This article was downloaded by: [186.148.4.

20] On: 07 December 2018, At: 14:08

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

Interfaces
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

Metro Uses a Simulation-Optimization Approach to

Improve Fare-Collection Shift Scheduling
Jaime Miranda, Pablo A. Rey, Antoine Sauré, Richard Weber

To cite this article:

Jaime Miranda, Pablo A. Rey, Antoine Sauré, Richard Weber (2018) Metro Uses a Simulation-Optimization Approach to Improve
Fare-Collection Shift Scheduling. Interfaces

Published online in Articles in Advance 06 Dec 2018

. https://doi.org/10.1287/inte.2018.0959

Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.

Copyright © 2018, INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 INTERFACES
Articles in Advance, pp. 1–14
http://pubsonline.informs.org/journal/inte/ ISSN 0092-2102 (print), ISSN 1526-551X (online)

Metro Uses a Simulation-Optimization Approach to Improve

Fare-Collection Shift Scheduling
Jaime Miranda,a Pablo A. Rey,b Antoine Sauré,c Richard Weberd
a
Department of Management Control and Information Systems, School of Economics and Business, Universidad de Chile, Santiago 8330015,
Chile; b Department of Industry, Faculty of Engineering, Universidad Tecnológica Metropolitana, Santiago 7800002, Chile; c Telfer School of
Management, University of Ottawa, Ottawa, Ontario K1N 6N5, Canada; d Department of Industrial Engineering, Faculty of Physical and
Mathematical Sciences, Universidad de Chile, Santiago 8370456, Chile
Contact: jmirandap@fen.uchile.cl, https://orcid.org/0000-0001-9682-9494 (JM); prey@utem.cl,
https://orcid.org/0000-0002-1337-4701 (PAR); asaure@uottawa.ca, https://orcid.org/0000-0002-0463-7602 (AS); rweber@dii.uchile.cl,
https://orcid.org/0000-0002-6351-1065 (RW)

Received: October 3, 2016
Revised: May 19, 2017; August 25, 2017;
March 24, 2018
Accepted: April 6, 2018
Published Online in Articles in Advance:
December 6, 2018
https://doi.org/10.1287/inte.2018.0959
Copyright: © 2018 INFORMS
Abstract. In addition to scheduling trains, drivers, and security personnel, many subway
companies worldwide are faced with the challenge of determining the staffing levels and
shift schedules required to operate their fare-collection systems. Subway companies
typically deal with a highly variable demand for fares, several operational requirements,
and increasing service expectations from users. In this paper, we describe a simulation-
optimization method for fare-collection shift scheduling in a subway network. We use an
integer programming model to generate shift schedules and a discrete-event simulation
model to evaluate the corresponding service levels in terms of the average number of users
waiting in line over time. If the expected service level at a particular station booth does not
meet a predefined service standard, then our method automatically generates an additional
constraint on the number of fare collectors at that station booth. The proposed approach
iteratively adds this new constraint to the integer programming model until either the current
schedule meets the service standard or the problem becomes infeasible. We use the proposed
solution approach to solve the fare-collection shift-scheduling problem faced by one of the
largest subway companies in Latin America, Metro de Santiago in Chile. The resulting shift
schedules show a significant improvement compared with Metro’s previous scheduling
practice in terms of service levels, operational costs, capacity utilization, and solution times.

History: This paper was refereed.

Funding: This research was partially funded by the Complex Engineering Systems Institute, ISCI [ICM-
FIC: P05-004-F; CONICYT: FB0816].

Keywords: subway system • fare collection • staffing • shift scheduling • integer programming • simulation

Similar to many other transportation companies, subway
companies face complex personnel scheduling decisions.
In addition to selecting duties for vehicles, crews,
and/or drivers, some companies must also determine
the staffing levels and personnel shift schedules required
to operate their fare-collection systems. To collect pas-
senger fares and provide users with tickets and prepaid
cards, they must allocate fare collectors and/or vending
machines at every station in their network.
Determining the appropriate number of fare collec-
tors to assign to each station booth in a network is not
a simple task. Subway companies face highly variable
demand for tickets and prepaid cards, several opera-
tional requirements, alternative fare-collection tech-
nologies, and increasing service expectations from
users. Subway systems typically connect suburbs to
city centers, drawing a large number of commuters
every day. Because most commuters travel at the same
time of day, morning demand peaks occur at suburban
stations, whereas evening demand peaks occur at city
1
center stations, with considerably lower demand levels
at intermediate stations.
Ideally, the number of fare collectors assigned to
a particular station booth should guarantee that the
system satisfies a predefined service standard in a cost-
effective manner. The definition of service standards is
generally with respect to either the average number of
users waiting in line or the average wait time for
service. Although subway companies may be able to
identify the most relevant factors that affect the de-
mand for tickets and prepaid cards, to the best of our
knowledge the majority of subway companies still
need a systematic way of generating and evaluating
different shift schedules—in terms of expected service
levels—to make better staff scheduling decisions. In the
absence of a systematic method for scheduling, subway
companies may end up understaffing some station
booths during peak hours and overstaffing some sta-
tion booths during off-peak hours. This can negatively
affect service levels and operational costs.

2 Interfaces, Articles in Advance, pp. 1–14, © 2018 INFORMS
In this paper, we describe a simulation-optimization
approach for fare-collection shift scheduling in a sub-
way network. The proposed approach determines daily
staffing levels and shift schedules for a week that satisfy
(when possible) a predefined service level and non-
negotiable operational conditions in a cost-effective
manner. First, for each station booth and workday, we
use an integer programming model to identify the
optimal number of fare collectors to schedule and their
respective shifts using aggregate demand estimates
and average transaction times. Second, a discrete-event
simulation model evaluates the performance of the
resulting shift schedules using disaggregated demand
estimates and stochastic service times. The simulation
model also considers users’ preference for type of fare
collector through a discrete probability distribution
that depends on the length of the wait lines and the
types of fare collectors available. If the performance of
the current schedule satisfies a predefined service
standard, then we have found a solution. Otherwise,
we generate a new constraint on the number of fare
collectors available during a specific period and in-
corporate it into the integer programming model. We
iteratively repeat this process until we find an ac-
ceptable solution or the model becomes infeasible. We
evaluate the benefits associated with the use of the
proposed approach using a real-world scenario based
on the operations of the largest subway company in
South America, Metro de Santiago in Chile.
We organize the remainder of this paper as fol-
lows. We begin by providing a detailed description of
the fare-collection shift-scheduling problem faced by
Metro de Santiago. Then we summarize the literature
relevant to our work. Subsequently, we describe the
proposed simulation-optimization approach and
the experimental results from its utilization. Finally,
we state our main conclusions and suggest possible
extensions.
Fare-Collection Staff Scheduling at Metro
Metro de Santiago (hereafter referred to as Metro) is
one of the largest public transportation companies in
Latin America. It operates the subway system that
serves the city of Santiago, Chile, providing nearly
661.2 million passenger rides in 2015 (Metro de
Santiago 2015). As of July 2016, Metro’s network was
the most extensive in South America, with five lines,
108 stations, and 103 kilometers of track (Figure 1).
Metro classifies fare-collection employees into four
types according to their position within stations and
fare-collection equipment. Depending on their posi-
tion, fare collectors are either static or mobile. Static fare
collectors work inside station booths, whereas mobile
fare collectors move freely outside. In terms of fare-
collection equipment, fare collectors are classified into
two classes: those with sales terminals and those without
Miranda et al.: Improving Fare-Collection Shift Scheduling at Metro
sales terminals. Fare collectors with sales terminals sell
tickets and reload prepaid cards, whereas fare collec-
tors without such terminals can only sell tickets. Tickets
and prepaid cards are the only two ways to obtain
access to Metro’s subway network. For operational
reasons, Metro only uses mobile fare collectors once it
has assigned all windows at a station booth to static
fare collectors.
In selecting a fare collector, users typically choose the
one with the shortest wait line. Furthermore, users
willing to buy a ticket typically select a fare collector
without a sales terminal. The actual probability func-
tion that we use to model users’ fare-collector choice
behavior depends on the particular station booth under
consideration. The general idea is as follows. If upon
arrival users find one or more suitable fare collectors
idle, then they choose with equal probability among
them. Otherwise, the probability of choosing a specific
fare collector is inversely proportional to the length of
the corresponding wait line.
Metro makes fare-collection shift-scheduling de-
cisions once per week, one week in advance. After
determining the shift schedules for all station booths in
its network, Metro communicates them to private
personnel companies, one per subway line. These
companies provide Metro with the required personnel
to operate its fare-collection system. Shift-scheduling
decisions for different station booths are independent
of each other; therefore, Metro deals with the fare-
collection shift-scheduling problem for each station
booth separately. The cost Metro pays to the private
personnel companies for a fare collector ranges from
14 to 91 U.S. dollars (USD), depending on the shift
length and the terms present in the respective con-
tracts. Shifts can be 2, 4, or 16 hours (two fare collectors
working 8 hours each). To provide this service, the
private personnel companies must determine the in-
dividual staff schedules for their employees. These
decisions are beyond the scope of the problem un-
der study.
Metro divides a typical working day, generally from
6:00 a.m. to 11:00 p.m., into 68 time blocks of 15 minutes
each. A team of well-trained professionals and data
analysts, all Metro employees, generates demand es-
timates through a process that is beyond the scope of
this paper. The number of users requesting a specific
type of transaction varies by hour of the day, day of the
week, week of the month, and month. Furthermore,
user arrival rates differ greatly across stations, with
very different patterns depending on the economic
activity around them. Figures 2–4 show the average
number of users accessing Metro’s subway platforms,
on a quarter-hour basis, for three stations. The average
number of users accessing subway platforms through
access gates is one of the main inputs of Metro’s de-
mand forecasting models. Figure 2 presents a typical
Miranda et al.: Improving Fare-Collection Shift Scheduling at Metro
Interfaces, Articles in Advance, pp. 1–14, © 2018 INFORMS
Figure 1. Diagram Showing Metro’s Subway Network as of July 2016
demand pattern for a terminal station that receives
users from feeder services. This pattern shows a clear
demand peak during the morning rush hour (people
traveling from home to work or school) and lower
demand values during the remainder of the day.
Figure 3 shows the opposite demand behavior at
a typical downtown station. The demand peak in this
case is during the evening rush hour and is associated
3
with a large number of people returning home. Figure 4
shows the typical demand behavior at an intermediate
subway station.
Literature Review
Researchers have extensively studied staff scheduling
problems as companies have increasingly sought to im-
prove service levels while reducing operational costs.
 Miranda et al.: Improving Fare-Collection Shift Scheduling at Metro
4 Interfaces, Articles in Advance, pp. 1–14, © 2018 INFORMS

Figure 2. Average Number of Users Accessing Metro’s Subway Network at the Las Rejas Terminal Station Between 6 a.m.
and 11 p.m.

Bechtold et al. (1991), Alfares (2004), Ernst et al. (2004a)
and (2004b), Brucker et al. (2011), and Van den Bergh
et al. (2013) provide comprehensive reviews on per-
sonnel scheduling by industry and application area.
The list of industries includes airlines, buses, railways,
mass transit (including underground rail), and others.
Unlike other personnel scheduling problems in
transportation, staff scheduling in fare-collection sys-
tems does usually not require moving people or other
resources. In addition, the exact number and type of
tasks that need to be performed is not known a priori.
These two characteristics make this problem similar to
that of determining staffing levels and shift schedules
for call centers. In both cases, capacity requirements
change over time; thus, shift schedules need to vary for
the system to efficiently achieve a predefined service
standard. However, the main difference is that in call
centers a system routes clients to qualified agents,
whereas in fare-collection systems the users choose
their own paths. This feature increases the complexity
of the system to be evaluated because we also have to
model each user’s fare-collector selection behavior.
Additionally, clients’ experiences waiting on hold in
call centers differ from those users encounter on physical
wait lines at subway stations. This is because call center
clients do not see others waiting for service, and un-
certain waits usually feel longer than known ones. In
general, call center clients are unaware of their progress
unless the call center provides such information. We
refer the reader to Gans et al. (2003) and Mandelbaum
(2004) for comprehensive reviews of the research on
call centers.
The first studies on staff scheduling decisions in the
literature date back to the 1950s. Edie (1954) proposed
the use of probability theory and statistics to calculate
the number of toll booths and thus the number of ca-
shiers required at toll plazas throughout a day. Dantzig
(1954) outlined a mathematical programming approach

Figure 3. Average Number of Users Accessing Metro’s Subway Network at the Plaza de Armas Downtown Station Between
6 a.m. and 11 p.m.
Miranda et al.: Improving Fare-Collection Shift Scheduling at Metro
Interfaces, Articles in Advance, pp. 1–14, © 2018 INFORMS
Figure 4. Average Number of Users Accessing Metro’s Subway Network at the San Joaquı́n Intermediate Station Between
6 a.m and 11 p.m.
to determine how many toll collectors are required to
keep the scheduled number of booths open. Following
these early efforts, staff scheduling problems have re-
ceived considerable attention both in business and in
the academic world. This has led to the development
of numerous applications in various industries, most
notably in the service sector, in which companies must
manage a large number of employees and the demand
for services varies significantly over the course of a day.
A few examples of areas for which staff scheduling
problems have been solved are these: call centers
(Church 1973, Lin 1999, Alfares 2007, Pot et al. 2008,
Hojati 2010, Dietz 2011), mail processing and distribution
centers (Bard 2004, Qi and Bard 2006, Bard et al. 2007),
restaurants (Hueter and Swart 1998, Choi et al. 2009),
manufacturing (Berman et al. 1997, Azmat and Widmer
2004, Lilly et al. 2007, Bard and Wan 2008, Sabar et al.
2008), hospitals (Agnihothri and Taylor 1991, Bard and
Purnomo 2005, Green 2005, Belien and Demeulemeester
2006, Bailyn et al. 2007, Brunner and Edenharter 2011),
and airports (Mason et al. 1998, Chu 2007, Thiel 2008,
Rong and Grunow 2009).
Although researchers have extensively studied per-
sonnel scheduling problems over the past two decades,
the community has paid little attention to staff sched-
uling problems faced by subway companies. Lezaun
et al. (2010) describe an applied study for Metro Bilbao,
in which the authors evaluate annual shift patterns for
station personnel to establish a more equitable work
allocation. Sodhi and Norris (2004) present a general
modeling approach to crew rostering and its application
to computer-assisted generation of rotation-based ros-
ters at the London Underground. Cavique et al. (1999)
use data from the Lisbon Underground train timetable
to test two approaches that are part of a decision support
system for developing crew schedules. They consider
contractual and operational constraints consisting pri-
marily of limits on work time and meal breaks. Other
5
transport systems have also used crew scheduling and
rostering methodologies. For example, Chu and Chan
(1998) describe the development of a crew-scheduling
decision support tool for Hong Kong Light Rail Transit.
More recently, Elizondo et al. (2010) proposed an evo-
lutionary and constructive approach for crew schedul-
ing at Metro de Santiago. Han and Li (2014) address
the crew-scheduling problem of the Taipei mass rapid
transit system. Chew et al. (2001) report on the de-
velopment of an optimization-based approach for
a computerized train-operator scheduling system that
was implemented at Singapore Mass Rapid Transit. The
system automates the train-operator scheduling process
and produces favorable schedules compared with
a manual process. It also handles the multiple objectives
inherent in crew-scheduling systems.
Regarding subway operation and maintenance ac-
tivities, Ding et al. (2013) propose a modeling frame-
work for analyzing train conditions recorded on both
train sensor network data and maintenance plans. The
authors consider the case of the Guangzhou Metro, the
metro system in the city of Guangzhou in the Guang-
dong Province of China, to illustrate the applicability of
their methodology. Chun and Suen (2014) describe
a system to plan, schedule, and optimize nightly engi-
neering tasks for both the commuter and rapid transit
lines in Hong Kong, where the MRT Corporation per-
forms maintenance, inspection, repair, and installation
activities along the rail lines during the four to five
nontraffic hours at night. The system optimizes the
allocation of resources to maximize the number of en-
gineering tasks performed simultaneously while en-
suring that all safety, environmental, and operational
requirements are met. Basra et al. (2007) describe how
the London Underground could use a multiple-agent
system to resolve scheduling issues. They describe the
development of a prototype system with the aim of
resolving scheduling issues in real time. Varakantham

6 Interfaces, Articles in Advance, pp. 1–14, © 2018 INFORMS
et al. (2013) and Lau et al. (2016) address the problem
of scheduling security teams to patrol mass rapid transit
systems. The main objective is to deploy security teams
to stations at varying periods subject to rostering and
security-related constraints. Both studies analyze the
case of the Singapore Mass Rapid Transit rail network.
Staff scheduling covers a wide variety of processes.
According to the classification proposed by Ernst et al.
(2004a), the fare-collection shift-scheduling problem
studied in this paper involves demand modeling (de-
termining how many staff members are needed at
different times over the scheduling horizon) and shift
scheduling (choosing which shifts are to be worked to-
gether with an assignment of the number of employees to
each shift to meet demand). The papers of Atlason et al.
(2008), Cezik and L’Ecuyer (2008), Avramidis et al. (2009),
Avramidis et al. (2010), and Helber and Henken (2010)
are closely related to our staffing and shift-scheduling
approach. Atlason et al. (2008) describe a simulation-
based methodology for optimizing call center staffing
levels in the single-skill case (single-call type and single-
agent type). The authors base their approach on adding
linear inequalities to a deterministic optimization prob-
lem intended to minimize the staffing costs and ensure
the satisfaction of predefined service levels per period.
Cezik and L’Ecuyer (2008) adapt the methodology of
Atlason et al. (2008) to the multiple-skill case (multiple-
call types and multiple-agent types). Avramidis et al.
(2009) solve the same multiple-skill problem by using
local search methods combined with an analytical
approximation of the service levels. More recently,
Avramidis et al. (2010) proposed a simulation-based
algorithm for solving the multiple-skill case that extends
the method of Cezik and L’Ecuyer (2008) by considering
constraints on the expected service level per call type
and period. Helber and Henken (2010) consider a system
with random arrival rates, skill-based routing, impatient
customers, and retrials. They develop a specific approach
in which a discrete-time model captures the dynamics of
the system and simultaneously optimizes staffing levels
and shift schedules over a set of different approximate
realizations of the underlying stochastic processes.
It is important to note that unlike many existing two-
step approaches in the call center literature (e.g., Gans
et al. 2003), the algorithms mentioned above, as well as
our solution approach, simultaneously optimize staffing
levels and shift schedules.
Simulation-optimization methods have become very
popular in many fields owing to the improved perfor-
mance and lower costs offered by new developments in
computers and simulation software. Researchers have
used the simulation-optimization approach, for exam-
ple, to solve personnel scheduling problems in customs
(Mason et al. 1998) and mail-processing and distri-
bution centers (Qi and Bard 2006). Others have applied
it to network-design problems in logistic services
Miranda et al.: Improving Fare-Collection Shift Scheduling at Metro
(Cheung et al. 2005), long-term care capacity planning
(Zhang et al. 2012), and production planning problems in
manufacturing (Hung and Leachman 1996, Byrne and
Bakir 1999, Kim and Kim 2001). Typically, authors use
optimization to identify cost-effective capacity con-
figurations and simulation to determine whether
a given capacity configuration meets predefined ser-
vice requirements. Optimization and simulation models
iterate until they find a feasible capacity configuration or
the optimization model becomes infeasible. We refer the
reader to Figueira and Almada-Lobo (2014) for a com-
plete overview of the full spectrum of current approaches
combining simulation and optimization.
Although researchers have used direct search methods
based on the interaction between an optimization and
a simulation model in other contexts (Amaran et al. 2016),
to the best of our knowledge this paper describes the
first application of this approach to supporting fare-
collection shift-scheduling decisions for a transportation
company.
Solution Approach
Our approach alternates between optimization and
simulation. For each station booth and day inde-
pendently, and using aggregate demand and average
service times, an integer programming model first
determines the optimal number of fare collectors of
each type and the corresponding shift schedules. Then,
a discrete-event simulation model with detailed in-
formation allows us to study the evolution of the wait
lines over time and thus evaluate the expected service
level provided by the proposed shift schedules. If the
expected average number of users waiting in line in
each of the 60-minute periods starting at the beginning
of each time block are all less than or equal to
a predefined service level, then we have found the final
staffing levels and shift schedules. Otherwise, we
generate an additional constraint on the number of fare
collectors on duty during a specific time block and
incorporate it into the integer programming model.
We iteratively repeat this process until the resulting
schedule satisfies the predefined service standard for
all time blocks or the model becomes infeasible. The
latter occurs when we reach the maximum number of
fare collectors allowed at a station booth and the cor-
responding schedule still does not meet the required
service conditions. In that case, we schedule the maxi-
mum number of fare collectors in those time blocks for
which there is insufficient service capacity. Figure 5 il-
lustrates the proposed solution approach for a given
station booth and day.
Optimization Model
The role of the integer programming model is two-
fold. First, it identifies initial staffing levels and shift
schedules using aggregate-demand estimates and
Miranda et al.: Improving Fare-Collection Shift Scheduling at Metro
Interfaces, Articles in Advance, pp. 1–14, © 2018 INFORMS
Figure 5. The Proposed Fare-Collection Shift-Scheduling
Approach Iteratively Uses an Optimization Model and
a Simulation Model Until It Finds a Feasible Solution or
Detects Infeasibility
Note. We solve Metro’s fare-collection shift-scheduling problem for
each day and station booth separately.
average transaction times. Second, it improves the
initial and subsequent fare-collection shift-scheduling
decisions using feedback from the discrete-event sim-
ulation model. The optimization model determines the
number of fare collectors of each type to schedule at
each station booth and the type of shifts to satisfy
service and operational requirements while minimiz-
ing the total staffing cost.
Among the operational requirements are upper and
lower limits on the staffing levels at each station booth
and constraints ensuring the allocation of mobile fare
collectors only after the assignment of all station booth
windows to static fare collectors. Further constraints
ensure that the aggregated service capacity provided
by the fare collectors on duty in each time block is
sufficient to cover the expected demand for service in
Figure 6. The Process Being Simulated Has Four Stages: User Arrival, Fare-Collector Selection, Service, and Exit
7
the corresponding time block. We also consider con-
straints that prevent the resulting schedules from
dictating abrupt changes in the total number of fare
collectors in consecutive periods. Finally, the optimi-
zation model includes lower bounds on the number of
fare collectors in each time block, which are updated
iteratively according to the feedback provided by the
discrete-event simulation model. We refer the reader to
the Appendix for a detailed description of the opti-
mization model.
Simulation Model
The role of the discrete-event simulation model is
threefold. First, it accounts for the uncertainty in demand
and service parameters to better represent the dynamics
of the fare-collection process. Second, it characterizes the
evolution of the fare-collection system, particularly wait
lines, over time and thus evaluates the staff schedules
generated by the integer programming model. Finally, it
determines the time blocks in which the current schedule
violates the predefined service standard.
Figure 6 portrays the process being simulated. Users
arrive at subway stations by using an escalator or
a staircase (stage 1). Those who have purchased fares in
advance directly access the subway network through an
access gate. Users who need to buy tickets or reload
prepaid cards approach a station booth, select a suitable
fare collector, and wait in the corresponding line until
those in front of them receive service (stage 2). Users
wishing to reload a prepaid card must select a fare
collector with a sales terminal. We consider that when
multiple suitable fare collectors are available, users will
choose the fare collector with the shortest wait line. After
receiving service (stage 3), users exit the system (stage 4).
In each iteration, the simulation model receives as
inputs the staffing levels and corresponding shift
schedules, the demand estimates for each time block, the
probability distributions associated with the different
transaction times, and the probability function used to
model users’ fare-collector selection behavior. It gener-
ates a number of data sets that describe the evolution of
the wait lines over time, one for each simulation run. We
average the average queue lengths over the simulation
runs to identify the time blocks in which the expected
number of users in line exceeds the predefined service
standard. We then generate an additional constraint on
the number of fare collectors based on the violations of
the pre-established service level and incorporate this
 Miranda et al.: Improving Fare-Collection Shift Scheduling at Metro
8 Interfaces, Articles in Advance, pp. 1–14, © 2018 INFORMS

Table 1. Metro Proposed Two Daily Staff Schedules for Each Station Booth, One for Weekdays and the Other for Saturday and
Sunday; In Contrast, We Used Our Approach to Produce Two to Seven Daily Staff Schedules for Each Station Booth

Plan Number of daily schedules Monday Tuesday Wednesday Thursday Friday Saturday Sunday

Metro’s 2 1 1 1 1 1 2 2
Two days 2 1 1 1 1 1 2 2
Three days 3 1 1 1 1 1 2 3
Four days 4 1 1 2 2 2 3 4
Five days 5 1 2 2 2 3 4 5
Seven days 7 1 2 3 4 5 6 7

Note. The numbers in the last seven columns indicate the days that belong to the same daily staff schedule.

constraint into the optimization model. We run the
simulation model for each station booth and day sep-
arately; see the Appendix for a detailed description.
A Practical Application
In this section, we present an application of the pro-
posed simulation-optimization approach to the par-
ticular fare-collection shift-scheduling problem faced
by Metro and show its benefits. We first provide details
regarding the evaluation setting, followed by the re-
sults for a particular evaluation period.
Evaluation Setting
We consider two sets of staffing plans for 55 station
booths in Metro’s subway network: one set gener-
ated by Metro and the second produced using our
simulation-optimization approach. Both sets correspond
to a representative week based on Metro’s demand es-
timates for that week. The staffing plan generated
by Metro for each station booth consists of two staff
schedules: one for weekdays and the second for Saturday
and Sunday. For staffing purposes, Metro treats holidays
the same as Sundays. We used our approach to produce
five staffing plans for each station booth. The first con-
sisted of two staff schedules: one for weekdays and the
second for the weekend. The second involved three staff
schedules: one for weekdays, one for Saturday, and one
for Sunday. The third consisted of four staff schedules:
one for Monday and Tuesday; one for Wednesday,
Thursday, and Friday; one for Saturday; and one for
Sunday. The fourth involved five staff schedules: one for
Monday; one for Tuesday, Wednesday, and Thursday;
one for Friday; one for Saturday; and one for Sunday.
The fifth consisted of a different staff schedule for each
day of the week. Table 1 summarizes the staffing-plan
configurations suggested to Metro.
We determined a different staff schedule for each
type of day; that is, all days associated with the same
daily plan had the same schedule. To guarantee the
desired service level on each day, we computed the
demand estimate for each time block within a given
type of day as the maximum demand estimate for that
time block over all days of the same type.
We used three metrics to quantify the main benefits
of using the proposed simulation-optimization ap-
proach: (1) the average weekly staffing cost, (2) the
average number of users waiting in line in the one-
minute intervals in which the system did not reach the
predefined service standard, and (3) the total number
of time blocks in which the system did not reach the
predefined service standard. We computed the last two
metrics using a predefined service standard of no more
than three users waiting in line. We analyzed the
performance of the proposed approach under the
various staffing plans by simulating the system evo-
lution over time using actual (i.e., observed) demand
data. We obtained transaction time distributions from
data collected at eight representative station booths.
Exponential distributions with means of 11 and 17
seconds provided the best fit for the service times for
ticket sales and prepaid card reloads, respectively.
We formulated the optimization model using GAMS
22.5 and solved it to optimality with CPLEX 10.2.
We implemented the simulation model in Java 1.7.
To compare the various staffing plans, we ran 100
replications of a simulation of the fare-collection pro-
cess with a simulation length of 17 hours, equivalent
to 68 time blocks. We used a computer with a 2.2-GHz
Dual Core central processing unit (CPU) and 8 GB of
random access memory for all experiments we discuss
in this paper.
Table 2 displays the number of variables, the number of
constraints, and the number of simulation-optimization
Table 2. Problem Size and Number of Simulation-
Optimization Iterations for the Largest and the Smallest
Instance of the Problem
Number of: Largest instance Smallest instance
Integer variables 1,768 544
Constraints 1,218 606
Iterations 9 4
CPU time (seconds) 17.4 9.9
Miranda et al.: Improving Fare-Collection Shift Scheduling at Metro
Interfaces, Articles in Advance, pp. 1–14, © 2018 INFORMS 9

iterations for the largest and the smallest station booth
in our analysis. It also provides the execution time
associated with the best staffing plan.
Results
The second column of Table 3 presents the average
weekly cost in USD of each staffing plan. We computed
the average staffing costs considering the 55 station
booths in this analysis. The third column in this table
provides the difference, in relative terms, between the
average weekly staffing costs associated with Metro’s
plan and those associated with the respective proposed
plans. Table 4 shows the average number of users
waiting in line in the one-minute intervals in which the
system exceeded the three-user service standard. It also
displays the total number of blocks in which the av-
erage number of users waiting in line exceeded three.
As shown, the staffing plans created through the
proposed simulation-optimization approach clearly
outperformed Metro’s staffing plans. We can also ob-
serve that the more disaggregated the staffing plan (i.e.,
the more day types), the lower the average staffing
cost. This makes demand forecasting a key factor in
the cost-effective determination of staffing levels and
shift schedules. The number of time blocks in which
the service level did not reach the pre-established
Table 3. Average Weekly Staffing Costs of the Staffing
Plans
Average weekly Difference in relation
Staffing plan staffing cost (USD) to Metro’s plan (%)
Metro’s 109,403
Two days 101,324
Three days 100,518
Four days 95,933
Five days 95,621
Seven days 92,889
service standard is higher for the seven-day plans.
However, the violation of the predefined service
standard in such cases was very small. In addition,
the seven-day plans reduced the average number of
users waiting in line by more than 60% for those time
blocks. Consequently, although not ideal, this situ-
ation does not prevent the use of seven-day (or five-
day) plans.
On the basis of these results presented in Table 3, we
estimate that Metro could save more than 800,000 USD
annually using our approach to generate seven-day
plans for the 55 station booths we considered in our
analysis (of 108 station booths in the entire network).
An alternative way of comparing shift schedules
generated by the proposed approach with those used
by Metro is through demand and capacity profiles.
Figures 7 and 8 show the total capacity (staffing levels),
in minutes, scheduled by the proposed approach on
Tuesday using a seven-day plan for two representative
station booths: Universidad de Santiago and Los Le-
ones, respectively. These two figures also display
Metro’s staffing levels and the total demand for service,
in minutes, for that day. The latter assumes transaction
times of 11 and 17 seconds for buying tickets and
reloading prepaid cards, respectively. As shown in
these two graphs, the staffing levels determined by the
proposed approach more closely follow the actual
demand patterns and require fewer fare collectors
during off-peak hours. The seven-day schedule for the
Los Leones station booth, however, assigns more fare
collectors than Metro’s schedule during the evening
demand peak, but only for a short period. Thus, the
seven-day schedule satisfies the short peak in demand
0.0 while maintaining lower staffing levels over the re-
−7.4
−8.1
mainder of the day. In general, schedules generated by
−12.3 the proposed simulation-optimization approach more
−12.6 closely follow demand and thus more efficiently utilize
−15.1 fare-collection personnel.

Conclusions
Table 4. Average Number of Users Waiting in Line in the In this paper, we present a simulation-optimization
One-Minute Intervals in Which the Three-User Service approach developed to support fare-collection shift-
Standard Was Exceeded; and the Total Number of Time scheduling decisions in a subway network. In addition
Blocks in Which the Average Number of Users Waiting in
to reductions in the total staffing costs and improved
Line Exceeded Three for the Staffing Plans
service levels, the proposed approach allows sensitivity
Staffing plan Average wait line (users) Number of time blocks analyses and a rapid response to new operational re-
quirements while reducing the risk, tediousness, and
Metro’s 9.4 904
uncertainties inherent in manual operation. It reduces
Two days 3.4 243
Three days 3.3 508 the time spent in fare-collection staff-planning tasks
Four days 3.4 808 from a couple of days to only a few hours. Given the
Five days 3.3 987 positive results and qualitative benefits, the approach
Seven days 3.4 1,523 was implemented at Metro.
Note. The total number of time blocks considered in this analysis was The approach was embedded into a stand-alone ap-
26,180. plication on a client-server architecture. The application,

10
Figure 7. Aggregate Demand and Seven-Day Staffing Levels Corresponding to Tuesday for the Universidad de Santiago
Station Booth
Figure 8. Aggregate Demand and Seven-Day Staffing Levels Corresponding to Tuesday for the Los Leones Station Booth
developed using Visual Basic.NET, consists of five
modules: user interface, data input module, optimization
module, simulation module, and report module. Metro’s
Commercial Management runs the application using its
GAMS/CPLEX license.
We believe that the proposed simulation-optimization
approach can be applied to any large-scale business
problem involving the determination of staff re-
quirements for a service with highly variable demand
over the course of a day. Thus, its potential applica-
tions include the scheduling of call center agents, bank
tellers, supermarket checkout cashiers, and hospi-
tal staff.
Acknowledgments
The authors thank all the engineers, managers, and clerks
at Metro de Santiago for their contributions in this project;
in particular they thank Crist óbal Montecino and José
Mosquera for the valuable insights provided throughout
this project.
Miranda et al.: Improving Fare-Collection Shift Scheduling at Metro
Interfaces, Articles in Advance, pp. 1–14, © 2018 INFORMS
Appendix
We consider a subway network consisting of a large number
of station booths. Users can perform different types of
transactions (e.g., buying tickets and reloading prepaid
cards). We classify fare collectors into a set of types based on
characteristics such as the types of transactions they can
execute (e.g., with or without sales terminal) and their
position within stations (e.g., static inside a station booth or
mobile outside). Each type of fare collector can execute
specific types of transactions. We divide working days into
time blocks of equal duration and define work shifts as
a number of consecutive time blocks. We consider a pre-
defined set of possible shift lengths.
Operational Requirements
A minimum number of fare collectors must be working at any
moment at each station booth. Additionally, a maximum
number of fare collectors of each type can be working si-
multaneously. The lower bound on the total number of fare
collectors results from minimum operational staffing levels
and security concerns. For example, at least one fare collector
Miranda et al.: Improving Fare-Collection Shift Scheduling at Metro
Interfaces, Articles in Advance, pp. 1–14, © 2018 INFORMS 11

{
must be on duty at any time during operating hours. The upper 1, if the number of fare collectors on duty
bounds on the number of fare collectors of each type generally in time block q is less than the one for
• yq 
follow space constraints. For example, a limited number of time block q − 1;
windows is available at each station booth. Moreover, additional 0, otherwise.
conditions exist on the priority of the fare-collector assignments.

{
For example, no fare collectors should be placed outside
a station booth unless all station booth windows have been filled 1, if the number of fare collectors on duty
in time block q is less than the one
previously. These conditions arise because some types of fare • zq 
for time block q + 1;
collectors are more secure, less expensive, and more efficient
0, otherwise.
than others. Furthermore, there is a need to avoid abrupt
changes in the resulting staffing levels by imposing constraints We determine the staffing levels and shift schedules for
on the number of fare collectors on duty in consecutive periods. a specific station booth and day by solving the following
These additional constraints typically arise from personnel integer programming model:
management and control difficulties.
 
Q
Notation min Cpl xpql (A.1)
p∈3 l∈+ q1
The notation for the mathematical formulation of the problem
is as follows. We omit day and station booth indices for the subject to
sake of simplicity because we solve the problem for each day 
npq  xpq′ l ∀p ∈ 3, q  1, . . . , Q.
and station booth separately. (A.2)
l∈+ q′ ∈4lq
We define the following sets and subsets:
npq ≤ Up wpq ∀p ∈ 3, q  1, . . . , Q. (A.3)
• 7: set of transactions types. 
• 3: set of fare-collector types. npq ≥ L q  1, . . ., Q. (A.4)
• 3t ⊆ 3: set of fare-collector types that can execute p∈3
 
transactions of type t.  
• 3p− ⊆ 3: set of fare-collector types that have higher np′ q ≥ Up′ wpq ∀p ∈ 3,
p′ ∈3p− p′ ∈3p− q  1, . . ., Q. (A.5)
priority than type p. We must schedule all fare collectors of
 
types in 3p− before scheduling a fare collector of type p. Ftpq npq ≥ Dtq q  1, . . . , Q. (A.6)
• +: set of possible shift lengths (in number of time blocks). t∈7 p∈3t t∈7
• 4lq : set of time blocks corresponding to a
shift of length   
np,q−1 − npq ≤ yq Up q  2, . . ., Q.
l ending in time block q; that is, 4lq  q′ ∈ {1, . . ., Q} : p∈3 p∈3 p∈3
(A.7)
q − l ≤ q′ ≤ q .   
np,q+1 − npq ≤ zq Up q  1, . . . , Q − 1. (A.8)
We use the following parameters:
p∈3 p∈3 p∈3
• Q: number of time blocks in a working day.
• Δ: duration of a time block (in minutes). yq + zq ≤ 1 q  2, . . ., Q − 1. (A.9)
• Cpl : cost of having a fare collector of type p ∈ 3 working yq+1 + zq−1 ≤ 1 q  2, . . ., Q − 1. (A.10)
a shift of length l ∈ +. 
npv ≥ gv ∀v ∈ 9 ⊆ {1, . . ., Q}. (A.11)
• Dtq : estimated number of users demanding a transaction p∈3
of type t ∈ 7 in time block q ∈ {1, . . . , Q}.
xpql ∈ Z+ ∀p ∈ 3, l ∈ +, q  1, . . ., Q.
• Ftpq : number of transactions of type t ∈ 7 that a fare
collector of type p ∈ 3 should be able to execute in time block q. npq ∈ Z+ , wpq ∈ {0, 1} ∀p ∈ 3, q  1, . . . , Q.
• L: minimum number of fare collectors that must be yq ∈ {0, 1} q  2, . . ., Q.
working at any moment. zq ∈ {0, 1} q  1, . . ., Q − 1.
• Up : maximum number of fare collectors of type p that
can be working simultaneously. This integer programming model minimizes the total staffing
cost in (A.1). Constraint (A.2) links the shift-scheduling variables
Optimization Model xpql and the staffing-level variables npq . Constraints (A.3) and
To model the fare-collection shift-scheduling problem for (A.4) impose upper and lower limits, respectively, on the
a given day and station booth, we consider the following staffing levels at the station booth. Constraint (A.5) guarantees
decision variables: that, for each time block q, the solution reaches the maximum
• xpql : the number of fare collectors of type p ∈ 3 to be staffing levels of all fare-collector types in set 3p− before
scheduled for a shift of length l ∈ + starting in time block q. assigning a fare collector of type p ∈ 3 to that time block.
• npq : the total number of fare collectors of type p ∈ 3 on Constraint (A.6) ensures that the aggregated service capacity
duty in time block q. provided by the fare collectors on duty in time block q is greater

{
1, if at least one fare collector of type
• wpq  p ∈ 3 is on duty in time block q;
0, otherwise.
To avoid changes in opposite directions in the total number of
fare collectors in consecutive periods, we define two addi-
tional variables for each time block q:
than or equal to the expected demand for service in that time
block. As defined above, Ftpq is an estimate of the number of
transactions of type t ∈ 7 that a fare collector of type p ∈ 3
should be able to execute in time block q. Constraints
(A.7)–(A.10) prevent the resulting staffing plan from exhibiting
an increase (decrease) immediately followed by a decrease
(increase) in the total number of fare collectors in consecutive

12
time blocks. Constraint (A.11) states that at least gv fare col-
lectors must be on duty in time block v ∈ 9 ⊆ {1, . . ., Q}. The set
9 begins empty and the values of gv equal to zero, but they are
iteratively updated on the basis of the feedback provided by
the simulation model.
User Behavior
We assume that users enter subway stations individually and
independently. We characterize the user arrival process at
a particular station booth on a given day by a nonhomogeneous
arrival rate and a discrete probability distribution for the type of
transaction. We model users’ fare-collector selection behavior as
a probability function that depends on the length of the wait
lines and the types of fare collectors available at the station
booth. Users requesting a transaction of type t ∈ 7 must select
a fare collector who is able to execute that type of transaction.
That is, they must select a fare collector of type p ∈ 3t . We
further assume that users do not balk and that there is an
unlimited waiting area.
As with the arrival process, we consider that users receive
service individually and independently. Transaction times
depend on the type of transaction and on the type of fare
collector performing the transaction. Once users receive
service, they exit the system. The fare collector then becomes
available to serve the next user in line.
Simulation Model
As with the optimization model, we run the discrete-event
simulation model for each station booth and day separately.
This model receives the current staffing plan from the opti-
mization model, the demand estimates for each time block,
the probability distributions associated with the different
transaction times, and the probability function we use to
model users’ fare-collector selection behavior. It generates
a set {ar }Rr1 of R data sets that describe the evolution of the
wait lines over time, one for each simulation run r (R is the
number of simulation runs). We assume that user arrivals are
Poisson with a constant rate within each time block. Users
requesting a transaction of type t in time block q arrive at the
station booth at a rate λ  60 × Dtq /Δ users per hour, t ∈ 7 and
q  1, . . ., Q.
To identify the time blocks in which the system does not
satisfy the predefined service standard S, we first compute the
average number of users waiting in line for the 60-minute
periods starting at the beginning of each time block q. We
perform this for each simulation run r as follows:
   Amaran S, Sahinidis N, Sharda B, Bury S (2016) Simulation opti-
1 x(q)+60 1 N(u)  r
αr (q)  ak (u) du
60 x(q) N(u) k1
q  1, . . ., Q, r  1, . . . , R. (A.12)
In Equation (A.12), x(q) is the starting time (in minutes) of
time block q, ark (u) is the length of the wait line in front of fare
collector k at time u for simulation run r, and N(u) is the total
number of fare collectors on duty at time u. We then average
the average lengths of the wait lines over the simulation runs
to determine the expected number of users waiting in line for
the 60-minute periods starting at the beginning of each time

block q, which we denote by ᾱ(q). That is, ᾱ(q)  R1 Rr1 αr (q),
q  1, . . ., Q. These values are the expected service levels we
Miranda et al.: Improving Fare-Collection Shift Scheduling at Metro
Interfaces, Articles in Advance, pp. 1–14, © 2018 INFORMS
ultimately use to generate additional constraints on the number
of fare collectors.
Because the service level offered in a given time block
depends on the staffing levels in previous time blocks, we
identify the earliest time block in which the system does not
satisfy the predefined service standard. For this time block,
we add a new constraint that ensures that the number of fare
collectors on duty is greater than the respective number in the
current staffing plan. More precisely, let q∗ be the first time
block for which the expected number of users in line exceeds
the service level S, and let N̄ q∗ be the number of fare collectors
on duty in time block q∗ according to the current staffing plan.

That is, q∗  min{q : ᾱ(q) > S} and N̄ q∗  p∈3 n̄pq∗ , where n̄pq is
the value of the variable npq in the current staffing plan. Then,
we update Constraint (A.11) in the optimization model by
defining 9  9 ∪ q∗ and gq∗  N̄ q∗ + 1. We perform this for
each iteration of the solution approach shown in Figure 5.
Additional Notes
As an alternative to imposing an additional constraint only on
the earliest time block with a service-level violation (or making
the corresponding constraint more restrictive), we could si-
multaneously generate and incorporate into the optimization
model multiple constraints to accelerate the iterative solu-
tion process. However, we observed in our experiments that
service-level violations typically exist in sequences of consec-
utive time blocks. This is the result of having a higher demand
in the first time block in the sequence, generating a long wait
line that takes fare collectors several time blocks to serve. In this
sense, avoiding the generation of a long wait line in the initial
time block would solve the problem for the entire sequence of
time blocks. Therefore, when considering the alternative ap-
proach of simultaneously generating and incorporating mul-
tiple constraints into the optimization model, we should
carefully choose the set of time blocks on which to impose ad-
ditional constraints to avoid adding unnecessary constraints at
a higher cost. Because solution times are generally sufficiently
small, we did not consider this alternative in our implementation.
References
Agnihothri S, Taylor P (1991) Staffing a centralized appointment
scheduling department in Lourdes Hospital. Interfaces 21(5):1–11.
Alfares H (2004) Survey, categorization, and comparison of recent
tour scheduling literature. Ann. Oper. Res. 127(3):145–175.
Alfares H (2007) Operator staffing and scheduling for an IT-help call
centre. Eur. J. Oper. Res. 1(4):414–430.
mization: A review of algorithms and applications. Ann. Oper.
Res. 240(1):351–380.
Atlason J, Epelman M, Henderson S (2008) Optimizing call center
staffing using simulation and analytic center cutting-plane
methods. Management Sci. 54(2):295–309.
Avramidis A, Chan W, L’Ecuyer P (2009) Staffing multi-skill call
centers via search methods and a performance approximation.
IIE Trans. 41(6):483–497.
Avramidis A, Chan W, Gendreau M, L’Ecuyer P, Pisacane O (2010)
Optimizing daily agent scheduling in a multiskill call center.
Eur. J. Oper. Res. 200(3):822–832.
Azmat C, Widmer M (2004) A case study of single shift planning and
scheduling under annualized hours: A simple three-step ap-
proach. Eur. J. Oper. Res. 153(1):148–175.
Miranda et al.: Improving Fare-Collection Shift Scheduling at Metro
Interfaces, Articles in Advance, pp. 1–14, © 2018 INFORMS
Bailyn L, Collins R, Song Y (2007) Self-scheduling for hospital nurses:
An attempt and its difficulties. J. Nursing Management 15(1):72–77.
Bard J (2004) Staff scheduling in high volume service facilities with
downgrading. IIE Trans. 36(10):985–997.
Bard J, Purnomo H (2005) Hospital-wide reactive scheduling of
nurses with preference considerations. IIE Trans. 37(7):589–608.
Bard J, Wan L (2008) Workforce design with movement restrictions
between workstation groups. Manufacturing Service Oper. Man-
agement 10(1):24–42.
Bard J, Morton D, Wang Y (2007) Workforce planning at USPS mail
processing and distribution centers using stochastic optimiza-
tion. Ann. Oper. Res. 155(1):51–78.
Basra R, Lu K, Skobelev P (2007) Resolving scheduling issues of
the London Underground using a multi-agent system. Internat.
J. Intelligent Systems Tech. Appl. 2(1):3–19.
Bechtold S, Brusco M, Showalter M (1991) A comparative evaluation
of labor tour scheduling methods. Decision Sci. 22(4):683–699.
Belien J, Demeulemeester E (2006) Scheduling trainees at a hospital
department using a branch-and-price approach. Eur. J. Oper. Res.
175(1):258–278.
Berman O, Larson R, Pinker E (1997) Scheduling workforce and
workflow in a high volume factory. Management Sci. 43(2):158–172.
Brucker P, Qu R, Burke E (2011) Personnel scheduling: Models and
complexity. Eur. J. Oper. Res. 210(3):467–473.
Brunner J, Edenharter G (2011) Long-term staff scheduling of phy-
sicians with different experience levels in hospitals using column
generation. Health Care Management Sci. 14(2):189–202.
Byrne M, Bakir M (1999) Production planning using a hybrid
simulation-analytical approach. Internat. J. Production Econom.
59(1–3):305–311.
Cavique L, Rego C, Themido I (1999) Subgraph ejection chains and
tabu search for the crew scheduling problem. J. Oper. Res. Soc.
50(6):608–616.
Cezik M, L’Ecuyer P (2008) Staffing multiskill call centers via linear
programming and simulation. Management Sci. 54(2):310–323.
Cheung W, Leung L, Tam P (2005) An intelligent decision support
system for service network planning. Decision Support Systems
39(3):415–428.
Chew K, Pang J, Liu Q, Ou J, Teo C (2001) An optimization based
approach to the train operator scheduling problem at Singapore
MRT. Ann. Oper. Res. 108(1–4):111–122.
Choi K, Hwang J, Park M (2009) Scheduling restaurant workers to
minimize labor cost and meet service standards. Cornell Hospital
Quart. 50(2):155–167.
Chu S (2007) Generating, scheduling and rostering of shift crew-
duties: Applications at the Hong Kong International Airport.
Eur. J. Oper. Res. 177(3):1764–1778.
Chu S, Chan E (1998) Crew scheduling of light rail transit in Hong
Kong: From modeling to implementation. Comput. Oper. Res.
25(11):887–894.
Chun A, Suen T (2014) Engineering works scheduling for Hong
Kong’s Rail Network. Proc. Twenty-Eighth AAAI Conf. Artificial
Intelligence (Association for the Advancement of Artificial In-
telligence, Palo Alto, CA), 2890–2897.
Church J (1973) SURE STAF: A computerized staff scheduling system
for telephone business offices. Management Sci. 20(4):708–720.
Dantzig G (1954) A comment on Edie’s “Traffic delays at toll booths.”
J. Oper. Res. Soc. Amer. 2(3):339–341.
Dietz D (2011) Practical scheduling for call center operations. Omega
39(5):550–557.
Ding J, Qin Y, Jia L, Zhu S, Yu B (2013) SOMS: A subway operation
and maintenance system based on planned maintenance model
with train state. J. Intelligent Learn. Systems Appl. 5(4):195–202.
Edie L (1954) Traffic delays at toll booths. J. Oper. Res. Soc. Amer. 2(2):
107–138.
Elizondo R, Parada V, Pradenas L, Artigues C (2010) An evo-
lutionary and constructive approach to a crew scheduling
13
problem in underground passenger transport. J. Heuristics
16(4):575–591.
Ernst A, Jiang H, Krishnamoorthy M, Sier D (2004a) Staff scheduling
and rostering: A review of applications, methods and models.
Eur. J. Oper. Res. 153(1):3–27.
Ernst A, Jiang H, Krishnamoorthy M, Owens B, Sier D (2004b) An
annotated bibliography of personnel scheduling and rostering.
Ann. Oper. Res. 127(1):21–144.
Figueira G, Almada-Lobo B (2014) Hybrid simulation–optimization
methods: A taxonomy and discussion. Simulation Model. Practice
Theory 46(August):118–134.
Gans N, Koole G, Mandelbaum A (2003) Telephone call centers:
Tutorial, review, and research prospects. Manufacturing Service
Oper. Management 5(2):79–141.
Green L (2005) Capacity planning and management in hospitals.
Brandeau ML, Sainfort F, Pierskalla WP, eds. Operations Research
and Health Care, International Series in Operations Research &
Management Science, vol. 70 (Springer, Berlin), 15–41.
Han A, Li E (2014) A constraint programming-based approach to the
crew scheduling problem of the Taipei mass rapid transit system.
Ann. Oper. Res. 223(1):173–193.
Helber S, Henken K (2010) Profit-oriented shift scheduling of inbound
contact centers with skills-based routing, impatient customers,
and retrials. OR Spectrum 32(1):109–134.
Hojati M (2010) Near-optimal solution to an employee assignment
problem with seniority. Ann. Oper. Res. 181(1):539–557.
Hueter J, Swart W (1998) An integrated labor-management system
for Taco Bell. Interfaces 28(1):75–91.
Hung Y, Leachman R (1996) A production planning methodology for
semiconductor manufacturing based on iterative simulation
and linear programming calculations. IEEE Trans. Semiconductor
Manufacturing 9(2):257–269.
Kim B, Kim S (2001) Extended model for a hybrid production
planning approach. Internat. J. Production Econom. 73(2):165–173.
Lau H, Yuan Z, Gunawan A (2016) Patrol scheduling in urban rail
network. Ann. Oper. Res. 239(1):317–342.
Lezaun M, Pérez G, de la Maza E (2010) Staff rostering for the station
personnel of a railway company. J. Oper. Res. Soc. 61(7):1104–1111.
Lilly M, Emovon I, Ogaji S, Probert S (2007) Four-day service-staff
work-week in order to complete maintenance operations more
effectively in a Nigerian power-generating station. Appl. Energy
84(10):1044–1055.
Lin C (1999) The development of a workforce management system for
a hotline service. Comput. Indust. Engrg. 37(1/2):465–468.
Mandelbaum A (2004) Call centers (centres): Research bibliography with
abstracts. Working paper, Technion-Israel Institute of Technology,
Haifa, Israel.
Mason A, Ryan D, Panton D (1998) Integrated simulation, heuristic
and optimisation approaches to staff scheduling. Oper. Res. 46(2):
161–175.
Metro de Santiago (2015) Annual report, Metro de Santiago, Santiago,
Chile.
Pot A, Bhulai S, Koole G (2008) A simple staffing method for mul-
tiskill call centers. Manufacturing Service Oper. Management 10(3):
421–428.
Qi X, Bard J (2006) Generating labor requirements and rosters for
mail handlers using simulation and optimization. Comput. Oper.
Res. 33(9):2645–2666.
Rong A, Grunow M (2009) Shift designs for freight handling per-
sonnel at air cargo terminals. Transportation Res. Part E: Logist.
Transportation Rev. 45(5):725–739.
Sabar M, Montreuil B, Frayret J (2008) Competency and preference based
personnel scheduling in large assembly lines. Internat. J. Comput.
Integrated Manufacturing 21(4):468–479.
Sodhi M, Norris S (2004) A flexible, fast, and optimal modeling
approach applied to crew rostering at London Underground.
Ann. Oper. Res. 127(1-4):259–281.

14
Thiel M (2008) Team-oriented airline crew rostering for cockpit
personnel. Hickman M, Mirchandani P, Voß S, eds. Computer-
aided Systems in Public Transport. Lecture Notes in Economics and
Mathematical Systems, vol. 600 (Springer, Berlin), 91–114.
Van den Bergh J, Beliën J, De Bruecker P, Demeulemeester E, De
Boeck L (2013) Personnel scheduling: A literature review. Eur. J.
Oper. Res. 226(3):367–385.
Varakantham P, Lau H, Yuan Z (2013) Scalable randomized patrolling
for securing rapid transit networks. Proc. Twenty-Fifth Conf. In-
novative Appl. Artificial Intelligence (Association for the Advancement
of Artificial Intelligence, Palo Alto, CA), 1563–1568.
Zhang Y, Puterman ML, Nelson M, Atkins D (2012) A simulation
optimization approach to long-term care capacity planning.
Oper. Res. 60(2):249–261.
Verification Letter
Jon Brintrup Store, Operations Programming Senior Man-
ager, Metro S.A., Av. Libertador Bernardo O’Higgins 1414,
Santiago, Chile, writes:
“Metro S.A. operates the subway system in Santiago, Chile.
Currently, our network has 108 stations, and on average we
manage approximately 2.3 million rides daily.
“One of the operational problems we were faced with was
the assignment of operators to our sales booths. Until the
year 2007, this assignment was performed manually twice
a month, and it took us several days to come up with an
acceptable solution that had to comply with two objectives:
low personnel costs and high service-levels, i.e. short wait-
ing lines.
“In 2008, we started a project with Operations Research
professionals from Universidad de Chile where the goal was
to develop a user-friendly decision support tool for operator
assignment in our sales booths.
“This tool has been developed in several stages during the
following years, and in early 2011, we installed the complete
system, which combines in an innovative way, optimization
algorithms with simulation tools.
“The results achieved have been very positive. Compared
with the base case, we could reduce operational costs by
3% on average maintaining standard queues of customers
waiting in front of our most demanded sales booths. The
process from data acquisition to shift assignment takes now
only 30 minutes of optimization plus a couple hours of data
Miranda et al.: Improving Fare-Collection Shift Scheduling at Metro
Interfaces, Articles in Advance, pp. 1–14, © 2018 INFORMS
preparation, compared with several days before we started
the mentioned project.
“We are very satisfied with the operations research solu-
tion developed by professionals from Universidad de Chile
and are looking forward to future collaborations.”
Jaime Miranda is an assistant professor at the Faculty of
Economics and Business at the University of Chile. He holds
an undergraduate degree in industrial engineering, a master’s
degree in operations management, and a PhD degree in engi-
neering systems from the same institution. He has worked on
various high-impact studies in industry and the public sector in
Chile and has participated in the development of numerous
decision-support systems based on optimization models.
Antoine Sauré is an assistant professor at the Telfer School
of Management at the University of Ottawa. His research
interests include stochastic modeling, dynamic optimization,
and decision-making under uncertainty. He has more than
twelve years of experience developing and applying ad-
vanced analytics techniques to large-scale problems in several
industries. He has worked on the development of numerous
planning and scheduling systems aimed to provide timely
access to quality cancer care.
Pablo A. Rey is an assistant professor at the Department of
Industry and an associate researcher of the Programa Institu-
cional de Fomento a la Investigación, Desarrollo e Innovación
at the Universidad Tecnológica Metropolitana, Chile. His
research interests include optimization, scheduling, and
transportation. He holds a BSc degree in mathematics from
the National University of Córdoba, Argentina, and a PhD
in electrical engineering from the Catholic University of Rio
de Janeiro, Brazil.
Richard Weber is an associate professor at the Department
of Industrial Engineering and director of the Department
of Graduate Studies at the University of Chile. His re-
search interests include data mining, dynamic data min-
ing, and computational intelligence. He is a senior member
of IEEE and a member of ACM and INFORMS. Richard
Weber holds a BSc in mathematics as well as an MSc and
a PhD, both in operations research, from Aachen Uni-
versity, Germany.