Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

This article was downloaded by: [University of California Santa Cruz]

On: 22 November 2014, At: 03:07


Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer
House, 37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Computer Integrated


Manufacturing
Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/tcim20

A multi-criterion genetic algorithm for order


distribution in a demand driven supply chain
Felix TS Chan & SH Chung
a
Department of Industrial and Manufacturing Systems Engineering , The University of
Hong Kong , Pokfulam Road, Hong Kong, People's Republic of China E-mail:
Published online: 19 Feb 2007.

To cite this article: Felix TS Chan & SH Chung (2004) A multi-criterion genetic algorithm for order distribution in
a demand driven supply chain, International Journal of Computer Integrated Manufacturing, 17:4, 339-351, DOI:
10.1080/09511920310001617022

To link to this article: http://dx.doi.org/10.1080/09511920310001617022

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of
the Content. Any opinions and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied
upon and should be independently verified with primary sources of information. Taylor and Francis shall
not be liable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other
liabilities whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or
arising out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
INT. J. COMPUTER INTEGRATED MANUFACTURING, JUNE 2004, VOL. 17, NO. 4, 339–351

A multi-criterion genetic algorithm for order


distribution in a demand driven supply chain
FELIX T. S. CHAN and S. H. CHUNG

Abstract. This paper develops a multi-criterion genetic Supply chain collaboration benefits include faster
optimization procedure, specifically designed for solving customer responsiveness, an increment of flexibility for
optimization problems in supply chain management. The
changing market conditions, an improvement in
Downloaded by [University of California Santa Cruz] at 03:07 22 November 2014

proposed algorithm is discussed with an order distribution


problem in a demand driven supply chain network. It customer service and satisfaction, and more customer
combines the analytic hierarchy process (AHP) with genetic retention (Cooke 1998, Stank et al. 1999, Horvath
algorithms. AHP is utilized to evaluate the fitness values of 2001). The first initiative of supply chain integration
chromosomes. The proposed algorithm allows decision- can be dated back to 1992 when 14 trade association
makers to give weighting for criteria using a pair-wise
comparison approach. The numerical results obtained from sponsors created a group named the ‘Efficient Con-
the proposed algorithm are compared with the one obtained sumer Response (ECR) Movement’ (Robins 1994,
from the multi-objective mixed integer programming ap- Barratt and Oliveira 2001). Three years later, five
proach. The comparison shows that the proposed algorithm is companies: the Benchmarking Partners, Warner-Lam-
reliable and robust. In addition, it provides more control and bert, Wal-Mart Stores, SAP, and Manaugistics worked on
information for the decision-makers to gain a better insight of
the supply chain network. the Collaborative Planning, Forecasting and Replenish-
ment (CPFR) project. The project goals were to
improve the business in the area of forecast accuracy,
store in-stock percentage, total inventory (for the
1. Introduction retailer and manufacturer combined), and cost (Cooke
1998). CPFR attempts to bring organizations (retailer
In supply chain management, optimization of multi- and manufacturer) together to make joint plans,
criterion problems may usually be encountered by including promotion sales, procurement, replenish-
decision-makers (DM), such as total costs, quality of ment, and logistics planning. Blair (1998) stated that
product, customer service level, inventory level, manu- CPFR has won the support of organizations in the drug,
facturing lead times, resources utilization, number of grocery, general merchandise, and apparel industries.
employees, material resources, etc. Each criterion may Stank et al. (1999) showed that there is a strong support
be limited with some constraints, highly interrelated, in utilizing automatic inventory replenishment in cross-
and influencing each other in nature. For instance, to organizational collaboration.
maximize the profit, minimize the total cost, and Barratt and Oliveira (2001) indicated that two
maximize the quality are conflicting aims in practice. significant barriers, encountered in the implementa-
The optimization complexity increases as supply chain tion of supply chain integration, are: (i) the lack of
collaboration becomes more prevalent in today’s visibility of true customer demand, and (ii) the
business model. An organization not only optimizes collaborative relationships in the area involved in joint
itself, but also collaborates with other organizations for decision-making. Within a supply chain partnership,
larger optimization planning. traditional competitive barriers between supply chain
members are mitigated to create mutually beneficial
relationships (Maloni and Benton 1997). Donselaar et
al. (2001) suggested that inventory reduction could be
achieved through the use of advanced demand
Authors: Felix T. S. Chan and S. H. Chung, Department of Industrial and
Manufacturing Systems Engineering, The University of Hong Kong, Pokfulam information. A superior supply chain strategy is the
Road, Hong Kong, People’s Republic of China. e-mail: ftschan@hkucc.hku.hk one that maximizes the values added by the internal
International Journal of Computer Integrated Manufacturing
ISSN 0951-192X print/ISSN 1362-3052 online # 2004 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/09511920310001617022
340 F. T. S. Chan and S. H. Chung

activities while developing solid partnerships leading to for supply chain management, should be capable of
high value external activities (Lakhal et al. 2001). New considering more than one single factor. In addition,
and Ramsay (1997) also stated that there is a danger of since different factors have different weightings in
unfair distribution of costs and benefits in practice. This various structures of the supply chain environment and
may be due to the inequality of power between the in different situations, consideration of weighting is also
organizations. Neuman and Samuels (1996) suggested critical.
that fairly sharing the gains from the supply chain is The remainder of this paper is divided into the
critical. Researchers have also investigated which following sections. Section 2 describes the problem
organization should be the central decision-maker to structure. Section 3 presents the optimization functions.
decide part or all of the coordination decisions among Section 4 discusses the proposed algorithm and section
the collaborating organizations (Pfohl and Buse 2001). 5 analyses the optimization results. Finally, section 6
Horvath (2001) suggested some fundamental concludes.
attributes of supply chain collaboration, including
open, low-cost connectivity, flexible, multimedia data
storage, systems and channel integration etc. Fourth Glossary
Downloaded by [University of California Santa Cruz] at 03:07 22 November 2014

party logistics (4PL) can provide a practical frame-


work for supply chain integration, and highlight avg fn average fitness value of the pool,
some characteristics of a central coordinator (UPS avg. UC average utilization of capacity,
Logistics Group 2001). Hoek and Chong (2001) Cdij delivery unit cost for each unit of product
presented the idea of 4PL from the experiences of transported from manufacturer (i) to
UPS Worldwide Logistics. 4PL is a supply chain customer (j),
service provider that participates in supply chain CMi maximum capacity of supply per week for
coordination rather than in supply chain operational Mi ,
services. The concept of Third-party Value-Added Cui production unit cost for Mi,
Networks (Third-party VAN) also illustrates the idea Dij quantity of Dj allocated to Mi,
of a central coordinator. First, a buyer sends a Dj demand from customer j,
number of purchase orders (POs) to the third-party EU equity on utilization ratios,
VAN. Second, the third-party VAN then sorts those fn fitness value,
received POs and transmits them to different Lij delivery day for each unit of product
suppliers (Ballou 1999). However, there is not much transported from manufacturer (i) to
research focused on the orders’ distribution metho- customer (j),
dology, which takes into consideration the problems mag fn magnified fitness value,
encountered in a supply chain. Mi manufacturer i,
Without randomness in behaviour and uncertain- MSE mean square error,
ties, and under a static and certainty condition, an RA relative value in chromosome A,
order distribution problem can be simplified as an RB relative value in chromosome B,
operations research problem such as resource alloca- TC total cost,
tion, transportation, or production. These problems TL total delivery lead times,
can be solved by linear programming, fractional UCi utilization of capacity for Mi.
programming, or multi-objective linear fractional pro-
gramming (Chakraborty and Gupta 2002). Similarly,
genetic algorithms (GAs) are widely adapted by recent 2. The problem structure
researchers on solving these problems. Traditionally,
optimization of the total cost is the most concerned Consider a hypothetical two-layer demand driven
factor. Hoek and Chong (2001) suggested that costs are supply chain with three manufacturers Mi = {M1, M2,
the traditional criterion in the development of logistics M3}, and ten customers Dj = {D1, D2, . . . , D10}, in which
services. In the supply chain, customer services are each customer will release one demand with a total of
more critical. ten orders. Each manufacturer is an individual organi-
In the supply chain environment, there are many zation, and they are coordinated/cooperate to imple-
decision factors that need to be considered. They are ment supply chain collaboration (in a horizontal
highly interrelated and influence each other. Most of manner). Suppose a central coordinator is introduced
the optimization tools that can only optimize one single between the customers and the manufacturers to
factor do so at the cost of other factors. A real implement the integration part as shown in figure 1.
intelligent distribution algorithm, especially designed This central coordinator gathers the customer orders,
A multi-criterion genetic algorithm 341
Table 1. The relative distance from manufacturers to the
customers.

Delivery day D1 D2 . . Dj . . D10

M1 L1,1 L1,2 . . L1,j . . L1,10


M2 L2,1 L2,2 . . L2,j . . L2,10
M3 L3,1 L3,2 . . L3,j . . L3,10

Figure 1. Central coordinator. becomes another crucial factor. The objective is to


secure the equity of sharing gains obtained from
collaborations in the chain. Similarly, this factor may
and then distributes them to the manufacturers. The affect the cost factor and the service level factor
central coordinator is not a manufacturer in itself. It is a either positively or negatively.
unique and distinct community, which focuses on Since those three criteria are conflicting in practice,
Downloaded by [University of California Santa Cruz] at 03:07 22 November 2014

supply chain coordination tasks only. The objective is optimization of one of the criteria may sacrifice the
to avoid bias for optimizing one’s own benefit and others. DMs are required to take a trade-off point for
encourage exerting more effort on the supply chain those three factors. Some DMs may insist that the cost
collaboration. Meanwhile, this can strengthen the trust factor should be slightly more/less important than the
of collaboration for the organizations participating in service level factor by X times and the equity of
the system. Products will be transported from manu- distribution factor by Y times. Meanwhile, the service
facturers directly to customers, and the relative level factor should be more/less important than the
distances from manufacturers to customers are shown equity of distribution factor by Z times. Again, different
in table 1. people will have different opinions on the weightings.
The problem encountered here is to determine In addition, they may even have other new factors that
an optimal solution, which indicates the allocation need to be considered in their situations. Indeed, an
schedule of the customer orders to suitable manu- optimal solution may be defined as a solution that has
facturers. Before an optimal solution can be deter- taken the desired factors into account under a clear
mined, it is necessary to define what an optimal declaration of the weighting relationship between those
solution is. Traditionally, cost is an important factor desired factors.
to be considered. Therefore, the optimal solution For the sake of explaining the proposed algo-
should give the lowest total cost for the system. Costs rithm’s procedure, three major attributes are se-
in a supply chain involve production cost, material lected – total cost, delivery days used, and equity of
cost, labour cost, inventory holding cost, storage cost, utilization ratios. Considerations under each attribute
delivery cost, quality control cost, loss and damage are simplified but representative. Total cost repre-
cost, warranty cost, delay penalty cost, information sents the cost attribute. It consists of production cost
system cost, etc. In supply chain management nowa- and delivery cost. The attribute of the delivery days
days, there is a trend to consider customer service used represents the customer service level. It is
level as more critical (Hoek and Chong 2001). calculated by the total number of days required to
Customer service level can be measured by customer deliver all the products to customers. A small total
response time, ability to respond to market changes, number of days means the suppliers can deliver
consistence of order cycle time, accuracy of order products to the customers faster, implying good
fulfilment rate, short delivery time, flexibility in customer service level. The equity of the utilization
order quantity, flexibility in product specification, ratios attribute represents the collaboration relation-
accuracy of information system, etc (Ballou 1999). ship. It is evaluated by the comparison of the
However, the critical optimization factors in the manufacturer’s utilization. An equal/closely equal
implementation of supply chain collaboration do not percentage of utilization between each manufacturer
only minimize the total operation cost and maximize implies no bias in favouring a particular manufac-
the customer service level. It is crucial that the turer.
coordinator should not be biased towards utilizing Some assumptions for the hypothetical supply chain
some of the suppliers (manufacturers in this can be stated as:
example). This is because bias may incur the
supplier’s dissatisfaction of the collaboration and . Manufacturers can produce the same product
disjoint. Therefore, a fair distribution of demands with equal quality.
342 F. T. S. Chan and S. H. Chung

. Order quantities will not change after release 3.2. Total delivery days (TL)
from the customer.
. The delivery day used is accurate, no early arrival This indicates the number of days utilized for
or delay is allowed. delivery in the optimal solution determined. A small
. No loss and damage during delivery. value of total delivery days implies that products can be
. Inventory is available and ready for delivery as delivered in a short time, thus representing a better
order arrived. service level.
. Shortage or backorder is not allowed. P
. Production unit cost will not change with quantity TL = (delivery days from manufacturer to custo-
to produce. mer per unit production)
. Delivery cost will not change with the quantity to XX
delivery. = Dij Zij Lij : (3)
i j

3. Optimization functions
Downloaded by [University of California Santa Cruz] at 03:07 22 November 2014

3.3. Equity on utilization ratios (EU)


This hypothetical supply chain can be structured as
a multi-objective mixed integer programming (MO- This indicates the equity of the capacity
MIP) model. In this paper, the objective is to minimize utilization ratio for manufacturers in the optimal
the total cost (TC) of the system, total delivery lead solution determined. The smaller the value, the
times (TL), and equity of utilization ratios (EU). closer the capacity utilization ratio for every manu-
facturer, thus ensuring the demands are fairly
Minimize Z ¼ TC þ TL þ EU : distributed.
X
The notations and equations utilized are shown as Utilization of capacity for Mi ðUCi Þ ¼ Dij =Cmi : ð4Þ
the following. j

Average utilization of capacity ðavg: UC Þ ¼


XX X
3.1. Costs
Dij = CMi ð5Þ
i j i
Total Cost (TC)
TC = Total production cost + Total delivery cost where CMi is the maximum capacity of supply.
XX XX EU is represented by mean square error (MSE).
= Dij Cui Zij þ Dij Cdij Zij : ð1Þ
i j i j
X
MSE ¼ ½ ðUCi  avg: UC Þ2 =n1=2 ð6Þ
Cui manufacturing cost required by Mi to produce i
one unit of product.
Dij quantity of Dj allocated to Mi. where n is the number of demands.
Zij 7 {0, 1} Zij = 1 if Dj is allocated to Mi, else = 0.

Delivery unit cost (Cdij) 3.4. Constraints


This is the delivery cost for each unit of product
transported from manufacturer (i ) to customer ( j ), 3.4.1. Demand allocation constraints. Total demand
allocated to all the manufacturers 4 Total capacity of
Cdij = delivery cost ratio 6 production unit cost 6 supply from all the manufacturers
delivery day XX X
= 0.2 CujLij. (2) Dij 4 CMi : ð7Þ
i j i

The delivery cost ratio varies from different types of


transportation, such as air to ship. Researchers reported Total demand allocated to a manufacturer 4 Max-
that logistics cost consumed an average 10.5% of the imum capacity of supply for that manufacturer
total sales for an organization (Ballou 1999). Therefore, X
in this paper, the delivery cost ratio will be set equal to Dij 4CMi ; 8i: ð8Þ
20% of the Cui. j
A multi-criterion genetic algorithm 343

Total demand allocated to all the manufacturers to the other potential manufacturer. The advantages of
= total demand this quantity distribution can fully utilize the capacity of
XX X one manufacturer. A larger order quantity may often
Dij 4 Dj : ð9Þ result in the reduction of production cost, handling
i j j
cost, and delivery cost. The representation of the
splitting process is defined in the extended segment,
which has the number of genes equal to the number of
4. Proposed algorithm GA + AHP available manufacturers.
For instance, in a hypothetical network with five
4.1. Genetic algorithms manufacturers and five orders, the largest maximum
capacity of supply is M3, followed by M5, as shown in
In the proposed algorithm, optimal solutions are table 2(b). In this example, D2 exceeds the capacity of
generated by GA, in which AHP is used to calculate the M3. Then, D2 needs to be split into two orders. One of
fitness value of chromosomes. the orders consumes all the capacity of M3, and
another order will be checked for the need of further
Downloaded by [University of California Santa Cruz] at 03:07 22 November 2014

4.1.1. Representation of chromosome. Each chromosome splitting. For example, if the remaining quantity is
represents a potential optimal solution. In this paper, it found to be larger than the capacity of M5, further
indicates which order should be supplied by which splitting will progress. The maximum capacity of M5
manufacturer. A chromosome is composed of genes. will then be fully utilized. Again, the remaining
Each gene contains a value. In this distribution quantity will be checked for further splitting. If no
algorithm, the value of a gene represents the manu- splitting is required, the remaining quantity will be
facturer, and the location of the gene represents the allocated randomly to any other manufacturer, such as
order number. For instance, gene 2 indicates that order manufacturer 1.
2 will be allocated to manufacturer 3 as shown in table
2(a). 4.1.3. Fitness value. The fitness value defines the
relative strength of a chromosome. It evaluates the
4.1.2. Scheduling of chromosome. A chromosome con- chromosome structure and returns a positive value. The
sists of a basic segment, and an extended segment. In larger the value, the stronger the chromosome. A
the basic segment, the number of genes is equal to the strong chromosome means that it is more desirable.
number of total orders, as shown in table 2(a). If a The fitness value is guided by an equation. This
single order is larger than the largest capacity of supply equation may formulate the total cost of the system,
among the manufacturers, this order will be divided and is considered by most researchers. Some research-
into two individual orders. One order will be in the ers formulate the equation to optimize the total lead
quantity of the largest absolute capacity of the times to increase the service level. In supply chain
manufacturer. The remaining quantity will be allocated management, it is crucial to consider the customer

Table 2. Representation chromosome structure.

(a) A sample of chromosome structure


Chromosome
:
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
8

M1 M3 M2 M3 M1
Position/Order 1 2 3 4 5

(b) A sample of chromosome structure with extended segment


Basic segment Extended segment for Order 2
:
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
8
:
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
8

M4 4 Max (M3). M2 M2 M1 *R (M1) E (M2) F (M3) E (M4) F (M5)


Gene Position 1 2 3 4 5 6 7 8 9 10

* R – remaining quantity
E – empty gene
F – fully utilized
344 F. T. S. Chan and S. H. Chung

service level, meanwhile the costs should be considered means human knowledge and experience can be
as well. In a situation such as supply chain cooperation, directly input into the system. Users can vary the
equal distribution of benefits should never be ne- weightings according to their desires. Different from
glected. In this paper, an innovative measurement for simply giving weightings to each criterion, weightings
the fitness value calculation is proposed by introducing in AHP are in pair-wise comparison. That is, the
AHP. This new approach for calculating the fitness decision-maker can specifically compare attribute A to
value will have the representations including the total attribute B and attribute C, and then specifically
cost of the system, total delivery day utilized, and the compare the attribute of B to C. This weighting
equality on manufacturer’s utilization ratio. approach allows much more detail on the weightings
for DM. These weightings provide guidelines for the
AHP to search and determine which potential solution
4.2. Analytic Hierarchy Process is the most optimal among all the options. For
instance, users may want an equal weighting on the
AHP was developed by Saaty (1980). It is a well- attribute of total cost and delivery day, and less
proven multi-attribute decision-making methodology, consideration for the utilization attribute, as shown
Downloaded by [University of California Santa Cruz] at 03:07 22 November 2014

especially powerful for those complex problems with a in table 3.


set of highly interrelated decision factors (Teltumbe Each chromosome represents a potential solution,
2000, Poh et al. 2001, Angelis and Lee 1996, Winkler and the corresponding values of total cost, total delivery
1990). With a pool of potential options, AHP helps to lead times, and capacity utilization of the manufacturer
determine which one is the most desirable. It can also can be evaluated. This evaluation process is examined
rank them from the most desirable to the most for all chromosomes in the pool. Before the numerical
undesirable. This makes AHP work perfectly in the process of the AHP, the values found are transformed
calculation of fitness values in GAs. into relative cost, relative day, and relative utilization
The principle of AHP is to organize the complex values. These relative parameters are now defined
problems into a hierarchy or a network structure. It
integrates all the criteria into a hierarchy of priorities. For total cost:
These priorities and the structure used to represent the Relative cost =
problem influence the resulting decision. AHP needs Total cost of the chromosome
ð10Þ
numbers (quantitative) and a modicum of mathematics Highest total cost of chromosome in the pool
to formalize judgements and make tradeoffs (Saaty
1994). It also has the ability to determine which For total delivery day:
objective factor outweighs the others. The ability to Relative day =
assign weighting to different factors is another advan- Total delivery day of the chromosome
tage of AHP because it can take into account human Longest total delivery day of chromosome in the pool
knowledge and experience in the decision-making ð11Þ
process.
The example problem in this paper can be For utilization:
modelled into two hierarchy levels. The first level Relative utilization =
represents all the attributes considered. The second MSE of the chromosome
ð12Þ
level represents the potential solutions, as shown in Highest MSE of chromosome in the pool
figure 2.
One of the useful features of AHP is the weighting A smaller relative value implies that chromosome
of the attributes that can be declared by users, which behaviour is better in that attribute. These relative
values will be input for the matrix computation in the
second level. For instance, in a pool of three chromo-

Table 3. A sample of weightings in the attribute matrix.

Attribute matrix TC TL EU

TC 1 1 40
TL 1 1 40
Figure 2. The hierarchy structure of the proposed algor- EU 0.025 0.025 1
ithm.
A multi-criterion genetic algorithm 345

somes as shown in table 4(a), the relative values are 4.3. Operators
presented in table 4(b).
The weightings for attributes in the attribute 4.3.1. Selection operator. The selection operator imple-
matrix are input by users. These weightings relate to ments the idea of the ‘survival of the fittest’. It is the
the importance of the attributes. Similarly, the process of selecting chromosomes from the pool to be
importance of chromosomes in respect of each the parent chromosomes for the next crossover process.
attribute should be related. The values of the This selection process is basically arbitrary. Every
weighting are calculated according to equations (13) chromosome may have a probability to be chosen.
and (14) as stated below. However, the stronger the chromosome, the better the
Each chromosome is compared with all the other chance that it will be selected. This mimics the idea of
chromosomes within the pool in a pair-by-pair ap- only the strong surviving, while the weaker will be
proach. The relative values under the same attribute eliminated. Indeed, every chromosome can appear
will be compared. Considering two chromosomes A more than once in the mating pool as the parent
(CA) and B (CB) in the relative cost values, if the chromosome. The number of chromosomes selected is
relative value in chromosome A (RA) is smaller than equal to the pool size initially defined.
Downloaded by [University of California Santa Cruz] at 03:07 22 November 2014

that in chromosome B (RB), it means chromosome A is In this paper, an innovative process is introduced to
better (with a lower cost) than chromosome B. The speed up the elimination process. The aim is to make
implication is that chromosome A should be weighted the pool mature faster; and to allow strong chromo-
more. Therefore, the value of weighting of CA to CB is: somes having more chances to do the crossover to
become even stronger. The approach is to magnify the
Weighting of CA to CB ¼ RB  RA þ 1: ð13Þ fitness values of those chromosomes with a fitness value
above the average. On the other hand, for those
In contrast, if RA is larger than RB, chromosomes which have their fitness values below the
average, their fitness values will be set equal to a
Weighting of CA to CB ¼ 1=ðRA  RB þ 1Þ: ð14Þ constant. This approach can ensure the strong chromo-
somes will be utilized more effectively. The fitness
The constant 1 in the equation serves two functions. values (fn) are converted into the corresponding
First it makes the weighting equal to 1 when RA equals magnified fitness values (mag fn).
RB. Second, it forces the weighting to be larger than 1.
Average fitness value ðAvg fnÞ ¼
The AHP value evaluated will be the fitness value of P
that chromosome in the pool. The larger the AHP value fn ð15Þ
returned, the more desirable the option. This is similar Total number of chromosomes
to the scenario of fitness value; the larger the fitness
value obtained, the stronger the chromosome. This . for those chromosomes with fn 4 Avg fn
means AHP works perfectly with GA. An example of
AHP to calculate the fitness values is shown in the ðmag fnÞk ¼ ½ðfnÞk  Avg fn  1000;
ð16Þ
appendix. k ¼ f1; 2; . . . ; ng

Table 4. An example of relative values calculation. where n is the number of chromosomes with
fitness values higher than Avg fn.
(a) A sample of three chromosomes, with the values of cost, . for those chromosomes with fn 4 Avg fn, their
day, and utilization fitness values will be set equal to a flat constant,
which is the smallest (mag fn) obtained from
Chromosome Cost Day Utilization
equation (16).
1 6730 10000 5.744
Flat constant ¼ min fðmagÞ fnÞk g;
2 6830 10000 2.52 ð17Þ
3 10000 20000 10 k ¼ f1; 2; . . . ; ng:
(b) The relative values calculated from table 4(a)

Chromosome Rel. Cost Rel. Day Rel. Utilization 4.3.2. Crossover operator. For this hypothetical model, it
adapts the multi-points crossover method. Because of
1 0.673 0.5 0.5744 the length of the chromosome in this problem size, a
2 0.683 0.5 0.252
two-point crossover is adapted, which is 20% of the
3 1 1 1
genes undergoing crossover. This low percentage of
346 F. T. S. Chan and S. H. Chung

crossover allows the chromosome to replace its weaker to the rounded off integer of pool size divided by ten –
genes gradually, and prevents the offspring chromo- approximately 10% of the chromosomes in the pool will
somes changing too rapidly from parent chromosomes. be replaced. With this low percentage, the strongest
The crossover operation is illustrated in figure 3(a). chromosomes can prevent domination in the mating
pool. When the number of strongest chromosomes is
4.3.3. Mutation operator. In this paper, the mutation of larger than the number of calculated insertions, the
chromosomes will occur only when a pair of chromo- number of insertions will be set equal to the number of
somes with an identical genes structure is selected to strongest chromosomes. Again, this ensures the survival
crossover. Both of them will mutate instead of cross- of all the strongest chromosomes. The insertion will be
over. Similar to crossover, the number of mutated genes spread out evenly, i.e. each insertion will be separated
is two, which is 20%. The mutation operation is by a space that is equal to the number of insertions
illustrated in figure 3(b). calculated.

4.3.4. Monitor operator. The idea of Elitist strategy is to


bring the best chromosomes from the previous stage to 5. Computational results and discussions
Downloaded by [University of California Santa Cruz] at 03:07 22 November 2014

the present stage without changing the gene structure


(DeJong 1975). This ensures the best chromosomes can The hypothetical supply chain presented in section
survive. Onwubolu and Kumalo (2001) proposed an 2 is optimized with the MOMIP by applying lexicogra-
‘Isbest strategy’, which ensures the best chromosomes phical goal programming. In order to minimize the
in the present stage are pushed into the next stage deviation from the optimal solution, the objective
(Onwubolu and Muting 2001). Meanwhile, the best function is modified and subject to additional con-
chromosomes are preserved, preventing loss of best straints as shown in the following.
chromosomes. In this paper, the function of a monitor
operator serves to monitor the performance of the Minimize
evolution of the pool. After each evolution, the
strongest chromosomes in the present stage will be Z ¼ aTC  þ b TL  þ gEU 
identified and recorded. These strongest chromosomes
will be compared with the overall strongest chromo- subject to
somes evaluated so far from the beginning. If these new
chromosomes are weaker than those overall strongest Z ¼ TC  TC þ þ TC 
chromosomes, the overall strongest chromosomes will l ¼ TL  TL þ þ TL 
be inserted back into the mating pool for the crossover m ¼ EU  EU þ þ EU 
or mutation processes. This allows the strongest
chromosomes to survive and have another chance to where
become stronger. The number of insertions is set equal
a, b, g = importance weighting,
Z = optimal value of TC,
l = optimal value of TL,
m = optimal value of EU.

Ten sets of experiments are run with different


parameters generated randomly. The optimization
result is shown in table 5, with a, b and g = 1. In table
5(a), the objective function of TC is set as a hard
constraint, which means TC7= 0. TL and EU are soft
constraints, while TL is set as a hard constraint in table
5(b), and EU is a hard constraint in table 5(c).
The ten experiments have also been optimized by
GA + AHP. The number of evolutions adapted is 200
times with a pool size of 20. The importance weighting
of TC is heavy in table 5(a), TL weights heavy in table
5(b), and EU weights heavy in table 5(c). The
Figure 3. Illustration of crossover and mutation operators. comparison of the optimization result indicates that
(a) A sample of two-point crossover. (b) A sample of mutation. GA + AHP is reliable and robust.
Table 5. Optimization results obtained by MOMIP.
Downloaded by [University of California Santa Cruz] at 03:07 22 November 2014

(a) TC as hard constraints.

TC Experiment 1 2 3 4 5 6 7 8 9 10

TC LP 2460 3590 3090 3090 3720 3750 4310 5236 3600 3632
GA + AHP 2460 3590 3090 3090 3720 3750 4310 5236 3600 3632
TL LP 15100 9500 6700 6700 13400 12400 10900 9800 13000 11200
GA + AHP 15100 9500 6700 6700 13400 12400 10900 9800 13000 11200
EQ LP 45.20 4.79 25.15 16.60 33.59 41.04 2.78 34.72 29.86 37.60

A multi-criterion genetic algorithm


GA + AHP 45.20 4.79 25.15 16.60 33.59 41.01 2.78 34.72 29.86 37.60

(b) TL as hard constraints.

TL Experiment 1 2 3 4 5 6 7 8 9 10

TC LP 4260 3750 3090 3090 4565 4358 4362 5236 5984 4536
GA + AHP 4260 3750 3090 3090 4565 4358 4362 5236 5984 4536
TL LP 6700 6700 6700 6700 6300 7400 7400 9200 9200 9100
GA + AHP 6700 6700 6700 6700 6300 7400 7400 9200 9200 9100
EQ LP 18.25 32.50 25.15 16.60 39.75 33.52 35.25 44.78 29.86 33.73
GA + AHP 18.25 32.50 25.15 16.60 39.75 33.52 35.25 44.78 29.86 33.73

(c) EU as hard constraints.

EQ Experiment 1 2 3 4 5 6 7 8 9 10

TC LP 3759 3660 3280 3220 4095 4115 4366 5300 4608 4096
GA + AHP 3759 3660 3280 3220 4095 4115 4366 5300 4608 4096
TL LP 8250 9500 9500 8500 9800 9200 10500 11000 11200 9900
GA + AHP 8250 9500 9500 8500 9800 9200 10500 11000 11200 9900
EQ LP 2.38 2.50 1.62 1.19 1.19 3.93 2.78 2.36 1.57 1.98
GA + AHP 2.38 2.50 1.62 1.19 1.19 3.93 2.78 2.36 1.57 1.98

347
348 F. T. S. Chan and S. H. Chung
Downloaded by [University of California Santa Cruz] at 03:07 22 November 2014

Figure 4. Pareto solution obtained by MOMIP for experi-


ment 1. (a) Pareto solution of TC (MOMIP). (b) Pareto
solution of TL (MOMIP). (c) Pareto solution of EU Figure 5. Pareto solution obtained by GA + AHP for experi-
(MOMIP). ment 1. (a) Pareto solution of TC (GA + AHP). (b) Pareto
solution of TL (GA + AHP). (c) Pareto solution of EU
(GA + AHP).

A set of Pareto solutions is determined for Experi- 6. Conclusions


ment 1, and plotted in figure 4. The characteristic of this
model is that the objective function of TC conflicts with In supply chain management, many decisions usually
that of TL. The value of a is increasing from 1 to 50, while involve more than one criterion. Each criterion may be
b and g remain equal to 1. TC, TL, and EU are set as soft limited by some constraints, and influence each other.
constraints. In figure 4(a), TC reduces along with the For instance, the cost and quality of a product may
increased importance. Meanwhile, TL becomes longer as conflict in nature. The supply chain collaboration
shown in figure 4(b), and EU fluctuates as shown in figure becomes more prevalent. The goals of collaboration
4(c). For the same model, a set of Pareto solutions is are to reduce the total cost of the supply chain and
determined by GA + AHP, and is plotted in figure 5. increase the customer service level. Meanwhile, it is
Similarly, the importance weightings of TL and EU are set crucial to maintain the collaboration relationship. An
to be equal, while the weighting of TC increases. These intelligent distribution methodology should be capable
graphs indicate that GA + AHP can also search the same of determining an optimal solution, which can take into
Pareto solution as the one obtained by LP. account multi-criterion decision-making process. In
A multi-criterion genetic algorithm 349

different situations, decision-makers may request differ- ONWUBOLU, G. C., and KUMALO, T., 2001, Optimization of
ent weightings on the importance of those criteria in multipass turning operations with genetic algorithms. Inter-
national Journal of Production Research, 39(16), 3727–3745.
their problems. For these purposes, this paper developed ONWUBOLU, G. C., and MUTING, M., 2001, Optimizing the
a multi-criterion genetic algorithm, which combines the multiple constrained resources product mix problem using
optimization methodology of genetic algorithms (GAs) genetic algorithms. International Journal of Production
with the decision-making technique of the analytic Research, 39(9), 1897–1910.
hierarchy process (AHP). AHP is used to evaluate the PFOHL, H. C., and BUSE, H. P., 2001, Inter-organizational logistics
systems in flexible production networks. An organizational
fitness values for chromosomes. The solution procedure capabilities perspective. International Journal of Physical
of the proposed algorithm is discussed with an example Distribution & Logistics Management, 30(5), 388–408.
problem of the order distribution problem. Through this POH, K. L., ANG, B. W., and BAI, F., 2001, A comparative analysis
example, this paper illustrated the evolution processes of of R&D project evaluation methods. R&D Management,
optimal solution, i.e. selection, crossover, mutation, and 31(1), 63–75.
ROBINS, G., 1994, Sailing into ECR’s uncharted waters. Stores,
monitoring. The analytical results indicate the proposed 76(10), 43–44.
algorithm is robust and reliable. SAATY, T. L., 1980, The Analytic Hierarchy Process (New York:
McGraw-Hill).
Downloaded by [University of California Santa Cruz] at 03:07 22 November 2014

SAATY, T. L., 1994, Fundamentals of Decision Making and Priority


References Theory with the Analytic Hierarchy Process (The Analytic
Hierarchy Process Series, 6).
ANGELIS, D. I., and LEE C. Y., 1996, Strategic investment analysis STANK, T. P., DAUGHERTY, P. J., and AUTRY, C. W., 1999,
using activity based costing concepts and analytical hier- Collaborative planning: supporting automatic replenish-
archy process techniques. International Journal of Production ment programs. Supply Chain Management: An International
Research, 34(3), 1331–1345. Journal, 4(2), 75–85.
BALLOU, R. H., 1999, Business Logistics Management, 4th edn TELTUMBE, A., 2000, A framework for evaluating ERP projects.
(Prentice Hall). International Journal of Production Research, 38(17), 4507–4520.
BARRATT, M., and OLIVEIRA, A., 2001, Exploring the experiences UPS LOGISTICS GROUP, 2001, United Parcel Service of America,
of collaborative planning initiatives. International Journal of Inc, http://www.upslogistics.com
Physical Distribution & Logistics Management, 31(4), 266–289. WINKLER L. R., 1990, Decision modeling and rational choice:
BLAIR, N., 1998, Minding the store: with inventory reduction AHP and utility theory. Management Science, 36(3), 247–275.
measures under way in the warehouse, executives are
eyeing similar strategies based on store-level data. Super-
market News, 48(3), 81–82.
CHAKRABORTY, M., and GUPTA, S., 2002, Fuzzy mathematical Appendix. Demonstration of AHP calculation for the
programming for multi objective linear fractional program- fitness values
ming problem. Fuzzy Sets and Systems, 125(3), 335–342.
COOKE, J. A., 1998, Into the great wide open. Logistics
Management and Distribution Report, 37(10), 84–87. For example, there are three chromosomes in the
DEJONG, K. A., 1975, An analysis of the behavior of a class of solution pool. The total cost, delivery days used, and the
genetic adaptive system. Doctoral dissertation, University of equity of utilization values are evaluated for those three
Michigan. chromosomes, as shown in table A1(a).
DONSELAAR, V. K., KOPCZAK, R. L., and WOUTERS, M., 2001, The
use of advance demand information in a project-based Table A1(a). A sample of three chromosomes.
supply chain. European Journal of Operational Research,
130(3), 519–538.
Chromosome Cost Day Utilization
HOEK, R. I. V., and CHONG, I., 2001, Epilogue: UPS logistics –
practical approaches to the e-supply chain. International
1 6730 10000 5.744
Journal of Physical Distribution & Logistics Management, 31(6),
2 6830 10000 2.52
463–468.
3 10000 20000 10
HORVATH, L, 2001, Collaboration: the key to value creation in
supply chain management. Supply Chain Management: An
International Journal, 6(5), 205–207.
LAKHAL, S., MARTEL, A., KETTANI, O., and ORAL, M., 2001, On the The relative values will be calculated as shown in
optimization of supply chain networking decisions. European table A1(b), where for instance (0.673 = 6370/10000).
Journal of Operational Research, 129(2), 259–270.
MALONI, J. M., and BENTON, W. C., 1997, Supply chain
partnerships: opportunities for operations research. Eur- Table A1(b). The relative values obtained from table A1(a).
opean Journal of Operational Research, 101(3) 419–429.
NEUMAN, J., and SAMUELS, C., 1996, Insights from industry supply Chromosome Rel. Cost Rel. Day Rel. Utilization
chain integration: vision or reality? Supply Chain Manage-
ment: An International Journal, 1(2), 7–10. 1 0.673 0.5 0.5744
NEW, S., and RAMSAY, J., 1997, A critical appraisal of aspects of 2 0.683 0.5 0.252
the lean chain approach. European Journal of Purchasing & 3 1 1 1
Supply Management, 3(2), 93–102.
350 F. T. S. Chan and S. H. Chung

For instance, the decision-maker decides to use the behaves better than chromosome 1. According to
attributes weighting as shown in table A2. equation (16),
The chromosome matrixes under each attribute
are shown in tables A3(a)–(c), for attribute cost, Cell(1,3) = (0.68 7 0.67) 6 10 + 1
delivery days used, and equity on utilization respec- = 1.1.
tively.
On the other hand, according to equation (17),

Calculation of weighting under the attribute matrix Cell(3,1) = 1/[(0.68 7 0.67) 6 10 + 1]


= 0.909091.
For instance, under the cost attribute (table A3(a)),
the cell (1,3) with value 1.1 is calculated from the Finally, the AHP values can be calculated as shown
equation (16). It is the comparison of the chromosome in table A4. These values will be the fitness values of
1 and chromosome 2. Since chromosome 1 (0.68) is the chromosomes in the pool. According to the
larger than chromosome 2 (0.67), chromosome 2 results, it indicates that chromosome 1 (0.4168) is
Downloaded by [University of California Santa Cruz] at 03:07 22 November 2014

Table A2. The attributes weighting.

(C) (D) (U) Normalized matrix RowS Average

Cost (C) 1 1 40 0.493827 0.493827 0.493827 1.481481 0.493827


Day (D) 1 1 40 0.493827 0.493827 0.493827 1.481481 0.493827
Uti (U) 0.025 0.025 1 0.012346 0.012346 0.012346 0.037037 0.012346
S 2.025 2.025 81

Table A3(a). The weighting under the cost attribute matrix.

C 1 2 3 Normalized matrix Row S Average

1 1 1.1 4.27 0.466574 0.470124 0.452331 1.389028 0.463009


2 0.909091 1 4.17 0.424158 0.427385 0.441737 1.293281 0.431094
3 0.234192 0.239808 1 0.109268 0.102491 0.105932 0.317691 0.105897
S 2.143283 2.339808 9.44

Table A3(b). The weighting under the delivery days used attribute matrix.

D 1 2 3 Normalized matrix Row S Average

1 1 1 1.5 0.375 0.375 0.375 1.125 0.375


2 1 1 1.5 0.375 0.375 0.375 1.125 0.375
3 0.666667 0.666667 1 0.25 0.25 0.25 0.75 0.25
S 2.666667 2.666667 4

Table A3(c). The weighting under the equity on utilization attribute matrix.

U 1 2 3 Normalized matrix Row S Average

1 1 0.236742 5.256 0.184697 0.174761 0.356678 0.716136 0.238712


2 4.224 1 8.48 0.780162 0.738189 0.575461 2.093812 0.697937
3 0.190259 0.117925 1 0.03514 0.087051 0.067861 0.190052 0.063351
S 5.414259 1.354667 14.736
A multi-criterion genetic algorithm 351
Table A4. The AHP values calculated.

Objectives C D U AHP value

Weights 0.493827 0.493827 0.012346

Chromosome
1 0.463009 0.375 0.238712 0.4168
2 0.431094 0.375 0.697937 0.4067
3 0.105897 0.25 0.063351 0.1765

the strongest. The next strongest is chromosome 2 utilization attribute. Therefore, although the equity on
(0.4067), and the weakest one is chromosome 3 utilization attribute of chromosome 2 is better than
(0.1765). In this example, both the cost attribute chromosome 1 and only a little bit weaker than
and delivery days used attributes are equally important chromosome 1 in the cost attribute, chromosome 1 is
Downloaded by [University of California Santa Cruz] at 03:07 22 November 2014

and are weighted much heavier than the equity on still stronger than chromosome 2.

You might also like