
In the multiperiod optimization model, we model the demand (both new and reman-
ufactured products) in each period based on the extended Bass model, given by Eqs.
(44) to (46). However, the production quantities of new and remanufactured prod-
∗ , q ∗ , q ∗ ), are expressed as a fraction of the market demand for new prod-
ucts (qn1 n2 r
ucts and remanufactured products, based on the pricing model, as given below:
1. During time periods when only the new products are produced (upto t < intro),
∗
we use Eq. (38) from Appendix 1 to determine qn1
2. During time periods when both new and remanufactured products are produced
t ≥ intro), we use Eq. (40) from Appendix 1 for determining qn2 ∗ , for new
products,
3. For remanufactured products, we use Eq. (42) from Appendix 1 for determining
qr∗ ,
The marketdemand, for new and remanufactured products in period t is given by
Eq. (44). Using marketdemand, the production quantities for new and remanufac-
tured products are determined as follows:
Let Qnewt be the production quantity of new product. Then,
∗ , we get:
(i) Using Eq. (38) from Appendix 1 for qn1
α1 + β1 − pn∗ α2 + β2 − pn∗
Qnewt = λ + (1 − λ) marketdemandt
1 − β1 1 − β2
(49)
for period t < intro, when only the new product is produced.
∗ , we get:
(ii) Using Eq. (40) from Appendix 1 for qn2
α1 + β1 − cn
Qnewt = λ min ,1 marketdemandt (50)
2β1
for period t, when both the new product and the remanufactured product are
produced.
(iii) Let Qremt be the weighted production quantity for the remanufactured product.
Qremt = qr∗ marketdemandt . Using Eq. (42) from Appendix 1 for qr∗ , we
get:
α2 + β2 − cr
Qremt = (1 − λ) min ,1 marketdemandt (51)
2β2
102 A. Dhamodharan and A. Ravi Ravindran
Appendix 4
References
Abbey, J. D., Meloy, M. G., Guide, V. D. R and Atalay, S. (2015aa), ‘Re-manufactured products
in closed-loop supply chains for consumer goods’, Production and Operations Management
24(3), 488–503
Abbey, J. D., Blackburn, J.D. and Guide, V.D.R. (2015bb) ‘Optimal pricing for optimal pricing for
new and remanufactured products’, Journal of Operations Management 24(3), 488–503.
Bass, F. M. (2004). Comments on “a new product growth for model consumer durables the bass
model”. Management Science, 50(12 supplement), 1833–1840.
Chen, J.-M., & Chang, C.-I. (2013). Dynamic pricing for new and remanufactured products in a
closed-loop supply chain. International Journal of Production Economics, 146(1), 153–160.
Debo, L. G., Toktay, L. B., & Wassenhove, L. N. V. (2006). Joint lifecycle dynamics of new and
remanufactured products. Production and Operations Management, 15(4), 498–513.
Ferrer, G., & Swaminathan, J. M. (2010). Managing new and differentiated remanufactured
products. European Journal of Operational Research, 203(2), 370–379.
Guide, V. D. R., Jr., & Li, J. (2010). The potential for cannibalization of new products sales by
remanufactured products. Decision Sciences, 41(3), 547–572.
Jayaraman, V. (2006). Production planning for closed-loop supply chains with product recovery
and reuse: An analytical approach. International Journal of Production Research, 44(5), 981–
998.
Hammond, D., & Beullens, P. (2007). Closed-loop supply chain network equilibrium under
legislation. European Journal of Operational Research, 183(2), 895–908.
Majumder, P., & Groenevelt, H. (2001). Competition in remanufacturing. Production and Opera-
tions Management, 10(2), 125–141.
Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 105
Min, H., Ko, C. S., & Ko, H. J. (2006). The spatial and temporal consolidation of returned products
in a closed-loop supply chain network. Computers & Industrial Engineering, 51(2), 309–320.
Ovchinnikov, A. (2011). Revenue and cost management for remanufactured products. Production
and Operations Management, 20(6), 824–840.
Pang, G., Casalin, F., Papagiannidis, S., Muyldermans, L., & Tse, Y. K. (2015). Price determinants
for remanufactured electronic products: A case study on ebay UK. International Journal of
Production Research, 53(2), 572–589.
Pazhani, S., & Ravindran, A. R. (2014). Design of closed loop supply chain networks. International
Journal of Business Analytics, 1(1), 43–66.
Savaskan, R. C., Bhattacharya, S., & Van Wassenhove, L. N. (2004). Closedloop supply chain
models with product remanufacturing. Management Science, 50(2), 239–252.
Toktay, L. B., & Wei, D. (2011). Cost allocation in manufacturing– Remanufacturing operations.
Production and Operations Management, 20(6), 841–847.
Supply Chain Risk Management in
Indian Manufacturing Industries:
An Empirical Study and a Fuzzy
Approach
The evolution of the free trade agreement during the past decade has facilitated
the movement of goods across the world (Moore and Moore, 2003). This enabled
companies to compete in the international markets with products produced for
the domestic markets without any trade barriers. Further, the liberalization and
the consequent economic reforms in India such as Foreign Direct Investment
(FDI) led to substantial investments in the manufacturing sector which not only
augmented as an opportunity for the revival of the Indian economy (Kumar, 2005)
but also led to increased competition. Globalization and competitive environments
necessitated companies to implement an aggressive and integrated enterprise-wide
approach towards risk management (Saeidi et al., 2019). Hence the implementation
of Supply Chain Risk Management, a critical strategic approach has become a
top priority. For managing and mitigating risk, effective SCRM systems con-
trol adverse outcomes through systematic implementation are required (Manuj
et al., 2014). This study provides a platform for researchers to study the concept
of SCRM and discover the underlying issues. Further, the practicing managers
can utilize the presented instruments to review and to redefine their SCRM
issues.
The study is done in two folds. First, the critical factors of SCRM are identified
and a framework is developed. The developed framework helps the researchers to
understand the relationship between the critical risk-related factors. The empirical
study is performed to gain the perceptions of the practicing managers about risks
perceived in their organization. In the second part, the critical factors of SCRM
are constituted as the states of the Fuzzy Cognitive Map, which represents the
dynamical system of the supply chain. The system helps researchers to identify
all plausible risks in the long run, given a risk observed from a point of time, and
suggests mitigation strategies for practicing managers. This chapter is organized
as follows: Sect. 2 describes the literature evidence of the supply chain risk
and mitigation strategies. Section 3 deals with the methodology of the research
frameworks and explains each of the methods used in the study. Then results are
presented in Sect. 4. The detailed discussion of the results can be seen in Sect. 5 and
the conclusion in Sect. 6.
2 Literature Review
Demand Side Risk1 (DSR) For good business performance, organizations need to
deliver the required quantities of the product at the pre-agreed time and place. Issues
such as volatile customer demand, inaccurate demand forecasting, competitive rival
products, poor understanding of customer preferences, short product life cycle, and
defects in the products are predominant demand side risks (Sodhi and Chopra, 2004;
Wagner and Bode, 2006; Oke and Gopalakrishnan, 2009; Trkman and McCormack,
2009).
Supply Side Risk1 (SSR) The supplier plays an important role in providing the
sub-assemblies to the original equipment manufacturers (OEMs) or (Tier-2 supplier
to Tier-1 supplier). The risks faced by the suppliers are due to market characteristics,
quality problems, financial instability, technological changes, and product design
issues (Wagner and Bode, 2006; Trkman and McCormack, 2009; Thun and Hoenig,
2011).
Logistic Risk (LR) Logistics service providers deliver products and services at the
predestined place at the appropriate time, which is a characteristic of an effective
and efficient supply chain. For efficient logistics, choice of a transport, the type of
vehicle, and their utilization play a major role (KPMG, 2010; Rogers et al., 2012).
Regulatory, Legal and Bureaucratic risk1 (RLB) Government, top management
and administration take part in important decisions pertaining to planning (long-
term, short-term), regulatory policies and reforms.While implementing a new plan,
policy, or reform, apart from confirming the overall cost-effectiveness, it has to be
ensured that every stakeholder in the supply chain gets an advantage (Hendricks and
Singhal, 2005; Wagner and Bode, 2006; Rogers et al., 2012).
Infrastructure Risk1 (IR) The infrastructure of the firm includes humans,
machines, buildings, and service systems, as they are all essential for the growth
and sustenance of the firm. So, it is mandatory for the firm to maintain cordial
Human Resource policies and perform regular maintenance (breakdown, preventive,
corrective) to machinery (FCCI, 2013).
Stock/Data Management risk1 (SDM) Stock information needs to be maintained
and shared across the supply chains Specifically for Vendor Managed Inventory
(VMI), it is vital to dispense the information across all the stakeholders of the supply
chain (Dath et al., 2009). Firms needs to maintain standard product identification,
EDI, CFAR, and customer preferences details to enable the vendors and suppliers
to know the expansions in the stock levels and replenish them suitably.
Environmental Risk1 (ER) Natural hazards, terrorist attacks on the establishment,
civil unrest, and epidemics disrupting the supply chain have been reported from the
1 Indicates the factors adapted appropriately and suitably from the work of Sodhi and Chopra (2004)
various parts of the world. Certain events may be unavoidable, but one needs to be
prepared to face unfavorable/undesirable events (Oke and Gopalakrishnan, 2009;
Thun and Hoenig, 2011).
Financial Risk1 (FR) Firms need to have sound financial stature to sustain in the
competitive environment. Hence, it vital to maintain the optimum cash flow into the
business from the customers on time. Also, financial activities such as diversification
of the funds, inappropriate drawing of loans to fund the business, over-leveraging of
the funds out of business in high-risk investments are undertaken to maximize the
returns that may prove fatal to the firms’ existence (NMCC, 2008; Dath et al., 2009;
CRISIL, 2010).
Top Management Commitment2 (TMC) SCRM needs are better understood by
Senior-level managers as they form the decision making team, which enables the
firm to be competitive. They should show their commitment, by guiding and leading
an effective SCRM implementation process (Dath et al., 2009; KPMG, 2010).
Mitigation Strategies and Risk Management Process1 (MS-RMP) In the com-
petitive and unstable environment, every firm inevitably faces risk. The risk in
the organization cannot be avoided, but it can be curtailed to a certain degree.
The organization can identify deviation, and implement appropriate strategies and
convert negative events into positive or near positive outcomes (Sodhi and Chopra,
2004; Oke and Gopalakrishnan, 2009; Tang and Nurmaya Musa, 2011; Radke and
Tseng, 2012; Arcelus et al., 2012; Chen et al., 2013; Sawik, 2013; Kırılmaz and
Erol, 2017; Tarei et al., 2020). In this study, the following sequence is followed for
alleviating risk:
• Risk Planning (RP) – establishment of plan of actions such as a business
continuity plan.
• Risk Monitoring (RM) – continuous assessment of control parameters, active
involvement of risk managers.
• Risk Avoidance (RA) – collaboration among trading partners, implementation of
forecasting and advance planning systems.
• Risk Sharing (RS) – equitable sharing of the impact of risk among the trading
partners as per agreeable established policies, insurance, and fund provisioning
to meet exigencies.
Performance measures2 (PERF) The effectiveness and efficiency of any system
can be estimated only when they are measured. It helps in benchmarking the
firms’ capabilities with the competing firms. The SCRM implementation warrants
inclusion of measures of performance from an internal perspective include the
inter-functional, customer and partner perspectives. From the perspective of the
manufacturer, the measures of performance are finance, customer perspective,
2 Indicates the factors adapted appropriately and suitably from the work of Dath et al. (2009). The
considered items are suitably modified and additional items added to every factor.
SCRM in Indian Manufacturing Industries 111
Conceptual Framework
The factors identified in Sect. 2.1 are portrayed through the conceptual model, as
shown in Fig. 1, from the perspectives of the manufacturers (including OEMs and
Suppliers) (Shenoi et al., 2016). However, they do not provide a comprehensive
questionnaire based on the conceptual framework. They hypothesized the frame-
work. In the model depicted below is a mapping from the SCRM constructs (the
Independent Variables (IVs)) and the performance measures, i.e., the Dependent
Variable(DV). The items in the middle are the mediator(s). The mediators cause
the relationship between the IV and DV to change the pathway depending on the
mediating variable’s participation in the relationship. In this study, it is evident
that the presence of the mediators (namely, risk planning, risk monitoring, risk
avoidance, and risk sharing described in Sect. 2.1) has a mediating effect on the
Fig. 1 The framework of SCRM with respect to Manufacturers (Shenoi et al., 2016)
112 V. Viswanath Shenoi et al.
association between the IVs and the DV taken one at a time (refer Fig. 1). Similar
to SCM, the SCRM also involves managing complex flow of information and
understanding the irrelevance. The SCRM gained importance due to the following
(Faisal et al., 2007):
• Focus on core-competencies.
• Elimination of geographical boundaries for establishing partnerships.
• Escalating disruptions in the supply chain due to man-made attacks and due to
natural disasters.
• Reduction in the supplier base attributed to JIT production.
• Turbulent nature of the economy due to strong linkage to commodities.
Lavastre et al. (2012) identified techniques for minimizing supply chain risks
by considering the attitudes of the managers towards risk and the tools used to
understand the risk.
The findings from the empirical study help us to develop the Fuzzy Cognitive Map
(FCM) needed to predict risks and suggest mitigation strategies. In the Sect. 2.1 and
section “Description of the Construct of SCRM: Manufacturer’s Perspective”, we
have identified the factors that affect the Supply chain. In the globalized environ-
ment, it is necessary to identify early warning signals and implement appropriate
strategic decisions to meet the exigencies. The objective is to develop a fuzzy model
to identify and predict all the plausible risks based on the instantaneous risk vector.
We utilize the responses of the empirical study for the construction of FCM. The
factors of SCRM represent the states of the FCM. FCM is used to represent the
overall behavior of the dynamic supply chain system. The instantaneous risk vector
is passed on to the dynamic system to identify all plausible risks until saturation.
The research contributions to managing supply chain risks, to date, are based on the
identification of sources of risk, and recommendation of corresponding mitigation
strategies. However, Sodhi and Chopra (2004) proposed that risks are dependent
on each other, and therefore, it is necessary to understand the interdependencies
between the risk constructs. Graph modeling seems to be an appropriate method
for understanding the interdependencies. Graph modeling approaches have been
applied to coordination, customer sensitivity, risk alleviation, and risk mitigation
(Kaur et al., 2006; Faisal et al., 2007). The other cognitive approach widely used is
neural regression.
SCRM in Indian Manufacturing Industries 113
Using fuzzy logic and Analytic Hierarchical Process (AHP), Kaur et al. (2007)
evaluated the extent of coordination among the trading partners in the supply chain.
An integrated approach using FCM and the fuzzy soft set model is proposed to
solve the supplier selection problem. The approach utilizes the feedback effect
among the criteria and considers the uncertainties in the decision-making process.
Micheli et al. (2014) developed an analytical model using stochastic integer linear
programming approach that incorporated supply chain managers, judgments using
utility functions, and the model employed fuzzy-extended pairwise comparisons for
choosing the mitigation strategy. However, the application of FCM in SCRM context
is sparsely available.
In our study, FCM is utilized to construct the dynamical system model of
risk present in the supply chain. Recurrent structures such as FCM help us to
establish cause-effect relationships, as the system variables are identified to the
graph nodes. The proposed model helps to determine all the plausible risks which
may appear in a supply chain and suggests strategies to mitigate risk through a fuzzy
approach.
The Fuzzy Cognitive Map concept, extended by Kosko (1986), used fuzzy causal
functions considering numbers in the interval [−1, 1]. Its application is extended to
decision making (Stylios et al., 2008), performance measurement (Glykas, 2013),
supplier selection (Shaw et al., 2012), and modeling of resilient supply chain
network design (Kristianto et al., 2014).
FCM provides a quick solution while modeling complex systems with lesser
computational effort. The relationships between variables obtained constitutes the
matrix of edges known as an adjacency model, representing the overall behavior of
the system. A non-linear activation function is used for transformation of the path
and is directed towards them into a value [0, 1] or [−1, 0, 1].
Cognitive mapping of FCMs helps the respondent to have an awareness of
their prediction model by drawing, capturing and transferring causal knowledge. It
enables a way to elicit, capture and transfer causal knowledge. Maps are constructed
in various ways, such as conducting interviews, analyzing texts, capturing partici-
pant perceptions through a questionnaire, and discussing in groups, to name a few.
Perceptions from varied and diverse sources are integrated to deal with limitations of
expert opinions. FCMs also allow quantitative analysis and interpretation of quasi-
dynamic behavior and support in making timely decisions. While adopting FCM,
modelers can use all combinations of input variables or suggest alternative system
descriptions while addressing complex problems or linking internally consistent
future state.
114 V. Viswanath Shenoi et al.
e14
A1 A4
e31
e43
R1 e12 R2
A2 A3
e23
Construction of FCM
“FCM is a directed graph with nodes referred to as policies, events, etc., and causal-
ities as edges. It represents the causal relationship between concepts” (Kandasamy
et al., 2007). From the Bivariate Correlation established in Tables 15 and 16, and the
rules defined for establishing the relationships, the FCM is developed. The directed
edge eij from causal concept Ai to concept Aj measures the extent of influence
SCRM in Indian Manufacturing Industries 115
R2
M1 R1 M2
A1 A3 A2 A4
of Ai on Aj . The edge eij takes values in the fuzzy interval [0,1]. If eij = 0
indicates no causality, whereas eij > 0 indicates causal influence. The influence
of concept Aj increases as Ai increases. if eij < 0, indicates causal decrease
or negative causality i.e., the influence of concept Aj decreases as Ai increases.
Consider the nodes A1 , A2 , . . . , An of the FCM. The edge eij represents the causal
relationship between causal concept Ai to concept Aj , its value is one when there is
significant correlation, otherwise it is zero. The adjacency matrix of FCM has been
constructed by observing Table 16 in respect of the presence/absence of significant
correlation;
⎧
⎨0 no significant correlation exists between causal concept Ai and concept Aj ;
eij =
⎩
1 significant correlation exists between causal concept Ai and concept Aj .
(1)
A symmetric Matrix E is obtained (see Table 3) and the entries in the main diagonal
are equal to 1. The relation represented is reflexive and symmetric. Hence, it is a
compatibility relation (Kandasamy et al., 2007). The diagonal entries are zero, as
they are self-contained.
eij = 0 if i = j
The relationship between the concept Ai and its risk factor-performance is captured
in r2 :
1 if Ai impacts the performance of the supply chain;
(Ai , r2 ) =
0 if Ai does not impact the performance of the supply chain.
(3)
n
if bi reduces disruption due to Ai and enhances
(r1 , bi ) = −1 ; (5)
performance
i=1
n
if bi reduces risk due to Ai and enhances perfor-
(r2 , bi ) = −1 ; (6)
mance
i=1
After augmenting the impact factors with the mitigation strategies, the dimension
of new relationship matrix M is (n + 2) × m.
In order to compare the results of the FCM approach, we consider the Fuzzy
TOPSIS approach. Identification of risks present in the supply chain and appropriate
mitigation strategies to curtail them is a Multi-Criteria Decision Making (MCDM)
problem. A solution to the MCDM problem can be achieved through TOPSIS
(Technique for Order Performance by Similarity to Ideal Solution) approach
(Rostamzadeh et al., 2018; Al Zubayer et al., 2019). This technique suggests the
best solution alternatives closest to the Positive Ideal Solution (PIS) and distant
from the Negative Ideal Solution (NIS). The benefit criteria is maximized and cost
criteria is minimized by PIS and NIS. This section deals with TOPSIS applied to
the fuzzy environment as proposed by Shemshadi et al. (2011) and Zadeh (1965) to
map the linguistic variables to the numerical variables.
118 V. Viswanath Shenoi et al.
Definition 2 The importance values for each criterion, the weighted normalised
fuzzy decision matrix is constructed as:
V = [vij ]m×n i = 1, 2, . . . , n j = 1, 2, . . . , m
m
Dj∗ = (vij , vj∗ ) j = 1, 2, . . . , m (11)
j =1
m
Dj− = (vij , vj− ) j = 1, 2, . . . , m (12)
j =1
SCRM in Indian Manufacturing Industries 119
Dj−
CCj = (13)
Dj− + Dj∗
Literature to date concerning India is majorly on supply chain flexibility (Khan and
Pillania, 2008), and alignment of supply chain strategy with the business strategy
(Dath et al., 2010). However, articles based on SCRM practices in Indian scenario
are rare. Hence, the present study provides insight into SCRM research in the Indian
context, considering the vibrant expansion of industries in India.
Researchers observed that SCRM is multifarious, involves numerous inter-
related risk factors, and needs a rigorous study to address it. Risk mitigation
strategies differ based on the perception of risk and the implementation strategy
that could extend for a short, medium or long term. Based on the literature review,
12 effective techniques widely employed by mangers for minimizing supply chain
risks have been identified. The techniques have been grouped into four categories,
namely risk planning, risk monitoring, risk avoidance, and risk sharing. The above
discussion strengthens the argument that the four mitigation strategy constructs have
a substantial effect on the relationship between the eight SCRM constructs and
the performance measures. The above observations lead us to 32 hypotheses from
the perspective of manufacturers. For example, H1a – corresponds to the effect of
the mitigation strategy, namely risk planning on the risk construct ‘Demand side
risk’. Similarly, other hypotheses can be developed based on mitigation strategies
with the risk constructs taken one at a time. Table 1 summarizes the hypotheses
conceived in this work. The hypotheses conceived shows the effect of each of the
individual mediators (namely, risk planning, risk monitoring, risk avoidance, and
risk sharing) on the relationship between Independent variables (IVs), taken one
at a time, and Dependent Variable (DV), namely, performance measures. Thus the
hypotheses H1a, H1b, H1c, H1d, H1e, H1f, H1g, and H1h correspond to the effect
of mediator variable, Risk avoidance (RA), on Demand Side Risk, Supply Side
Risk, Logistic Risk, Regulatory, Legal and Bureaucratic Risk, Infrastructure Risk,
Stock/Data Management risk, Environmental Risk, and Financial Risk respectively.
The hypotheses H2, H3, and H4 correspond to the effect of mediating variables, Risk
Planning, Risk Monitoring, Risk Avoidance and Risk Sharing, on the individual risk
constructs.
Further, the FCM is developed by using the bivariate correlation among the risk
variables. The study also illustrates practitioners and researchers applying concepts
such as FCM and Fuzzy TOPSIS for the identification and ranking of strategies.
120 V. Viswanath Shenoi et al.
3 Methods
This study helps practitioners understand the likelihood of risks and their severity in
the practitioners’ respective companies. The responses to the questionnaire3 from
the practicing managers serves as the data for our study.
The population considered for the study includes OEMs and suppliers who are
the manufacturers in India. The final product manufacturers are the OEMs in
various sectors, namely, the automotive (e.g., manufacturers of two and four-
wheeler), the heavy engineering (e.g., manufactures of heavy machinery), home
appliances (e.g., manufacturers of washing machines), and general engineering.
The respondents considered for our study includes Vice-Presidents and Senior
personnel in the purchasing and materials management departments. Numerous
respondents were contacted, and the questionnaire was shared with around 120
shortlisted respondents. Around 40 respondents are from OEMs, and 80 respondents
3 https://tinyurl.com/scrmquestionnaire
SCRM in Indian Manufacturing Industries 121
are suppliers to OEMs. The contacted respondents work in various industrial sectors
mentioned in the sample profile summarized in Table 2 below.
Similarly, for the supplier perspective studies, sub-assembly suppliers to OEMs
in the above-mentioned sectors were shortlisted. The supplier respondents are
manufacturers of fasteners, brake linings, wheels, electrical components such as
coils, motors, etc. In the supplier perspective study, out of 80 suppliers contacted,
52 suppliers responded. A majority of the respondents (44) are from the organi-
122 V. Viswanath Shenoi et al.
zations that supply sub-assemblies and parts to OEM’s of two and four-wheeler
vehicles, large capacity vehicles and earth-moving equipment manufacturers. The
questionnaires were circulated to the respondents in three different forms, namely,
direct interview, E-mail and web forms.
The supply chain managers underestimate risk in terms of their likelihood and
impact (Thun and Hoenig, 2011). The importance of development of instruments
is emphasized by the researchers for theory building (Dath et al., 2010). The survey
instruments help the managers to identify the aspects of SCRM and empirically
validate the SCRM constructs from the perspective of the manufacturers. The
exhaustive review of the literature resulted in the development of the instru-
ment.
An instrument with 1 indicating ‘very low’ and 5 indicating ‘very high’ (referred
to as five point Likert scale) has been adopted to study the likelihood of the risk and
its severity in this work.
As discussed in Sect. 2.2, the adjacency matrices are computed for the risk con-
structs. The correlation matrix is given in Table 16. In this study, Ai ’s are considered
as risk constructs, namely, QDSR, QSSR, QLR, QRLB, QIR, QSDM, QER, and
QFR. ‘Q’ is prefixed with the constructs label to indicate the severity of the impact
due to that risk. The entries in Table 3 corresponds to the significant correlations
between the variables as observed from Table 16 along with equation 1, 2 and 3.
‘Q’ is prefixed with the mitigation strategies to indicate the impact of implementing
the alleviation strategy. Mi ’s are considered as QRP, QRM, QRA, and QRS. The
relationship matrix presented in Table 4 is also generated from Table 16, along with
SCRM in Indian Manufacturing Industries 123
equations 4, 5 and 6. Figure 5 represents the construction of the FCM model of risk
constructs. Figure 6 represents the relationship between risk constructs, mitigation
strategies, and the nature of the impact on the supply chain.
QDSR
QSSR QFR
QLR R1
QRLB R2
QIR QER
QSDM
observed variables, reliable indicators for each of the constructs, the CFA technique
has been adopted.
On freezing the scale of measurement, the researchers need to ensure the construct
validity of the instrument and gain confidence that the inferences derived based
on the proposed model reflects reality. The fundamental prerequisite for construct
validity is the unidimensionality of the measure. CFA is done for all the constructs.
A value of 0.9 or higher for the Comparative Fit Index (CFI) suggests that the model
framework has an adequate fit, and unidimensionality exists (Sureshchandar et al.,
2001). The Root Mean Error of Approximation (RMSEA) is an additional measure
for model fit adequacy. The maximum value for RMSEA deemed to be acceptable
is 0.1 (Hair et al., 2006).
The consistency of the instrument is a measure of Reliability, also referred to
as internal consistency to measure the intended issue. The measure of internal
SCRM in Indian Manufacturing Industries 125
Ai
QDSR
QSSR Mi
QRP
QLR
QRM
QRLB R1
QRA
R2
QIR
QRS
QSDM
QER
QFR
consistency is Cronbach’s alpha (α) (Cronbach, 1951), and a value of 0.7 and above
is acceptable.
Face and content validity indicates the extent to which the research instrument
depicts the concepts (Sureshchandar et al., 2001) and determines the face value of
a good representation of the construct (Kaplan and Saccuzzo, 2012). Review of the
questionnaire by expert professors and industry personnel ensures face validity. The
126 V. Viswanath Shenoi et al.
extent to which the factors converge is indicated by convergent validity and a value
of 0.9 and above for Bentler Bonet Fit Index (BBI) satisfies the requirement.
In the present study, the instrument is developed for manufacturers, which
include suppliers and OEMs. The items contained in instruments are drawn from
the literature.
4 Results
4.1 Factor Loadings
Table 5 presents the Factor loadings containing the values of various indices
such as BBI, CFI, Chronbach’s alpha and RMSEA. The values are above the
limits satisfying the required validity and reliability requirements. Top Management
Commitment (TMC) and Mitigation strategies (MS) have RMSEA values above
0.1, but the values of other indices such as BBI and CFI are in range. Tables 5, 15
and 16 indicate the presence of discriminant validity (Crocker and Algina, 1986).
Table 5 continued.
Item Item
Constructs number loadings BBI CFI α RMSEA
Mitigation Strategies and Risk Management Process (MS-RMP)
Risk Planning (RP) 1 0.92 1.00 1.00 0.783 0.00
2 0.61
3 0.69
4 0.64
Risk Monitoring (RM) 1 0.88 1.00 1.00 0.74 0.00
2 0.64
3 0.63
Risk Avoidance (RA) 1 0.81 1.00 1.00 0.86 0.00
2 0.73
3 1.00
Risk Sharing (RS) 1 0.84 1.00 1.00 0.87 0.00
2 0.86
3 0.80
Financial Performance 1 0.61 0.90 0.91 0.83 0.12
2 0.67
3 0.89
4 0.56
5 0.54
6 0.44
Customer perspective
1 0.70 0.97 0.98 0.93 0.11
performance
2 0.73
3 0.79
4 0.79
5 0.86
6 0.54
Trading partner perfor- 1 0.79 0.95 0.96 0.95 0.09
mance
2 0.89
3 1.00
4 0.83
5 0.87
Internal business
1 0.68 0.98 1.00 0.92 0.07
perspective performance
2 0.85
3 0.75
4 0.77
5 0.76
6 0.79
Innovation and learning
1 0.77 1.00 1.00 0.93 0.00
perspective performance
2 0.69
3 0.86
Notes: Acceptable limits: CFA Index: BBI > 0.9; CFI > 0.9; α > 0.6; RMSEA< 0.1
SCRM in Indian Manufacturing Industries 129
Table 10 The effect of mediating variables on Regulatory, Legal and Bureaucratic risk(RLB)
SOBEL TEST
Risk-R MV R → MV(a) Sa R&MV → P(b) Sb TS SE p-value
RLB RP 4.08 0.18 0.44 0.11 3.93 0.45 0.00
RLB RM 3.89 0.23 0.49 0.07 6.46 0.29 0.00
RLB RA 4.14 0.24 0.36 0.08 4.35 0.34 0.00
RLB RS 4.47 0.23 0.41 0.08 4.95 0.36 0.00
performed for risk construct and performance. The Bivariate correlation helps the
researchers to understand the inter-dependency between the variables (Tables 15
and 16).
Table 15 Bi-Variate Correlation between the likelihood occurrence of risk constructs and performance
* IDSR ISSR ILR IRLB IIR ISDM IER IFR IRP IRM IRA IRS PERF
IDSR 1
ISSR 0.61a 1
ILR 0.68a 0.60a 1
IRLB 0.31a 0.47a 0.35a 1
IIR 0.47a 0.56a 0.65a 0.71a 1
SCRM in Indian Manufacturing Industries
The risk is categorized into four types the low likelihood and low impact, the low
likelihood and high impact, the high likelihood and low impact, and high likelihood
and high impact. In the proposed model, the normalized weights for each criterion
is obtained from the likelihood of a risk. The ranking of the alternatives and the
weights of the criteria are calculated based on Fuzzy TOPSIS approach. The theory
of fuzzy sets reconciles the alternatives evaluation to deal with the uncertainty. The
triangular fuzzy numbers used for the evaluation are represented in Fig. 7.
In this problem, there are eight risk factors and four mitigation strategies as
alternatives. The desire is to rank the alternatives using risk and to reduce the risk
present in the supply chain. The steps involved in the ranking of the alternatives as
follows:
1. There are four alternatives, namely risk planning, risk monitoring, risk avoidance
and risk sharing.
2. Risk criteria considered are:
• Demand Side Risk (DSR)
• Supply Side Risk (SSR)
• Logistic Risk (LR)
• Regulatory, Legal and Bureaucratic risk (RLB)
• Infrastructure Risk (IR)
• Stock/Data Management risk (SDM)
• Environmental Risk (ER)
• Financial Risk (FR)
3. The weights of the criteria are calculated (see Table 20)
4. The alternatives are evaluated using the linguistic variables. The linguistic values
are converted to appropriate equivalent fuzzy numbers (see Tables 21 and 22).
5. The weighted decision matrix for alternatives are calculated (see Tables 23, 24,
and 25).
6. FPIS and FNIS are calculated (see Table 26 for results)
7. The distance of each alternative is calculated from FPIS and FNIS to determine
the closeness coefficient for each of the alternatives (see Table 27).
8. Rank of the alternatives with respect to their closeness coefficient is thus obtained
and is shown in Table 28.
From the manufacturer’s perspective, the effect of each mediator in the relationship
between IVs is taken one at a time, and DV is individually studied through
regression analysis using SPSS. Some of the items are eliminated during the
validation process. The complete final instrument can be had from link.3
Tables 7, 8, 9, 10, 11, 12, 13, and 14 give the results of the effect of individual
mediators on individual IVs taken one at a time. The Sobel test is done to analyze the
136
DSR (0.13) SSR (0.14) LR (0.13) RLB (0.13) IR (0.12) SDM (0.12) ER (0.11) FR (0.12)
RP 0.33 0.44 0.56 0.33 0.44 0.56 0.11 0.22 0.33 0.11 0.22 0.33 0.33 0.44 0.56 0.11 0.22 0.33 0.11 0.22 0.33 0.33 0.44 0.56
RM 0.78 0.89 1 0.78 0.89 1 0.78 0.89 1 0.78 0.89 1 0.78 0.89 1 0.78 0.89 1 0.78 0.89 1 0.78 0.89 1
RA 0.33 0.44 0.56 0.11 0.22 0.33 0.11 0.22 0.33 0.11 0.22 0.33 0.56 0.67 0.78 0.11 0.22 0.33 0.33 0.44 0.56 0.33 0.44 0.56
RS 0.56 0.67 0.78 0.56 0.67 0.78 0.56 0.67 0.78 0.33 0.44 0.56 0.56 0.67 0.78 0.56 0.67 0.78 0.56 0.67 0.78 0.56 0.67 0.78
137
138
mediating effect. It is observed that if the p-value falls below the established α level
of 0.05, indicating the association between the pair of IV and DV (i.e., individual
risk construct and performance), risk is significantly reduced by the inclusion of
the mediating variable (namely, risk planning, risk monitoring, risk avoidance, risk
sharing). From the results obtained, it is inferred that the mitigation strategies have
significantly reduced the risk. Hence, all hypotheses H1 through H4 are accepted.
The policymakers should train the practicing managers in risk management. The
process of SCRM involves the design of strategy and framework. An instrument
is essential to validate the framework pertaining to the industry. Based on the
framework, functional managers implement operational decisions. Managers in all
the levels of the supply chain need to know their business and be aware of types and
levels of risks perceived. Risk awareness leads to efficient monitoring of the supply
chain and enables the successful implementation of the SCRM.
Instantaneous vector Q1 represents the risk perceived during an instant of time. The
instantaneous vector Q1 is passed on to the adjacency represented in Table 3. The
resultant vector is threshold to one and is passed on to the adjacency matrix until the
equilibrium.
Equilibrium is attained in the third iteration (refer to Table 29). The equilibrium
vector obtained is passed on to the relationship matrix M . The results in Table 30
suggest that the risks are mitigated, and the impact of risks is alleviated to a
greater extent, as the results show negative values or zeroes. The equilibrium vector
obtained is passed on to the relationship matrix M. The results in Table 31 suggest
The problem of selecting the best mitigation strategy is evaluated through two
approaches. The proposed model is based on the decision makers’ perception of
risk. The ranking of the mitigation strategy is comparatively the same using both the
approaches (refer to Tables 28 and 31). Both the methods scale well in the selection
of mitigation strategies in this case.
The FCM approach interacts with every other risk factors before attaining the
equilibrium. On the other hand, FTOPSIS helps us to evaluate the ranking of the
alternatives to mitigate the risk factors. The FCM approach not only suggests the
level of application of alternatives but also predicts the risk faced by the firm using
the instantaneous vector (refer to Table 30).
6 Conclusions
The study is performed to identify the constructs of SCRM from the manufactur-
ers’ perspective. Validation of constructs is done through a series of tests to capture
the various aspects of SCRM. The present work reinforces that the organizations
needs to have a long-term association with the partners of the supply chain. It is
evident that every firm encounters risks that arise from various entities of the supply
chain. Hence, the supply chain managers need to be proactive in identifying the
risk as and when it surfaces in the supply chain. If ignored, it may lead to grave
consequences. Practicing managers can use the proposed methodology to identify
the risk, which also suggests an appropriate mitigation strategy that reduces the
impact of risk on the supply chain, resulting in improved performance.
FCM approach predicts the future risk and also suggests the mitigation strategies.
FTOPSIS ranks the alternate mitigation strategies. From our study, it is concluded
that risks have to be identified as soon as they surface. In our study, both the
approaches scale well in finding the appropriate mitigation strategies. Accordingly,
one needs to implement an appropriate mitigation strategy to reduce the impact
of risk and improve the supply chain performance. Our findings emphasize that
monitoring supply chain is essential for sustained performance. Also, training of
managers is essential to be risk-sensitive. Broader insights into the models are
SCRM in Indian Manufacturing Industries 143
possible only through a larger data set. Due to operational constraints in the present
study, the survey is restricted to the samples available. Future studies can also look
into specific risk aspects based on different types of manufacturing and service
industries.
Acknowledgments The authors are most grateful to the reviewers of the questionnaire, who have
suggested constructive improvements and augmentative comments. The authors also thank the
respondents for taking their valuable time in participating in the empirical study. The authors are
grateful to the editors and reviewers, who have given us a chance to improve the original version
of the book chapter.
References
Al Zubayer, M. A., Ali, S. M., & Kabir, G. (2019). Analysis of supply chain risk in the
ceramic industry using the topsis method under a fuzzy environment. Journal of Modelling
in Management, 14, 792–815.
Arcelus, F. J., Kumar, S., & Srinivasan, G. (2012). Risk tolerance and a retailer’s pricing and
ordering policies within a newsvendor framework. Omega, 40(2), 188–198.
Chen, J., Sohal, A. S., & Prajogo, D. I. (2013). Supply chain operational risk mitigation: a
collaborative approach. International Journal of Production Research, 51(7), 2186–2199.
CRISIL. (2010). Rating methodology for manufacturing companies. CRISIL Rating Report (pp. 1–
4).
Crocker, L., & Algina, J. (1986). Introduction to classical and modern test theory. Fort Worth:
Harcourt Brace Jovanovich College Publishers, Orlando.
Cronbach, L. J. (1951). Coefficient alpha and the internal structure of tests. Psychometrika, 16(3),
297–334.
Dağdeviren, M., Yavuz, S., & Kılınç, N. (2009). Weapon selection using the ahp and topsis
methods under fuzzy environment. Expert Systems with Applications, 36(4), 8143–8151.
Dath, T. N. S., Rajendran, C., & Narashiman, K. (2009). A conceptual framework for supply
chain managment with specifice reference to a developing economy. International Journal of
Logistics Systems and Management, 5(5), 473–524.
Dath, T. N. S., Rajendran, C., & Narashiman, K. (2010). An empirical study on supply chain
management in india: The perspective of original equipment manufacturers and suppliers.
European Journal of Industrial Engineering, 4(1), 2–39.
Faisal, M., Banwet, D. K., & Shankar, R. (2007). Management of risk in supply chains: SCOR
approach and Analytic Network Process. Supply Chain Forum: An International Journal, 8(2),
66–79.
FCCI. (2013). FCCI report on power transmission. http://www.ficci.com/spdocument/20311/
power-transmission-report_270913.pdf. [Online; accessed 25 Dec 20014].
Ghadge, A., Dani, S., Chester, M., & Kalawsky, R. (2013). A systems approach for modelling
supply chain risks. Supply Chain Management: An International Journal, 18(5), 523–538.
Glykas, M. (2013). Fuzzy cognitive strategic maps in business process performance measurement.
Expert Systems with Applications, 40(1), 1–14.
Hair, J. F., Black, W. C., Babin, B. J., Anderson, R. E., & Tatham, R. L. (2006). Multivariate data
analysis (Vol. 6). Upper Saddle River: Pearson Prentice Hall.
Hendricks, K. B., & Singhal, V. R. (2005). Association between supply chain glitches and
operating performance. Management Science, 51(5), 695–711.
Kandasamy, W. B. V., Smarandache, F., & Ilanthenral, K. (2007). Elementary fuzzy matrix theory
and fuzzy models for social scientists (Infinite Study). Los Angeles: Automaton.
144 V. Viswanath Shenoi et al.
Kaplan, R., & Saccuzzo, D. (2012). Psychological testing: Principles, applications, and issues.
Andover: Cengage Learning.
Kaur, A., Kanda, A., & Deshmukh, S. G. (2006). A graph theoretic approach for supply chain
coordination. International Journal of Logistics Systems and Management, 2(4), 321–341.
Kaur, A., Kanda, A., & Deshmukh, S. G. (2007). Coordination in supply chains: An evaluation
using fuzzy logic. Production Planning & Control, 18(5), 420–435.
Khan, A., & Pillania, R. K. (2008). Strategic sourcing for supply chain agility and firms’
performance: A study of Indian manufacturing sector. Management Decision, 46(10), 1508–
1530.
Kırılmaz, O., & Erol, S. (2017). A proactive approach to supply chain risk management: Shifting
orders among suppliers to mitigate the supply side risks. Journal of Purchasing and Supply
Management, 23(1), 54–65.
Kosko, B. (1986). Fuzzy cognitive maps. International Journal of Man-Machine Studies, 24(1),
65–75.
KPMG. (2010). KPMG Analysis Global Transport, Logistics in India, 2010. KPMG White Paper.
Kristianto, Y., Gunasekaran, A., Helo, P., & Hao, Y. (2014). A model of resilient supply chain
network design: A two-stage programming with fuzzy shortest path. Expert Systems with
Applications, 41(1), 39–49.
Kumar, N. (2005). Liberalisation, foreign direct investment flows and development: Indian
experience in the 1990s. Economic and Political Weekly, 40, 1459–1469.
Lavastre, O., Gunasekaran, A., & Spalanzani, A. (2012). Supply Chain Risk Management in
French companies. Decision Support Systems, 52(4), 828–838.
Liu, H., Chen, W., Kang, Z., Ngai, T., & Li, Y. (2005). Fuzzy multiple attribute decision making
for evaluating aggregate risk in green manufacturing. Tsinghua Science & Technology, 10(5),
627–632.
Manuj, I., Esper, T. L., & Stank, T. P. (2014). Supply chain risk management approaches under
different conditions of risk. Journal of Business Logistics, 35(3), 241–258.
Micheli, G. J. L., Mogre, R., & Perego, A. (2014). How to choose mitigation measures for supply
chain risks. International Journal of Production Research, 52(1), 117–129.
Moore, M., & Moore, M. K. (2003). A world without walls: Freedom, development, free trade and
global governance. Cambridge: Cambridge University Press.
NMCC. (2008). Current global financial crisis and its impact on India-measures for ensuring
growth of the real sector of the economy. National Manufacturing Competitiveness Council
Report (pp. 1–9).
Oke, A., & Gopalakrishnan, M. (2009). Managing disruptions in supply chains: A case study of a
retail supply chain. International Journal of Production Economics, 118(1), 168–174.
Preacher, K. J. (2010). Calculation for the Sobel test: An interactive calculation tool for Mediation
tests. http://quantpsy.org/sobel/sobel.htm. [Online; accessed 25 Dec 2014].
Preacher, K. J., & Hayes, A. F. (2008). Asymptotic and resampling strategies for assessing and
comparing indirect effects in multiple mediator models. Behavior Research Methods, 40(3),
879–891.
Radke, A. M., & Tseng, M. M. (2012). A risk management-based approach for inventory planning
of engineering-to-order production. CIRP Annals-Manufacturing Technology, 61(1), 387–390.
Rogers, H., Srivastava, M., Pawar, K., & Hartmann, E. (2012). A review of supply chain
risk management research from an Indian perspective. In: Proceedings of European Risk
Conference.
Rostamzadeh, R., Ghorabaee, M. K., Govindan, K., Esmaeili, A., & Nobar, H. B. K. (2018).
Evaluation of sustainable using an integrated fuzzy topsis-critic approach. Journal of Cleaner
Production, 175, 651–669.
Saeidi, P., Saeidi, S. P., Sofian, S., Saeidi, S. P., Nilashi, M., & Mardani, A. (2019). The impact
of enterprise risk management on competitive advantage by moderating role of information
technology. Computer Standards & Interfaces, 63, 67–82.
Sawik, T. (2013). Selection of resilient supply portfolio under disruption risks. Omega, 41(2),
259–269.
SCRM in Indian Manufacturing Industries 145
Shaw, K., Shankar, R., Yadav, S. S., & Thakur, L. S. (2012). Supplier selection using fuzzy AHP
and fuzzy multi-objective linear programming for developing low carbon supply chain. Expert
Systems with Applications, 39(9), 8182–8192.
Shemshadi, A., Toreihi, M., & Shirazi, H. (2011). Supplier selection based on supplier risk: An
anp and fuzzy topsis approach. Journal of Mathematics and Computer Science, 2, 111–121
Shenoi, V. V., Dath, T. N. S., & Rajendran, C. (2016). Supply chain risk management in the
Indian manufacturing context: A conceptual framework. International Journal of Logistics
and Systems Management, 25(3), 313–335.
Sodhi, M. S., & Chopra, S. (2004). Avoiding supply chain breakdown. MIT Sloan Management
Review, 46(1), 53–62.
Stylios, C. D., Georgopoulos, V. C., Malandraki, G. A., & Chouliara, S. (2008). Fuzzy cognitive
map architectures for medical decision support systems. Applied Soft Computing, 8(3), 1243–
1251.
Sureshchandar, G. S., Rajendran, C., & Anantharaman, R. N. (2001). A holistic model for total
quality service. International Journal of Service Industry Management, 12(4), 378–412.
Tang, O., & Nurmaya Musa, S. (2011). Identifying risk issues and research advancements in supply
chain risk management. International Journal of Production Economics, 133(1), 25–34.
Tarei, P. K., Thakkar, J. J., & Nag, B. (2020). Benchmarking the relationship between supply
chain risk mitigation strategies and practices: An integrated approach. Benchmarking: An
International Journal, 27, 1683–1715.
Thun, J. H., & Hoenig, D. (2011). An empirical analysis of supply chain risk management in the
German automotive industry. International Journal of Production Economics, 131(1), 242–
249.
Trkman, P., & McCormack, K. (2009). Supply chain risk in turbulent environments a conceptual
model for managing supply chain network risk. International Journal of Production Eco-
nomics, 119(2), 247–258.
Wagner, S. M., & Bode, C. (2006). An empirically-based and more fine-grained investigation into
supply chain risks. In Proceedings of 4th Worldwide Research Symposium on Purchasing and
Supply Chain Management, San Diego.
Wagner, S. M., & Bode, C. (2008). An empirical examination of supply chain performance along
several dimensions of risk. Journal of Business Logistics, 29(1), 307–325.
Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8(3), 338–353.
Improving Service Supply Chain
of Internet Services by Analyzing Online
Customer Reviews
1 Introduction
Service supply chain, a series of activities involving the flow of service, includes
customers, service providers and service producers, as the main entities (Sampson
2000). Among these components, customers and their opinion are considered to be
essential factors in the production process (Shahin 2010). In particular, services,
such as wireless telecommunications and internet service providers, involves direct
customer-provider interaction, and hence, it is necessary to investigate such services
for quality improvement. More specifically, the wireless communications industry
is proliferating with a reported growth of 4.1% from 2013 to 2018 and an expected
growth of 3.5% from 2018 to 2023 in the United States (Ibis World Report 2020).
This expected industry growth will lead to an estimated revenue increase of $114.6
billion (Ibis World Report 2020). Several components factor into this anticipated
advancement, including faster connectivity speeds and better service, which leads
to an overall desire for increased internet usage and the idea that internet access is a
necessity rather than viewing it was a luxury (Ibis World Report 2020). The wireless
communications industry is highly competitive, as approximately two-thirds of the
adults use broadband internet at home (Kohut 2018).
Despite this level of competition, in the United States, the industry is dominated
by very few noticeable players, as the five biggest companies combine to control
over 68% of total market share—AT&T (24.9%), Comcast (17.2%), Charter
(12.9%), Verizon (7.4%), and CenturyLink (5.8%) (Ibis World Report 2020). In
such a competitive environment and a quickly growing industry, Internet Service
Providers (ISPs) must ensure that their efforts are targeted towards attracting and
retaining customers for the continued growth of their business. Nearly 78% of
revenue originates in the ISP-to-customer sector, representing the largest market in
which ISPs operate and compete. This sector includes both residential and business
services with the most common activities involving downloading multimedia con-
tents, browsing the internet, and using basic internet applications such as accessing
email accounts (Ibis World Report 2020). ISPs in this domain compete on many
factors, namely, availability of service, price, speed of connectivity, and customer
service (Ibis World Report 2020).
In order to attract and retain the most customers possible, ISPs must determine
the client’s expectations and tailor their offerings to meet consumer needs and
wants. One of the best ways for ISPs to identify what their users perceive about their
services is to analyze online reviews posted by customers. Several social networking
sites allow consumers to discuss their experiences with various ISPs and detail
their levels of satisfaction or dissatisfaction. As Web 2.0 has gained traction and
popularity, this style of content that is generated by users has increased rapidly (Guo
et al. 2017). Reviews can be either negative or positive, and since negative reviews
circulate five to six times faster than positive reviews, their impact on the business
can be monumental (Salehan and Kim 2016).
As social media usage continues to grow, enterprises experience additional road-
blocks since they must examine OCR to understand how customers are interacting
with their products and to determine the performance of their products and services
(Sen and Lerman 2007). Recently, studies have also investigated online customer
reviews (OCR) to find the strengths and weaknesses associated with the products
and services offered by a business (Salehan and Kim 2016; Srinivas and Rajendran
2019; Rajendran 2020). By using OCR to understand the strengths and weaknesses
of their offerings, any organization, including ISPs, can discover more ways to
improve and enhance the experiences that customers have with their products and
services.
Insights that ISPs may extract from online customer reviews include deciding
which of their services to improve in order to satisfy their clients who have been
disappointed in the past or determining the areas in which they perform very well
and developing marketing plans to highlight those fields of service to attract more
customers (Rajendran and Pagel 2020a). As studies, such as Cheung and Lee (2012),
have shown that new customers place heavy emphasis on the opinions of existing
clients and that OCR weigh heavily into a consumer’s view on which company to
choose, it is extremely likely that potential ISP consumers might check OCR to
gain insight into the experiences that prior users have had with each ISP that they
are considering.
Each ISP adds value to customers in a different way tailored to its own business
model. Many companies find success through economies of scale by bundling
phone, internet, and television services into one package for consumers. Other ISPs
find a niche market and focus on providing the best service in a particular sector,
Improving Service Supply Chain of Internet Services by Analyzing Online. . . 149
such as rural areas (Ibis World Report 2020). Regardless of what an ISP’s business
model is, OCR carry abundant information that can be beneficial to ISPs as they
seek to provide the most value for their customers. The purpose of this paper is
to identify the strengths, weaknesses, opportunities, and threats of an ISP utilizing
knowledge mining to improve customer satisfaction based on information gained
from online consumer reviews. Each consumer review explores various topics
(e.g., speed, connection, customer service, etc.), and conveys an attitude (positive,
negative, or neutral) towards each topic. In this study, we use text analytic methods
to extract and analyze thousands of reviews for identifying those key topics and
propose managerial recommendations based on our results.
The rest of the paper is organized as follows. Section “Literature Review”
covers a review of several prominent studies discussed in the literature. A thorough
discussion of the proposed methodological approach is given in Section “Method-
ology”. Section “Case Study” includes the description of the case study as well
as an analysis of the results obtained. Lastly, Section “Conclusions” presents the
conclusions and the scope of later work.
2 Literature Review
As discussed earlier, the focus of this paper is to capture the voice of customers
using online review mining and enable ISPs to provide better service. Therefore, in
this section, we present some prominent studies related to wireless service providers
and the impact of word of mouth on companies.
One of the very first studies on customer expectations in the wireless service
industry was pioneered by Zander (1997). The author explored the idea that
customers of wireless communication services expect to receive the same services
as they would receive from a fixed network. The paper analyzed both a “universal”
coverage scenario providing wideband services at all locations and a “hot-spot”
scenario where coverage is offered, but only in select geographical areas. The study
showed companies the benefits of presenting wireless services to all consumers and
but also indicated that as bandwidth increases, the infrastructure cost of a wireless
system would also increase. Bar and Park (2005) looked at public Wi-Fi networks
in the United States and the policy issues that the new trends raise. Metropolitan
cities already have the existing infrastructure which supports the Wi-Fi networks,
and they can help provide connectivity for city employees while making urban areas
more desirable for residents and businesses.
With the rapid development in technology, broadband connectivity has become
more widespread, and voice services are more mobile; as a result, ISPs are bundling
150 S. Rajendran and J. Fennewald
Erkan and Evans (2016) examined the influence of eWOM in social media on
consumers’ purchasing intentions. Several hypotheses were considered, and the
authors proved that eWOM is positively related to consumers’ purchasing intention,
the usefulness of eWOM information is positively associated with the adoption
of eWOM information, and the quality of eWOM is positively associated with
usefulness. Wang et al. (2016) showed several differences between traditional WOM
and eWOM, including the anonymity of eWOM messages, the fact that multiple
people can access eWOM at any time, and that eWOM is more persistent and
measurable. Their research indicated that about 70% of 28,000 internet users in
56 countries rely on online consumer reviews for recommendations.
Text analytic tools for examining eWOM have not only been used in product
improvement, but also in the service sector (Rajendran and Pagel 2020a). For
instance, Liang et al. (2018) studied the effect of eWOM on hospitality and tourism
management. Their study showed that eWOM is highly influential due to its speed,
amplitude, and convenience. This idea, coupled with the fact that WOM reduces
a company’s ability to influence consumers through traditional marketing and
advertising, shows how vital it is for companies to become actively involved in
online consumer communities (Rajendran and Pagel 2020b).
While the studies mentioned above examined the relationship between online
reviews and customer purchasing intentions, Jalilvand et al. (2011) investigated the
reason for the extensive spread of WOM. Their research showed that consumers
spread WOM for several reasons, including extreme satisfaction or dissatisfaction,
Improving Service Supply Chain of Internet Services by Analyzing Online. . . 151
commitment to the firm, length of the relationship with the firm, and novelty of
the product. Additionally, users may be motivated to share their experiences due to
satisfaction, pleasure, or sadness. The authors showed that the main antecedents of
WOM influence are demographic similarity, and perceptual affinity, and that WOM
could influence product evaluations.
Balaji et al. (2016) examined how negative WOM (NWOM) on social network-
ing sites (SNS) can adversely impact companies. Customer complaints can lead to
brand dilution, volatile stock prices, and on a grand scale can even be the cause for
public relations crises. Approximately 77% of online shoppers rely on user reviews
to make purchasing decisions, and over one million people read product or service
reviews every week on SNS, and more than 80% of these reviews are negative. Their
study looked at the cognitive dissonance and social support theories to understand
the determinants of NWOM on SNS. Since sharing opinions is a social activity,
the social support theory shows that NWOM sharing may be due to the desire for
social interaction. Additionally, consumers may engage in NWOM communications
in order to reduce their cognitive dissonance levels because they are frustrated with
a product or service.
This proposed research is one of the first to analyze online customer reviews
posted for internet service providers using text analytics. We propose insights and
managerial recommendations based on strategic planning tool (SWOT) and six
sigma techniques (root cause analysis) to capture the voice of the customers. Though
the literature discusses few studies that are conducting the SWOT analysis for
internet service providers, to the best of the authors’ knowledge, this study is the
first to leverage customers’ reviews and understand the voice of the customers for
this purpose.
3 Methodology
The consumer reviews of the ISPs under study are scraped from several online
customer review sites, and since the customer feedback might not be particularly
centered around any one single aspect of the corporation, reviews are separated into
individual sentences, which are analyzed as independent comments. The reviews
are cleaned prior to doing the text analysis by removing incomplete, duplicate,
and irrelevant reviews. Subsequently, the sentences are tokenized, special characters
and non-English words are eliminated, inflected words are stemmed, characters are
converted to lowercase, and stop and infrequent words are removed.
After scraping and pre-processing, the most frequent topics discussed in reviews
need to be identified, which is achieved by analyzing the frequent words, bigrams,
and trigrams. A bigram is a combination of two words that are used consecutively in
Improving Service Supply Chain of Internet Services by Analyzing Online. . . 153
relevance to the same topic. Examples pertaining to our study would be “customer
service” or “connectivity issues”. A trigram is a group of three words that are used
consecutively, and examples would be “good customer service” and “connectivity
issues rural”.
This section presents an analysis of bigrams and trigrams, as discussed by
Jurafsky and Martin (2014). If x1 , x2 , . . . , xN are a set of words, then the probability
that these words occur in a sequence is denoted by P(X) as shown in Eq. (1).
P (x1 , x2 )
P (x2 |x1 ) = (2)
P (x1 )
Applying the conditional probability chain rule to the sequences of words under
consideration, Eqs. (8) and (9) give the probability of N words occurring in a
sequence.
P (x1 . . . xN ) = P (x1 ) × P (x2 |x1 ) × P x3 |x12 . . . P xN |x1N−1 (8)
N
P (x1 . . . xN ) = P xN |x1N−1 (9)
i=1
In Eq. (10), we consider the Markovian assumption for the bigram model.
P xN |x1N−1 = P (xN |xN−1 ) (10)
154 S. Rajendran and J. Fennewald
ν (xN−1 xN )
P (xN |xN−1 ) = (11)
x ν (xN−1 x)
Constraint (12) uses the trigram model to predict the conditional probability of
the Nth word in a sequence.
ν (xN−2 xN−1 xN )
P (xN |xN−1 , xN−2 ) = (12)
x ν (xN−2 xN−1 x)
For each ISP, the reviews are separated based on their ratings as negative (≤2 stars),
neutral (3-stars), and positive (≥4 stars), following which commonly co-occurring
bigrams and trigrams are identified for each category. These bigrams and trigrams
are then used to conduct a SWOT analysis for individual ISPs and the industry as a
whole.
SWOT is a planning tool most commonly used for recognizing the strategic
factors (both internal and external) that are important to the development of an
enterprise (Kurttila et al. 2000). The strengths and weaknesses are identified by
analyzing the internal characteristics of the organization, while opportunities and
threats constitute the external elements, such as competition. The basic framework
of SWOT analysis is shown in Fig. 2. Such a SWOT assessment highlights the
Strengths Weaknesses
ISP offerings that enrich customer’s Qualities of the ISP that hurt customer
Internal
Opportunities Threats
Potential changes companies can adapt to Obstacles that the ISP faces due to
External
critical factors that influence the viewpoint of the corporation and can aid the ISP in
their planning decisions.
4 Case Study
In this study, we have chosen eight companies that control over 70% of the total
market share in the ISP domain. We have de-identified and referred to the ISPs
under consideration as ISP#1 through #8. In total, 23,145 online customer reviews
are considered with the following breakdown of reviews for each company: ISP#1-
2407, ISP#2-4208, ISP#3-937, ISP#4-3446, ISP#5-989, ISP#6-3324, ISP#7-5019
and ISP#8-2815.
Table 1 shows an overview of the ISPs considered and the primary services
offered. As it can be seen, wireless internet is a primary service for every company
except ISP#3 and ISP#6; and ISP#2 and ISP#3 are the only two companies that offer
entertainment media as their primary services.
Table 2 shows the data features and an explanation of each field. The two fields
that are considered in this study are “Rating” and “Review”. Figure 3 shows the
distribution of reviews for the different internet service providers under considera-
tion. We can see that the majority of the time, all companies are being negatively
perceived by customers. More specifically, over 80% of the total reviews that are
posted for ISP#1, ISP#3, ISP#4 and ISP#8 are given less than a 2-star rating. On the
other hand, relatively more positive reviews are observed for ISP#5, whereas neutral
reviews are comparatively more in ISP#6.
156 S. Rajendran and J. Fennewald
Once the reviews are extracted, the data is cleaned using the procedure discussed
in Sect. “Web Scraping and Text Pre-Processing”. This data pre-processing is done
with the modules in the Python natural language toolkit (NLTK). In addition to the
default English stop words in the NLTK toolkit, we also created a custom set of stop
words to filter unnecessary words that do not add value to our study.
Table 3 gives an overview of the topics that occur most frequently in the
OCR, a description for each topic, and the significant words associated with each
topic. Clearly, ‘customer service’, ‘connectivity’ and ‘speed’ are the key topics
that customer reviews are emphasized on. For the purpose of illustration, the most
frequently occurring bigrams and trigrams related to the ‘service’ Topic is shown
in Fig. 4. Many of these word combinations revolve around customer service or
whether the actual services (wireless internet, TV, phone) are good or bad.
Table 4 shows the list of most positively- and negatively-discussed topics based
on bigram and trigram analyses. Topics related to customer service, connectivity,
compatibility with other devices, billing and payments are several of the most
common.
Improving Service Supply Chain of Internet Services by Analyzing Online. . . 157
BIGRAMS
(‘customer’, ‘service’), (‘cancel’, ‘service’), (‘phone’, ‘service’), (‘service’, ‘call’),
(‘service’, ‘slow’), (‘tv’, ‘service’)
TRIGRAMS
(‘worst’, ‘customer’, ‘service’), (‘terrible’, ‘customer’, ‘service’), (‘contact’, ‘customer’,
‘service’), (‘poor’, ‘customer’, ‘service’), (‘horrible’, customer’, ‘service’), (‘call’,
‘customer’, ‘service’), (‘customer’, ‘service’, ‘rep’), (‘call’, ‘cancel’, ‘service’)
Fig. 4 Frequently Occurring Bi-grams and Tri-grams for Negative Opinions Related to the Topic
‘Service’
From the topics discussed in Table 4, SWOT analysis for each ISP can be developed.
A sample SWOT Analysis for ISP#1 is given in Table 5. Following that, we also
propose several managerial insights based on the observations made.
The managerial insights derived from the SWOT analysis pertaining to a specific
company (ISP#1) is presented below. A similar approach can be adopted for other
providers as well.
• Customized Bundle Plans: Develop a bundling plan that can be tailored to
individual consumer’s preferences. In other words, the most effective package
will be the one which allows consumers to pay for specific services which they
value and nothing more (e.g., fast speeds to play games online but limited TV
channels, wide range of TV channels without high-speed internet).
• Employee Training: Ensure that all customer service representatives are well
trained on how to best serve their customer base. This may include (i) educating
them on products, policies, and packages that the company offers, and (ii) clearly
communicating with customers on billing dates and plans.
158 S. Rajendran and J. Fennewald
Table 4 Positively and negatively perceived topics by customers for each ISP
ISP Topics based on positive reviews Topics based on negative reviews
ISP#1 Fast speeds internet in cities; versatility; Slow service in rural areas; changing
easy installation billing dates and switching plans without
prior notification; routers are not durable;
slow speed while streaming media
ISP#2 Helpful technician; good streaming for Wrongful billing; increasing service costs
playing online games
ISP#3 Triple play package; fast speed; wide TV Hidden fees; disrupted TV streaming
channel variety
ISP#4 Works well even during bad weather; Slow installation; overseas call centers;
good upload and download speeds fixed 2-year contract; poor transition
from one generation to its next
ISP#5 Good family plan; cost effective; service Unauthorized payments and charges; hard
representative speaks in English fluently to cancel contract
ISP#6 Binge data streaming; good service in Offshore call centers; incompatible with
Europe; service representatives give good certain android devices
instructions
ISP#7 Consistent internet service; helpful Poor international service; call center
salespeople representatives do not speak English
fluently; poor signal in rural areas; slow
4G LTE
ISP#8 Helpful technicians; wide TV channel Services are priced high; DVR is
selection; easy and quick troubleshoots; expensive
providing house calls services
Figure 5 presents a positive link analysis representing the overall topics that are
commonly discussed in the positive reviews across all ISPs. A larger-sized node
indicates that the topic is spoken more often by customers. Clearly, we can see
that ‘Bundling’, ‘Connectivity’, and ‘Technical Service’ were three of the most
repeated topics in the positive review domain. Under the ‘bundling’ topic, customers
praise the triple and quadruple play packages as well as the family plans due to
their cost savings. The topic ‘Connectivity’ is observed to be positively spoken by
customers of certain companies and negatively for other ISPs. Thus we interpret
that connectivity is a key issue that companies must pay attention to. Customers
prioritize consistent connection over speed, while good online gaming streaming is
also a factor that they consider. During a technical issue, customers expect proper
assistance either by visiting them at home or solving the problem over the phone
(i.e., easy troubleshooting).
Figure 6 depicts the results of the root cause analysis through a fishbone diagram
highlighting the overall causes that lead to customer dissatisfaction in the internet
service sector. The primary causes are ‘Contracts’, ‘Connectivity’, ‘Cell Phone
Service’, and ‘Call Center’. The associated secondary causes are also shown in Fig.
6. Clearly, customers are highly unsatisfied when they are charged with additional
fees without prior notification. Interrupted service and slow loading speed in rural
areas are also some issues that users experience. Moreover, they are not comfortable
with automation, especially studies have proven that the older population is resistant
to automation changes (Vaportzis et al. 2017).
Based on the bigram and trigram analyses, we suggest the following recommenda-
tions to the internet service providers as a whole.
• Streamline customer service call process: When customers attempt to contact
ISPs over the phone, they often complain that it is difficult to speak to a human
and that processes are automated. Our recommendation would be for ISPs to
have online live chat programs that can achieve a balance between customer
satisfaction and cost.
• Provide positive work environment for customer service representatives: Many
online reviews mention frustrations with customer service representatives,
specifically for being “rude”, “unhelpful” or “not caring” about the user
problems. Prior studies have proven a positive relationship between employee
job stress, job satisfaction, and customer loyalty (Loveman 1998; Hansemark
and Albinsson 2004; Hill and Alexander 2017). Therefore, we theorize that
the rude behavior of customer service agents is due to the burnout that these
employees experienced from answering customers’ calls monotonously. Hence
it is recommended that ISPs provide a more positive work environment offering
free/discounted health club memberships and conduct regular medical checkups.
Also, ISPs can develop a handbook for call-center representatives providing a
set on instructions on how to handle different types of customer grievances.
• Payment communication: Many consumers are upset that ISPs did not com-
municate clearly about contracts, extra fees, change in billing dates, etc. Our
recommendation for ISPs is to interact and update customers by different modes
of communication (such as, phone calls, emails, and text messages).
Improving Service Supply Chain of Internet Services by Analyzing Online. . . 161
5 Conclusions
In recent years, due to the increasing popularity of service sectors, service supply
chains are being widely investigated to improve service quality. Studies have shown
that online consumer reviews heavily influence consumer’s decisions across many
industries, including ISPs. This research is one of the first to provide managerial
insights for ISPs to improve the wireless services offered to their customers
by analyzing online user reviews. While very few research has been conducted
involving SWOT analysis from an internet provider’s perspective, examining online
customer reviews enables ISPs to understand their consumer needs better. Over
23,000 online reviews were first collected using web scraping, reviews were
subsequently separated based on their star ratings, following which commonly co-
occurring words were identified in each category, and a SWOT and root cause
analyses were finally conducted.
We identified six meaningful topics using our approach and conducted a SWOT
analysis to showcase the critical factors that impact consumers’ choice of ISP. We
also proposed the following managerial recommendations to the ISP sector, based
on the results obtained.
• Streamline customer service call process
• Provide a positive work environment for customer service representatives
• Interact with customers using different means of communication
• Establish proper infrastructure in rural areas to improve connectivity
• Make design changes to ensure device compatibility
• Provide language training for representatives
162 S. Rajendran and J. Fennewald
References
Balaji, M. S., Khong, K. W., & Chong, A. Y. L. (2016). Determinants of negative word-of-mouth
communication using social networking sites. Information & Management, 53(4), 528–540.
Bar, F., & Park, N. (2005). Municipal Wi-Fi networks: The goals, practices, and policy implications
of the US case. Communications and Strategies, 61, 107–126.
Cheung, C. M., & Lee, M. K. (2012). What drives consumers to spread electronic word of mouth
in online consumer-opinion platforms. Decision Support Systems, 53(1), 218–225.
Erkan, I., & Evans, C. (2016). The influence of eWOM in social media on consumers’ purchase
intentions: An extended approach to information adoption. Computers in Human Behavior, 61,
47–55.
Guo, Y., Barnes, S. J., & Jia, Q. (2017). Mining meaning from online ratings and reviews: Tourist
satisfaction analysis using latent dirichlet allocation. Tourism Management, 59, 467–483.
Hansemark, O. C., & Albinsson, M. (2004). Customer satisfaction and retention: The experiences
of individual employees. Managing Service Quality: An International Journal, 14(1), 40–57.
Hill, N., & Alexander, J. (2017). The handbook of customer satisfaction and loyalty measurement.
New York: Routledge.
Ibis World Report (2020). https://www.ibisworld.com/industry-statistics/market-size/wireless-
telecommunications-carriers-united-states/
Jalilvand, M. R., Esfahani, S. S., & Samiei, N. (2011). Electronic word-of-mouth: Challenges and
opportunities. Procedia Computer Science, 3, 42–46.
Jurafsky, D., & Martin, J. H. (2014). Speech and language processing (p. 3). London: Pearson.
Kohut, A. (2018). Demographics of internet and home broadband usage in the United States.
Pew Research Center: Internet, Science & Technology. Retrieved February 5, 2018, from
www.pewinternet.org/fact-sheet/internet-broadband/
Kurttila, M., Pesonen, M., Kangas, J., & Kajanus, M. (2000). Utilizing the analytic hierarchy
process (AHP) in SWOT analysis—A hybrid method and its application to a forest-certification
case. Forest Policy and Economics, 1(1), 41–52.
Liang, L. J., Choi, H. C., & Joppe, M. (2018). Exploring the relationship between satisfaction, trust
and switching intention, repurchase intention in the context of Airbnb. International Journal of
Hospitality Management, 69, 41–48.
Loveman, G. W. (1998). Employee satisfaction, customer loyalty, and financial performance: An
empirical examination of the service profit chain in retail banking. Journal of Service Research,
1(1), 18–31.
Markendahl, J., & Makitalo, O. (2007, June). Analysis of business models and market players for
local wireless internet access. In 6th Conference on Telecommunication Techno-Economics (pp.
1–8). IEEE.
Quach, T. N., Thaichon, P., & Jebarajakirthy, C. (2016). Internet service providers’ service
quality and its effect on customer loyalty of different usage patterns. Journal of Retailing and
Consumer Services, 29, 104–113.
Rajendran, S. (2020). Improving the performance of global courier & delivery services industry by
analyzing the voice of customers and employees using text analytics. International Journal of
Logistics Research and Applications, 2020, 1–21.
Rajendran, S., & Pagel, E. (2020a). Recommendations for emerging air taxi network operations
based on online review analysis of helicopter services. arXiv preprint arXiv:2006, 10898.
Rajendran, S., & Pagel, E. (2020b). Insights on next-generation manufacturing of smart devices
using text analytics. Heliyon, 6(7), e04491.
Salehan, M., & Kim, D. J. (2016). Predicting the performance of online consumer reviews: A
sentiment mining approach to big data analytics. Decision Support Systems, 81, 30–40.
Sampson, S. E. (2000). Customer-supplier duality and bidirectional supply chains in service
organizations. International Journal of Service Industry Management, 11, 17.
Sen, S., & Lerman, D. (2007). Why are you telling me this? An examination into negative consumer
reviews on the web. Journal of interactive marketing, 21(4), 76–94.
Improving Service Supply Chain of Internet Services by Analyzing Online. . . 163
Seo, D., Ranganathan, C., & Babad, Y. (2008). Two-level model of customer retention in the US
mobile telecommunications service market. Telecommunications Policy, 32(3–4), 182–196.
Shahin, A. (2010). SSCM: Service supply chain management. International Journal of Logistics
Systems and Management, 6(1), 60–75.
Srinivas, S., & Rajendran, S. (2019). Topic-based knowledge mining of online student reviews for
strategic planning in universities. Computers & Industrial Engineering, 128, 974–984.
Vaportzis, E., Giatsi Clausen, M., & Gow, A. J. (2017). Older adults perceptions of technology and
barriers to interacting with tablet computers: A focus group study. Frontiers in Psychology, 8,
1687.
Wang, Y., Wallace, S. W., Shen, B., & Choi, T. M. (2015). Service supply chain management: A
review of operational models. European Journal of Operational Research, 247(3), 685–698.
Wang, T., Yeh, R. K. J., Chen, C., & Tsydypov, Z. (2016). What drives electronic word-of-mouth
on social networking sites? Perspectives of social capital and self-determination. Telematics
and Informatics, 33(4), 1034–1047.
Zander, J. (1997, May). On the cost structure of future wideband wireless access. In IEEE 47th
Vehicular Technology Conference. Technology in Motion (Vol. 3, pp. 1773–1776). IEEE.
An Integrated Problem of Production
Scheduling and Transportation
in a Two-Stage Supply Chain
with Carbon Emission Consideration
1 Introduction
Sequencing and scheduling involve decision-making that plays a vital role in various
manufacturing and service industries (Pinedo 2012, pp. 1–5). In practice, many of
these industries strive to ensure the effective use of available resources to meet
the organization’s specific objectives by improving the efficiency in scheduling
and sequencing. Supply chain management involves the reduction in operational
costs, improving customer service levels, and establishing a competitive advantage
by integrating and coordinating different functions of the supply chain (Chopra
et al. 2016, pp. 1–10). The scheduling problem in a supply chain may face several
constraints arising from various levels of the chain. Therefore, scheduling theory
assumes great significance in integrated supply chain planning as well as logistic
planning activities in organizations aiming to improve customer service (Azadian
et al. 2015). In this study, an integrated logistics and scheduling problem with the
consideration of carbon emissions in a supply chain system is described, and a
mixed-integer linear programming (MILP) model is proposed.
B. C. Jos ()
Department of Mechanical Engineering, Mar Athanasius College of Engineering,
Kothamangalam, Kerala, India
e-mail: bobincherian@mace.ac.in
C. Rajendran
Department of Management Studies, Indian Institute of Technology Madras, Chennai, India
e-mail: craj@iitm.ac.in
H. Ziegler
School of Business, Economics and Information Systems, University of Passau, Passau, Germany
e-mail: hans.ziegler@uni-passau.de
In the present scenario, customers being the focus of the organization, retaining
them becomes the primary objective of any organization. Retaining customers can
be achieved by meeting and sometimes exceeding their expectations. One of the
main reasons for the loss of customer goodwill is the non-conformance to due
dates (Unlu and Mason 2010). Many times, either due to the non-availability of
enough capacity to process all the jobs or the non-availability of the resource for
manufacturing, many manufacturing firms outsource some of their jobs or orders
to potential subcontractors located geographically away from the manufacturer.
Transportation delay and the associated transportation costs are additional aspects
that need to be considered while preparing the production schedule of the jobs when
outsourcing is chosen as an option (Qi 2008).
In order to reduce the tardiness of jobs, subcontractors usually increase the
number of shipments from their locations. Energy and environmental factors need
to be considered in many such cases. Most of the manufacturing firms focus on
minimizing the cost of production, sometimes by ignoring the resultant carbon
emissions. Because of the growing concerns in climate change, manufacturers
are now forced to reduce carbon emissions. Hence this study also focuses on
minimizing carbon emissions related to the transportation schedules while preparing
the operations plan in a two-stage supply chain with outsourcing options.
2 Literature Review
these parallel identical production resources/outsourced centers with its own design
and planning functions. The authors quoted Nike as an example. They proposed
heuristics to solve this problem and conducted computational experiments.
Alidaee and Li (2014) studied a similar problem with an option of choosing a
subset of subcontractors from a set of subcontractors, and to schedule the operations
in all the chosen subcontractors to minimize the sum of costs involving the cost
of holding machines, cost of total machine usage, and total tardiness costs. They
considered identical processing times and a common due date for all jobs, and
proposed dynamic programming algorithms. Mokhtari et al. (2012) presented a
job scheduling problem with outsourcing options and proposed a MILP model
to minimize the cost of mean weighted flow time and cost of processing the
outsourced orders. However, their model did not consider the transportation of
finished products. The study by Wang and Alidaee (2018) considered unrelated
parallel machines and scheduling of jobs with the objective of minimizing the total
workload. Since the objective was not correlated with minimizing tardiness, the
transportation concerns were not significant, and hence they presented mathematical
models and heuristics by ignoring the logistics schedule.
Research on the reduction in carbon emissions has attained great attention
recently because of the growing concerns on climate change. In the report of
the United Nation’s Inter-Governmental Panel on Climate Change (UN IIPC),
Kahn Ribeiro et al. (2007) states that, by 2030, energy use and carbon emissions
from transport are predicted to be 80 percent higher than the current levels. The
transportation sector is responsible for approximately 23% of the global energy
related carbon dioxide emissions (Kahn Ribeiro et al. 2007). Logistics are expected
to make a large contribution to the drastic reductions in CO2 emissions that will be
required by 2050 to contain the global temperature increase within 2 ◦ C by 2100
(McKinnon 2010).
Ritchie and Roser (2017) have consolidated various studies related to CO2
emissions. The statistics presented in their studies on global CO2 emissions show the
significance of reducing such emissions to prevent further increases in carbon emis-
sions. Because of the growing concerns on global warming and climate changes,
many countries have implemented a ‘carbon emissions policy’ as mandatory. From
Fig. 1, it is evident that the transport sector is the second major contributor to
CO2 emissions, following the ‘electricity and heat’ sector. Hence, manufacturers
now include ‘reduction in carbon emissions’ as a key criterion while preparing the
operations plan. In our study, we also consider it as one of the key criteria while
preparing the production schedule.
Few of the recent studies in scheduling problems related to ‘reduction in carbon
emissions’ are given in this section. The reduction of potential carbon emission in an
integrated scheduling problem of ‘picking up and delivering customers to airport’
was studied by Yu et al. (2016). Their study illustrated that integrated scheduling
could reduce carbon emissions to a significant level. The problem of vehicle routing
and scheduling in a similar environment was studied by Yang and Tang (2013)
with the objective of minimizing carbon emissions. They presented a nonlinear 0–1
integer programming model. They assumed that the carbon emission (E) of a vehicle
168 B. C. Jos et al.
14
Transportaon
12
Manufacturing/Const
rucon
10 Building
Bunker Fuels
4
Other Fuel
2 Combuson
Fugive Emissions
0
1990 1995 2000 2005 2010 2015
Source: CAIT Climate Data Explorer
3 Problem Description
The integrated logistics and job scheduling problem in a supply chain with non-
identical parallel manufacturing facilities is described as follows. At the beginning
of the planning period, the manufacturer (also called central organization) receives
a set of orders or jobs J = {1, 2, 3, . . . , N} from its customers. It is assumed that all
jobs are available for processing at the beginning of the planning period and need
to be processed either by in-house or outsourced facilities of the organization. The
central manufacturing center is equipped with Min parallel machines, which we refer
to as in-house facilities. In order to minimize the tardiness of jobs (to make an early
delivery), they can outsource some of the jobs to subcontractors, each equipped with
a single machining facility. We consider them as the outsourced manufacturing facil-
ities of the central organization. Machining facilities of subcontractors need not be
identical and are located geographically away from the central organization. Hence,
outsourcing will incur transportation delay and transportation costs, in addition to
the fixed cost (hiring cost) of outsourcing the facility and the associated variable
cost of production. There are Mout potential subcontractors willing to process the
jobs as per the design and schedule given by the central organization. We consider
the availability of Min in-house machines and the Mout subcontractors’ machines
(with each subcontractor having a single machine as the manufacturing facility)
170 B. C. Jos et al.
In-house machines
1
2 {12- 1-7}
3
4 {10-3-11}
5
6 {8-2}
7
8
Outsourced Centers
9
10 {4-6-9}
11
12 {φ}
13
14 {5-14}
15
{13-15}
Fig. 2 Supply chain system with parallel manufacturing facilities both in-house and outsourced
cost), the transportation time required to ship the outsourced jobs to the central
manufacturing facility (Tm ) and the associated cost per trip for transportation (Cm )
are also known in advance. There is a fixed amount of carbon emissions per
shipment (Em ) while transporting the finished jobs back to the central facility from
the outsourced facility m. Em is proportional to fuel consumption (Yu et al. 2016),
and we assume Em is, in effect, proportional to the transportation time taken by
the trucks. Hence it is calculated as Em = r × Tm , where r is the carbon emission
index parameter. The trucks available at facility m are assumed to be homogeneous
and unlimited, and each has a finite capacity, defined in terms of Bm , the number
of jobs. Transportation time, transportation cost, fixed hiring cost, and carbon
emissions per shipment are assumed to be zero for in-house facilities. Note that
the machine usage cost per unit time Cm ∗ is a constant for a given manufacturing
facility, and is independent of the jobs. Hence the total machine usage cost of each
manufacturing facility is proportional to the sum of the corresponding setup times
and the processing times of jobs assigned to that facility. In other words, the total
machine usage cost is given by the product of the machine usage cost per unit time
and the sum of appropriate setup times and processing times of all jobs assigned
to that manufacturing facility. We also assume that a manufacturing facility with
better process capability can process the jobs in lesser time but are associated with
a higher machine usage cost (both cost per unit time and fixed hiring cost) than a
machine with lesser processing capability. However, machine usage cost per unit
time of in-house machines is assumed to be less in comparison to all outsourced
facilities, irrespective of the processing capability.
The decisions to be taken are: (i) to determine the manufacturing facilities (both
in-house and outsourced) to be chosen and engaged for processing; (ii) to determine
the subset of jobs to be processed on the selected manufacturing facilities (both
in-house and outsourced); (iii) to prepare the production schedule of jobs on each
manufacturing facility; and (iv) to prepare the transportation schedule of completed
jobs at each outsourced manufacturing facility.
We present this problem with the objective of minimizing the sum of fixed cost of
hiring manufacturing facilities, their usage cost (variable cost), tardiness penalties
(variable cost), and transportation costs incurred while shipping finished jobs from
172 B. C. Jos et al.
outsourced centers (related to fixed cost per trip from each outsourced center to the
central manufacturing firm).
A Mixed Integer Linear Programming (MILP) model is proposed to solve the supply
chain problem under study, and is described in this section.
Parameters:
Pi, m Processing time for job i on manufacturing facility m.
Di Due date for job i.
Wi Tardiness penalty per unit time for job i.
Tm Transportation time required to transport a job from manufacturing facility
m to the central manufacturing firm /* Tm = 0 if m ∈Min (the set of in-house
machines) */.
Cm Fixed transportation cost required to transport a job from manufacturing
facility m to the central manufacturing firm /* Cm is non-decreasing with
respect to Tm ; Cm = 0 if m ∈ Min */.
Cm Fixed hiring cost associated with manufacturing facility m /* Cm = 0 if m
∈Min */.
Cm∗ Machine usage cost per unit time for manufacturing facility m.
S0, i, m Setup time for job i on manufacturing facility m, where job i is the first job
to be processed on facility m /* fictitious job i = 0 represents the beginning
of processing on a facility and S0, i, m is a positive integer */.
Si ,i,m Setup time for job i processed immediately after job i on manufacturing
facility m.
i, i Indices for jobs, i ∈ J and i ∈ J.
Bm Capacity restriction in a truck or Maximum batch size (given in terms of
the maximum number of jobs transported per trip) from a given facility
m /* it assumes that the capacity of vehicle available at an outsourced
facility can be different from the capacity of vehicle available at a different
outsourced facility */.
Em Carbon emission associated with each trip from manufacturing facility m
to the central facility.
EUB Upper bound on the total carbon emission, EUB = N × maxm (Em )
EL Desired limit for the total carbon emissions set by the decision makers.
EL = fr × EUB where fr denotes a fraction (0 < fr ≤ 1)
Fm Capacity of manufacturing facility m (in terms of the number of jobs that
can be processed during the scheduling period).
G A sufficiently large positive integer, where
G≥ N i=1 maxm pi,m +
N
i=1 maxi ,m si ,i,m +
N
i=1 maxm s.0,i,m
An Integrated Problem of Production Scheduling and Transportation in a Two-. . . 173
Note: n{J} = N and n{P} = M, where J is the set of jobs and P is the set of
available manufacturing facilities (both in-house and outsourced facilities included).
Binary Variables:
Δm a binary variable that takes the value 1 if manufacturing facility m
is selected for processing jobs; 0 otherwise.
Φ i, m a binary variable that takes the value 1 if manufacturing facility m
is selected and job i is assigned to facility m; 0 otherwise.
γi ,i,m a binary variable that takes the value 1 if job i is processed
immediately after job i on the manufacturing facility m to which
both jobs are assigned; 0 otherwise.
γN+m,i a binary variable that takes the value 1 if job i is processed as the
first job on the selected manufacturing facility m; 0 otherwise.
γ i, N + M + m a binary variable that takes the value 1 if job i is processed as the
last job on a selected manufacturing facility m; 0 otherwise.
θ i, b, m a binary variable that takes the value 1 if order i is assigned to
facility m and transported to the central facility in batch shipment
b; 0 otherwise.
∂ b, m a binary variable that takes the value 1 if any job is assigned to
facility m and transported to the central facility in batch shipment
b; 0 otherwise.
Decision variables:
si start time of job i.
oi occupancy time of job i /* the sum of the appropriate machine- and sequence-
dependent setup time and the processing time of job i on its assigned
machine at a manufacturing facility */.
ti completion time of job i.
t i, m time at which job i is shipped from machine m (equals its completion time ti
if processed in-house; t i, m ≥ ti for outsourced facility m).
∗
tb,m completion time of the last job assigned to batch b of machine m /* time of
shipment of batch b from facility m */.
t i time at which job i is delivered at the central facility.
μm total machine usage time of a manufacturing facility m.
τi tardiness of job i.
Objective function:
N
M
N
M
M
∗
Min Z = Wi τi + Cm m + Cm ∂b,m + Cm μm (1)
i=1 m=1 b=1 m=1 m=1
M
i,m = 1 ∀i ∈ J (2)
m=1
174 B. C. Jos et al.
N
γi ,i,m + γN+m,i = i,m ∀i ∈ J, ∀m ∈ P (3)
i =1, i =i
N
γi,i ,m + γ i,N+M+m = i,m ∀i ∈ J, ∀m ∈ P (4)
i =1, i =i
N
γN+m,i ≤ m ∀m ∈ P (5)
i=1
N
γi,N+M+m ≤ m ∀m ∈ P (6)
i=1
N
i,m ≤ N × m ∀m ∈ P (7)
i=1
N
i,m ≥ m ∀m ∈ P (8)
i=1
N
N
i,m ≤ N γN+m, i ∀m ∈ P (9)
i=1 i=1
N
N
i,m ≤ N γi,N+M+m ∀m ∈ P (10)
i=1 i=1
M
si ≥ ti − G 1 − γi ,i,m ∀i ∈ J, i ∈ J and i = i (11)
m=1
si ≤ G 1 − γN+m, i ∀i ∈ J, ∀m ∈ P (12)
N
M
M
M
oi = γi ,i,m Si ,i,m + γN+m, i S0,i,m + Pi,m i,m ∀i ∈ J
i ,i =i m m m
(13)
ti = si + oi ∀i ∈ J (14)
M
ti,m ≥ ti ∀i = J (15)
m=1
N
θi,b,m = i,m ∀i = J, ∀m ∈ P (16)
b=1
An Integrated Problem of Production Scheduling and Transportation in a Two-. . . 175
∗
tb,m ≤ G × ∂b,m for b = 1, .., N, ∀m ∈ P (17)
N
∂b,m ≤ N × m ∀m ∈ P (19)
b=1
N
m ≤ ∂b,m ∀m ∈ P (20)
b=1
N
∂b,m ≤ θi,b,m ∀m ∈ P , ∀b = 1, .., N (21)
i=1
N
θi,b,m ≤ N × ∂b,m ∀m ∈ P , ∀b = 1, .., N (22)
i=1
∗
ti,m ≥ tb,m − G 1 − θi,b,m ∀i = J, ∀b = 1, .., N, m ∈ P (23)
∗
ti,m ≤ tb,m + G 1 − θi,b,m ∀i = J, ∀b = 1, .., N, m ∈ P (24)
ti,m ≤ G × i,m ∀i = J, ∀m ∈ P (25)
M
M
ti =
ti,m + Tm i,m ∀i = J (26)
m=1 m=1
∗ ∗
tb,m ≥ t(b−1),m − G 1 − ∂b,m ∀b = 2, .., N, ∀m ∈ P (27)
∗
tb,m ≤ G × ∂b,m ∀b = 2, .., N, ∀m ∈ P (28)
τi ≥ t i − Di ∀i ∈ J (29)
μm ≥ ti,m ∀i ∈ J, ∀m ∈ P (30)
μm ≤ G × m ∀m ∈ P (31)
N
θi,b,m ≤ Bm ∀m ∈ P , ∀b = 1, .., N (32)
i=1
176 B. C. Jos et al.
N
i,m ≤ Fm ∀m ∈ P (33)
i=1
N
M
Em ∂b,m ≤ E L ∀m ∈ P , ∀b = 1, .., N (34)
b=1 m=1
i,m ∈ {0, 1} ; γi ,i,m ∈ {0, 1} ; γN+m,
i ∈ {0, 1} ; γ i,N+M+m ∈ {0, 1} ;
m ∈ {0, 1} ; ∂b,m ∈ {0, 1} ; θi,b,m ∈ {0, 1} ; ∀m ∈ P , ∀i ∈ J, ∀i ∈ J and
i = i; and all other variables are ≥ 0.Fm and Bm are integers.
(35)
Equation (1) minimizes the total costs linked with the tardiness of jobs, the total
fixed hiring costs of the manufacturing facilities selected, the total transportation
costs incurred while shipping the jobs processed at outsourced facilities back to the
central organization in batches, and the total machine usage costs of the facilities
selected. Constraint (2) reflects that a job can be assigned to at most only one
manufacturing facility. The feasibility of the sequence is maintained by Eqs. (3) and
(4). Constraints (5) and (6) guarantee that each manufacturing facility, if selected,
can have at most one start job as well as one end job. These jobs are assumed as
fictitious jobs and are named as (N + j) and (N + J + j) respectively. Constraints
(7) and (8) warrant that if at least one job is allocated to a manufacturing facility,
then the facility is active; and that a manufacturing facility is inactive if no job is
allocated to it. Constraints (9) and (10) ensure the allocation of fictitious jobs to a
manufacturing facility if any of the jobs are allocated to that facility. Constraint (11)
guarantees that the start time of any job is greater than or equal to the completion
time of its predecessor. Constraint (12) assigns the start time of the first job on each
manufacturing facility to zero. Equations (13) and (14) compute the completion time
of each job as the sum of the respective start time, setup time, and processing time.
The shipment of any job from a manufacturing facility is possible only after
the completion of that job at the respective facility. This condition is ensured by
constraint (15). Constraint (16) states that shipment of a job in any batch from
a manufacturing facility is possible only if the job is processed in that facility.
Constraint (17) ensures that completion time of the last job assigned to any batch
shipment from a facility can be considered as a positive integer, only if at least one
job is assigned to that batch shipment from the respective facility. ‘Compression
constraint’ (18) ensures that a positive batch-quantity shipment cannot follow a zero
batch-quantity shipment. Constraints (19) and (20) ensure that the batch shipment
from a manufacturing facility is possible only if that facility is active. Constraints
(21) and (22) ensure that a batch shipment from a manufacturing facility is active
only if jobs are assigned to that batch from the respective facility. Constraints (23)
and (24) state that the shipment time of a job from a manufacturing facility is equal
to its corresponding batch’s shipment time from the respective facility. Constraint
(25) ensures that the shipment time of any job from a facility is a positive integer
only if that job is allocated to that facility. Constraint (26) states that the arrival time
An Integrated Problem of Production Scheduling and Transportation in a Two-. . . 177
of a job at the central manufacturing facility is the sum of shipment time from the
manufacturing facility and the associated transportation time. Constraint (27) and
(28) ensure that the completion times of different batches from a facility are in the
non-decreasing order of batch numbers.
Expression (29) determines the tardiness of all jobs. Expressions (30) and
(31) define the total machine usage time of each manufacturing facility selected.
Constraints (32) and (33) maintain the batch capacity and manufacturing capacity
restrictions, respectively. Constraint (34) is used to keep the total carbon emission
from transportation within the desired upper limit. Finally, Expression (35) defines
all the binary variables and continuous variables.
5 A Numerical Example
In this section, we present a numerical example of the problem under study and
its solution using the proposed solution technique. Consider a supply chain system
with number of jobs (N) = 8 and number of machines (M) = 2, where in-house
machines Min = 1 and outsourced centers Mout = 1. This example is meant only for
the illustration purpose, and hence units of the input parameters are not specified.
Processing time of each job i on machine j (Pi, j ), due date associated with job i (Di ),
and tardiness cost/unit time (Wi ) are given in Table 1. Sequence-dependent setup
times (Si ,i,m ) on machines m = 1 and 2 are given in Tables 2 and 3, respectively.
Other input parameters are given in Table 4. The maximum capacity of a truck
from facility m (Bm ) is assumed as 3 jobs and is same for all m, and the maximum
production capacity (Fm ) at a manufacturing facility is 6 jobs and is same for all
m. The carbon emission limit is calculated as 80 using the equation EL = fr × EUB
where fr is assumed to be 0.5. The values of the different parameters are generated
randomly, and is described in section 6.
The given problem instance is solved using the proposed MILP model and using
IBM CPLEX optimization studio 12.7.1. The objective function value of the optimal
Table 1 Input parameters: processing time, due date, and tardiness cost
Processing times on facilities (Pi, m )
Job no. (i) Pi, 1 Pi, 2 Due date (Di ) Tardiness Cost (Wi )
1 11 12 38 12
2 6 6 39 11
3 9 10 36 15
4 10 11 37 12
5 9 10 34 10
6 3 3 34 10
7 6 6 37 12
8 9 10 41 12
178 B. C. Jos et al.
Table 2
Succeeding job (i)
Sequence-dependent setup
time (Si ,i,m ) on facility Preceding job (i ) 1 2 3 4 5 6 7 8
m=1 0 4 2 2 6 4 2 3 2
1 0 3 5 5 5 5 2 3
2 4 0 2 6 5 6 2 4
3 4 5 0 5 4 2 5 2
4 2 2 4 0 6 5 6 4
5 3 5 4 5 0 3 4 6
6 4 5 2 6 3 0 2 4
7 4 5 3 2 3 2 0 2
8 3 2 3 3 5 5 3 0
Table 3
Succeeding job (i)
Sequence-dependent setup
time (Si ,i,m ) on facility Preceding job (i ) 1 2 3 4 5 6 7 8
m=2 0 4 2 2 6 4 2 3 2
1 0 3 5 5 5 5 2 3
2 4 0 2 6 5 6 2 4
3 4 5 0 5 4 2 5 2
4 2 2 4 0 6 5 6 4
5 3 5 4 5 0 3 4 6
6 4 5 2 6 3 0 2 4
7 4 5 3 2 3 2 0 2
8 3 2 3 3 5 5 3 0
6 Computational Experimentation
2 8 2
1 3 6 7 4 1 5
5 10 15 20 25 30 35 40 45 50 55 60
Time
Fig. 3 Gantt chart representation of the optimal solution
to generate the problem sets, and then demonstrate the results of the computational
analysis obtained using the proposed solution technique.
The values of the different parameters for the integrated logistics and job
scheduling problem in a supply chain with non-identical parallel manufacturing
facilities are generated as given below. For each order i ∈ J, the processing time
Pi, m on each facility m ∈ P is an integer generated uniformly in the interval [1,
20]. The sequence-dependent setup times Si ,i,m and initial setup times S0, i, m are
also integers generated uniformly in the interval [1, 10]. Triangular inequality is
maintained while generating setup times. We assume that Pi, m depends on the
machining facility at a manufacturing center and a machine performance factor fm
determines the relative performance of the facilities, thereby the processing time of
a job. fm also influences the sequence-dependent setup times at each manufacturing
facility (Chen 2009). We consider the machine performance factor fm as a uniform
random number in the interval [2/3, 3/2]. The following equations are used to
incorporate the machine performance factor while generating the processing times
and setup times to ensure that the parameter values are within the interval
specified
earlier: Pi, m = fm × Pi, v and Si,j,m = fm × Si ,i,v where Pi,v = U 2 × 1, 32 × 20
3
and Si ,i,v = U 32 × 1, 32 × 10 . Here, v denotes a fictitious machine, Pi, v denotes
the processing time of job i on machine v, and Si ,i,v denotes the appropriate setup
times on the machine v.
The due date (Di ) is set to maxi ,m Si ,i,m + max slack, maxm Pi,m +
avgm (Tm ) , where slack is generated uniformly in the interval Pa [1 − Π − R/2
1 − Π + R/2] where Pa = N i=1
M 2
m=1 Pi,m /M (Liaw et al. 2003). The job due-
180 B. C. Jos et al.
date is designed to avoid tardiness if a given job is assigned as the first job on any
facility without the consideration of transportation time. The parameter Π is called
tardiness factor and R is called due-date range, and both parameters are chosen from
the values 0.3 and 0.7 (Potts and Van Wassenhove 1985).
For each facility m available for processing
the given jobs (i.e., m ∈ P),
the
machine usage cost per unit time Cm ∗ , the fixed hiring cost of a facility C ,
m
the transportation time necessary for jobs to ship from the outsourced center to
the central organization (Tm ) and the associated cost for transportation (Cm ) are
generated as described. We also assume that manufacturing facilities with advanced
processing techniques can process the jobs in lesser time, but with a higher per
unit machining cost and holding cost. However, machine usage costs of in-house
facilities are assumed to be the least, irrespective of the machine performance.
The transportation time, transportation cost, and the fixed holding cost are
assumed to be zero for the in-house machines. The usage cost of in-house machine
∗ ) is an integer generated uniformly in the interval [1, 5]. The parameter values
(Cin
associated with the outsourced facilities are integers and are calculated as follows.
Transportation time (Tm ) to ship the finished jobs to the central organization is
generated as a uniform random number in the interval [1, 20], corresponding
transportation cost (Cm ) is calculated as Ω × Tm and the usage cost of outsourced
∗ ) is calculated as C ∗ × ∗ . For a specific problem instance, Ω and
facility (Cout in fm
Ω ∗ are constants and are generated randomly in the interval [1, 3] and [2, 5],
respectively. This ensures the production cost to be always higher at the outsourced
centers than the in-house center, and also in the increasing order of the machine
performance factor. Note that the usage cost per unit time Cm ∗ is a constant for
CPLEX. For each problem instance, we observe the objective function value and the
execution time required for the solver to converge optimally under the given system
environment. If the problem instance has not been converged to optimum within the
stipulated time, we observe the % optimality gap reported by the solver.
Computational analysis of the proposed MILP model is reported in three
dimensions, (i) the average execution time taken to solve the test instances optimally
within the given time limit of 7200 seconds, (ii) the number of problem instances
solved to optimality within the given time limit in a set of 3 problem instances with
similar parameters, and (iii) the average optimality gap (%) reported by the solver
during termination. Results are compiled and given in Table 5.
An analysis of the experimental results given in this section shows the effective-
ness of the proposed MILP model. In small problem instances, the proposed model
could give optimal solutions within a short time. When the job size is increased to
12, the model could not converge to an optimum value within the stipulated time
7 Summary
Scheduling theory assumes great significance in the supply chain planning activities
of an organization involved with in-house and outsourced manufacturing facilities.
In this chapter, we have studied an integrated problem of logistics and scheduling
of jobs with options of outsourcing facilities and additional consideration of
transportation decisions related to transportation cost, transportation time, and
carbon emissions.
The present study has integrated three different issues in a supply chain: the
selection of subcontractors, the scheduling of jobs, and the scheduling of logistics.
The main contribution of the study is the proposed solution methodology that
integrates these three different issues which have not been considered together in
the literature so far. In addition, this study considers ‘carbon emission aspects’
associated with the shipment of finished jobs that is necessary for making the supply
chain eco-friendly.
In this supply chain system, we need to choose a set of subcontractors (also called
outsourced manufacturing facilities) from the available subcontractors to process a
part of its orders/jobs, in addition to the manufacturing facilities located in-house, to
maximize the supply chain profitability while considering carbon emissions. All the
completed jobs at the outsourced centers need to be transported back to the central
manufacturing firm. The orders/jobs to be processed are having different processing
times and due dates. We assume that the manufacturing facilities, both outsourced
and located in-house, can be non-identical and are working in parallel.
We have proposed a MILP model capable of giving optimum solutions for small
problem instances (up to job size 10) within a reasonable time. Precedence-based
binary variables are used in the proposed MILP formulation to denote the sequence
of jobs in a manufacturing facility. A numerical example and its solution have been
presented to describe the problem scenario. Computational analysis with sample
problem instances is conducted to test the robustness of the proposed solution
An Integrated Problem of Production Scheduling and Transportation in a Two-. . . 183
References
Alidaee, B., & Li, H. (2014). Parallel machine selection and job scheduling to minimize
sum of machine holding cost, total machine time costs, and total tardiness costs. IEEE
Transactions on Automation Science and Engineering, 11(1), 294–301. https://doi.org/10.1109/
TASE.2013.2247757.
Azadian, F., Murat, A., & Babu, R. (2015). Integrated production and logistics planning: Contract
manufacturing and choice of air/surface transportation. European Journal of Operational
Research, 247(1), 113–123. https://doi.org/10.1016/j.ejor.2015.06.004.
Cai, X., Lee, C. Y., & Vairaktarakis, G. L. (2005). Optimization of processing and delivery deci-
sions involving third-party machines. Nonlinear Analysis, Theory, Methods and Applications,
63(5–7), 2269–2278. https://doi.org/10.1016/j.na.2005.03.067.
Cao, D., Chen, M., & Wan, G. (2005). Parallel machine selection and job scheduling to minimize
machine cost and job tardiness. Computers & Operations Research, 32(8), 1995–2012. https://
doi.org/10.1016/j.cor.2004.01.001.
Chen, J. F. (2009). Scheduling on unrelated parallel machines with sequence- and machine-
dependent setup times and due-date constraints. International Journal of Advanced Manufac-
turing Technology, 44(11–12), 1204–1212. https://doi.org/10.1007/s00170-008-1917-3.
Chopra, S., Meindl, P., & Kalra, D. V. (2016). Supply chain management: Strategy, planning, and
operation (6th ed.). Harlow: Pearson.
Fang, K., Uhan, N., Zhao, F., & Sutherland, J. W. (2011). A new approach to scheduling in man-
ufacturing for power consumption and carbon footprint reduction. Journal of Manufacturing
Systems, 30(4), 234–240. https://doi.org/10.1016/j.jmsy.2011.08.004.
Kahn Ribeiro, S., Kobayashi, S., Beuthe, M., Gasca, J., Greene, D., Lee, D. S., Y. Muromachi, P.
J. Newton, S. Plotkin, D. Sperling, R. & Wit, R. (2007). Transport and its infrastructure. In:
Netz, B., Davidson, O. R., Bosch, P. R., Dave, R., & Meyer, L. A. (eds) Climate change 2007:
mitigation. Cambridge University Press, Cambridge, UK, pp 324–385.
184 B. C. Jos et al.
Liaw, C.-F., Lin, Y.-K., Cheng, C.-Y., & Chen, M. (2003). Scheduling unrelated parallel machines
to minimize total weighted tardiness. Computers and Operations Research, 30, 1777–1789.
https://doi.org/10.1016/S0305-0548(02)00105-3.
McKinnon, A. (2010). The role of government in promoting green logistics. In A. McKinnon,
S. Cullinane, M. Browne, & A. Whiteing (Eds.), Green logistics: Improving the environ-
mental sustainability of logistics, Kogan Page, London, 341–360. doi:https://doi.org/10.1080/
01441647.2010.537101.
Mokhtari, H., Nakhai, I., Abadi, K., & Amin-Naseri, M. R. (2012). Production scheduling
with outsourcing scenarios: A mixed integer programming and efficient solution procedure.
International Journal of Production Research, 50(19), 5372–5395. https://doi.org/10.1080/
00207543.2011.627687.
Oğuz, C., Salman, F. S., & Yalçin, Z. B. (2010). Order acceptance and scheduling decisions
in make-to-order systems. International Journal of Production Economics, 125(1), 200–211.
https://doi.org/10.1016/j.ijpe.2010.02.002.
Pinedo, M. L. (2012). Scheduling: Theory, algorithms, and systems (4th ed.). New York: Springer.
Potts, C. N., & Van Wassenhove, L. N. (1985). A branch and bound algorithm for the Total
weighted tardiness problem. Operations Research, 33(2), 363–377.
Qi, X. (2008). Coordinated logistics scheduling for in-house production and outsourcing. IEEE
Transactions on Automation Science and Engineering, 5(1), 188–192.
Ritchie, H., & Roser, M. (2017). CO2 and greenhouse gas emissions. Our World in Data. Retrieved
June 22, 2020, from https://ourworldindata.org/co2-and-other-greenhouse-gas-emissions
Ruiz-Torres, A. J., López, F. J., Ho, J. C., & Wojciechowski, P. J. (2008). Minimizing the average
tardiness: The case of outsource machines. International Journal of Production Research,
46(13), 3615–3640. https://doi.org/10.1080/00207540601158799.
Unlu, Y., & Mason, S. J. (2010). Evaluation of mixed integer programming formulations for
non-preemptive parallel machine scheduling problems. Computers and Industrial Engineering,
58(4), 785–800. https://doi.org/10.1016/j.cie.2010.02.012.
Wang, H., & Alidaee, B. (2018). Unrelated parallel machine selection and job scheduling with the
objective of minimizing total workload and machine fixed costs. IEEE Transactions on Automa-
tion Science and Engineering, 15(4), 1955–1963. https://doi.org/10.1109/TASE.2018.2832440.
Wang, J., Yao, S., Sheng, J., & Yang, H. (2019). Minimizing total carbon emissions in an integrated
machine scheduling and vehicle routing problem. Journal of Cleaner Production, 229, 1004–
1017. https://doi.org/10.1016/j.jclepro.2019.04.344.
Yang, P., & Tang, J. (2013). Minimizing carbon emissions through vehicle routing and scheduling
in the shuttle service of picking up and delivering customers to the airport. Acta Automatica
Sinica, 39(4), 424–432. https://doi.org/10.1016/S1874-1029(13)60042-7.
Yu, Y., Tang, J., Li, J., Sun, W., & Wang, J. (2016). Reducing carbon emission of pickup and
delivery using integrated scheduling. Transportation Research Part D, 47, 237–250. https://
doi.org/10.1016/j.trd.2016.05.011.
A Simulation-Based Evaluation of Drone
Integrated Delivery Strategies for
Improving Pharmaceutical Service
1 Introduction
Over the past decade, the market has seen increased growth owing to direct-to-
consumer (D2C) deliveries; more specifically, e-commerce has grown between 7
and 10 percent in developed markets, and 300% in developing ones (Dayarian
et al. 2020). Shopping in e-commerce is primarily focused on speed, flexibility,
security, and cost of delivery. Notably, most online shoppers are focused on the
speed of delivery while remaining price sensitive. It is estimated that around 50%
of U.S adults are more likely to purchase if same-day delivery is offered, and
40% expect the option to be available (Epps and Jordan 2015). However, 88% of
consumers prefer free shipping over faster shipping (Deloitte 2018). As consumers
balance their demand for cheap and quick delivery, retailers must bear most of
the costs of delivery. This has caused retailers to search for cheap alternatives to
provide a speedy delivery service (Dayarian et al. 2020). Furthermore, this sort
of environment provides a near perfect opportunity for drone applications that can
serve the commercial needs of retailers and consumers.
Traditional delivery systems that employ trucks are subject to various factors
that hurt consumers and limit their effectiveness. Trucks are limited by a region’s
A. Jackson ()
Department of Industrial and Manufacturing Systems Engineering, University of Missouri,
Columbia, MO, USA
e-mail: atjb47@mail.missouri.edu
S. Srinivas
Department of Industrial and Manufacturing Systems Engineering, University of Missouri,
Columbia, MO, USA
Department of Marketing, Trulaske College of Business, University of Missouri, Columbia, MO,
USA
e-mail: srinivassh@missouri.edu
© Springer Nature Switzerland AG 2021 185
S. Srinivas et al. (eds.), Supply Chain Management in Manufacturing and Service
Systems, International Series in Operations Research & Management Science 304,
https://doi.org/10.1007/978-3-030-69265-0_7
186 A. Jackson et al.
Table 1 Summary of drone Drone company Payload (kg) Range (km) Speed (km/h)
specifications
Matternet 2 10 40
DHL Parcel 4 65 130
Zipline 1.75 80 100
infrastructure as their only way to deliver their products is by road. This increases
delivery times as there is often traffic or congestion blocking the roads. Secondly,
trucks require a high investment cost to purchase the vehicles as a typical delivery
truck costs $50,000 (Hawes 2016). Thirdly, trucks are burdened by high mainte-
nance costs for operation as the average cost of operating a truck is $0.08 cents per
mile (Barnes and Langworthy 2004); Furthermore, they require hundreds of dollars
in terms of fuel costs. Lastly, trucks also have high labor costs associated with them
as each vehicle requires a human operator. UPS drivers are paid up to $25 an hour
(Welch 2015), which translates to $52,000 per year. All these costs can be saved by
considering alternate delivery methods, especially drones.
Drones made by companies such as Matternet are unmanned and battery-
powered flying vehicles that use multiple rotors, but drones made by DHL Parcel
and Zipline, while unmanned and battery-powered, are fixed winged and designed
to be like small planes. Their sizes are typically small but can vary depending
on their type and use. Recently, drones have been typically used for surveillance,
photography, entertainment, civilian, and military purposes. Innovations in drone
hardware, software, and networks have allowed for the continued exploration of
drone utilization. Furthermore, the different payloads, range, and speeds of different
types of drones currently in use, as shown in Table 1, adapted from Scott and Scott
(2017), Edenhofer (2018), Zipline (2020), have allowed these advances to serve
market demand.
Drones have the potential to lower labor costs, decrease delivery times, and
reduce maintenance costs as, in general, they are less expensive to maintain and can
operate autonomously (Dorling et al. 2016). Furthermore, drones are not limited to
traditional infrastructure such as roads; hence, face less obstacles in their delivery
networks. This allows for a speedy and cost-effective way for goods to be delivered.
Although commercial delivery with drones has many uses, commercial utilization
is not the only service that drones can provide to society.
Medication delivery for healthcare can greatly benefit from the implementation
of drone technology. Especially during emergencies, drones can deliver vital
supplies such as food, water, and medications to disaster areas by flying over
damaged infrastructure (Dorling et al. 2016). The timely delivery of medications,
blood, and vaccines is critical in healthcare, and some innovative organizations
have begun to use drones for healthcare services. Matternet is one such company
that uses autonomous drones to deliver medical supplies (Scott and Scott 2017).
Although emergencies are a viable application for drones, they do not occur enough
to warrant the quick implementation of drones. Pharmacy prescriptions, on the other
hand, offer a unique opportunity for possible drone delivery as quick delivery can
A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 187
better serve the community, and 47% of consumers are interested in a drone-based
delivery service (ReportLinker 2016). The most significant challenge facing drone
distribution networks is the optimization of different drone delivery strategies. Every
region has different variables that can affect the strategy implementation for drone
deliveries such as infrastructure, urban density, and terrain. However, one strategy
utilized by Matternet is the “truck-drone tandem” delivery system.
A truck-drone tandem delivery system focuses on the drone and delivery truck
working together to deliver their goods. Since the drone has a limited battery life
and range, the delivery truck enables the drone to be launched from the truck at
locations from where the drones can effectively complete their deliveries. This
method allows for long-distance deliveries as it relies on the truck for the initial
leg of the delivery, and only allows the drone to take off from the truck for the
“last-mile” of the delivery. Furthermore, this method allows for more packages to
be delivered farther away from the distribution centers as the truck can hold multiple
packages. However, many issues also arise from the tandem method. The initial leg
of the trip is still subjugated to all the problems of a traditional delivery method
such as poor infrastructure and traffic, high maintenance and labor costs, and high
delivery times. Therefore, the tandem method may be optimal in certain scenarios,
but other methods need to be considered to have the most effective delivery system
with regards to delivery time and cost.
A drone delivery method that solely utilizes drones for the delivery of goods can
also be simulated. The Drone-Only method bypasses the limitations of the truck
entirely; therefore, it can be more optimal in certain regions. A Drone-Only method
will eliminate the limitation of infrastructure, long delivery times, and reduce the
capital requirements, maintenance, and labor costs associated with trucks. However,
a greater number of drones will be required to serve market demand, and the delivery
distances will be restricted by the drone’s limited battery life and range.
This paper seeks to determine the most optimal drone delivery method for
the delivery of medical prescription drugs with regards to delivery time and
cost. A discrete-event simulation is utilized for a variety of scenarios, and the
different delivery methods are compared. Furthermore, the optimal number of
drones required for different areas are determined, and a cost-benefit analysis is
performed against traditional delivery trucks. This will provide the information
needed for municipalities to determine the validity of a drone integrated delivery
system.
2 Literature Review
Research has been conducted on multiple fronts for the implementation and applica-
tion of drone technology for delivery services. Drone routing studies such as Rabta
et al. (2018), Ha et al. (2018) focus on developing mathematical models to explore
the delivery networks where drones work in tandem with trucks. While they focus
on different aspects of the Truck-Tandem strategy, they also build upon research
that has not been conducted before. Furthermore, Rabta et al. (2018) focuses on the
188 A. Jackson et al.
3 Problem Statement
Prescription drugs are life-saving medications that treat people with a variety
of diseases, and, currently, prescriptions filling requires customers to come to
the pharmacy to pick up the drugs. However, different regions have different
infrastructure and variables that affect commute times for customers and can lead to
customer dissatisfaction. While customer dissatisfaction may not be blamed on the
pharmacies themselves, a pharmacy offering prescription drug delivery may attract
new customers or increase customer loyalty. Belfast, Northern Ireland of the United
Kingdom, was named UK’s worst city for traffic congestion in 2016 as citizens
spend an extra 87% of their actual travel time stuck in traffic (Index, TomTom Traffic
2016). Therefore, Belfast provides a unique opportunity for pharmacies to service
A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 189
4 Methodology
Start
Prescriptions
are filled
No
Is it time for a
Yes batched to be
released
Release
drones to Is location Deliver
No Yes
deliver a demand Prescription Yes
Last-Mile point
No
Does vehicle
hold more
prescriptions
Yes
No
Vehicle
returns to
pharmacy
No
Yes
End
Daily Orders
trucks depreciate at 8.33% a year, and drones lose roughly 50% of their value each
year (Australian Taxation Office 2019), and so drones must be replaced frequently,
increasing their costs.
Multiple functions are in place to ensure that the model works properly. “Model
trace” provides the live steps the model takes in its logic processing. The trace allows
the user to follow along as the model runs its processes and can precisely pinpoint
where errors occur. Secondly, a “pause” function is built within the model to pause
the model whenever a minor error occurs. These errors only interrupt a small portion
of the model, and the model can circumvent them to continue operation. Lastly, the
model has a “stop” function that completely stops the model when a major error
occurs. These errors affect the model as a whole and typically involve flaws in the
logic processes. However, all errors can be pinpointed through the model trace, and
a solution can be implemented accurately.
Validation
6 Alternative Scenarios
As shown in Fig. 3, the demand points are randomly generated through the software
within the limitations of the city. The trucks start from the pharmacy and take
the shortest path to visit every demand point before returning to the pharmacy.
Before that, the trucks wait for the prescriptions to be batched together. When a
batch is ready, the prescriptions are loaded onto a truck, and the truck departs for
194 A. Jackson et al.
delivery. The truck navigates the road to its appropriate demand points. Once the
truck reaches a demand point, 1.5 min is allotted for the truck to make the delivery,
and then the truck calculates the next demand point that needs service and sets
out to make its delivery. The truck repeats the process until it has delivered all the
prescriptions and then returns to the pharmacy. After returning to the pharmacy, the
truck checks to see if more prescriptions in the batch need to be delivered; if not,
the truck waits for the next batch to be ready.
As shown in Fig. 4, the demand point clusters are randomly generated through the
software within the limitations of the city. The trucks start at the pharmacy and wait
for the prescriptions to be batched together to make their deliveries. When a batch
is ready, the truck travels to a designated cluster center where it launches the drone.
The drone makes the last-mile flight to the demand point to complete the delivery.
Once the drone takes off, the drone flies above obstacles and in a direct line to a
cluster’s demand points for quick delivery, where 1.5 min is allotted. The drone
delivers the product at the demand location and returns to the truck. After the drone
A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 195
has returned to the truck, the truck proceeds to calculate the next cluster that it needs
to serve. The truck determines the quickest route to the next cluster center point and
travels by road to reach it. Upon reaching the center point, the drone delivery process
is repeated until all demand points in the cluster are served. The truck continues this
operation until it has delivered all its prescriptions and then returns to the pharmacy.
After the truck has returned to the pharmacy, it determines if more prescriptions
from the batch need to be delivered. If so, it will set out again to deliver the order,
but if not, it will wait until the next batch is ready.
As shown in Fig. 5, the demand point locations are randomly generated within the
limitations of the drone range. The drones begin at the pharmacy and fly a direct
196 A. Jackson et al.
route to each demand point location to complete the delivery before returning to the
pharmacy. Before leaving the pharmacy, the drones wait for the prescriptions to be
batched together to make their deliveries. When a batch is ready, the prescriptions
are loaded onto the drones, and the drones depart for delivery. Drones will be loaded
only with 1 prescription to be delivered to a demand point. The drone then takes off
and travels in a straight path to the demand point. Upon arrival, the drone will deliver
the prescription and then return to the pharmacy to pick up another prescription. The
process is repeated until all prescriptions are delivered in a batch.
7 Results
The parameters of the models are implemented in Simio OptQuest with a warm-up
period of 100 h. 300 simulations with 5 replications are used to achieve the results
with a 95% confidence interval. OptQuest experimental setup in Simio considered
1–15 trucks with capacities ranging from 1–80 prescriptions for scenario 1, 1–15
A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 197
trucks with capacity ranging from 1–80 prescriptions and 1–5 drones for Scenario
2, and 1–30 drones for Scenario 3 to achieve its results. Note that the capacity of
drones is considered as 1 prescription per delivery.
As shown in Table 6, the delivery method with 4 drones proves to be the cheapest
feasible option. The batch delivery time for 4 drones is 2.69 h and requires an
investment of $20,000. However, yearly cost accounts for $49,602.5, mostly due
to the depreciation of the drones.
for each scenario to illustrate the cost increase and batch delivery time reduction
with additional vehicles.
The initial analysis determined the least cost option while remaining feasible and the
most effective options given additional vehicles. However, as demand grows with
time, the impact on delivery time due to demand variation needs to be determined.
Figure 6, below, shows the impact on batch delivery time for both the most effective
and low-cost option for each scenario and includes a 3.25-h limitation to ensure that
the daily orders can be completed. For the most effective option, variation below
the current average has little to no impact. However, as the demand increases, batch
delivery times for Truck-Tandem increase dramatically, and the current settings is
no longer the most effective. The option now requires an adjustment in the number
of trucks and truck capacity to become the most effective again. Furthermore, as the
demand increases, the batch delivery times for all scenarios in the low-cost option
increase steadily. However, both Truck-Only and Drone-Only become infeasible and
will require an additional truck or drone.
Lastly, as shown in Fig. 7, the impact on batch delivery times due to variation
in the number of batches a pharmacy decides to operate with is determined. For the
most effective option, an increase in the number of batches above the current setting
has little impact. However, decreasing the number of batches, vastly increases the
delivery times for both Truck-Only and Truck-Tandem. For the low-cost option,
decreasing the number of batches vastly increases the delivery time for Truck-Only
and Drone-Only, and will require additional trucks or drones to remain effective.
However, increasing the number of batches, steadily decreases the batch delivery
2.5 3.5
3
2 2.5
1.5 2
1 1.5
1
0.5 0.5
0 0
215 265 315 365 415 215 265 315 365 415
Daily Demand Daily Demand
Fig. 6 Comparing the impact of daily demand variation on batch delivery times with the 3.25-h
limitation for the 3 scenarios. The red line represents the upper limitation of batch delivery time to
complete the daily delivery
200 A. Jackson et al.
Fig. 7 Comparing the Impact of the number of batches on batch delivery time with 3.25-h
limitation between the 3 scenarios. The red lines illustrate the batch delivery time limitation for
each number of batches
times for all scenarios. Nevertheless, the low cost option for Truck-Only no longer
becomes feasible as the trucks are no longer able to deliver the entire batch before
the next delivery batch time.
7.6 Discussion
In this section, the different scenarios are compared with respect to their cost, ease
of implantation, performance, risk, and coverage. These comparisons are formed
into a ranking system that a pharmacy manager can utilize for the decision-making
process. As each scenario has potential benefits and problems, it is to the pharmacy
manager to decide which is best for their pharmacy.
Cost Each scenario requires costs to operate, and scenario 3 requires the least. Due
to the relatively low costs of drones, scenario 3 is the cheapest option, followed by
scenarios 2 and 1. Scenario 1 requires the most investment as trucks require a high
investment and accrue higher costs as they are used. Scenario 2 is a combination of
a truck and drone delivery system, and it serves as a middle ground between the two.
Scenario 2 requires fewer capital expenditures as its assets are shifted from trucks
and into drones. Furthermore, its costs are reduced as its labor and fuel costs are cut,
and maintenance cost is reduced.
Ease of Implementation Each scenario will require planning to implement, and
scenario 1 contains the least number of obstacles to overcome. Trucks are vehicles
that have been a part of society for decades, and a Truck-Only system will be the
easiest to implement. Scenario 1 can be implemented almost immediately; however,
scenario 3 will be the most difficult. As drone technology is relatively new, politics
A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 201
and regulations offer substantial hurdles for companies to overcome. There are
many concerns regarding drones that have led regulatory government bodies such
as the Federal Aviation Administration to prevent the implementation of drones
until further research could be completed. Scenario 2 again would serve as a middle
ground between scenario 1 and 3 as it is a mixture of the two. However, scenario
2 will run into the same regulatory hurdles as Scenario 3; therefore, its ease of
implementation is still behind Scenario 1.
Performance Each scenario can deliver batch orders promptly; however, scenario
2 is the quickest with regards to its low-cost option that remains feasible. However,
scenario 3 becomes the best performing if there are large amounts of drones.
As drones are not limited by infrastructure or traffic, they can fly over many
obstacles that would normally inhibit trucks. Therefore, scenario 3, with many
drones, provides the quickest delivery times for a pharmacy. Due to the relatively
low cost of operating drones, significantly more drones can be purchased, reducing
the delivery times. Furthermore, scenario 2 is limited by space in the truck and
cannot hold a large number of drones needed to increase its performance, but it will
remain more stable than scenario 1 as the demand grows. Lastly, scenario 1 requires
more trucks to be competitive with scenario 2.
Risk Scenario 3 contains significant amounts of risk as drones are still developing,
and the risk is exacerbated with additional drones. There is a safety risk as drones
can cause mass damage if not operated appropriately, especially around airports.
Furthermore, citizens are concerned about the risk drones pose to their privacy as
drones are capable of circumventing traditional barriers. Moreover, it also faces
a loss of property risk as armed citizens may destroy the drones they suspect of
wrongdoing, or drones could malfunction, causing a crash. Political risk is abundant
as regulatory agencies are unsure of allowing drones to operate. Lastly, operational
risk is prevalent as technology and code are prone to malfunction, and packages
may be delivered to the wrong location. However, many of these risks can be
mitigated as technology and public education improves, and regulations become
up-to-date. Scenario 1 bears the least amount of risk as trucks are a well-regulated
technology and are confined to a predetermined path with its own regulations.
However, scenario 1 bears some risk as driving conditions may be unpredictable
and human error too is a factor. Scenario 2 bears the most risk as it contains both
trucks and drones; therefore, combining the risks.
Coverage Scenario 3 covers the least amount of territory as drones are limited in
range. Due to their battery lives, drones are restricted to serve a small area around
a pharmacy and require frequent recharges. Scenario 1 and 2 operate in a more
extensive territory as the ranges are dependent on trucks. However, Scenario 2
covers the most range as it can launch a drone if a demand location is beyond the
max range of a truck.
202 A. Jackson et al.
The model scenarios provide insights into how a drone delivery service can be
implemented at pharmacies. However, it is up to the pharmacy managers to deter-
mine which scenario is more suitable for them. Pharmacy managers are assumed
to know what is best for their pharmacy, especially as they may have knowledge of
unknown variables not incorporated into this model. However, rankings can be given
for the manager to use to support them in decision making. As shown in Table 8, no
scenario is dominant. Each scenario has its benefits and drawbacks, and it is up to
the pharmacy manager to prioritize the rankings.
Managers prioritizing a simple fix can utilize scenario 1 and avoid scenario 3.
Scenario 1 requires the least number of obstacles to overcome, and scenario 3 will
require the most effort to implement. Furthermore, if a manager requires the least
amount of risk, scenario 1 can be utilized, and scenario 2 should be avoided. As
trucks are the simplest to implement, they contain the least amount of risk. However,
trucks bear some risk, and Scenario 2 adds that to the significant amount of risk
drones carry. If a manager wishes to cover the most amount of territory, Scenario
2 can be utilized, and scenario 3 should be avoided. The combination of truck and
drone allows Scenario 2 to cover the most territory, and the limitations of drone
range confine Scenario 3 to a small area. Lastly, if a manager has a limited budget
or wants the best performing system, Scenario 3 should be adopted, and Scenario 1
needs to be avoided. Drones perform the best within their range, and their low costs
allow multiple drones to be purchased while remaining under budget. However,
trucks require high capital expenditures and accrue high costs due to labor, fuel,
and maintenance.
with regards to delivery time and cost. The model suggests that a Drone-Only
system is the most effective if there are a high number of drones, and the system
still remains cheap and feasible. While each scenario can fulfill the demand, the
different scenarios offer unique benefits and drawbacks that balance each other out.
A pharmacy manager will be required to prioritize the rankings and determine which
scenario is the best fit for them.
This study has its limitation, and future work can improve the current model.
Factors considering truck routing and drone features are not considered. Truck
routing can be optimized with regards to the shortest path between demand points to
calculate the true delivery times considering truck routing strategies. Furthermore,
partnering with a pharmacy to gather real patient locations can make the model
reflect real-world data more adequately. Next, factors such as package weight can
be researched to determine how the weight will affect drone speed and battery life.
Furthermore, different batteries can also be researched to determine the interval time
between recharges.
References
Australian Taxation Office. (2019). TR 2019/5 - Income tax: effective life of depreciating assets
(applicable from 1 July 2019).
Barnes, G., & Langworthy, P. (2004). Per Mile Costs of Operating Automobiles and Trucks.
Transportation Research Record, 1864(1), 71–77.
Dayarian, I., Savelsbergh, M., & Clarke, J.-P. (2020). Same-day delivery with drone re-supply.
Transportation Science, 54(1), 229–249.
Deloitte. (2018). Deloitte study: Holiday Shoppers Hooked on fast and free shipping.
Dorling, K., et al. (2016). Vehicle routing problems for drone delivery. IEEE Transactions on
Systems, Man, and Cybernetics: Systems, 47(1), 70–85.
Edenhofer, A. (2018). DHL Parcelcopter, Deutsche Post DHL Group.
Edwards, R. (2011). In struggle to cut expenses, hospitals eye the pharmacy. Hospitals & Health
Networks, 85(11), 28.
Eisenbach Consulting. Electricity rates by state (updated April 2020)”, © 2002–2017 Eisenbach
Consulting, LLC.
Epps, N., & Jordan, E. (2015). Coldwell Banker commercial affiliates survey reveals consumer
expectations for same-day delivery, use of mobile wallet shaping retail experience, Coldwell
Banker Commercial.
Figliozzi, Mi. (2018). How sustainable are drone (UAV) deliveries? TREC Friday Seminar Series
149.
Ha, Q. M., et al. (2018). On the min-cost traveling salesman problem with drone. Transportation
Research Part C: Emerging Technologies, 86, 597–621.
Hawes, C. (2016). UPS orders 200 workhorse group hybrid delivery trucks, Trucks.com.
Hong, I., Kuby, M., & Murray, A. (2017). A deviation flow refueling location model for continuous
space: A commercial drone delivery system for urban areas. Advances in Geocomputation,
2017, 125–132.
Index, TomTom Traffic. (2016). Tomtom.com.
Kim, E. (2016). The most staggering part about Amazon’s upcoming drone delivery service.
Retrieved from http://www.businessinsider.de/cost-savings-from-amazon-dronedeliveries-
2016-6
204 A. Jackson et al.
Matternet: delivery drones that are delivering now. (2015). Nanalyze.com. Retrieved December 3,
2015, from http://www.nanalyze.com/2015/12/matternet-delivery-drones-that-are-delivering-
now/.Published
Murray, C. C., & Chu, A. G. (2015). The flying sidekick traveling salesman problem: Optimization
of drone-assisted parcel delivery. Transportation Research Part C: Emerging Technologies, 54,
86–109.
Rabta, B., Wankmüller, C., & Reiner, G. (2018). A drone fleet model for last-mile distribution in
disaster relief operations. International Journal of Disaster Risk Reduction, 28, 107–112.
ReportLinker. (2016). From warehouse to front door: Drones deliver the goods, Consumer
Surveys, Latest Articles & Indicators, Transportation Industry.
Scott, J., & Scott, C. (2017). Drone delivery models for healthcare. In Proceedings of the 50th
Hawaii international conference on system sciences.
United Kingdom. (2019). GP prescribing data, Government License.
United Postal Service. (2017). UPS expands investment in vehicle collision mitigation technology,
Press Room.
United States Energy Information Administration. (2019). Gasoline and diesel fuel update.
Wei, F., & Figliozzi, M. (2012). Conventional vs electric commercial vehicle fleets: A case study
of economic and technological factors affecting the competitiveness of electric commercial
vehicles in the USA. Procedia – Social and Behavior Sciences, 39, 702–711.
Welch, A. (2015). A cost-benefit analysis of Amazon Prime Air.
Zipline. (2020). How-it-works. Retrieved September 11, 2020, from https://flyzipline.com/how-it-
works
Pro-Active Strategies in Online Routing
Stefan Bock
In this paper, we summarize the basic ideas of recent approaches for the efficient
controlling of urgent delivery processes in real-time. As in real-world applications
the requests to be serviced the same day are usually unknown and will occur
dynamically, such delivery processes possess a considerable degree of dynamism
that has to be handled by a real-time approach. For this purpose, the transportation
plan that is already in execution is continuously adapted by applying suitable
optimization approaches according to all decisions that are not implemented. Due
to the assumed urgency of the considered requests, the aim of these approaches is to
minimize the total weighted request response times in order to minimize resulting
customer inconveniences. Through the exploitation of past request data, stochastic
knowledge about future request occurrences is derived that enables significant
reductions of the request response times if the given data possesses a certain degree
of diversity. Furthermore, empirical analyses reveal that the degree of structural
diversity enables reliable forecasts concerning the positive impact of integrating
stochastic knowledge derived from the given past request data. More recently, this
approach has been extended for identifying recurring request patterns in real time
in order to improve the quality of forecasts concerning future request arrivals. In
addition to that, it has been shown that frequent en route diversions executed during
iterative plan adaptations can be strongly limited without significantly worsening
the attained request response times. Note that an unlimited exhaustive usage of en
route diversions may cause an increase of accidents due to driver distractions.
S. Bock ()
Business Computing and Operations Research, Schumpeter School of Business and Economics,
University of Wuppertal, Wuppertal, Germany
e-mail: sbock@winfor.de
1 Introduction
This paper considers the well-known Vehicle Routing Problem (VRP). To the best
of the author’s knowledge (see also Pillac et al. (2013)), the VRP was originally
introduced as a generalization of the Traveling Salesman Problem by Dantzig
and Ramser (1959). Usually, it is defined on a graph G = (V, E, C), with V =
, . . . , vn} (the
{v0 , v1 set of all customer vertices
v1 , . . . , vn and the depot v0 ),
E = vi , vj | vi , vj ∈ V × V ∧ i = j (the set of transportation edges), and
C : E → R (the costs of using a transportation edge). For reasons of simplicity, the
cost function C is frequently defined as a matrix Ci,j (with 0 ≤ i, j ≤ n and i = j) such
that Ci,j gives the costs for traveling from vi to vj . In order to service all n customer
vertices, a fleet of K homogeneous vehicles is available. Hence, the aim of the VRP
is to find K cyclical tours that start and end at the central depot v0 while together
visiting all customer vertices at minimum total transportation costs. The literature
provides various extensions of this basic formulation (see Pillac et al. (2013) or
Ferrucci (2013), pp. 32–38). For instance, the Capacitated-VRP (CVRP) extends the
VRP by limiting the loading of each vehicle by imposing a maximum capacity while
each customer has a predetermined demand. Moreover, the urgency or availability
of some customers is mapped by integrating time windows at the customer vertices.
As a consequence, a vehicle tour cannot service a customer before the opening of
the respective window while servicing a customer after the end of the time window
is either forbidden (hard time windows) or penalized (soft time windows). Hence,
the respective variants of the VRP are denoted as the VRPTW or as the VRPSTW.
This paper considers routing problems that deal with the delivery of urgent
goods or services. Hence, material or immaterial goods (e.g., spare parts, repair
services, specific technical support, or a combination of these examples) have to be
delivered while an immediate delivery is of utmost importance to the customer. For
instance, such deliveries become necessary after a customer complaint due to an
unsuccessful first delivery or if this first delivery contains erroneous parts. In these
exemplary cases, only an instant subsequent delivery of the needed goods or services
in real time may limit the occurred customer inconvenience. Hence, in the literature
these processes are frequently denoted as the distribution of perishable goods and
the respectively considered real-world processes are named Real-Time Distribution
of Perishable Goods (RDOPG) applications (see Ferrucci (2013), pp. 1–2). As a
consequence, Ferrucci et al. (2013) propose the usage of a specifically defined
objective function that pursues the finding of tour plans ζ minimizing the customer
inconvenience z(ζ ). For each request i with a predefined weight wi , this customer
inconvenience is operationalized as a function F(ti ) of the resulting request response
time ti . This response time ti measures the time between request occurrence and
request delivery. According to the results published by Davis and Maggard (1990)
and by Kristensen et al. (1992), Ferrucci et al. (2013) model customer inconvenience
as a linear as well as a quadratic function of the resulting request response times.
The exemplary courses of both functions are sketched by Fig. 1.
Pro-Active Strategies in Online Routing 207
Fig. 1 Illustration of the two customer inconvenience functions (linear2X and quadratic) modelled
and analyzed by the approach of Ferrucci et al. (2013) (see Ferrucci (2013), p. 82)
with (x) = 1 for x > 0 and 0 otherwise see Ferrucci et al. (2013 p. 132).
Another important issue, whose integration allows for covering various practical
applications by the Vehicle Routing Problem, deals with the restricted or delayed
availability of information. In contrast to the standardized deterministic VRP where
all information is available, static, and reliable throughout the ongoing transporta-
tion process, dynamic variants of the VRP (abbreviated as DVRP) distinguish
themselves by the fact that a considerable amount of data is not reliably given
beforehand, but becomes available or changes during the execution of the transports.
This complicates an efficient execution of the process as unexpected events like
new requests or congestions may occur and have to be handled in real time and
simultaneous to the execution of the ongoing transports. As illustrated in Fig. 2,
the service of dynamically incoming requests has to be integrated into the existing
vehicle tours that are already in execution. For this purpose, the existing and
208 S. Bock
Fig. 2 Illustration of the Dynamic Vehicle Routing Problem with newly incoming requests
Fig. 3 Possible information flow within a real-time control approach (see Bock (2010), p. 734)
already executed tour plan has to be adapted. Note that the significant progress of
communication and information technologies makes it possible to control a fleet of
vehicles in real time with a centralized approach that is continuously informed about
the current situation in the network in order to handle incoming events by necessary
plan adaptions.
These adaptations are in turn communicated to the respective vehicles. The
possible information flow within a real-time control concept is depicted by Fig.
3.
The task of the central database component is to map the current situation of the
controlled transportation process. For this purpose, recent events affecting the state
of the vehicle fleet or of the transportation network are transferred by respective
services and are directly integrated into the mapped situation. As a result, a static
optimization problem is derived that defines the current state of the ongoing process
by respective parameters (e.g., current vehicle positions, vehicle states, or known
requests), whereas all changeable decisions constitute the variables of this model.
In order to handle incoming events and to ensure an efficient process execution, the
Pro-Active Strategies in Online Routing 209
real-time control has to continuously adapt the plan that is already in execution.
Note that the concurrency of plan execution and adaptation requires a sophisticated
update handling as – due to the complexity of the considered optimization problem –
plan adaptations cannot be done in zero time, but require substantial computational
time.
One main characterizing aspect of each DVRP-instance that significantly drives
the complexity of efficiently controlling the considered routing process is the
ratio of unexpectedly occurring events and the time pressure under which these
events have to be handled. Depending on the specific setting of the respectively
considered DVRP, the literature provides different so-called degrees of dynamism
(dod) measuring this aspect. According to Lund et al. (1996), the dod is originally
defined by δ = nd /ntot . In this computation, nd is the number of unexpected
dynamically occurring requests and ntot defines the total number of all requests
(static and dynamic ones) to be satisfied by the tour plan. In order to additionally
integrate the time pressure under which the dynamically incoming requests have to
be handled, Larsen (2000) extended the dod δ to the effective dod that is defined by
nd ria
δe = i=1 T ·ntot . In this calculation, ri ≤ T gives the arrival time of request
a
i, while T provides an upper bound of all request arrival times. If, in addition,
time windows exist, Larsen (2000) recommends applying the effective dod with
time windows δ etw instead. This measure sums up the difference between T and
the remaining response time before the closing of the respective time window ril
T − r l −r a
over all dynamically incoming requests, i.e., it holds δ et w = ntot
1
· ni=1
d i
T
i
.
Hence, this definition measures the time pressure for handling each event by the
maximum remaining time before reaching the end of the respective time window.
This section briefly sketches a reactive real-time approach that is also used (in a
slightly modified variant) for the pro-active concept introduced later. It applies a
basic concept that has been evaluated in terms of being able to efficiently control
various dynamic, unreliable real-world processes in production and logistics (see
Gendreau et al. (1999), Ichoua et al. (2000), Bock et al. (2006), Bock (2010),
Ferrucci and Bock (2014)). Before starting the delivery process, a first transportation
plan (denoted as the first relevant plan P0r ) for servicing the initially known requests
is generated by applying a suitable optimization procedure. Subsequently, at time 0,
this first relevant plan P0r is set in execution. In what follows, we denote Pτr as the
relevant plan executed at time τ . As the Vehicle Routing Problem is an N P-hard
problem, the finding of efficient tour plans is a complex task and cannot be done
in zero or almost zero time. Hence, simultaneous to the execution of the ongoing
transportation process, the real-time control approach continuously tries to improve
all not implemented decisions of a currently stored theoretical plan Pτt . During this
improvement process, the best performing theoretical plan Pτbt is kept for a possible
210 S. Bock
exchange with the current relevant plan. In other words, the entire execution time
of the ongoing process is used for improving all still theoretical plan decisions. Due
to the concurrency of plan execution and possible plan adaptations, the real-time
concept ensures a correct plan update by iteratively solving static snapshots of the
ongoing process within predetermined time limits. These time limits are denoted as
anticipation horizons and separate the continuous time lapse into discrete intervals
of customizable length ta .
Specifically, with the beginning of a new anticipation horizon at time τ , all
requests that have occurred and buffered during the preceding horizon are integrated
into the three stored plans, namely into the relevant plan, the theoretical plan, and
into the best theoretical plan. Subsequently, the best performing plan of these three
plans is taken as the new relevant plan. In order to ensure that the theoretical
transportation plan that is modified by the applied optimization procedure during the
considered anticipation horizon starting at time τ is applicable after the elapse of this
horizon at time τ + ta , all decisions that will be taken by the relevant plan during this
horizon are pre-simulated. Hence, after this pre-simulation, the theoretical plan only
contains decisions that will be taken after time τ + ta . Therefore, the optimization
procedure is applied to a static problem instance that maps a situation in the delivery
network that will occur in the future, namely at time τ + ta . As a consequence, with
the beginning of a new anticipation horizon, all decisions to be taken during this new
horizon become fixed and are assumed to be taken as defined by the relevant plan.
For further details of the applied plan adaptation methods, we refer to Bock (2010)
and Ferrucci (2013). Figures 4 and 5 illustrate the described update handling of a
Fig. 4 Continuously adapting the current relevant plan by iteratively solving static snapshots of
the ongoing process at the adaptation level (cf. Bock 2010)
Pro-Active Strategies in Online Routing 211
Fig. 5 Continuously adapting the current relevant plan by iteratively solving static snapshots of
the ongoing process at the adaptation level (cf. Bock 2010)
be mainly attributed to the fact that the problem instances of the extended Pickup
and Delivery Problem considered in Bock (2010), which integrates transshipments
and external service providers, are considerably more complex. Therefore, more
substantial adaptations are frequently necessary to attain efficient process executions
and to allow for an adequate handling of the various incoming events. This makes
a continuous application of the optimization process reasonable. In contrast to this,
most of the considered dynamic VRP-instances were efficiently adapted in a single
adaptation horizon by the applied Tabu Search algorithm.
However, Ferrucci (2013) reports that the continuous optimization outperforms
the event-oriented one for larger sized instances. In order to reduce the time
effort necessary for conducting the various computational simulations, all real-time
simulations conducted by Ferrucci et al. (2013) apply solely the event-oriented
variant.
Note that the reactive real-time approach applies a so-called idle vehicle waiting
strategy. This strategy requires that each idle vehicle, i.e., a vehicle that does not
possess an assigned non-serviced request, has to wait at its current location of the
network until a new request is assigned to be serviced. As a consequence, if the
majority of requests occur regularly in specific local areas, this positioning of idle
vehicles is quite reasonable as future requests can be expected close by their current
locations. If, however, a vehicle possesses further requests in its tour (determined by
the current relevant tour plan) this vehicle starts traveling to the next request location
directly after completing the service of the preceding one.
Empirical tests reveal that the aforementioned reactive real-time approach is able
to ensure an efficient execution of the routing processes even if a high degree of
dynamism has to be handled. This can be ascribed to the fact that by continuously
optimizing all remaining decisions of the current relevant plan in execution, this
approach is able to attain the necessary adaptability. As a consequence, increased
response times, caused by necessary instant plan adaptations conducted due to
incoming events, can be frequently mitigated due to a later finding of improved
plans (see Bock (2010), Ferrucci (2013), pp. 220–224). Specifically, Ferrucci (2013)
reports that, due to the fast improvements attained by the applied Tabu Search
procedure, its optimization process can be restricted to 10 s after each dynamic
event as, after this time limit, the applied meta heuristic is able to attain almost
optimal tour plans.
However, although these results are promising, closer analyses also reveal that
sometimes vehicles have to conduct considerable detours to service dynamically
incoming requests that are located far away while the tours of other vehicles have
visited these areas earlier. Hence, if it would be possible to forecast some future
Pro-Active Strategies in Online Routing 213
requests and integrate this information into the tour plan, response times may
be further reduced. For this purpose, Ferrucci et al. (2013) propose to exploit
available past request data in a preceding offline step that derives forecasts of
future requests. By doing so, the reactive real-time approach is turned into a
proactive one. However, in contrast to many other pro-active approaches, this
method does not assume or apply any knowledge about existing distributions or
patterns defining the dynamism of the routing processes without proving their
existence. Rather, stochastic knowledge is derived solely from past data while all
the required attributes needed for deriving the respective forecasts are checked.
In order to derive stochastic knowledge about the occurrence of future requests, past
data is separated into segments. As illustrated in Fig. 6, each segment s represents all
incoming requests that occur in a quadratic spatial area with side length DCse during
a specific time interval of length DCte . The incoming requests of each segment are
modeled by a time-space Poisson process ps with a request arrival rate λ(ps ) that is
calculated by the average number of requests arriving in the spatial-time assignment
of segment s during the past nf days.
Ferrucci et al. (2013) report that the applied setting DCse = 2.5 km and
DCte = 1 minute leads only to low rate values λ(ps ). These low rates are frequently
not significant enough to justify the rerouting of a vehicle at a specific time to the
defined spatial area. Hence, after generating the segments, neighboring segments
are iteratively combined into so-called clusters in order to fulfill given restrictions.
Specifically, clusters are generated by the combination of segments that are located
close to each other such that the following criteria are met:
• The cluster does not exceed a predetermined maximum spatial (DCmse ) and tem-
poral extension (DCmte ), as otherwise, its forecast information is not significant
(the occurrence area is too large).
• The sum of the rate values of the assigned segments is at least DCminλ , i.e., a
minimum total request occurrence probability is guaranteed in this cluster.
• Further quality criteria are fulfilled (see, for details, Sect. 3.2).
Clearly, one may assume that the finding of such clusters can be done by applying
one of the well-known standardized clustering approaches such as, for instance, one
k-means clustering variant (see Arthur and Vassilvitskii (2007) or Jain et al. (1999)).
However, these approaches require the predetermination of the number of clusters
to be built. Moreover, the fulfillment of the itemized quality measures cannot be
controlled in these standardized approaches. Hence, Ferrucci et al. (2013) develop
a new specifically designed clustering method that solves a defined Mixed Integer
Problem (MIP) by applying the CPLEX standard solver. In order to map the entire
set of possible clustering alternatives, all clusters that are theoretically generatable
and fulfill the defined quality criteria are built and integrated into the defined MIP.
Note that this includes overlapping and therefore mutually excluding clusters.
After being generated, each (still theoretical) cluster ctemp is checked as to whether it
fulfills the following two quality criteria: (i) a maximum average travel time inside
the cluster that is necessary for servicing the respective requests and (ii) that the
request arrivals fulfill, at least to a customizable extent, basic requirements of a non-
homogeneous time-space Poisson process. Note that the fulfillment of the second
criterion is assumed and used throughout the approach. In what follows, some
details concerning both criteria are depicted (cf. Ferrucci (2013), pp. 175–177):
To criterion (i): As the actual location of a dynamically incoming request may differ
from the derived cluster location, travel times inside a cluster are an issue of
assessing the quality of a generated cluster. Namely, as the intention of deriving
a cluster is to pro-actively guide a vehicle to the corresponding request-likely
area at the specified time in order to reduce request response time, the travel time
inside a cluster has to be kept strongly limited. For this purpose, the connectivity
of the barycenter (derived for the requests occurring in considered cluster ctemp
during the last nf days) inside the real-world street network is checked first.
This is done by applying the Dijkstra algorithm (cf. Dijkstra (1959)) to find,
within a predetermined travel time radius DCradiusTT , the nearest location on
a road of a class possessing a maximum speed limit (i.e., a road categorized
in the street network allowing for the fastest travel). If such a location cannot
be found within the prescribed time distance, a location in this area is chosen
instead that is located on a road with maximum speed limit. In both cases, this
node is denoted as nctemp and represents the location the corresponding dummy
customer is assigned to if the respective cluster is chosen for being integrated
into the stochastic knowledge. After determining the location nctemp (i.e., the
avgT T
location that vehicles are pro-actively rerouted to) the average travel time ct emp
is computed to all past request locations that occurred in ctemp within the last
avgT T
nf days. If ct emp exceeds the predefined threshold DCmaxAvgTT , the temporary
cluster ctemp is discarded.
Pro-Active Strategies in Online Routing 215
To criterion (ii): Each generated potential cluster (that may become a dummy
request) is checked as to whether it represents basic attributes of a non-
homogeneous time-space Poisson process, at least, to a certain extent. Note
that these attributes are used for the updating of the remaining occurrence
probabilities of future requests in the generated clusters that are needed by the
applied Tabu Search approach to weight the dummy requests throughout the
ongoing computations. For this purpose, N(t) is denoted as a stochastic counting
process that measures the incoming requests in the considered cluster ctemp until
time t. According to Ross (2019), the following three criteria have to be checked
for N(t). First, it holds that N(0) = 0. As this criterion is obviously fulfilled, it
does not have to be checked for cluster ctemp . Second, for time length t, N(t) is
Poisson distributed with mean t · λ(ctemp ) as λ(ctemp ) denotes the rate parameter
of the Poisson process. Third, the numbers of incoming requests in disjoint time
periods are independent. In order to check the second criterion for a possible
cluster ctemp, Ferrucci et al. (2013) apply the Chi-Square Goodness-of-Fit test
with a customizable type I error α according to the guidelines described by
Kvam and Vidakovic (2007, p. 156). Specifically, by this test, it is analyzed
whether substantial deviations exist between the actual number of occurred
requests in the respective cluster ctemp during the past nf days and those that
would be expected by applying a Poisson distribution with the derived parameter
λ(ctemp ). For this purpose, the sample X1 , . . . , Xnf is considered as the number
of observed request arrivals in cluster ctemp during the past nf days. For the details
of this test, the reader is referred to Ferrucci (2013, pp. 175–177). The authors
applied the Chi-Square Goodness-of-Fit test for different values of α. Although
in the literature very small values of α = 0.05 are frequently proposed (cf., e.g.,
Israel (2008, p. xxii) or Sapsford and Jupp (2006, p. 217), Ferrucci (2013) reports
that, in the considered application, these quality thresholds were much too low to
guarantee the presence of Poisson process characteristics with a sufficiently high
reliability. Hence, best results were achieved by applying the threshold α = 0.4.
This value leads to an average Poisson quality in the clusters of more than 0.6 for
the generated data sets as well as for the real-world data. Ferrucci (2013) states
that although further increasing α improved this value, setting α to 0.4 guaranteed
an efficient compromise between the conflicting goals of ensuring an acceptable
minimum Poisson quality in the clusters and reducing erroneous rejections of
clusters that possess a sufficient Poisson quality. The third criterion of Poisson
processes that requires that the requests occur independently for disjoint time
intervals (cf. Ross 2019) can also be checked by using an appropriate Chi-
Square Goodness-of-Fit test. However, due to the small temporal extension of a
cluster, Ferrucci et al. (2013) refrain from performing this test. This is because the
required accuracy in request arrival times is frequently not fulfilled by practical
request data sets and therefore would lead to an unfavorable rejection of clusters
(for further details, see Ferrucci (2013, pp. 175–177).
Only if both checks are successfully completed, is the considered cluster ctemp
kept for being integrated into the MIP formulation for finding a maximum number
216 S. Bock
As P Sλ(ctemp ) (0) provides the probability that no request will occur in ctemp , the
defined weight-parameter wi gives the probability that at least one request will occur
in ctemp. As this parameter discounts the contribution (the customer inconvenience
as a rated response time) of the request, this weight parameter gives the priority
of the dummy request within the reactive real-time control. Moreover, the derived
service time for traveling and servicing the expected number of requests in ctemp is
used for estimating the time that a vehicle is kept at the respective location of the
derived dummy request nctemp of cluster ctemp.
As stated above, the preceding steps have led to the constitution of a set of all
potential (not necessarily non-overlapping) clusters C that are considered for being
chosen to define additional dummy requests to be additionally integrated into the
tour plans by the reactive real-time approach. In order to compute a suitable set of
non-overlapping clusters out this set C, the following MIP is defined. In this MIP,
the binary variable xi for i ∈ C is used that is one (xi = 1) if and only if the cluster
i ∈ C is selected and zero otherwise. Furthermore, the parameter cist art T L gives the
temporal threshold at which cluster ci starts and M defines a big number dominating
all other values.
Maximize M· xi − cist art T L · xi
i∈C
s.t. (1) ∀s ∈ SC : ci,s · xi ≤ 1 and (2) xi ≤ C max.
i∈C i∈C
The above MIP pursues the maximization of the number of generated clusters.
However, if this number is identical, solutions with earlier starting clusters are
preferred. Restriction (1) ensures that chosen clusters are non-overlapping. This is
ensured by using the parameter ci, s that is one if and only if segment s ∈ SC belongs
to cluster i. Hence, if there is some overlapping in the cluster selection, there exists
a segment that is assigned to more than one chosen cluster. This is excluded by
restriction (1). Moreover, restriction (2) defines an upper bound of the number of
Pro-Active Strategies in Online Routing 217
chosen clusters. This bound is initially set to C max = s∈SC λ (ps ) /DC minλ
with SC defining the set of all segments belonging to the clusters of set C. Thus,
a maximum number of clusters can be obtained by the total number of expected
requests over all segments divided by the minimum number of expected requests
required for building a cluster. The bound defined in restriction (2) is used for
technical reasons as it enables the applied CPLEX solver to derive tighter bounds.
These tighter bounds enable a reduction of the solution space that has to be explored
by the conducted examination process. By solving the MIP through the application
of the CPLEX solver, the set of generated clusters is determined. Each of these
clusters defines a dummy request with the aforementioned parameter setting that is
added to the set of real requests. Therefore, it has to be integrated into the tour plans
computed and controlled by the reactive real-time approach.
The applied reactive real-time approach is able to attain a high adaptability that
handles incoming requests in an efficient way even when a high degree of dynamism
is given. However, response times of urgent requests can be further reduced if these
requests can be reliably forecasted and this stochastic knowledge is integrated into
the tour scheduling process. For this purpose, the generated clusters allow for the
additional definition of so-called dummy requests possessing a specific parameter
setting that distinguishes them from ordinary requests. While ordinary requests have
the weight one (occurrence is deterministic), each dummy request is discounted
with its occurrence probability riw ≤ 1 (defined in Sect. 3.2). Each ordinary request
has a uniform fixed service time of 60 s, whereas the service time of a dummy
request is set to the total time necessary on average for servicing all requests that
are expected in the respective time-spatial area (see Sect. 3.2). Each dummy request
i being defined by cluster ci is integrated into the set of requests that have to be
delivered by the reactive real-time control approach. However, while the ordinary
requests require a delivery (or a pick-up) of the respective goods (or persons), the
service of a dummy request by the reactive approach has a modified intention. It
reroutes some vehicle to the defined spatial area at the specified time cist art and keeps
it there until time ciend . Both dummy request dependent parameters result from the
segments belonging to ci . Note that the integration of dummy requests requires a
special handling by the reactive real-time control approach during the execution
of the transportation process. As long as this derived interval is not reached by
the current time τ + , all aforementioned parameters of the dummy request i remain
unchanged. Specifically, if τ + ≤ cist art holds, we have wi (τ + ) = wi and si (τ + ) = si .
If, however, the end of the current anticipation horizon τ + exceeds cist art , i.e., if it
holds that τ + > cist art , the parameter values of the corresponding dummy requests
218 S. Bock
change dynamically with the lapse of time. For this purpose, Ferrucci et al. (2013)
apply the following update handling for the dummy requests:
The beginning of the time window (ei ):
∞
1
n=1 P Sλ(ci ,τ + ) (n)· n+1
ei τ + = cist art τ + + ciend − cist art τ + ·
1 − P Sλ(ci ,τ + ) (0)
The idea behind this updating scheme is to compute at time τ + the expected
arrival time of the first future request if altogether n = 1, 2, . . . , ∞ additional
requests will occur in the remaining segments (i.e., assigned segments that start at
time τ + or later) of cluster ci . If n > 0 requests occur, there are n + 1 intervals
before and after these arrivals that are assumed to be equally distributed over the
entire remaining interval. Thus, ∞ 1
n=1 P Sλ(ci ,τ + ) (n)· n+1 gives the proportion of
the remaining interval ciend − cist art τ + that is expected to elapse until the first
request arises. As this calculation assumes that at least one additional request will
occur, it has to be divided by the probability that no request occurs, namely, it has
to be divided by 1 − P Sλ(ci ,τ + ) (0).
By multiplying the sum of the constant service time Rst and the average travel
at time τ + in the cluster ci ,
avgT T
time ci with the number of requests still expected
+ st
· λ ci , τ +
avgT T
the updated service time si is obtained. si τ = R + ci
As each dummy request is weighted (i.e., is discounted) with the probability that
at least one additional request will occur in the remaining segments of cluster ci
and this probability continuously changes when τ + > cist art holds, this weight wi is
updated by the formula:
wi τ + = P (X ≥ 1) = 1 − P Sλ(ci ,τ + ) (0)
for a timely arrival at nci . Clearly, during this waiting time, the respective vehicle is
available for servicing other requests.
5 Computational Evaluation
All experiments simulate the transportation process in real time on the real-
world road network of the city of Dortmund (a medium-sized German city). For
this purpose, a discrete-event based simulator was developed and applied. The
algorithms were implemented in Delphi and the experiments were conducted on
Personal Computers equipped with Pentium D 2.8 GHz CPU and 2.5 GB Ram.
The considered region covers an area of 22.5 km × 20 km = 450 km2 . In order
to test different scenarios with individual capacity demands, fleet sizes of 8, 10,
and 12 were simulated. At the beginning of each experiment, an initial tour plan
was generated by applying the Tabu Search procedure for 120 seconds. This plan
assigns all known real requests and dummy requests to vehicle tours.
In what follows, we describe the main aspects of the instance generation process
that is applied by Ferrucci et al. (2013) to constitute the set SGEN . Note that this
generation is comparable to the method applied by Ichoua et al. (2006). The used
parameter values in all generated instances of SGEN are empirically derived in
preliminary tests and given in Table 1.
In order to simulate various settings with individual distributions of dynamically
incoming requests, the considered service area is divided into a set of P disjunctive
uniformly-sized quadratic subregions denoted as a1 , . . . , aP . Moreover, the time
lapse is separated into Q time slices of equal uniform length. For each defined
220 S. Bock
Table 1 Values of the main parameters applied during the computational experiments
Parameter Brief description Value/values
nf Number of past days used 60 days
DCse × DCse Spatial extension of each segment 2.5 km × 2.5 km
DCte Temporal extension of each segment 1 min
DCmse Maximum spatial cluster extension in both 2 segments
directions
DCtse Maximum temporal cluster extension 15 min
DCradiusTT Maximum travel time radius within a 300 s
cluster
DCmaxAvgTT Maximum average travel time radius 650 s
within a cluster
DCminλ Minimum occurrence probability of a 1.0, 1.2, 1.5,1.8,2.0
request controlled by a predetermined (i.e., probabilities are 0.6321,
minimum λ-value 0.6988, 0.7769, 0.8347, 0.8647)
DCλrem Threshold for removing a request 0.25, 0.5 (i.e., probabilities are
0.2212, 0.3945)
RD RegionDiversity 0, 0.25, 0.5, 0.75, 1.0
TD TimeDiversity 0, 0.25, 0.5, 0.75, 1.0
Q Number of time slices 5
P Number of subregions 18
nd /ntot Proportion of dynamically incoming 0.9
requests
Srd Size of each request data set 30 instances
Rpen Penalty value for late requests 100
Rmax Maximum response time 3600 s
F(Rmax ) Largest inconvenience value for a non-late 1.0
request
Fig. 7 Linear averaging exemplarily applied to the individual request arrival probabilities within
the subregions a1 , . . . , aP . (see Ferrucci (2013), p. 213)
structural diversity in a general way, but allow for a relative comparison of generated
instances. While instances of set SGEN generated with the setting RD/TD = 0 do not
possess any diversity (i.e., the dynamism does not change spatially nor with the
time lapse) a considerable dynamism can be expected from instances of set SGEN
generated with the setting RD/TD = 1.0.
As depicted in Table 1, during the experiments, the number of subregions P
was set to 18 while the number of time slices Q was set to 5. Further details of
the instance generation process executed with one of the 24 settings are given by
Ferrucci (2013, pp. 211–217).
The proposed reactive and the pro-active real-time approach were applied to 30
instances of the real-world data class SREAL. For comparison reasons, a Greedy
heuristic was also applied that handles dynamically incoming requests by finding a
least cost insertion position in the currently executed transportation plan. Note that
this heuristic is also applied by the sophisticated reactive real-time approach when-
ever new requests occurred during the preceding anticipation horizon. However,
to find further plan improvements, the reactive real-time approach subsequently
applies the Tabu Search heuristic for 10 seconds. Table 2 gives the average
improvement rates attained by the reactive real-time approach in direct comparison
with the greedy heuristic.
Comparable to the results reported by the studies of Gendreau et al. (1999),
Ichoua et al. (2000), and Bock (2010), the reactive approach substantially outper-
forms the greedy heuristic by between 9% (fleet of 12 vehicles) and 32.45% (fleet of
8 vehicles) for the linear objective function and by between 19% (12 vehicles) and
almost 52% (8 vehicles) for the quadratic objective function. These improvements
underline that an efficient control of such kind of dynamic processes requires the
application of a sophisticated real-time approach that is able to continuously adapt
Table 2 Improvements attained for data class SGEN by the reactive real-time control approach in
direct comparison with the greedy heuristic according to the average customer inconvenience and
the number of late requests
Objective function Linear2X Improvement vs. Greedy heuristic
8 vehicles 32.45% / 11 vs. 175 late
10 vehicles 14.33% / 0 vs. 9 late
12 vehicles 9.21% / 0 vs. 0 late
Objective function quadratic Improvement vs. Greedy heuristic
8 vehicles 51.96% / 14 vs. 158 late
10 vehicles 30.65% / 0 vs. 5 late
12 vehicles 19.28% / 0 vs. 1 late
Pro-Active Strategies in Online Routing 223
Table 3 The average number of clusters that are attained by applying CPLEX for solving the
MIP to optimality (see Sect. 3.1) in dependence of the tested DCminλ -values (see Ferrucci (2013,
p. 232)
Number of clusters generated in dependence of the quality threshold DCminλ
DCminλ = 1.0 DCminλ = 1.2 DCminλ = 1.5 DCminλ = 1.8 DCminλ = 2.0
67.7 44.6 24.4 13.1 9.4
the ongoing transportation plan according to the incoming dynamic events. This is
particularly true if, due to a smaller fleet size, the number of available vehicles is
strongly limited. Obviously, such adaptability is not reachable by applying a simple
least cost insertion greedy heuristic. In order to additionally apply and evaluate the
pro-active approach, stochastic knowledge has to be exploited from SREAL in an
offline step as described in Sect. 3.1. For this purpose, aside from the 30 days used
for the performance assessment, a further 60 days are selected. These exploitations
were repeated with different threshold values of DCminλ determining the minimum
request occurrence probability in a cluster. For these evaluated DCminλ -values, Table
3 gives the average number of clusters (and therefore dummy requests) generated
by applying CPLEX. It is worth mentioning that the finding of optimal solutions
of the MIP are reasonable as these solutions provide a substantially increased
number of generated clusters. Ferrucci (2013) reports that, in direct comparison
with an alternatively applied heuristic, optimal solutions of CPLEX provide between
7.7% (for DCminλ = 2.0) and 13.4% (for DCminλ = 1.0) more clusters. Note
that the number of generated non-overlapping clusters constitutes the stochastic
knowledge of a predetermined quality exploited from the given data set. Therefore,
the improvements attainable by the application of the pro-active approach solely
depend on them.
By analyzing the results attained by the pro-active real-time approach in com-
parison to the reactive one (see Table 4), no clear dominance can be identified. The
pro-active approach uses the stochastic knowledge efficiently only in cases of larger
fleet sizes (10 or 12 vehicles) and by applying the quadratic customer inconvenience
objective function.
In these cases, and in contrast to the linear customer inconvenience function, the
quadratic customer inconvenience function prevents larger delays of urgent requests
in favor of dummy requests that require larger detours. Moreover, due to the larger
fleet sizes, falsely forecasted requests have a milder effect as the available vehicle
capacity is not that scarce. However, the positive effects of pro-actively guiding
vehicles to uncovered areas where forecasted requests actually occur, were observed
only rarely. This can be ascribed to the fact that further analyses of the instances
of the data set SREAL reveal that these instances possess a limited diversity. For
instance, Ferrucci (2013) shows that the request arrival rates in the 72 subregions
are quite stationary in so far that the subregions obtaining the majority of incoming
requests remain quite stable during the day. Hence, as the reactive approach keeps
idle vehicles at their last customer location, scenarios are quite unlikely where
these positions are far away from future request occurrences. Due to this limited
224 S. Bock
Table 4 The average improvement rates attained by the pro-active approach in comparison with
the reactive one
Results of the pro-active real-time approach in comparison with the reactive one forSREAL
Vehicles Predetermined minimum quality of stochastic knowledge (Parameter DCminλ )
1.0 1.2 1.5 1.8 2.0
Linear customer inconvenience objective function
8 −11.5% (28) ↑ −7.66% (21) ↑ −3.12% (18) ↑ −0.27% (18) ↑ 1.6% (10) ↑
10 −4.86% (22) −1.42% (20) −0.68% (17) 1% (16) 0.01% (16)
12 −0.67% (15) −0.14% (16) 1.7% (13) 0.64% (11) 0.88% (11)
Quadratic customer inconvenience objective function
8 −19.91% (27) ↑ −7.52% (21) ↑ 0.21% (14) ↑ 1.17% (15) ↑ 1.13% (13) ↑
10 −3.99% (20) 0.7% (12) 2.76% (9) 2.79% (10) 0.71% (14)
12 6.59% (9) 7.93% (5) 6.78% (9) 7.31% (8) 6.08% (8)
Values in bold print are the results for the best choices of DCminλ . The value in brackets gives the
number of instances with worse results. Entries with “↑” indicate that the number of late requests
increased by applying the pro-active approach (see Ferrucci (2013), p. 236)
5.3 Measured Results for the Instances of the Data Class SGEN
Fig. 8 The attained improvement rates of the pro-active approach for the linear customer
inconvenience objective function and a fleet size of 10 vehicles
Fig. 9 The attained improvement rates of the pro-active approach for the quadratic customer
inconvenience objective function and a fleet size of 10 vehicles
This can be ascribed to the fact that the increased time and spatial variety in the
data allows for the finding of clusters that fulfill the required quality issues. Hence,
if, for instance, the maximum values DCminλ = 1.8 or DCminλ = 2.0 are applied, no
cluster is found for the settings of lower diversity RD = TD = 0, RD = TD = 0.25,
and RD = TD = 0.5. In contrast to this, clusters of this higher quality are only found
for the diversity settings RD = TD = 0.75 and RD = TD = 1. In accordance with
these results are the measured improvement rates attained by the pro-active real-time
approach and highlighted by Figs. 8 and 9 for the two simulated objective functions,
respectively. These two figures indicate that the pro-active approach outperforms its
reactive counterpart for all 30 instances of SGEN if the full diversity (generated by
226 S. Bock
Note that this definition omits the movement to the position of the barycenter
y
of the first time period, i.e., the movement to b1x , b1 . This is done as the initial
position would be the central depot where all vehicle tours start from. Therefore,
this first time period is used to obtain instance-dependent vehicle positions. But,
as the number of requests occurring during the first period n1 is kept small in
the considered experiments, the omission of the first movement does not have a
significant influence on the computed dosd-values (for details, see Ferrucci (2013),
pp. 249–254). Ferrucci et al. (2013) evaluate the practicability of the proposed
dosd by analyzing the attained improvements rates for all conducted experiments
(possessing a comparable number of serviced requests) depending on the dosd-
228 S. Bock
value calculated for the respectively given request data set. In order to keep the
numbers of serviced requests per vehicle comparable, these analyses were generated
for different sizes of the used vehicle fleets. In order to assess best or almost best
results for all experiments, the highest improvement rates over all tested parameter
settings (i.e., the most suitable value for DCminλ ) were taken as the respective result.
In order to derive the correlation between the attained improvement rates and
the dosd of the request data, Ferrucci et al. (2013) apply six linear regression
models for each of the applied inconvenience functions. By applying the Pearson
product-moment correlation coefficient r according to the description given by
Rodgers and Nicewander (1988), convincing values between 0.9161 and 0.9647
were attained depending on the fleet sizes or assumed inconvenience functions
(linear or quadratic). This proves a strong correlation between the dosd and the
reduction of the resulting response time attainable by using stochastic knowledge
exploited from the available request data. Ferrucci et al. (2013) therefore conclude
that the diversity in a request data set with a dosd ≥ 0.6 enables the proposed
pro-active approach to attain considerable reductions of the request response time.
Clearly, these conclusions also depend on the applied exploitation methods (applied
for finding the clusters) and on the definition of the dosd. But, as the reactive
approach keeps idle vehicles at the location of their last service, a successful
use of pro-active vehicle relocation strategies basically depends to some extent
on a minimum existing diversity in the data. Therefore, it can be stated that the
derived dosd constitutes a suitable offline measure that improves the applicability
of the proposed real-time approach. Namely, it allows reliable conclusions for
suitably customizing the pro-active methods and whether significant reductions of
the attainable response times of the reactive real-time approach are possible by
integrating stochastic knowledge concerning future request occurrences. For further
details, we refer to the more detailed evaluations and descriptions provided by
Ferrucci (2013).
Many real-time approaches that adapt an existing tour plan simultaneous to its
execution use en route diversions. Such an en route diversion defines a situation
where a vehicle is diverted from its current travel to a request destination by
a plan modification conducted and communicated in real time. This en route
measure is possible due to advances in communication technologies integrated in the
exemplary information flow of a real-time control depicted by Fig. 3. Specifically,
these systems allow for locating vehicles in real time as well as for an instant
communication of new orders to the drivers (see Giaglis et al. (2004), Larsen et
al. (2008), or Ferrucci and Bock (2015)). By doing so, the adaptability of a real-
time control and therefore the efficiency of the tour execution can be considerably
improved. For instance, if the delivery or pick-up location of a newly incoming
request is in the direct vicinity of the position of a vehicle currently traveling to
Pro-Active Strategies in Online Routing 229
Thus, both kinds of en route diversion are interpreted as negative and should be kept
limited.
In order to be able to control the number of conducted en route diversions in
the real-time approach, a penalty cost function D(ζ ) is integrated into the pursued
inconvenience objective z(ζ ). By charging a customizable penalty rate p for each
conducted en route diversion, a considered tour plan ζ that causes n(ζ ) en route
diversions has to account for an additional penalty D(ζ ) = n(ζ ) · p. Note that n(ζ ) is
computed dependent upon the assumed driver profile (JE or CA). As a consequence
of this extension, the reduced costs attained by the best of the two tour plans Pτt and
Pτbt that are generated during a considered anticipation horizon (starting at time τ )
has to compensate for the additional costs being charged for a potentially increased
number of planned en route diversions. Otherwise, the real-time approach does not
change the relevant plan Pτr that is currently in execution. In other words, depending
on the customizable cost rate p, additional en route diversions are only integrated
into the tour plan if the resulting cost reductions are substantial enough that they
exceed these penalties. The attained cost reductions are reductions of the resulting
response time and can be computed depending upon the given objective function
(linear or quadratic) and the defined cost rate p (see Table 6). This allows for a
customizable control of using en route diversions through determining a threshold
of necessary response time reductions. For practical applications, the cost rate p may
be derived from a required minimum reduction of the resulting response time that
should compensate for the each additionally caused en route diversion. In order to
evaluate suitable values for p and to analyze the consequences of resulting en route
diversions and customer inconvenience values, Ferrucci and Bock (2015) conduct
various computational experiments.
For a detailed depiction of the measured results, we refer the reader to the paper
of Ferrucci and Bock (2015). However, the main results of this study can be made
clear by considering Tables 7 and 8 that show the effect of the different penalty cost
rates p for the two simulated objective functions (linear and quadratic, respectively)
and the driver profile “change avoidance”.
Table 6 Response time thresholds and penalty cost rates for en route diversions (see Ferrucci
and Bock (2015), p. 82)
Response time threshold (s) Value for p dependent upon the pursued objective function
Linear Quadratic
0 (unrestricted case) 0 0
36 0.01 0.0001
72 0.02 0.0004
180 0.05 0.0025
360 0.1 0.01
720 0.2 0.04
1800 0.5 0.25
3600 1.0 1.0
Complete prohibition 1,000,000 1,000,000
Pro-Active Strategies in Online Routing 231
Table 7 Percentage changes of the total linear customer inconvenience attained in comparison to
the unrestricted case by applying various values for the penalty cost rate p > 0 and using a fleet
size of 10 vehicles
Profile Values Percentage reductions/increases with the penalty cost rate p=
CA p=0 0.01 0.02 0.05 0.1 0.2 0.5 1.0 1,000,000
Obj. val. 0.2698 −1.4 −0.6 0.4 2.2 8.1 18.2 25.6 27.2
#Dr 54.2 −23.4 −33.5 −57.9 −74.1 −87.6 −97.8 −99.8 −100
#Dd 27.6 −8.8 −15.1 −27.3 −37.4 −61.4 −90.1 −98.1 −100
Distance 609,074 0.2 0.4 1.9 4.3 6.1 8.4 9.6 9.7
The simulated driver profile is CA
Table 8 Percentage changes of the total quadratic customer inconvenience attained in comparison
to the unrestricted case by applying various values for the penalty cost rate p > 0 and using a fleet
size of 10 vehicles
Profile Values Percentage reductions/increases with the penalty cost rate p=
CA 0 0.0001 0.0004 0.0025 0.01 0.04 0.25 1.0 1,000,000
Obj. Val. 0.094 0.1 0 −0.1 2.0 3.7 29.7 56 61.8
#Dr 52.3 −1 −0.5 −12.4 −36.4 −69.8 −96.4 −99.8 −100
#Dd 30.1 −0.7 −5.8 −2.3 −12.5 −26.0 −76.7 −98.5 −100
Distance 620,405 0.2 −0.4 0.2 1.8 3.3 7.6 9.8 9.8
The simulated driver profile is CA
reduces the resulting customer inconvenience, but allows only a very moderate
number of diversions.
Due to the derived cognitions of the computational evaluation, it can be stated
that the extension of the real-time approach proposed by Ferrucci and Bock (2015)
allows for a suitable control of en route diversions for the first time. Hence, the
aforementioned negative consequences of this measure can be suitably covered.
distance of each request occurring in the respective time period to the respective
barycenter of the day (or group of days) is computed.
In order to compute the sought profiles as groups of similar past days, the
k-means++ approach proposed by Arthur and Vassilvitskii (2007) is applied.
By using a sophisticated seeding, this approach provides a competitive heuristic
solution to this NP-hard clustering problem within a short time limit. Moreover,
by incorporating some randomness, it produces various profiles in repeated runs.
Hence, Ferrucci and Bock (2016) apply this clustering algorithm 1000 times.
Subsequently, the best assignment is selected that possesses the minimum total
distance of days to their assigned profiles.
The main idea of the approach of Ferrucci and Bock (2016) is to improve the
quality of the derived and applied stochastic knowledge by filtering out the past
days that belong to a profile whose request occurrence structure is similar to the one
observed for the ongoing day. In other words, dummy requests should be generated
solely from those past days that are assumed to possess an identical or comparable
pattern as that observed for the current day. Hence, after generating the profiles
in an offline step, the real-time approach of Ferrucci and Bock (2016) analyzes
the structure of incoming requests in order to identify similar known profiles, i.e.,
the respectively assigned days, in real time. Only these days should be used to
derive suitable dummy requests. However, due to the significant computational
complexity of the MIP that has to be solved for deriving these dummy requests (see
Sec. 3), this computation is done beforehand in an effortful offline step. Therefore,
after generating the respective profiles as sets of past days, the dummy requests
are computed for each subset of existing profiles. Specifically, for each subset of
234 S. Bock
profiles, to derive the respective dummy requests, the stochastic knowledge building
procedure is applied that is described in Sect. 3. Hence, for k existing profiles, this
procedure is applied 2k − 1 times by the offline procedure. Clearly, this exponential
complexity limits the number of profiles which can be generated.
One may think that a significant drawback of applying a k-means clustering
algorithm for identifying existing profiles in the data set consists of the fact that this
approach requires a predetermination of the number of profiles k to be built. Indeed,
this is a significant prerequisite of the clustering algorithm. However, after starting
with a two-group clustering by setting k :− 2, this can be handled by iteratively
increasing the parameter k by one (setting k :− k + 1) until the new clustering with
k + 1 groups no longer significantly outperforms the former one possessing only k
groups.
In order to identify in real-time the subset of profiles that are relevant in time slice
t (note that these time slices are identical with the time slices used in the profile
definition) for the ongoing day, Ferrucci and Bock (2016) propose two different
rules:
• Standardized Rule (SR): This rule keeps a profile p ∈ P as relevant if its total
distance to the current day over the elapsed times slices (i.e., the time slices
1, . . . , t) does not exceed by more than 20% the maximum distance of a day
within profile p to the profile itself over the time slices 1, . . . , t. I.e., SR keeps a
profile for the current day at time slice t if it holds that Δ (d c , p) < 1.2· Δmaxt,p =
1.2· maxΔt (i, p). In this computation, the abbreviation Dp gives the set of days
i∈Dp
assigned to profile p ∈ P and t (i, j) defines the total sum of distance measure
values for the periods 1, . . . , t for two sets of days i and j. Note that a factor of
1.2 is assumed to cover some additional acceptance tolerance.
• Extended Rule (ER): While the standardized rule rates all profiles independently
of each other, the extended rule applies a rating that is relative to the profile
that best matches the current day at the considered time. For this purpose,
at first, for the periods 1, . . . , t, a best matching profile p∗ is identified,
i.e., a profile fulfilling Δt (d c , p∗ ) = minΔt (d c , p). Then, accept the sub-
" p∈P
#
set of profiles p ∈ P | Δt (d c , p) < (1.5 − 0.05· t) · pt · Δmax t,p with pt =
c ∗ max
Δt (d , p ) /Δt,p∗ . The extended rule becomes restrictive if profile p∗ matches a
current day well as, in this case, pt becomes small. This fast orientation towards
a well matching profile is the main intention of the rule. By subtracting 0.05 · t
pt · Δmax
t,p from the threshold 1. 5 pt · Δt,p , this orientation is strengthened when
max
dependent request arrival probabilities) are built analogously to the days of the data
class SGEN introduced in Sect. 5.1. Following the findings of the study of Ferrucci et
al. (2013) concerning specific structures of request arrivals supporting a successful
application of the pro-active approach, all patterns are generated such that they
possess a high diversity. In other words, it is ensured that the regions that possess
the most incoming requests change from time slice to time slice. Hence, the arrival
probabilities of the 18 subregions differ significantly from time slice to time slice,
while the expected request arrival rate in each time slice is identical in all patterns.
Subsequently, by applying each generated pattern, 60 days are respectively built.
In what follows, the abbreviations 2p, 3p, and 4p are used and denote a data set of
120 days based on two patterns (2p), a data set of 180 days based on three patterns
(3p), and a data set of 240 days that is based on four patterns (4p).
In what follows, the original approach (proposed in Ferrucci et al. (2013)) that
uses only a single profile is identified with an extension “SP”, while the extended
version that applies 2, 3, 4, or more profiles obtains the extension “MP”. In order
to decide about the decision rule used for filtering out the relevant profiles in real
time, Table 9 gives the improvement rates attained by replacing the pro-active-
SP-approach by the pro-active-MP approach for different settings of the number
of generated patterns, assumed profiles, and applied objective functions (linear or
quadratic customer inconvenience).
These improvement rates indicate a substantial outperformance of SR by ER
that is further increased for settings being generated by using more patterns. This
clear dominance of ER can be attributed to the fact that ER is able to identify the
relevant profiles much faster than SR. Specifically, by comparing each profile with
the best matching profile, potentially existing outliers in some profiles do not harm
the identification process. This can be particularly observed while applying SR to
settings where the number of profiles coincides with the number of patterns (i.e.,
if k = 2 is applied to 2p, if k = 3 is applied to 3p, or if k = 4 is applied to 4p).
Table 9 Averagely attained improvement rates by applying the extended profile acceptance rule
(ER) (instead of SR) within the pro-active approach for 2, 3, or 4 profiles, best DCminλ -values, and
a fleet size of 10 vehicles
Number of patterns Number of generated profiles
k=2 k=3 k=4 k=5 k=6
Linear objective function – objective values
2p 4.8% 4.5% 3.9% 3.1% 3.5%
3p 8.8% 12.5% 4.1% 4.0% 2.7%
4p 2.0% 7.7% 12.7% 8.9% 9.2%
Quadratic objective function – objective values
2p 12.6% 9.9% 8.0% 13.2 9.3%
3p 18.4% 25.0% 12.2% 12.0% 10.5%
4p 4.7% 16.0% 23.5% 18.1% 19.0%
Values in bold print are the largest improvement rates attained by ER for two, three, and four
patterns, respectively (see Ferrucci and Bock (2016), p. 367)
236 S. Bock
Table 10 Average improvement rates in comparison to the standardized aproach with a single
profile attained by applying k = 2, 3, 4, 5, and 6 profiles for the linear inconvenience objective
function (see Ferrucci and Bock (2016), p. 366)
Linear Number of generated profiles
Number of
vehicles Applied patterns k=2 k=3 k=4 k=5 k=6
8 2p 6.7% 5.0% 4.4% 6.2% 4.4%
3p 5.2% 7.6% 8.0% 9.3% 7.8%
4p 0.7% 4.2% 7.6% 6.9% 9.3%
10 2p 7.0% 6.0% 6.0% 7.5% 6.4%
3p 10.0% 13.9% 14.4% 13.9% 13.5%
4p 0.9% 6.8% 12.4% 11.9% 12.8%
12 2p 8.6% 8.8% 7.2% 9.2% 7.9%
3p 10.7% 17.5% 16.9% 18.0% 16.1%
4p 0.8% 9.4% 17.6% 16.3% 16.9%
Table 11 Average improvement rates in comparison to the standardized aproach with a single
profile attained by applying k = 2, 3, 4, 5, and 6 profiles for the quadratic inconvenience objective
function (see Ferrucci and Bock (2016), p. 366)
Quadratic Number of generated profiles
Number of
vehicles Applied patterns k=2 k=3 k=4 k=5 k=6
8 2p 10.7% 9.8% 10.2% 13.7% 8.5%
3p 14.2% 20.6% 19.8% 17.8% 15.3%
4p 1.9% 6.5% 16.0% 14.3% 15.8%
10 2p 11.4% 11.5% 8.6% 12.3% 12.2%
3p 15.8% 23.1% 22.0% 22.5% 20.2%
4p 4.0% 15.4% 23.9% 23.7% 25.6%
12 2p 16.9% 15.8% 14.2% 17.5% 18.0%
3p 19.7% 31.8% 31.1% 31.6% 29.4%
4p 2.6% 16.0% 30.1% 30.9% 30.4%
patterns in the past data to derive stochastic knowledge of higher quality. While the
original approach is not able to derive suitable dummy requests from the past data
due to interfering effects, the fast identification of a relevant subset of useful profiles
done by the pro-active-ER approach leads to substantial reductions of the resulting
request response times. This is particularly true if the number of assumed profiles
k reaches the actual number of patterns. Increasing k to this value enables the pro-
active-ER approach to delete interferences by non-relevant days almost completely.
Moreover, changes of the request arrival structure that may occur throughout the
ongoing day are efficiently handled by updating the set of relevant profiles for
each new time slice. As the attained improvement rates are quite stable when k is
further increased, an efficient practical application of the approach does not require
to exactly estimate the number of existing patterns.
Rather, it is sufficient not to significantly underestimate this number while
overestimation (within a reasonable range) does not significantly reduce the attained
performance. This improves the practicability of the proposed approach.
8 Brief Summary
This paper provides a brief overview of some recent pro-active real-time routing
approaches that are designed for the delivery processes of urgent goods or services.
These pro-active real-time approaches are characterized by the fact that stochastic
knowledge about future request arrivals that is used for guiding vehicles into
request-likely areas is derived from past data during a prior offline step. By combin-
ing time-spatial segments of past days, future request occurrences are forecasted.
By doing so, no unverified assumptions are made concerning the availability of
a specific distribution. This improves the practical applicability of the proposed
approaches. In comparison to a reactive real-time approach that keeps each idle
vehicle at its last service location, the pro-active approach of Ferrucci et al. (2013)
guides vehicles to request-likely areas and keeps them there for a specific amount of
time. It is shown that this approach leads to substantial improvements of the request
response times if the request arrival data possess a minimum degree of diversity.
By defining the so-called degree of structural diversity as a general measure, it is
possible to reliably estimate the attainable performance of additionally using the
derived stochastic knowledge. Subsequently, extensions of this pro-active approach
are described. In order to efficiently control en route diversions that may increase the
number of accidents occurring, the approach of Ferrucci and Bock (2015) integrates
a customizable penalty cost value into the objective function. By doing so, additional
en route diversions have to compensate the defined penalty costs with corresponding
reductions of the request response times. It turns out that, depending on the pursued
objectives, reasonable definitions of the penalty cost value can be derived. Finally,
the pro-active approach can be extended to be able to identify multiple patterns in
past request data. By doing so, the approach is capable of filtering out past days
that are relevant for the current ongoing day in real time. As these selections avoid
238 S. Bock
the unwanted interfering effects of past days that are irrelevant to the current day,
the quality of the derived stochastic knowledge is frequently significantly improved.
This may lead to further substantial reductions of request response times.
Acknowledgments I thank DDS Digital Data Services GmbH in Karlsruhe, Germany, for
generously providing us with excellent real-world road network data.
References
Arthur, D., & Vassilvitskii, S. (2007). k-means++: The advantages of careful seeding. In
Proceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms, pp.
1027–1035 (PDF).
Bock, S. (2010). Real-time control of freight forwarder transportation networks by integrating mul-
timodal transports and multiple transshipments. European Journal of Operational Research,
200(3), 733–746. https://doi.org/10.1016/j.ejor.2009.01.046.
Bock, S., Rosenberg, O., & van Brackel, T. (2006). Controlling mixed-model assembly lines in
real-time by using distributed systems. European Journal of Operational Research, 168(3),
880–904. https://doi.org/10.1016/j.ejor.2004.07.035.
Dantzig, G., & Ramser, J. (1959). The truck dispatching problem. Management Science, 6(1), 80–
91. https://doi.org/10.1287/mnsc.6.1.80.
Davis, M. M., & Maggard, M. J. (1990). An analysis of customer satisfaction with waiting times in
a two-stage service process. Journal of Operations Management, 9(3), 324–334. https://doi.org/
10.1016/0272-6963(90)90158-A.
Dijkstra, E. W. (1959). A note on two problems in connexion with graphs. Numerische Mathematik,
1, 269–271. https://doi.org/10.1007/BF01386390.
Ferrucci, F. (2013). Pro-active dynamic vehicle routing: Real-time control and request-
forecasting – Approaches to improve customer service. PhD dissertation at the University
of Wuppertal. Springer Science&Business Media. https://doi.org/10.1007/978-3-642-33472-6,
ISBN: 978-3-642-33472-6.
Ferrucci, F., & Bock, S. (2014). Real-time control of express pickup and delivery processes in
a dynamic environment. Transportation Research Part B Methodological, 63, 1–14. https://
doi.org/10.1016/j.trb.2014.02.001.
Ferrucci, F., & Bock, S. (2015). A general approach for controlling vehicle en-route diversions in
dynamic vehicle routing problems. Transportation Research Part B Methodological, 77, 76–87.
https://doi.org/10.1016/j.trb.2015.03.003.
Ferrucci, F., & Bock, S. (2016). Pro-active real-time routing in applications with multiple request
patterns. European Journal of Operational Research, 253(2), 356–371. https://doi.org/10.1016/
j.ejor.2016.02.016.
Ferrucci, F., Bock, S., & Gendreau, M. (2013). A proactive real-time control approach for dynamic
vehicle routing problems dealing with the delivery of urgent goods. European Journal of
Operational Research, 225(3), 130–141. https://doi.org/10.1016/j.ejor.2012.09.016.
Gendreau, M., Guertin, F., Potvin, J.-Y., & Taillard, E. (1999). Parallel Tabu search for real-
time vehicle routing and dispatching. Transportation Science, 33(4), 381–390. https://doi.org/
10.1287/trsc.33.4.381.
Giaglis, G. M., Minis, I., Tatarakis, A., & Zeimpekis, V. (2004). Minimizing logistics risk through
real-time vehicle routing and mobile technologies. International Journal of Physical Distribu-
tion & Logistics Management, 34(9), 749–764. https://doi.org/10.1108/09600030410567504.
Ichoua, S., Gendreau, M., & Potvin, J.-Y. (2000). Diversion issues in real-time vehicle dispatching.
Transportation Science, 34(4), 426–438. https://doi.org/10.1287/trsc.34.4.426.12325.
Pro-Active Strategies in Online Routing 239
Ichoua, S., Gendreau, M., & Potvin, J.-Y. (2006). Exploiting knowledge about future demands for
real-time vehicle dispatching. Transportation Science, 40(2), 211–225. https://doi.org/10.1287/
trsc.1050.0114.
Israel, D. (2008). Data analysis in business research: A step-by-step nonparametric approach. Los
Angeles, CA: Response Books SAGE, ISBN: 9788178298757.
Jain, A. K., Murty, M. N., & Flynn, P. J. (1999). Data clustering: A review. ACM Computing
Surveys, 31(3), 264–323. https://doi.org/10.1145/331499.331504.
Kaber, D. B., Liang, Y., Zhang, Y., Rogers, M. L., & Gangakhedkar, S. (2012). Driver performance
effects of simultaneous visual and cognitive distraction and adaptation behavior. Transportation
Research Part F, 15, 491–501. https://doi.org/10.1016/j.trf.2012.05.004.
Kristensen, K., Kanji, G. K., & Dahlgaard, J. J. (1992). On measurement of customer satisfaction.
Total Quality Management & Business Excellence, 3(2), 123–128. https://doi.org/10.1080/
09544129200000013.
Kvam, P. H., & Vidakovic, B. (2007). Nonparametric statistics with applications to science
and engineering (Wiley series in probability and statistics). Hoboken, NJ: Wiley, ISBN:
0470081473.
Larsen, A. (2000). The dynamic vehicle routing problem. PhD thesis at the Technical University of
Denmark. IMM-PHD, No. 2000-73.
Larsen, A., Madsen, O. B. G., & Solomon, M. M. (2008). Recent developments in dynamic
vehicle routing systems. In B. L. Golden, S. Raghavan, & E. A. Wasil (Eds.), The vehicle
routing problem: Latest advances and new challenges (Operations research/Computer science
interfaces series) (Vol. 43, pp. 199–218). Heidelberg: Springer. https://doi.org/10.1007/978-0-
387-77778-8_9.
Lund, K., Madsen, O. B. G., & Rygaard, J. M. (1996). Vehicle routing problems with varying
degrees of dynamism. Technical report 1/96 at the Institute of Mathematical Modelling (IMM).
Owens, J. M., McLaughlin, S. B., & Sudweeks, J. (2010). Driver performance while text messaging
using handheld and in-vehicle systems. Accident Analysis & Prevention, 43, 939–947. https://
doi.org/10.1016/j.aap.2010.11.019.
Park, S. K., & Miller, K. W. (1988). Random number generators: Good ones are hard to find.
Communications of the ACM, 31(10), 1192–1201. https://doi.org/10.1145/63039.63042.
Park, S. K., Miller, K. W., & Stockmeyer, P. K. (1993). Technical correspondence: Remarks on
choosing and implementing random number generators. Communications of the ACM, 36(7),
105–110. https://doi.org/10.1145/159544.376068.
Pillac, V., Gendreau, M., Guéret, C., & Medaglia, A. L. (2013). A review of dynamic vehicle
routing problems. European Journal of Operational Research, 225(1), 1–11. https://doi.org/
10.1016/j.ejor.2012.08.015.
Rodgers, J. L., & Nicewander, W. A. (1988). Thirteen ways to look at the correlation coefficient.
The American Statistician, 42(1), 59–66. https://doi.org/10.2307/2685263.
Ross, S. M. (2019). Introduction to probability models (12th ed.). Amsterdam, Boston: Academic.
ISBN: 9780128143469.
Sapsford, R., & Jupp, V. (2006). Data collection and analysis (2nd ed.). London: Sage,
9780761943631.
Young, K. L., Salmon, P. M., & Cornelissen, M. (2013). Distraction-induced driving error: An on-
road examination of the errors made by distracted and undistracted drivers. Accident Analysis
& Prevention, 58, 218–225. https://doi.org/10.1016/j.aap.2012.06.001.
Prescriptive Analytics for Dynamic Real
Time Scheduling of Diffusion Furnaces
1 Introduction
et al. 2011; Mathirajan and Sivakumar 2006). This study addresses a scheduling
problem in the manufacturing of wafer fab, which is the first phase process in SM.
Scheduling is very important in wafer fab due to complex operations involving
multiple types of expensive machines, re-entrant systems, and time-consuming
processes. Thus, this study is concerned about scheduling in wafer fab, particularly
diffusion operation. The diffusion operation is carried out in a diffusion furnace
(DF), which is the batch processor (BP). In BP, more than one job is processed
simultaneously (with the same start and finish times) as a batch subject to the
capacity of the BP. The efficient scheduling of diffusion furnaces would significantly
improve the performance of the SM industry’s supply chain, as the diffusion process
is the lengthiest process (9–10 h) in the wafer fab area. Because of the long
processing time required for the diffusion operation, the production rate, as well
as due-date compliance of the product, are highly dependent on the efficient way of
scheduling DF(s).
Further, in reality, there are real-time events (RTE) dynamically occurring
associated with job(s) (such as change in due-date, change in arrival time, change
in priority, job cancellation) and/or resource(s) (e.g., machine breakdown, operator
illness, tool failure, shortage of material, defective material) (Wang and Fei 2014).
Due to this, in general, the production managers are finding difficulty in meeting
the commitment given to the customer for delivering the product with the existing
schedule. So, the production managers not only need to generate high-quality
solutions, but they also have to address the real-time dynamic event in a cost-
effective manner. Hence, this study focuses on dynamic real-time scheduling
(DRTS) of diffusion furnace(s). Further, as the processing time required at the
diffusion furnace accounts for nearly 75% of the processing time required for the
wafer fab, we considered different due-date based objectives.
The organization of the chapter is as follows. The description of the research
problem is given in the next section. Section 3 reviews the relevant previous research
done in the scheduling of diffusion furnaces. In Sect. 4, the proposed mathematical
model is discussed. Seven different Apparent Tardiness Cost (ATC) based greedy
heuristic algorithms (GHA) are proposed and the same are discussed in Sect. 5.
Analyses of the proposed seven different ATC-GHA are presented in Sect. 6. In the
final section, the conclusion and potential future research directions are discussed.
The diffusion furnace available for scheduling may be a single diffusion furnace,
identical parallel diffusion furnaces, or non-identical parallel diffusion furnaces.
Further, due to technical reasons, some jobs can be processed only in a particular
diffusion furnace available in the shop floor (this refers to machine eligibility
restriction in scheduling). This study focuses on both single diffusion furnace and
non-identical parallel diffusion furnaces with machine eligibility restrictions. The
diffusion furnace simultaneously processes 6 to 12 standard jobs as a batch.
Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 243
There are ‘N’ jobs available. Each job belongs to a different family, and
each family has a different processing time. However, the jobs having the same
processing time requirement belong to the same family. Due to the distinct nature
of the chemical process in each family (incompatible-job families), a batch can be
formed by considering a set of jobs from the same family. Further, jobs are having
different arrival time, different due-date, and they are non-agreeable. That is, if the
arrival time of job ‘i’ is less than the arrival time of job ‘j’ then it is not-implied
that due date of job ‘i’ is less than the due date of job ‘j’. Further, the random
occurrences of real time events with respect to job(s)/resources(s) while scheduling
diffusion furnace(s) are considered.
With this problem description, the objective of this study is to optimize eight
different customer perspectives objectives/criteria: Total Tardiness (TT), Total
Weighted Tardiness (TWT), Number of Tardy (NT) Jobs, Weighted NT (WNT)
Jobs, On Time Delivery (OTD) rate, Total Earliness and lateness (TEL), Total
Weighted Earliness and Lateness (TWEL), and Maximum Lateness (Lmax). These
criteria are defined as follows:
N
TT = Tj
j=1
N
TWT = Wj ∗ Tj
j=1
N
NT = UPj
j=1
N
WNT = Wj ∗ UPj
j=1
N
OTD Rate = OTDj/N
j=1
N
TEL = CTj − DDj
j=1
N
TWEL = Wj ∗ CTj − DDj
j=1
Where,
CTj : Completion time of job ‘j’
DDj : Due-date of job ‘j’
Tj : Tardiness of job ‘j’ = (CTj - DDj ) if CTj > DDj ; 0 Otherwise
Wj : Weight of job ‘j’
UPj : Unit penalty of job ‘j’ = 1 if CTj > DDj ; 0 Otherwise
OTDj : On time delivery of job ‘j’ = 0 if CTj > DDj ; 1 Otherwise
Lj : Lateness of job ‘j’ = (CTj - DDj )
This study makes the following assumptions for developing prescriptive analytics
model(s)/methodologies:
• Other than RTE, which occurs randomly during the schedule planning period, we
assumed the deterministic situation.
• Every job must pass through the diffusion operation, and it is not dependent on
other jobs
• For each diffusion furnace, the number of jobs in a batch should be less than or
equal to the maximum limit of the corresponding diffusion furnace.
• Preemption is not allowed
Among the four phases of SM, wafer fab is the highly complex and capital-
intensive phase (Li et al. 2017). The importance of scheduling in wafer fab has been
increasing steadily over the past few decades (Sarin and Shikalgar 2001). This area
takes a total of 3–15 weeks in comparison with the required overall processing time
of 8–30 weeks for SM. The longer processing times often belong to batch processing
operations (diffusion, oxidation, deposition, etc.,) in wafer fab. Scheduling of batch
processing machines causes higher machine utilization, lower work-in-process
inventory, shorter cycle time, and greater customer satisfaction (Pinedo 2012). This
study particularly focuses on the scheduling of diffusion furnaces, a batch processor,
where the diffusion operation takes up to 10 h (Mönch et al. 2006). Furthermore,
around 30% of the total WIP (Work in Process) in a wafer fab lie in diffusion stages,
due to long process times (Jung et al. 2014).
Based on the type of data, diffusion furnace(s) scheduling can be grouped into
deterministic and stochastic. This study focuses on the deterministic scheduling of
diffusion furnace(s) as data tracking is not a difficult one in the SM industry, where
the whole shop floor is computerized. However, readers interested in stochastic
scheduling of BPM can refer to Park and Banerjee (2011).
Some of the researchers (e.g. Pirovano et al. 2020) in deterministic scheduling of
diffusion furnace(s) consider one or more upstream and/or downstream operations
along with the diffusion operation. This study concerns about single bottleneck
operation: diffusion operation only. The diffusion furnace(s) considered for research
Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 245
in wafer fab may be grouped into single diffusion furnace, identical parallel
diffusion furnaces, non-identical parallel diffusion furnaces, diffusion furnaces with
machine eligibility restrictions, and diffusion furnaces without machine eligibility
restrictions. This study focuses on both single and non-identical parallel diffusion
furnaces with machine eligibility restrictions.
Further, diffusion furnace(s) scheduling can be grouped into static scheduling
(e.g. Mönch and Roob 2018) and dynamic scheduling. Various DF studies address-
ing the dynamic scheduling of DF(s) consider the future arrival of jobs related
to the dynamic situation. However, in reality, there is another situation related to
the dynamic situation, i.e., RTE. Hence, this study considers both the dynamic
situations: dynamic job arrival and the occurrences of RTE randomly.
Based on objectives, the studies on deterministic and dynamic scheduling of
diffusion furnace(s) are further classified into (i) completion time-based objectives
(related to organization perspective), and (ii) due-date based objectives (related
to customer perspective). Because the present study considers customer perspec-
tives objective, the organizational perspectives objectives (completion time-based
objectives) (e.g. Rocholl et al. 2018) are not discussed here. Finally, based on the
solution methodologies, the existing research can be broadly classified into mathe-
matical programming-based approaches, greedy heuristic approaches, metaheuristic
approaches, and simulation.
Based on the classifications presented here, a brief review of closely related
earlier studies are summarized in Table 1. Table 1 clearly indicates that the
research problem configuration considered (last row of Table 1) in this study:
“dynamic real-time scheduling of (a) single diffusion furnace, and (b) non-identical
parallel diffusion furnaces with machine eligibility restrictions, dynamic job-
arrivals, incompatible-job families, non-agreeable release times & due-dates, and
unexpected RTE, to optimize eight different due-date based scheduling objectives”
is a new scheduling problem associated with batch processor in general, particularly
diffusion furnaces.
The proposed mathematical models for dynamic scheduling (DS) of single dif-
fusion furnace (SDF) and non-identical parallel diffusion furnaces with machine
eligibility restrictions (NPDF-MER) to optimize eight different objectives/criteria
are explained in this section. In the proposed mathematical models, only the
future arrival of the jobs is considered under dynamic scheduling, whereas the
randomly occurring real-time events and dynamic arrival of jobs are considered in
the proposed ATC based greedy heuristic algorithms. Further, this section discusses
the validation of the proposed mathematical model.
Table 1 A closely related review
Author Diffusion furnace as Dynamic Scheduling Release time & due-date Solution methodologies Scheduling objective
With With Job
With With job resource and Wt. of
future related related resource Greedy Meta on-
Identical Parallel Non-identical arrival RTE RTE related Non- Mathematical Heuris- Heuris- time OTD
Machine Parallel machine of jobs Only Only RTE Agreeable agreeable Programming tic tic jobs TT TWT NT WNT rate TEL TWEL Lmax
Without With Without With
machine machine machine machine
Single Eligibil- Eligibil- Eligibil- Eligibil-
machine ity ity ity ity
√ √ √ √ √
Uzsoy
(1995)
√ √ √ √ √
Kurz
(2003)
√ √ √ √ √ √
Mönch
et al.
(2005)
√ √ √ √ √ √
Dirk and
Monch
(2006)
√ √ √ √ √
Mönch
et al.
2006
√ √ √ √ √ √
Malve
and
Uzsoy
(2007)
√ √ √ √ √ √
Cheng et
al.
(2008)
√ √ √ √ √
Chiang
et al.
(2008)
√ √ √ √ √
Kurz and
Mason
(2008)
(continued)
Table 1 (continued)
√ √ √ √ √ √
Li and Qiao
(2008)
√ √ √ √ √ √
Li et al. (2009)
√ √ √ √ √
Bar-Noy et al.
(2009)
√ √ √ √ √ √
Chiang et al.
(2010)
√ √ √ √ √ √
Guo et al. (2010)
√ √ √ √ √
Kim et al. (2010)
√ √ √ √ √ √
Li et al. (2010)
√ √ √ √ √
Mathirajan and
Vimalarani
(2012)
√ √ √ √ √ √
Jia et al. (2013a)
√ √ √ √ √
Chen et al.
(2013)
√ √ √ √ √ √
Jia et al. (2013b)
√ √ √ √ √ √ √
Bilyk et al.
(2014)
√ √ √ √ √
Mansoer and
Koo (2015)
√ √ √ √ √ √ √
Vimala Rani and
Mathirajan
(2015)
√ √ √ √ √ √
Vimala Rani and
Mathirajan
(2016a)
√ √ √ √ √ √
Vimala Rani and
Mathirajan
(2016b)
(continued)
Table 1 (continued)
Author Diffusion furnace as Dynamic Scheduling Release time & due-date Solution methodologies Scheduling objective
With With Job
With With job resource and Wt. of
future related related resource Greedy Meta on-
Identical Parallel Non-identical arrival RTE RTE related Non- Mathematical Heuris- Heuris- time OTD
Machine Parallel machine of jobs Only Only RTE Agreeable agreeable Programming tic tic jobs TT TWT NT WNT rate TEL TWEL Lmax
Without With Without With
machine machine machine machine
Single Eligibil- Eligibil- Eligibil- Eligibil-
machine ity ity ity ity
√ √ √ √ √ √ √ √ √ √
Vimala Rani
and
Mathirajan
(2016c)
√ √ √ √ √ √ √
Fidelis and
Arroyo
(2017)
√ √ √ √ √ √ √ √ √
Vimala Rani
and
Mathirajan
(2020a)
√ √ √ √ √ √ √ √
Vimala Rani
and
Mathirajan
(2020b)
√ √ √ √ √ √ √ √ √ √ √ √ √ √ √
Proposed
research
study
Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 249
The notations used for the proposed mathematical model for DS-SDF are discussed
as follows.
Notations:
Sets:
J Jobs
F Families
B Batches
Index:
j 1 . . . N for jobs
b 1 . . . K for batches
f 1 . . . G for families
Parameters:
A First time availability of the DF
BC Batch capacity of DF
RTj Release time of a job ‘j’
DDj Due-date of a job ‘j’
Wj Priority (or) Weight of a job ‘j’
PTf Processing time of a family ‘f’
FAjf Family association for a job, equals 1 if job ‘j’ belong to family ‘f’; 0
otherwise
Decision variables:
Xjb 1 if job ‘j’ is processed in a batch ‘b’; 0 Otherwise
Yfb 1 if family ‘f’ is processed in a batch ‘b’; 0 Otherwise
Dependent variables:
CTBb Completion time of a batch ‘b’
RTBb Release time of a batch ‘b’
PTBb Processing time of a batch ‘b’
CTj Completion time of job ‘j’
Tj Tardiness of job ‘j’
ELj Earliness lateness of job ‘j’
UPj Unit penalty of job ‘j’
TT Total tardiness of ‘N’ jobs
TWT Total weighted tardiness of ‘N’ jobs
NT Number of tardy jobs among ‘N’ jobs
WNT Weighted number of tardy jobs
OTDRate On time delivery rate among ‘N’ jobs
TEL Total earliness lateness among ‘N’ jobs
250 M. Vimala Rani and M. Mathirajan
Minimize T T (1)
Subject to
G
Yf b ≤ 1 ∀b ∈ [1, K] (2)
f =1
G
G
Yf,b−1 ≥ Yf b ∀b ∈ [2, K] (3)
f =1 f =1
Xj b ∗ FAjf ≤ Yf b ∀j ∈ [1, N] ; ∀b ∈ [1, K] ; ∀f ∈ [1, G]
(4)
N
Xj b ∗ FAjf ≥ Yf b ∀f ∈ [1, G] ; ∀b ∈ [1, K] (5)
j =1
K
Xj b = 1 ∀j ∈ [1, N] (6)
b=1
N
Xj b ≤ BC ∀b ∈ [1, K] (7)
j =1
RTB b ≥ RTj ∗ Xj b ∀j ∈ [1, N] ; ∀b ∈ [1, K] (9)
G
P TB b = PTf ∗ Yf b ∀b ∈ [1, K] (11)
f =1
CTj ≥ CTB b − (BigM) ∗ 1 − Xjb ∀b ∈ [1, K] ; ∀j ∈ [1, N]
(13)
N
TT = Tj (17)
J =1
N
T WT = Wj ∗ Tj (18)
j =1
UPj ≥ min 1, Tj ∀j ∈ [1, N] (19)
N
NT = UPj (20)
j =1
N
W NT = Wj ∗ UPj (21)
j =1
⎛ ⎞
N
OT D Rat e = ⎝N − UPj ⎠ /N (22)
j =1
ELj ≥ CT j − DD j ∀j ∈ [1, N] (23)
N
T EL = ELj (24)
j =1
N
T W EL = Wj ∗ ELj (25)
j =1
The objective function (Eq. 1) seeks to minimize the total tardiness of ‘N’ jobs.
Constraint (2) ensures that no more than one family is assigned to a batch. Constraint
(3) takes a responsibility to construct a batch sequentially. Constraint (4) states that
a job can be processed in a batch only if the corresponding family is processed in
that batch. Constraint (5) conditions that empty family (that is, a family without any
jobs) cannot be assigned to any batch. Constraint (6) is used to assign the job to any
one batch without splitting it. Constraint (7) ensures that the number of jobs in any
batch does not exceed the batch capacity of the diffusion furnace.
Constraint (8) guarantees that the first batch can start after the first-time
availability (that is the given, ‘Ath’ hour) of the diffusion furnace. Moreover, the
starting time of every batch should be greater than the release time of jobs in that
batch and the completion time of the previous batch. Constraint (9) and (10) ensure
this.
Constraint (11) states that the processing time of a batch is equal to the processing
time of the corresponding family. Constraint (12) computes the completion time of
each batch by adding its starting time and processing time. Constraint (13) computes
the completion time of the job. Constraint (14) and Constraint (15) assign binary
values to decision variables. Constraint (16) and (17) are used to compute tardiness
and total tardiness, respectively.
For primary scheduling objective represented in constraint (1), Constraints (18)
to (26) are introduced to compute the value of other scheduling objectives/criteria.
Accordingly, Constraint (18) calculates total weighted tardiness. Constraint (19)
assigns ‘1’ to jobs that are completed after their due-date. Constraint (20) and
(21) calculate the total number of tardy jobs and the total weight of all tardy
jobs, respectively. Constraint (22) is used to calculate on time delivery rate. A
Constraint (23) calculates the earliness lateness as the absolute difference between
the completion time of a job and its due-date. Constraint (24) calculates the
total earliness lateness, and Constraint (25) calculates the total weighted earliness
lateness of the job. Finally, Constraint (26) calculates maximum lateness among ‘N’
jobs.
Like minimizing total tardiness (TT) of ‘N’ jobs as a primary scheduling
objective, we can optimize any one the following scheduling criterion/objective as
a primary scheduling criterion/objective as follows:
1. To minimize TWT, we need to have the Constraints (2) to (16) and Constraint
(18) along with Constraint (1) as a new objective function: Min TWT
2. To minimize NT, we need to have the Constraints (2) to (16), Constraint (19),
and Constraint (20) along with Constraint (1) as a new objective function: Min
NT
3. To minimize WNT, we need to have the Constraints (2) to (16), Constraint (19),
and Constraint (21) along with Constraint (1) as a new objective function: Min
WNT
4. To maximize OTD rate, we need to have the Constraints (2) to (16), Constraint
(19) and Constraint (22) along with Constraint (1) as a new objective function:
Max OTD rate
Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 253
5. To minimize TEL, we need to have the Constraints (2) to (15), Constraint (23),
and Constraint (24) along with Constraint (1) as the new objective function: Min
TEL
6. To minimize TWEL, we need to have the Constraints (2) to (15), Constraint (23),
and Constraint (25) along with Constraint (1) as the new objective function: Min
TWEL
7. To minimize Lmax, we need to have the Constraints (2) to (16) and Constraint
(26) along with Constraint (1) as the new objective function: Min Lmax
Validation of the Proposed (0-1) MILP Model for DS-SDF:
To validate the proposed (0-1) MILP model, numerical data (Table 2) represent-
ing the research problem is generated. Further, this study assumed that maximum of
two jobs could be assigned to a batch, and the furnace is available from the second
hour onwards. This numerical example is given as input to the LINGO set code for
generating the proposed (0-1) MILP model and solved using the LINGO solver. The
optimal solution obtained is presented in Table 3. The analysis of this table confirms
the right mathematical representation of the DS-SDF.
The additional/modified notations required towards the proposed model for DS-
NPDF-MER are given here.
Additional Notations
Sets:
M Machines (Diffusion furnaces)
Index:
m 1 . . . L for machines
Parameters:
254 M. Vimala Rani and M. Mathirajan
Table 3 (continued)
Batch (B) Job
Batch Job RT PT CT DD W T E L WT UPJ WNT EL WEL
B6 J5 31 20 51 31 3 3
B7 J3 51 16 67 23 3 3
B8 J10 67 20 87 26 1 1
Weighted number of tardy jobs 7
B1 J1 2 4 6 26 6 0
B2 J4 6 10 16 16 6 0
B3 J8 16 10 26 30 1 0
B4 J6 26 2 28 29 2 0
J7 28 6 0
B5 J2 28 2 30 31 2 0
J9 36 2 0
B6 J5 30 20 50 31 3 1
J10 26 1 1
B7 J3 50 16 66 23 3 1
On time delivery rate (10- unit penalty)/10 0.7
B1 J4 2 10 12 16 6 4 0 4
J8 30 1 18 0 18
B2 J5 12 20 32 31 3 0 1 1
J10 26 1 0 6 6
B3 J2 32 2 34 31 2 0 3 3
J6 29 2 0 5 5
B4 J7 34 2 36 28 6 0 8 8
J9 36 2 0 0 0
B5 J1 36 4 40 26 6 0 14 14
B6 J3 40 16 56 23 3 0 33 33
Total earliness lateness 92
B1 J4 2 10 12 16 6 4 0 24
J8 30 1 18 0 18
B2 J3 12 16 28 23 3 0 5 15
B3 J6 28 2 30 29 2 0 1 2
J7 28 6 0 2 12
B4 J1 30 4 34 26 6 0 8 48
B5 J2 34 2 36 31 2 0 5 10
J9 36 2 0 0 0
B6 J5 36 20 56 31 3 0 25 75
J10 26 1 0 30 30
Total weighted earliness lateness 234
B1 J1 2 4 6 26 6 0
B2 J4 6 10 16 16 6 0
J8 30 1 0
B3 J3 16 16 32 23 3 9
(continued)
256 M. Vimala Rani and M. Mathirajan
Table 3 (continued)
Batch (B) Job
Batch Job RT PT CT DD W T E L WT UPJ WNT EL WEL
B4 J5 32 20 52 31 3 21
J10 26 1 26
B5 J6 52 2 54 29 2 25
J7 28 6 26
B6 J2 54 2 56 31 2 25
J9 36 2 20
Maximum lateness 26
Minimize T T (27)
Subject to
G
Yf bm ≤ 1 ∀b ∈ [1, K] ; ∀m ∈ [1, L] (28)
f =1
Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 257
G
G
Yf,b−1,m ≥ Yf bm ∀b ∈ [2, K] ; ∀m ∈ [1, L] (29)
f =1 f =1
Xj bm ∗ F Ajf ≤ Yf bm ∀j ∈ [1, N ] ; ∀f ∈ [1, G] ; ∀b ∈ [1, K] ; ∀m ∈ [1, L]
(30)
N
Xj bm ∗ FAjf ≥ Yf bm ∀m ∈ [1, L] ; ∀f ∈ [1, G] ; ∀b ∈ [1, K]
j =1
(31)
K
L
Xj bm ∗ MAj m = 1 ∀j ∈ [1, N] (32)
m=1
b=1
N
Xj bm ≤ BC m ∀b ∈ [1, K] ; ∀m ∈ [1, L] (33)
j =1
RT B bm ≥ Am b = 1; ∀m ∈ [1, L] (34)
RT B bm ≥ RT j ∗ Xj bm ∀b ∈ [1, K] ; ∀m ∈ [1, L] ; ∀j ∈ [1, N]
(35)
G
P T B bm = P T f ∗ Yf bm ∀b ∈ [1, K] ; ∀m ∈ [1, L] (37)
f =1
CT j ≥ CT B bm − (BigM) ∗ 1 − Xj bm ∀j ∈ [1, N] ; ∀b ∈ [1, K] ; ∀m ∈ [1, L]
(39)
Tj ≥ CT j − DD j ∀j ∈ [1, N] (42)
258 M. Vimala Rani and M. Mathirajan
Validation of the Proposed (0-1) MILP Model for DS-NPDF-MER: The same
numerical problem presented in Table 2 is used to validate the model for DS-NPDF-
MER. In addition, three non-identical parallel diffusion furnaces: MC1, MC2 and
MC3 are considered for the numerical example with capacity as 2 jobs, 3 jobs and
4 jobs, respectively. This study assumes that the first-time availability of these three
non-identical parallel diffusion furnaces is to be 1st hour, 2nd hour, and 3rd hour,
respectively. Further, jobs that belong to family 3 (f3) are processed only in MC2,
out of three non-identical parallel diffusion furnaces considered here. The same
numerical example, given in Table 2, along with required additional data presented
here, is given as input to the LINGO set code for generating the proposed (0-1)
MILP model and solving using the LINGO solver. The optimal solution reports,
similar to Table 3, are prepared. Due to the brevity of this chapter, the reports are
not presented here. However, by analyzing the reports prepared, the proposed model
for DS-NPDF-MER has been validated.
Since the mathematical model for DS-SDF to minimize TWT is computationally
intractable (Vimala Rani and Mathirajan 2016a), the proposed mathematical model
for DS-NPDF-MER is also computationally intractable. So, we focus on GHA based
on dispatching rules, particularly Apparent Tardiness Cost (ATC) rule, and the same
is presented in the next section.
then modify the DF’s available time as T = T+ delay time due to resource related
RTE.
Step 2: At the time of deciding a batch for scheduling a DF, capture all the
characteristics of the jobs which are waiting for diffusion operation. Further,
capture the DF capacity (B) and the available time (T).
Step 3: Cluster the jobs waiting in front of the DF based on their family.
Step 4: In each cluster, calculate “Job-Priority-Index” for every job and sort the
same.
Step 5: For each cluster, create a “temporary batch” by picking ‘B’ jobs from the
top. If the number of jobs in any of the temporary batch is not equal to ‘B’, then
check the WIP for whether the jobs of the same family are coming for diffusion
operation in the future. If it is true, then wait for those jobs to form a full batch.
Otherwise, form a partially filled batch. Then modify the starting time of the
temporary batch of that cluster as max (T, longest release time of all jobs in that
temporary batch).
Step 6: Compute and compare the starting time of each temporary batch,
If any temporary batch has a completion time strictly less than the starting time
of all other temporary batches, then select it and go to Step 9
Else go to the next step
Step 7: Compute “Batch-Priority-Index” for each temporary batch formed in Step
5.
Step 8: Choose the temporary batch that has the greatest “Batch-Priority-Index”.
Step 9: If any of the job(s) related RTEs occur, then modify the corresponding job-
data on the work-in-process (WIP), else if any of the resource(s) related RTEs
occur, then modify the DF’s available time as T = T+ delay time due to resource
related RTE and go to Step 2.
Else, go to the next step.
Step 10: Assign the selected batch to the diffusion furnace.
Step 11: Compute TWT, NT, WNT, OTD rate, TEL, TWEL, and Lmax of assigned
batch (es).
Step 12: Remove the assigned jobs from WIP and modify the DF’s available time
“T” as the completion time of the assigned batch in Step 10
Step 13: Repeat Step 1 until all jobs are scheduled.
This study proposed and implemented seven different above GHA in Turbo C.
All these seven different GHA differ only in two steps (Step 4 and Step 7), and the
same is presented in Table 4.
The details of ATC based GHA for DRTS of NPDF-MER are given below:
Step 1: At the time of deciding a batch (batch selection) for scheduling a DF, (a) if
any of the job(s) related RTEs occur, then modify the corresponding job-data on
260 M. Vimala Rani and M. Mathirajan
the work-in-process (WIP), and (b) if any of the resource(s) related RTEs occur
then modify the corresponding DF’s (say, DFi) available time as Ti = Ti + delay
time due to resource related RTE.
Step 2: At the time of deciding a batch for scheduling a DF, capture all the
characteristics of the jobs, which are waiting for diffusion operation and the
available time for each of the diffusion furnaces (say ti , ti > 0, where i = 1 to
M and M indicates the number of diffusion furnace).
Step 3: Select the diffusion furnace, which is available earlier.
(i) If a tie occurs in terms of DF available time, then select the DF, which has the
maximum capacity.
(ii) If a tie occurs in terms of DF available time and maximum capacity, then
select the DF, which has machine eligibility restriction.
(iii) If a tie occurs in terms of DF available time, maximum capacity and machine
eligibility restriction, then select the DF arbitrarily.
Step 4: Store the selected DF available time and its capacity to ‘B’ & ‘T’
respectively.
Step 5: Cluster the jobs waiting in front of the selected DF based on their family.
Step 6: For each of the qualified cluster (that is, families, which are possible to
process in a selected diffusion furnace), calculate “Job-Priority-Index” for every
job and sort the same.
Step 7: For each of the qualified clusters, create a “temporary batch” by picking ‘B’
jobs from the top. If the number of jobs in any of the temporary batch in any of
the qualified clusters is not equal to ‘B’ then check the WIP for whether the jobs
of the same family are coming for diffusion operation in the future. If it is true,
then wait for those jobs to form a full batch. Otherwise, form a partially filled
batch. Now, modify the starting time of the corresponding temporary batch as
max (T, longest release time of all jobs in that temporary batch).
Step 8: Compute and compare the starting time of each temporary batch,
Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 261
If any temporary batch has a completion time strictly less than the starting time
of all other temporary batches, then select it and go to Step 11
Else go to the next step
Step 9: Compute Batch-Priority-Index for each temporary batch formed in Step 7.
Step 10: Choose the temporary batch that has the greatest “Batch-Priority-Index”.
Step 11: If any of the job(s) related RTEs occur, then modify the corresponding
job-data on the work-in-process (WIP), and (b) if any of the resource(s) related
RTEs occur, then modify the corresponding DF’s (say, DFi) available time as
Ti = Ti + delay time due to resource related RTE and go to Step 2.
Else, go to the next step.
Step 12: Assign the selected batch to the selected diffusion furnace.
Step 13: Compute TT, TWT, NT, WNT, OTD rate, TEL, TWEL, and Lmax of
assigned batch(es)
Step 14: Remove the assigned jobs from WIP and modify the selected DF’s
available time “Ti ” as the completion time of the allocated batch in Step 12.
Step 15: Repeat Step 1 until all jobs are scheduled.
This study proposed seven different ATC-GHA for DRTS-NPDF-MER by
varying Step 6 and Step 9, as per Table 4 and implement all these seven different
ATC-GHA in Turbo C.
different ATC-GHA, this study computes deviation as per the Eq. (43), for each
problem instances. To know the extent of this deviation w.r.t. benchmark solution
(BSi ), this study divides the deviation of each different ATC-GHA for each problem
instance (Dij ) by corresponding benchmark solution (BSi ) and converts it into
percentage as per equation in (44).
Dij = F S ij − BS i (43)
RP D ij = Dij /BS i ∗ 100 (44)
Where
i: Problem instances and i ∈ [1,450]
j: ATC-GHA and j ∈ [1,7]
FSij : Feasible solution obtained from ‘jth ’ ATC-GHA for ‘ith ’ problem instance
BSi : Benchmark solution for ‘ith ’ problem instance
Dij : Deviation of ‘jth ’ ATC-GHA for ‘ith ’ problem instance
RPDij : Relative percentage deviation of ‘jth ’ ATC-GHA for ‘ith ’ problem instance
Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 263
Then, for each different ATC-GHA, the score on ARPD for each problem
configuration as well as for 450 problem instances are computed as per Eq. (45)
N
ARP D j = RP D ij /N (45)
i=1
Where
ARPDj : Average relative percentage deviation of ‘jth ’ ATC-GHA
N = 10 when ARPD is computed considering problem configuration wise
N = 450 when ARPD is computed considering entire problem instances
Further, for each of the different ATC-GHA, this study computes IRANK, which
is proposed by Mathirajan et al. (2004), for each of the objectives considered in this
study as per the Eq. (46).
Maxrank Maxrank
IRANKj = {N (r, j ) ∗ r} / N (r, j ) (46)
r=1
r=1
Where
r: Rank and r ∈ [1, 7]
N (r, j): Number of times the ‘jth ’ ATC-GHA in rank ‘r’
Maxrank: Maximum Rank possible (Maxrank = 7 = Number of proposed ATC-
GHA).
IRANKj : Integrated rank of ‘jth ’ ATC-GHA
Performance Measures for Statistical Analysis For performance evaluation using
statistical analyses, first we compute descriptive statistics: mean, median, standard
deviation, and 95% confidence interval. As normality test is failed over the obtained
RPD results, Kruskal-Wallis non-parametric test are conducted to compare the
proposed different ATC-GHA (Beldar and Costa 2018).
Empirical analyses of seven different ATC-GHA for DRTS of (i) SDF, and (ii)
NPDF-MER w.r.t. the objectives: TT, TWT, NT, WNT, OTD rate, TEL, TWEL,
and Lmax are discussed here.
Performance Analyses of ATC-GHA for DRTS-SDF: The first ninety problem
instances (small-scale data) are solved by both mathematical model as well as seven
different ATC-GHA for DRTS-SDF in order to obtain the optimal solution and
feasible solution, respectively, for the objective: TT, TWT, NT, WNT, OTD rate,
264 M. Vimala Rani and M. Mathirajan
Fig. 1 Performance of seven different ATC-GHA for DRTS-SDF with respect to optimal: TT,
TWT, NT, WNT, OTD rate, TEL, TWEL, and Lmax: over 90 problem instances
TEL, TWEL, and Lmax. Using the solutions obtained on small scale data, this study
computes RPD for each objective as per the Eq. (44) by considering the optimal
solution as a benchmark solution. Then, for each of the scheduling objectives and
each of the different ATC-GHA, this study calculates the average value of ninety
small-scale problems’ RPD values. ARPD score of seven different ATC-GHA for
each of the objectives are presented in Fig. 1. From this figure, it is observed based
on solving small-scale problems that, by and large, ATC-GHA5, and ATC-GHA6
are performing better for the majority of the objectives considered here.
The reason for the outperforming ATC-GHA may be due to the following:
(a) The ATC rule (given by Li et al. (2010)) considered in ATC-GHA5 helps to
compute exact slack time. This helps to avoid getting high priority for low
Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 265
priority jobs. In addition, the ATC rule (given by Vimala Rani and Mathirajan
(2016a)) considered in ATC-GHA6 ensures that if the slack is negative or the
job is tardy, the ATC rule reduces not only to WSPT rule but also to weighted
shortest tardy.
(b) The BATC index rule (given by Mönch et al. (2006)) considered in ATC-GHA5
and ATC-GHA6 ensures that unavailable jobs are assigned low priority.
Furthermore, this study evaluates seven different ATC-GHA w.r.t. estimated opti-
mal solution (EOS) as (i) the mathematical models are extremely time-consuming,
which is not suitable for large-scale data, and (ii) the findings observed based on
small-scale data may not be generalized to say that this is true for large-scale data
also. Accordingly, EOS is computed as per Rardin and Uzsoy (2001) for every 450
large scale data, and for each of the objective functions.
By considering these estimated optimal solutions as a benchmark solution,
RPD is computed as per the Eq. (44) for each problem instance and for each
objective. Then, for every 45 problem configurations, ARPD is calculated using the
corresponding RPD score obtained over ten problem instances per configuration,
as per Eq. (45) for the objectives: TT, TWT, NT, WNT, OTD rate, TEL, TWEL &
Lmax. Due to the brevity of the study, these results are not presented. However, the
result obtained for the objective TT is given in Table 6, as an example. However,
the average of the RPD values over entire problem instances is computed as per
Eq. (45) by considering ‘N’ as 450 for each objective and is presented in Fig. 2. It
is observed from Table 6 and Fig. 2 that, for large-scale data, by and large, ATC-
GHA5 and ATC-GHA6 are outperforming ATC-GHA for most of the scheduling
objectives. It endorses the observation obtained from small-scale data.
This study triangulates the findings observed from the ARPD analysis by using
another performance measure, IRANK, over 450 problem instances. Accordingly,
a [450×7] ranking matrix, indicating a rank for each of the seven different ATC-
GHA, is constructed for every objective considered here. Using the ranking matrix
[450×7], a frequency ranking matrix [7×7] indicating how many times each of
the proposed ATC-GHA performed to have a particular rank is computed for
every objective. A sample frequency ranking matrix for the scheduling objective
(minimizing TT) is given in Table 7. Using the computed frequency ranking matrix
[7×7] of each of the scheduling objectives, the IRANK of seven different ATC-
GHA is computed using the Eq. (46).
The computed IRANK of seven different ATC-GHA w.r.t. to every objective is
presented in Table 7. This Table indicates that ATC-GHA5 and ATC-GHA6 are
outperforming ATC-GHA for most of the objectives. This is clearly endorsing the
observation obtained from the analysis of ARPD score wr.t. large-scale data.
Performance Analyses of ATC-GHA for DRTS-NPDF-MER Due to computational
intractability in scheduling DRTS-NPDF-MER for real-life sized problems and the
highly questionable nature of generalizability of better performing variants observed
based on small-scale problem, the performance analyses are carried out only
on large-scale data considering both performance measures: ARPD and IRANK.
Similar to single diffusion furnace, for every objective, problem instance wise EOS
Table 6 Performance of seven different ATC-GHA for DRTS-SDF w.r.t. EOS
ARPD score of seven different ATC-GHA in comparison with EOS
S.No Problem configuration ATC-GHA1 ATC-GHA2 ATC-GHA3 ATC-GHA4 ATC-GHA5 ATC-GHA6 ATC-GHA7
1 J1.A1,D1 274.05 32.39 18.05 −0.74 −0.25 −0.25 −0.25
2 J1.A1,D2 861.21 103.36 126.04 127.91 126.04 126.04 −2.62
3 J1,A1,D3 1248.84 5.80 84.06 84.06 84.06 84.06 104.06
4 J1,A2,D1 275.68 228.64 228.64 228.64 228.64 228.64 426.86
5 J1,A2,D2 149.78 0.00 0.00 44.00 0.00 0.00 906.67
6 J1,A2,D3 46.44 22.59 −3.41 71.92 −3.41 −3.41 10.86
7 J1,A3,D1 765.59 76.29 42.73 60.98 42.73 42.73 502.26
8 J1,A3,D2 47.90 0.45 −6.22 0.45 −6.22 −6.22 35.16
9 J1,A3,D3 −13.33 0.00 0.00 0.00 0.00 0.00 0.00
10 J2.A1,D1 105.90 7.65 4.77 4.77 4.77 4.77 33.36
11 J2.A1,D2 261.05 44.00 11.12 8.12 11.12 11.12 8.00
12 J2,A1,D3 1693.88 −2.47 −1.62 −0.76 −1.62 −1.62 57.57
13 J2,A2,D1 31.12 13.48 0.23 8.20 0.23 0.23 11.80
14 J2,A2,D2 972.17 873.77 −5.74 859.24 −5.74 −5.74 866.85
15 J2,A2,D3 404.08 0.92 −1.67 48.71 −1.67 −1.67 41.57
16 J2,A3,D1 82.26 5.48 −0.39 1.37 −0.39 −0.39 3.96
17 J2,A3,D2 64.84 53.17 3.42 39.54 3.42 3.42 50.75
18 J2,A3,D3 416.12 118.91 375.32 142.30 375.32 375.32 142.30
19 J3,A1,D1 102.05 55.73 44.09 58.18 44.41 44.41 61.37
20 J3,A1,D2 294.71 46.08 19.23 54.48 18.32 20.59 73.82
21 J3,A1,D3 142.79 29.44 38.85 55.72 38.85 40.50 58.73
22 J3,A2,D1 56.80 10.11 4.31 7.47 4.67 4.67 17.92
23 J3,A2,D2 110.47 49.76 22.76 39.79 22.76 22.76 42.07
(continued)
Table 6 (continued)
24 J3,A2,D3 351.55 141.99 79.43 144.26 77.62 77.94 71.82
25 J3,A3,D1 99.88 30.27 31.57 21.68 31.57 32.08 31.87
26 J3,A3,D2 169.85 78.57 65.15 76.41 67.30 67.54 67.19
27 J3,A3,D3 172.80 51.21 33.41 63.67 33.41 33.41 60.41
28 J4.A1,D1 59.78 24.94 31.10 29.32 32.15 32.17 47.45
29 J4.A1,D2 50.37 18.29 33.48 32.61 30.61 31.40 35.17
30 J4,A1,D3 85.67 39.23 59.25 46.41 58.34 61.46 72.62
31 J4,A2,D1 43.18 27.19 25.02 27.86 27.42 27.35 40.22
32 J4,A2,D2 36.14 8.22 5.40 7.74 9.00 8.46 13.46
33 J4,A2,D3 58.09 34.34 30.63 30.89 35.04 37.17 46.05
34 J4,A3,D1 38.08 20.26 13.51 23.77 12.43 12.79 21.44
35 J4,A3,D2 55.54 40.92 32.75 38.17 32.90 33.59 34.55
36 J4,A3,D3 94.98 23.03 17.71 16.78 18.01 22.71 48.84
37 J5,A1,D1 8.88 7.94 −3.27 4.09 −2.61 −2.91 9.12
38 J5,A1,D2 23.71 25.64 36.26 40.59 32.25 29.74 45.23
39 J5,A1,D3 24.09 23.07 13.33 16.32 19.80 28.19 28.03
40 J5,A2,D1 37.56 23.61 32.34 38.72 33.20 26.31 41.10
41 J5,A2,D2 35.70 27.30 24.87 27.36 29.07 29.92 40.76
42 J5,A2,D3 23.96 24.24 18.22 17.62 18.94 23.36 31.69
43 J5,A3,D1 38.40 22.13 16.40 19.31 17.31 20.79 17.18
44 J5,A3,D2 35.95 16.40 11.60 18.21 16.73 20.91 26.71
45 J5,A3,D3 46.70 34.61 10.51 34.19 14.42 15.95 21.05
268 M. Vimala Rani and M. Mathirajan
Fig. 2 Performance of seven different ATC-GHA for DRTS-SDF w.r.t. estimated optimal: TT,
TWT, NT, WNT, OTD rate, TEL, TWEL, and Lmax: over 450 problem instances
is computed using seven objective values, and then using these EOS as benchmark
solution, RPD is computed as per Eq. (44). Then, objective-wise, the score on ARPD
for (i) 10 problem instances over each of the problem configuration, and (ii) 450
problem instances are computed for seven different ATC-GHA for DRTS-NPDF-
MER. To keep this chapter brief, the results developed from the solution obtained,
similar to Table 6 and Fig. 2, are not presented here.
Similar to single diffusion furnace, IRANK performance score, while scheduling
DRTS-NPDF-MER over 450 problem instances, was computed to triangulate the
analysis with ARPD score. To keep this chapter brief, the computed IRANK of
seven different ATG-GHA for DRTS-NPDF-MER is not presented here.
Table 7 Performance of seven different ATC-GHA for DRTS-SDF, over 450 instances, w.r.t. IRANK
Number of times the proposed ATC-GHAs for
DRTS-SDF yielded relatively best TT with the The ranking position, based on IRANK (over 450 problem instances),
rank position of of seven different ATC-GHA for DRTS-SDF, the scheduling objective
Proposed ATC-GHAs 1 2 3 4 5 6 7 TT TWT NT WNT OTD rate TEL WTEL Lmax
ATC-GHA1 77 79 80 63 64 56 31 7 7 7 7 7 6 7 3
ATC-GHA2 153 102 65 51 33 30 16 4 3 3 3 3 4 3 4
ATC-GHA3 203 92 68 47 22 14 4 1 4 4 4 4 1 5 1
ATC-GHA4 135 115 81 42 35 25 17 5 5 5 5 5 5 4 4
ATC-GHA5 174 110 64 51 34 15 2 2 1 2 2 2 2 1 2
ATC-GHA6 178 98 58 49 42 16 9 3 2 1 1 1 3 2 5
ATC-GHA7 145 87 72 41 42 32 31 6 6 6 6 6 7 6 6
270 M. Vimala Rani and M. Mathirajan
However, from the analyses of the ARPD scores with respect to configuration-
wise (10 problem instances per configuration) and irrespective of the configuration
(i.e. overall 450 instances), it is observed that by and large, ATC-GHA5 and ATC-
GHA6 are continuously performing better for the majority of the objectives in the
large-scale data. This observation was further endorsed by analyzing the IRANK
score. The better performing ATC-GHA, observed based on the empirical analyses
for both cases of DRTS-SDF and DRTS-NPDF-MER, are further confirmed by
statistical analysis, and the same is discussed in the next section.
In addition to empirical analyses presented here, this study conducts statistical anal-
ysis using SPSS software. Accordingly, this study computes descriptive statistics:
mean, median, standard deviation, and 95% confidence interval, for seven different
ATC-GHA for DRTS of SDF w.r.t. every objective considered here, using 450
objective values obtained corresponding to 450 problem instances and is given in
Table 8. It is observed from Table 8 that ATC-GHA5 is continuously outperforming
for all the objectives. Moreover, ATC-GHA6 is also performing better for most of
the objectives.
Further, for each objective, this study desires to check whether the distribution of
RPD across seven different ATC-GHA for DRTS-SDF is the same or not. For that,
this study conducts Kruskal-Wallis non-parametric test, and the results are presented
in Table 9a, b. From Table 9a, it is observed that a statistically significant difference
is there in the distribution of RPD values across seven different ATC-GHA, as the P-
Values is 0.000, which is less than 0.05. In addition, Table 9b clearly says that ATC-
GHA5 and ATC-GHA6 are relatively outperforming heuristic algorithms among
seven different ATC-GHA for DRTS-SDF to optimize various objectives considered
here.
Similar to statistical analyses carried out on seven different ATC-GHA for DRTS
of SDF were performed on seven different ATC-GHA for DRTS-NPDF-MER.
Due to the brevity of the report, the statistical results obtained are not presented
here. However, the inference obtained from the statistical results endorses the same
inferences obtained with respect to seven different ATC-GHA for DRTS-SDF to
optimize various objectives. From both the analyses, this study detected that by
and large ATC-GHA5, and ATC-GHA6 are continuously performing better for both
DRTS-SDF and DRTS-NPDF-MER to optimize most of the objectives considered
in this study. This could be due to the job-priority-index and batch-priority-index
computed by these ATC-GHA that could yield an exact value of the slack. And this
helps to avoid getting high priority for low priority jobs or unavailable jobs. Further,
for the tardy jobs, these rules give a high priority to the job, which has the largest
weighted tardy.
Table 8 Descriptive statistics of the proposed seven different ATC-GHA for DRTS-SDF
Descriptive Statistics
Proposed ATC-GHA Scheduling Objective Mean Median Standard Deviation 95% Confidence Interval
ATC-GHA1 TT 1392.99 407.00 2005.15 (1207.73,1578.25)
ATC-GHA2 1248.85 248.50 1879.97 (1075.15,1422.54)
ATC-GHA3 1209.36 239.00 1840.64 (1039.29,1379.42)
ATC-GHA4 1263.90 268.00 1919.33 (1086.57,1441.24)
ATC-GHA5 1225.41 239.00 1862.93 (1053.28,1397.53)
ATC-GHA6 1240.50 239.00 1876.87 (1067.09,1413.91)
ATC-GHA7 1322.84 278.50 2002.29 (1137.84,1507.84)
ATC-GHA1 TWT 7617.35 2169.00 11048.96 (6596.58638.2)
ATC-GHA2 5245.97 1235.50 7769.11 (4528.16,5963.79)
ATC-GHA3 5314.14 1169.00 7996.89 (4575.28,6053)
ATC-GHA4 5422.12 1393.50 8125.54 (4671.37,6172.87)
ATC-GHA5 5195.25 1137.00 7834.27 (4471.41,5919.09)
ATC-GHA6 5253.20 1137.00 7889.45 (4524.26,5982.13)
ATC-GHA7 5462.17 1405.00 8115.24 (4712.38,6211.97)
ATC-GHA1 NT 30.61 17.50 32.44 (27.61,33.61)
ATC-GHA2 25.98 13.00 29.17 (23.28,28.67)
ATC-GHA3 25.99 13.00 29.30 (23.28,28.69)
ATC-GHA4 26.42 13.00 29.51 (23.69,29.14)
ATC-GHA5 25.80 13.00 28.97 (23.12,28.48)
ATC-GHA6 25.46 13.00 28.58 (22.82,28.1)
ATC-GHA7 26.46 13.00 29.42 (23.75,29.18)
(continued)
Table 8 (continued)
Descriptive Statistics
Proposed ATC-GHA Scheduling Objective Mean Median Standard Deviation 95% Confidence Interval
ATC-GHA1 WNT 168.45 97.00 178.90 (151.92,184.98)
ATC-GHA2 137.35 68.00 154.93 (123.03,151.66)
ATC-GHA3 138.69 70.00 156.85 (124.2153.18)
ATC-GHA4 140.84 70.00 157.83 (126.26,155.42)
ATC-GHA5 137.12 70.00 154.15 (122.88,151.37)
ATC-GHA6 135.48 70.00 152.96 (121.34,149.61)
ATC-GHA7 141.47 74.50 157.89 (126.88,156.06)
ATC-GHA1 OTD rate 0.34 0.26 0.27 (0.32,0.37)
ATC-GHA2 0.46 0.40 0.26 (0.44,0.49)
ATC-GHA3 0.47 0.40 0.26 (0.44,0.49)
ATC-GHA4 0.46 0.40 0.26 (0.43,0.48)
ATC-GHA5 0.47 0.40 0.26 (0.45,0.49)
ATC-GHA6 0.47 0.40 0.25 (0.45,0.5)
ATC-GHA7 0.45 0.40 0.26 (0.42,0.47)
ATC-GHA1 TEL 1470.75 460.00 2010.13 (1285.03,1656.48)
ATC-GHA2 1399.99 452.00 1908.26 (1223.68,1576.3)
ATC-GHA3 1370.57 428.00 1869.04 (1197.88,1543.25)
ATC-GHA4 1413.53 440.00 1939.35 (1234.35,1592.72)
ATC-GHA5 1384.05 429.00 1889.40 (1209.48,1558.62)
ATC-GHA6 1402.19 431.50 1909.40 (1225.78,1578.61)
ATC-GHA7 1472.40 438.00 2027.01 (1285.12,1659.69)
(continued)
Table 8 (continued)
ATC-GHA1 TWEL 8049.60 2613.00 11078.13 (7026.06,9073.15)
ATC-GHA2 6119.09 2274.50 7951.30 (5384.44,6853.74)
ATC-GHA3 6245.64 2338.50 8194.66 (5488.51,7002.78)
ATC-GHA4 6283.41 2386.00 8260.75 (5520.17,7046.65)
ATC-GHA5 6108.99 2270.50 8017.37 (5368.24,6849.74)
ATC-GHA6 6181.25 2270.50 8090.68 (5433.72,6928.77)
ATC-GHA7 6317.45 2319.50 8277.90 (5552.62,7082.27)
ATC-GHA1 Lmax 70.20 55.00 56.54 (64.97,75.42)
ATC-GHA2 70.53 50.00 62.41 (64.76,76.29)
ATC-GHA3 68.33 49.00 61.50 (62.64,74.01)
ATC-GHA4 70.32 50.00 62.20 (64.58,76.07)
ATC-GHA5 69.12 49.00 62.19 (63.37,74.86)
ATC-GHA6 70.52 49.00 63.19 (64.69,76.36)
ATC-GHA7 73.04 55.00 63.59 (67.16,78.91)
Table 9 Kruskal-Wallis test results on seven different ATC-GHA for DRTS-SDF
(a): Test Statistics for the Objective
TT TWT NT WNT OTD Rate TEL TWEL Lmax
Kruskal- 213.202 293.054 165.270 160.027 243.242 70.504 147.988 62.566
Wallis
H
Degrees of 6 6 6 6 6 6 6 6
freedom
Asymp. 0 0 0 0 0 0 0 0
Sig.
(b): Mean Rank for the seven different ATC-GHA w.r.t the Objective
Proposed ATC-GHA N TT TWT NT WNT OTD Rate TEL TWEL Lmax
ATC-GHA1 450 2060.41 2204.68 2032.64 2042.05 2165.02 1822.38 2032.40 1768.72
ATC-GHA2 450 1590.35 1511.50 1534.63 1525.30 1512.24 1561.60 1504.32 1608.47
ATC-GHA3 450 1350.76 1409.86 1451.84 1455.17 1432.73 1431.27 1476.61 1420.86
ATC-GHA4 450 1575.73 1567.92 1551.39 1558.45 1531.01 1562.18 1546.32 1581.76
ATC-GHA5 450 1362.28 1345.14 1423.02 1421.52 1391.54 1443.44 1406.56 1432.56
ATC-GHA6 450 1392.81 1373.55 1410.03 1419.16 1390.06 1488.08 1451.08 1496.05
ATC-GHA7 450 1696.16 1615.85 1624.96 1606.86 1605.89 1719.55 1611.20 1720.09
Total 3150
Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 275
7 Conclusion
A new research problem on DRTS of DF(s), considering the job and/or resource
related RTE along with different job-arrival time, machine eligibility restriction,
incompatible-job families, and non-agreeable release time & due-date to optimize
one at a time the objectives: TT, TWT, NT, WNT, OTD rate, TEL, TWEL, and Lmax
is considered in this study. (0-1) MILP models for DS-SDF, as well as DS-NPDF-
MER, considering each of the objectives, are proposed. LINGO Set Code for all
the (0-1) MILP models is also developed. All the proposed models are validated by
solving ninety small-scale data.
Since the mathematical models are not amenable to solve large-scale data, this
study proposed seven different ATC-GHA for (i) DRTS-SDF, and (ii) DRTS-NPDF-
MER. The performance evaluation of seven different ATC-GHA for (i) DRTS-SDF,
and (ii) DRTS-NPDF-MER is carried out in comparison with (a) the optimal value
for 90 small-scale data, and (b) the estimated optimal value for 450 large scale data,
which are generated based on the experimental design, using ARPD and IRANK
measure.
In addition to the empirical analyses, this study employs statistical analyses:
descriptive statistics and Kruskal Wallis test. From both the analyses, this study
concludes that, by and large, ATC-GHA5 and ATC-GHA6 are outperforming GHA
for DRTS-SDF as well as DRTS-NPDF-MER to optimize the objectives: TT, TWT,
NT, WNT, OTD rate, TEL, TWEL, and Lmax, one at a time. This could be because
the job-priority-index and batch-priority-index computed by these ATC-GHA that
could provide an accurate value of slack. This helps to avoid getting high priority
for low priority jobs or unavailable jobs. Further, for the tardy jobs, these rules give
a high priority to jobs which have large weighted tardy.
The important managerial implication stems from identifying better performing
ATC-GHA that would help the production manager to generate an efficient schedule
even if any unexpected real time events occur. Further, it helps to deliver the
product on or before its due date, which helps to improve the performance of the
semiconductor manufacturing industry.
The bench mark solution procedure: estimated optimal solution procedure has its
own limitations. Moreover, the problem instances used in this study are generated
not collected from SM industry. Considering (a) completion time-based scheduling
objectives, and (b) upstream/downstream operation(s) while doing DRTS of dif-
fusion furnace(s) without the limitation listed above would be the future research
direction in this area.
References
Balasubramanian, H., Monch, L., Fowler, J., & Pfund, M. (2004, January 01). Genetic algorithm
based scheduling of parallel batch machines with incompatible job families to minimize total
weighted tardiness. Int J Prod Res, 42(8), 1621–1638.
276 M. Vimala Rani and M. Mathirajan
Bar-Noy, A., Guha, S., Katz, Y., Naor, J. S., Schieber, B., & Shachnai, H. (2009, March
01). Throughput maximization of real-time scheduling with batching. ACM Transactions on
Algorithms, 5(2), 1–17.
Beldar, P., & Costa, A. (2018, January 01). Single machine batch processing problem with release
dates to minimize total completion time. International Journal of Industrial Engineering
Computations, 9(3), 331–348.
Bilyk, A., Mönch, L., & Almeder, C. (2014, October 23). Scheduling jobs with ready times
and precedence constraints on parallel batch machines using meta heuristics. Computers &
Industrial Engineering, 78, 175–185.
Chen, L., Lu, C., Hui, X., & Li, L. (2013, December 01). Learning-based adaptive dispatching
method for batch processing machines. In Proceedings of the 2013 Winter Simulation Confer-
ence, pp. 3756–3765.
Cheng, H. C., Chiang, T. C., & Fu, L. C. (2008, December 01). A memetic algorithm for
parallel batch machine scheduling with incompatible job families and dynamic job arrivals.
Proceedings of IEEE International Conference on Systems, Man and Cybernetics, 2008, 541–
546.
Chiang, T.-C., Cheng, H.-C., & Fu, L.-C. (2008, November 01). An efficient heuristic for mini-
mizing maximum lateness on parallel batch machines. Proceedings of the Eighth International
Conference on Intelligent Systems Design and Applications, 2, 621–627.
Chiang, T.-C., Cheng, H.-C., & Fu, L.-C. (2010, December 01). A memetic algorithm for
minimizing total weighted tardiness on parallel batch machines with incompatible job families
and dynamic job arrival. Computers & Operations Research, 37(12), 2257–2269.
Dirk, R., & Monch, L. (2006, January 01). Multiobjective scheduling of jobs with incompatible
families on parallel batch machines. Lecture Notes in Computer Science, 3906, 209–221.
Fidelis, M. B., & Arroyo, J. E. C. (2017, October 01). Meta-heuristic algorithms for scheduling on
parallel batch machines with unequal job ready times. Proceedings of the IEEE International
Conference on Systems, Man and Cybernetics, 2017, 542–547.
Guo, C., Jiang, Z., & Hu, H. (2010, December 01). A hybrid ant colony optimization method for
scheduling batch processing machine in the semiconductor manufacturing. IEEE International
Conference on Industrial Engineering & Engineering Management (IE&EM), 2010, 1698–
1701.
Hildebrandt T, Heger J, Scholz-Reiter B (2010, July 07). Towards improved dispatching rules
for complex shop floor scenarios—A genetic programming approach. In Proceeding of
the GECCO’10, July 7–11, Portland, Oregon, USA. Retrieved April 3, 2020, from http://
www.genetic-programming.org/hc2010/6-Hildebrandt/Hildebrandt-Paper.pdf
Jia, W., Jiang, Z., & Li, Y. (2013a, August 01). Closed loop control-based real-time dispatching
heuristic on parallel batch machines with incompatible job families and dynamic arrivals.
International Journal of Production Research, 51(15), 4570–4584.
Jia, W., Jiang, Z., & Li, Y., (2013b, August 01). A job-family-oriented algorithm for re-entrant
batch processing machine scheduling. In IEEE International Conference on Automation
Science and Engineering, pp. 1022–1027.
Jung, C., Pabst, D., Ham, M., Stehli, M., & Rothe, M. (2014, August 01). An effective problem
decomposition method for scheduling of diffusion processes based on mixed integer linear
programming. IEEE Transactions on Semiconductor Manufacturing, 27(3), 357–363.
Kim, Y. D., Joo, B. J., & Choi, S. Y. (2010, January 01). Scheduling wafer lots on diffusion
machines in a semiconductor wafer fabrication facility. IEEE Transactions on Semiconductor
Manufacturing, 23(2), 246–254.
Koo, P.-H., & Moon, D. H. (2013, January 01). A review on control strategies of batch processing
machines in semiconductor manufacturing. IFAC Proceedings, 46(9), 1690–1695.
Kurz, M.E. (2003, May 22). On the structure of optimal schedules for minimizing total weighted
tardiness on parallel, batch-processing machines. In Proceedings of 10th IE Research Confer-
ence, Portland, pp. 1–5.
Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 277
Kurz, M. E., & Mason, S. J. (2008, January 01). Minimizing total weighted tardiness on a batch-
processing machine with incompatible job families and job ready times. International Journal
of Production Research, 46(1), 131–151.
Li, L., &Qiao, F. (2008, August 01). ACO-based scheduling for a single batch processing machine
in semiconductor manufacturing. In 4th IEEE International Conference on Automation Science
and Engineering, pp. 85–90.
Li, L., Qiao, F., & Wu, Q.D. (2009, August 01). ACO-based scheduling of parallel batch
processing machines to minimize the total weighted tardiness. In 5th Annual IEEE Conference
on Automation Science and Engineering, pp. 280–285.
Li, L., Qiao, F., & Pan, G. (2010, August 17). ACO-based multi-objective scheduling of
identical parallel batch processing machines in semiconductor manufacturing. INTECH Open
Access Publisher. Retrieved April 3, 2020, from http://www.intechopen.com/books/future-
manufacturing-systems/aco-based-multi-objective-scheduling-of-identical-parallel-batch-
processing-machines-in-semiconducto
Li, Y., Jiang, Z., & Jia, W. (2017, January 02). API-based two-dimensional dispatching decision-
making approach for semiconductor wafer fabrication with operation due date-related objec-
tives. International Journal of Production Research, 55(1), 79–95.
Malve, S., & Uzsoy, R. (2007, January 01). A genetic algorithm for minimizing maximum lateness
on parallel identical batch processing machines with dynamic job arrivals and incompatible job
families. Computers and Operations Research, 34(10), 3016–3028.
Mansoer, P., & Koo, P.-H. (2015, January 01). A batching strategy for batch processing machine
with multiple product types. Journal of Industrial and Intelligent Information, 3(2), 138–142.
Mathirajan, M., & Vimalarani, M. (2012, December 13). Scheduling a BPM with
incompatible job-families and dynamic job-arrivals. In Proceedings of the 2012 IEEE
International Conference on Industrial Engineering and Engineering Management,
pp. 622–626. Retrieved April 3, 2020, from https://www.researchgate.net/publication/
265597365_Scheduling_a_BPM_with_Incompatible_Job-Families_and_Dynamic_Job-
Arrivals
Mathirajan, M., & Sivakumar, A. I. (2006, July 01). A literature review, classification and simple
meta-analysis on scheduling of batch processors in semiconductor. The International Journal
of Advanced Manufacturing Technology, 29, 990–1001.
Mathirajan, M., Sivakumar, A. I., & Kalaivani, P. (November, 2004). An efficient simulated
annealing algorithm for scheduling burn-in oven with non-identical job-sizes. The International
Journal of Applied Management and Technology, 2(2), 117–138.
Mehta, S. V., & Uzsoy, R. (1998, February 01). Minimizing total tardiness on a batch processing
machine with incompatible job families. IIE Transactions, 30(2), 165–178.
Mönch, L., & Roob, S. (2018, January 01). A matheuristic framework for batch machine
scheduling problems with incompatible job families and regular sum objective. Applied Soft
Computing, 68, 835–846.
Mönch, L., Balasubramanian, H., Fowler, J. W., & Pfund, M. E. (2005, November 01). Heuristic
scheduling of jobs on parallel batch machines with incompatible job families and unequal ready
times. Computers & Operations Research, 32(11), 2731–2750.
Mönch, L., Zimmermann, J., & Otto, P. (2006, April 01). Machine learning techniques for
scheduling jobs with incompatible families and unequal ready times on parallel batch machines.
Engineering Applications of Artificial Intelligence, 19(3), 235–245.
Monch, L., Fowler, J., Dauzère-Pérès, S., Mason, S., & Rose, O. (2011, January 07). A
survey of problems, solution techniques, and future challenges in scheduling semiconductor
manufacturing operations. Journal of Scheduling, 14(6), 583–599.
Mönch, L., Fowler, J., & Mason, S. J. (2013). Production planning and control for semiconductor
wafer fabrication facilities: Modeling, analysis, and systems. New York, NY: Springer.
Mönch, L., Chien, C.-F., Dauzère-Pérès, S., Ehm, H., & Fowler, J. W. (2018, July 03). Modelling
and analysis of semiconductor supply chains. International Journal of Production Research,
56(13), 4521–4523.
278 M. Vimala Rani and M. Mathirajan
Park, H., & Banerjee, A. (2011, December 01). A new dynamic scheduling for batch processing
systems using stochastic utility evaluation function. Proceedings of the 2011 Winter Simulation
Conference, pp. 2302–2319.
Pinedo, M. (2012). Scheduling: Theory, algorithms, and systems. New York: Springer.
Pirovano, G., Ciccullo, F., Pero, M., & Rossi, T. (2020, January 01). Scheduling batches with time
constraints in wafer fabrication. International Journal of Operational Research, 37(1), 1.
Rardin, R. L., & Uzsoy, R. (2001, January 01). Experimental evaluation of Heuristic optimization
algorithms: A tutorial. Journal of Heuristics, 7, 261–304.
Rocholl, J., Monch, L., & Fowler, J. W. (2018, December 01). Electricity power cost-aware
scheduling of jobs on parallel batch processing machines. In Proceedings - Winter Simulation
Conference, pp. 3420–3431.
Sarin, S. C., &Shikalgar, S. T. (2001, August 25). Reduction of average cycle time at a wafer
fabrication facility. GaAs MANTECH, Inc, Grado Department of Industrial and Systems
Engineering, Virginia Tech. Retrieved April 3, 2020, from https://pdfs.semanticscholar.org/
d3fd/2c0e192b6bcde5d2166ac7d91a2e81ac0917.pdf
SGSR: Semiconductor Global Sales Report. (2020). Retrieved August 28, 2020, from https://
www.statista.com/statistics/266973/global-semiconductor-sales-since-1988/
Sun, C., & Rose, T. (2015, January 01). Supply Chain complexity in the semiconductor industry:
Assessment from system view and the impact of changes. IFAC PapersOnline, 48(3), 1210–
1215.
Uzsoy, R. (1995, October 01). Scheduling batch processing machines with incompatible job
families. International Journal of Production Research, 33(10), 2685–2708.
Vimala Rani, M., & Mathirajan, M. (2015, December 01). Dynamic scheduling of diffusion
furnace in semiconductor manufacturing with real time events. IEEE International Con-
ference on Industrial Engineering and Engineering Management. https://doi.org/10.1109/
IEEM.2015.7385617.
Vimala Rani, M., & Mathirajan, M. (September 02, 2016a). Performance evaluation of ATC based
greedy heuristic algorithms in scheduling diffusion furnace in wafer fabrication. Journal of
Information and Optimization Sciences, 37(5), 717–762.
Vimala Rani, M., & Mathirajan, M. (2016b, May 05). Dynamic scheduling of diffusion furnace
in semiconductor manufacturing with resource related real time events. In Analytics in
operations/supply chain management (pp. 381–396). New Delhi: I.K International Publishing
House.
Vimala Rani, M., & Mathirajan, M. (2016c, January 01). Multi objective dynamic real-time
scheduling of batch processing machine. International Journal of Operations and Quantitative
Management, 22(1), 53–73.
Vimala Rani, M., & Mathirajan, M. (2020a, June 01). Performance evaluation of due-date based
dispatching rules in dynamic scheduling of diffusion furnace. OPSEARCH, 57(2), 462–512.
Vimala Rani, M., & Mathirajan, M. (2020b, January 01). Meta-heuristics for dynamic real
time scheduling of diffusion furnace in semiconductor manufacturing industry. International
Journal of Industrial and Systems Engineering, 34(3), 365–395.
Wang, R., & Fei, S. (2014, January 01). Rescheduling: External environment-related real-time
events. IFAC Proceedings Volumes, 47(3), 10743–10747.