International Series in Operations Research

& Management Science

Volume 232

Series Editor
Camille C. Price
Stephen F. Austin State University, TX, USA

Associate Series Editor

Joe Zhu
Worcester Polytechnic Institute, MA, USA

Founding Series Editor

Frederick S. Hillier
Stanford University, CA, USA
More information about this series at http://www.springer.com/series/6161
H. A. Eiselt • Vladimir Marianov
Editors

Applications of Location
Analysis
Editors
H. A. Eiselt Vladimir Marianov
Faculty of Business Administration Department of Electrical Engineering
University of New Brunswick Pontificia Universidad Católica de Chile
Fredericton Macul
New Brunswick Chile
Canada

ISSN 0884-8289 ISSN 2214-7934 (electronic)

International Series in Operations Research & Management Science
ISBN 978-3-319-20281-5 ISBN 978-3-319-20282-2 (eBook)
DOI 10.1007/978-3-319-20282-2

Library of Congress Control Number: 2015948175

Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made.

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media
(www.springer.com)
Preface

This book is designed to highlight some of the applications of location analysis in a

variety of different fields. In some sense, it is a companion piece of the “Foundations
of Location Analysis” volume that was published in 2011. While the latter book
focused on the problems that can be solved, this book describes scenarios, in which
location techniques have been used to solve actual problems.
It is well known that the techniques that are usually described in books and taught
in many location courses—medians, centers, covering problems—are but prototypes
that may be applied to real situations when suitably modified. In addition, many lo-
cation problems comprise multiple objectives, which makes tools from multicriteria
decision making an attractive choice. As was stated elsewhere, the fact that loca-
tion problems are typically found on the strategic level, means that more often than
not, simple single-objective optimization models may very well be a good starting
point, but have to be supplemented by other tools that allow the inclusion of more
sophisticated features, such as nonlinear choice rules and multiple objectives.
This book features spotlights on some important applications of location models
and the different tools used in the decision-making process. There are three main
classes of applications: applications by businesses, those that deal with public ser-
vices, and applications that deal with law enforcement and first responders. Overall,
it is noticeable that the public sector appears to be the main user of location models.
The first application in the private sector is described by Başar et al. The au-
thors describe the location of bank branches in Istanbul, Turkey. The tool is a
mixed-integer programming problem, whose objective is to maximize the net profit,
based on the expected transaction volume at the branches as well as branch opening
and closing penalties. One of the main issues here is a description of the banking
behavior of individual and commercial users.
In the second chapter, Bhadury et al. discuss the potential location of a logistics
park (a.k.a. freight village) in the southeastern part of North Carolina. The authors
delineate a four-step procedure that incorporates the main phases of the location
process: choice of the general region, identification of suitable sites within the cho-
sen region, assessment of the sites that were identified in the previous step, and a

v
vi Preface

comparison of the sites. In the end, the situation had changed, and the logistics park
was not built and a new process started. While disappointing, this is actually not
uncommon: the case of the sewage treatment plant in Halifax, Canada, was, at least
in that respect, similar.
In the third chapter, Gunn focuses on sustainable forest management. His main
objective is the maximization of discounted profit. Again, the tool are mixed-integer
optimization formulations. Important features include a concentration of the entire
supply chain that comprises the forest, sawmills, pulpmills, and different products
delivered to different customers. The constructions of roads to bring equipment to
the parcels that are to be harvested is an important cost factor. Also important are
neighborhood restrictions that ensure that the planners avoid denuding large swaths
of land, which would foster erosion.
Chapter 4 by Arnolds and Nickel looks at layout problems in a hospital. It is
well known that layout problems are much more difficult than location problems,
as the former have to respect the shape of the facilities (e.g., rooms), whereas lo-
cation problem essentially locate points in some space. The objective in this piece
is to locate wards, walk-in clinics, and operating rooms so as to minimize travel
distances and times. It is well-known that quadratic assignment problems, one of
the cornerstones of layout problems, are already among the most difficult problems
in integer programming. Any additional model features that make the formulations
more realistic will further increase the degree of difficulty, so that heuristics have to
be applied.
The first chapter dealing with public services is Chap. 5. Here, Church et al. deal
with the design of habitats for wildlife in California under special consideration of
the spotted owl. The tool of choice is a mixed-integer programming formulation of
an anti-covering model. In general, the problem of designing a habitat is a difficult
one, in part due to the different requirements of individual species: for instance,
while mountain lions will thrive in large, contiguous areas, bald eagles will prefer
a large number of small protected zones. If preserves for multiple species are to be
designed, suitable compromises will have to be found.
Chapter 6 deal with the control of forest fires. Church et al. set up a multiobjective
mixed-integer optimization model that includes fuel removal, the minimization of
damage in the transition region, known as wildland-urban interface, the minimiza-
tion of the variation in year-to-year workload of the crews, and the maximization of
adjacency in some period and subsequent periods. The main concern is scheduling,
and the authors set up a decision support system for that purpose.
Chapter 7 by Geetla et al. discusses the location of intelligent sensors along
highways to enable emergency vehicles to respond in a timely fashion. The au-
thors devise an explicit-implicit model, which is succeeded by an optimization that
includes single and double coverage. It is noteworthy that the model maximizes cov-
erage from nodes as well as coverage from paths for an area in a part of Buffalo,
New York. A variety of sensitivity analyses establishes the robustness of the system.
In Chap. 8, Verter and Zhang describe the problem of locating breast cancer
screening centers in Montreal, Quebec, Canada. The authors formulate a mixed-
integer programming problem, whose objective is the maximization of the level
Preface vii

of participation in the program. Again, proximity and patient behavior are main
features of the model.
Chapter 9 by Yezer and Gillula is rooted in economic theory. The paper performs
an economic analysis for the locations of post offices. Among other concerns, in-
cluded in the model are revenue, the number of windows, where service is offered,
and the size of the facility. The objective is the maximization of net revenue per
unit area of the facility. The authors’ model allows the analysis of an existing retail
network and permits decisions regarding the closures of some of the facilities.
In Chap. 10, Giesen et al. investigate the location of schools in a region of Brazil.
The objective function minimizes the sum of transportation, operating, and penalty
costs. Complicating features include the existence of different types of schools (pub-
lic and charger schools), which operate under different rules (children must attend
the school to which they are assigned by a central authority, or the child’s parents
have the choice), and there may be a choice between multigrade schools and single-
grade schools. Communities, politicians, unions, and others have views that will
have to be included for a successful implementation. A significant increase in the
efficiency is possible when the system is adopted. One of the key recommenda-
tions for a successful implementation includes a smooth transition from the present
solution to the optimized solution.
The third part of this book deals with applications of enforcement and first re-
sponders. Chapter 11 by Murray discusses the well-known problem of locating fire
stations. The objective function for the problem in a city in California minimizes the
sum of fixed and operating costs, while constraints ensure that a certain percentage
of the demand is actually satisfied. The author’s model is a generalization of the
well-known maximum capture problem. After the recommendations were made to
the authorities, the city in question merged with another population agglomeration,
which, in conjunction with a general economic downturn, resulted in the fact that
the system was not implemented. The lessons of the process, though, remain for
future users.
In Chap. 12, Gentili and Mirchandani describe a model for the location of sensors
for travel time information. These sensors can be used to detect changes in traffic
patterns to avoid congestion as well as plan future traffic networks. The application
maximizes the distance that can be monitored by the sensors. The model is applied
in an area of Texas. Tradeoff curves are determined for the location of additional
sensors.
In Chap. 13, Bucarey et al. consider the problem of police districting. Their
model is a largely extended p-median formulation that locates centers of districts,
so as to minimize a linear convex combination of three objective functions, viz.,
the sum of center-customer distances, a penalty for odd shapes of the district, and
prevention demand. Based on the complexity of this mixed integer programming
problem, the authors design a location—allocation heuristic to solve the problem.
The model is applied to Chile, resulting in major benefits as compared to the present
solution.
In Chap. 14, Marianov analyzes the location of jails in Chile. A mixed-integer
optimization problem is formulated, which minimizes the costs of opening new
viii Preface

facilities, the costs of expanding existing facilities, a penalty cost for inmates’ trans-
portation, and the last term is another penalty cost of overpopulation in the jail.
The author outlines some of the major difficulties with this subject, particularly
changes in demographics, changes in the law, and major trends, such as the increase
in drug-related offenses.
Chapter 15 by Pelot et al. deals with the location of Coast Guard vessels off
the Canadian Atlantic Coast, in order to be able to efficiently respond to incidents,
whose severity is measured in different categories. The model also uses differ-
ent workload capacities, and the different versions are variations on max cover
problems.
In the last chapter of the book, Bell surveys military applications of location
analysis throughout different periods of recent history. The type of models used in
the military spans a wide range from weapons positioning, the selection of training
sites, hospital locations, locations for search and rescue aircraft, the location of spare
aircraft engines, the closure of military bases, and many others.
Finally, it is our great pleasure to express our sincere thanks to our authors and
their cooperation, as well as to Camille Price, Matthew Amboy, Christine Crigler,
and Neil Levine. Without their assistance and encouragement, the task of putting
this book together would have been a lot less pleasant. Many thanks to all of them.

H. A. Eiselt
Vladimir Marianov
Contents

1 Location Analysis in Practice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

H. A. Eiselt, Vladimir Marianov and Joyendu Bhadury

Part I Business

2 Location Analysis in Banking: A New Methodology and Application

For a Turkish Bank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Ayfer Başar, Özgür Kabak, Y. İlker Topçu and Burçin Bozkaya
3 Location Modeling for Logistics Parks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Joyendu Bhadury, Mark L. Burkey and Samuel P Troy
4 An Introduction to Industrial Forestry from a Location Perspective . . 85
Eldon Gunn

5 Layout Planning Problems in Health Care . . . . . . . . . . . . . . . . . . . . . . . . 109

Ines Arnolds and Stefan Nickel

Part II Public Services

6 Modeling the Potential for Critical Habitat . . . . . . . . . . . . . . . . . . . . . . . . 155

Richard L. Church, Matthew R. Niblett and Ross A. Gerrard

7 Saving the Forest by Reducing Fire Severity: Selective Fuels

Treatment Location and Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Richard L. Church, Matthew R. Niblett, Jesse O’Hanley,
Richard Middleton and Klaus Barber
8 Locating Intelligent Sensors on a Transportation Network to
Facilitate Emergency Response to Traffic Incidents . . . . . . . . . . . . . . . . 191
Tejswaroop Geetla, Rajan Batta, Alan Blatt, Marie Flanigan
and Kevin Majka

ix
x Contents

9 Location Models for Preventive Care . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

Vedat Verter and Yue Zhang
10 Modeling the Location of Retail Facilities: An Application to the
Postal Service . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
Anthony M. Yezer and James W. Gillula

11 Rural School Location and Student Allocation . . . . . . . . . . . . . . . . . . . . . 273

Ricardo Giesen, Paulo Rocha E Oliveira and Vladimir Marianov

Part III Enforcement and First Responders

12 Fire Station Siting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293

Alan T. Murray
13 Locating Vehicle Identification Sensors for Travel Time
Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
Monica Gentili and Pitu B. Mirchandani

14 Shape and Balance in Police Districting . . . . . . . . . . . . . . . . . . . . . . . . . . . 329

Victor Bucarey, Fernando Ordóñez and Enrique Bassaletti

15 Location and Sizing of Prisons and Inmate Allocation . . . . . . . . . . . . . . 349

Vladimir Marianov

16 Vessel Location Modeling for Maritime Search and Rescue . . . . . . . . . 369

Ronald Pelot, Amin Akbari and Li Li

17 Military Applications of Location Analysis . . . . . . . . . . . . . . . . . . . . . . . . 403

John E. Bell and Stanley E. Griffis

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
Contributors

Amin Akbari Department of Industrial Engineering, Dalhousie University,

Halifax, NS, Canada
Ines Arnolds Discrete Optimization and Logistics at the IOR, Karlsruhe Institute
of Technology, Karlsruhe, Germany
Klaus Barber US Forest Service Region 5, Vallejo, CA, USA
Ayfer Başar Industrial Engineering Department, Istanbul Technical University,
Maçka Campus, İstanbul, Turkey
Enrique Bassaletti Department of Criminal Analysis, Carabineros de Chile,
Santiago, Chile
Rajan Batta Department of Industrial and Systems Engineering, University at
Buffalo (SUNY), Buffalo, NY, USA
John E. Bell Department of Marketing & Supply Chain Management, University
of Tennessee, Knoxville, TN, USA
Joyendu Bhadury Bryan School of Business and Economics, University of North
Carolina at Greensboro, Greensboro, NC, USA
Alan Blatt Center for Transportation Injury Research, CUBRC, Buffalo, NY, USA
Burçin Bozkaya Sabanci School of Management, Sabanci University, Tuzla,
İstanbul, Turkey
Victor Bucarey Department of Industrial Engineering, Universidad de Chile,
Santiago, Chile
Mark L. Burkey School of Business and Economics, North Carolina A & T State
University, Greensboro, NC, USA
Richard L. Church University of California, Santa Barbara, CA, USA

xi
xii Contributors

H. A. Eiselt Faculty of Business Administration, University of New Brunswick,

Fredericton, NB, Canada
Marie Flanigan Center for Transportation Injury Research, CUBRC, Buffalo, NY,
USA
Tejswaroop Geetla Department of Industrial and Systems Engineering, University
at Buffalo (SUNY), Buffalo, NY, USA
Monica Gentili Department of Mathematics, University of Salerno, Fisciano,
Salerno, Italy
Ross A. Gerrard USDA Forest Service, PSW Research Station, Davis, CA, USA
Ricardo Giesen Department of Transport Engineering and Logistics, Pontificia
Universidad Católica de Chile, Santiago, Chile
James W. Gillula IHS Economics, Washinton, D.C., USA
Stanley E. Griffis Department of Supply Chain Management, Michigan State
University, East Lansing, MI, USA
Eldon Gunn Department of Industrial Engineering, Dalhousie University, Halifax,
NS, Canada
Özgür Kabak Industrial Engineering Department, Istanbul Technical University,
Maçka Campus, İstanbul, Turkey
Li Li Department of Industrial Engineering, Dalhousie University, Halifax, NS,
Canada
Vladimir Marianov Department of Electrical Engineering, Pontificia Universidad
Católica de Chile, Santiago, Chile
Kevin Majka Center for Transportation Injury Research, CUBRC, Buffalo, NY,
USA
Richard Middleton Los Alamos National Laboratory, Los Alamos, NM, USA
Pitu B. Mirchandani School of Computing, Informatics and Decision Systems
Engineering, Arizona State University, Tempe, AZ, USA
Alan T. Murray Center for Spatial Analytics and Geocomputation, College of
Computing and Informatics, School of Public Health, Drexel University, Philadel-
phia, PA, USA
Matthew R. Niblett University of California, Santa Barbara, CA, USA
Stefan Nickel Discrete Optimization and Logistics at the IOR, Karlsruhe Institute
of Technology, Karlsruhe, Germany
Jesse O’Hanley Kent Business School, Kent University, Canterbury, UK
Paulo Rocha E Oliveira IESE Business School, University of Navarra, Pamplona,
Spain
Contributors xiii

Fernando Ordóñez Department of Industrial Engineering, Universidad de Chile,

Santiago, Chile
Ronald Pelot Department of Industrial Engineering, Dalhousie University,
Halifax, NS, Canada
Y. İlker Topçu Industrial Engineering Department, Istanbul Technical University,
Maçka Campus, İstanbul, Turkey
Samuel P Troy Bryan School of Business and Economics, University of North
Carolina at Greensboro, Greensboro, NC, USA
Vedat Verter Desautels Faculty of Management, McGill University, Montreal,
QC, Canada
Anthony M. Yezer George Washington University, Washinton, D.C., USA
Yue Zhang College of Business and Innovation, University of Toledo, Toledo, OH,
USA
Chapter 1
Location Analysis in Practice

H. A. Eiselt, Vladimir Marianov and Joyendu Bhadury

1.1 Location Problems and Its Features

Location choices are as old as mankind. Individuals have always chosen their
place of residence, starting with the appropriate cave, which would have to be in
reasonable proximity to places in which food, and later work, could be found).
Municipalities located central places to conduct business, such as the Greek agora
or the Roman forum (or arenas for gladiator fights, for that matter), while leaders of
countries located places for their administration, their armies, &, eventually, their
burial sites (e.g., pyramids).
An interesting look at one of these early situations is provided by Wilamowsky
et al. (1995), who demonstrated that Joshua’s (of Bible fame) choice of locations for
three cities for “unintentional murderers” were actually optimal sites. Along similar
lines, ReVelle and Rosing (2000) showed Emperor Constantine the Great’s choices
for the locations of his armies in the fourth century AD. More specifically, there
were eight regions of the Roman Empire, in which four existing field armies had to
be positioned. The authors of the study used the concept of covering models to locate
the field armies, as to require the smallest number of steps of relocation, in case an
attack on any of the regions ensued. Actually, the emperor’s choice was not optimal
with respect to ReVelle and Rosing’s proximity criterion; it appears that stationing
two field armies in Rome was more of a political than strategic decision. Even in

H. A. Eiselt ()
Faculty of Business Administration, University of New Brunswick, Fredericton,
NB E3B 5A3, Canada
e-mail: [email protected]
V. Marianov
Department of Electrical Engineering, Pontificia Universidad Católica de Chile,
Santiago, Chile
e-mail: [email protected]
J. Bhadury
Bryan School of Business and Economics, University of North Carolina at Greensboro,
Greensboro, NC 27402-6170, USA
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 1

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_1
2 H. A. Eiselt et al.

literature we can find location problems, albeit hidden somewhat. A good example
is Lewis Carroll’s (1880–5) Tangled Tales, where in Knot 2: Eligible apartments,
the author has one of his protagonists state: “‘One day-room and three bedrooms,’
said Balbus, as they returned to the hotel. ‘We will take as our day-room the one
that gives us the least walking to do to get to it.”’ In location analytic terms, we are
looking for the location of a median (the concept will be defined below).
Since that time, a large number of contributions have described possible, realistic,
real, and implemented location models. The contribution by Eiselt (1992) surveys
some of the applications of the 1980s as a lead-in to a special issue of applications in
location analysis. A survey that investigates and summarizes contributions related to
agriculture is found in Lucas and Chhajed (2004). In addition to the many facilities
that have been located throughout the decades, e.g., warehouses, blood banks, fire
stations (a good taxonomy is provided by Başar et al. 2012), ambulances, etc., (many
of these real location problems since 1990 are surveyed in Table 1.1 in this paper),
there are some unusual location problems. Among them are the optimal location of
disinfection stations for water supply networks (Tryby et al. 2002) and the prime
locations for street robbers (and, presumably, cops, who will try to foil attempts of
the robbers to ply their craft), see Bernasco et al. (2013).
This discussion leads us immediately to the fact that many real location deci-
sions will not rely on a single quantitative criterion, such as distances, but rely
on other, often, intangible criteria. In personal location decisions (e.g., choosing
a place to live), this could be the perceived beauty of the area, the perceived quality
of the neighborhood including the type of school district, and many others. In pub-
lic location decisions, public support/opposition, environmental impact, and similar
concerns are often found. Private location problems, on the other hand, are typically
simpler, as profit or cost are the main concerns. These criteria are easily quantifiable
and fit very well into a mathematical model.
This chapter will analyze some of the foundations of location models, and deter-
mined classes of applications and their main features. In order to do so, we will very
loosely and generally model a location problem, in which we have a metric space,
in which customers (we use the term here in the widest possible sense), at least in
the short to medium run, occupy fixed locations or move along known paths, and in
which a decision maker will locate facilities based on known or perceived interac-
tions between the facilities and the customers. These interactions will typically be
transportation, either from a facility to a customer or vice versa.
Weber (1909) was among the first to look at actual transportation problems from
a mathematical point of view. More specifically, the appendix of his book written by
Georg Pick, minimizes transportation costs among three customers and one facility,
which are assumed to be proportional to Euclidean distances between the points.
Equivalently, this can be seen as finding the point in a triangle, such that the sum of
distances between the new point and the vertices of the triangle are as small as pos-
sible. This problem was first posed by Fermat (formulated before 1640), solved by
Torricelli in 1645, and for which a well-known iterative algorithm was described by
Weiszfeld (1937). Since Weber’s seminal work, geographers and economists have
debated the location of facilities (mostly firms in the early days). It is interesting to
1 Location Analysis in Practice 3

note that while many facilities that deal with the first sector of the economy (i.e.,
extraction, such as agriculture and mining) are located close to the source, many
facilities of the third sector of the economy (the service sector) are located close to
the customers. (We wish to note at this point that this assertion holds true for brick-
and-mortar facilities, we do not include online firms or facilities in our discussion).
There are many reasons for this phenomenon: Consider a firm in the first sector,
e.g., a gold mine. Typically, we may expect about one ounce of gold in three tons
of ore. Clearly, given the very low output-to-input ratio, nobody would transport the
ore to sites close to the customer to extract the gold at that location. On the other
hand, facilities in the service sector have extensive dealings with customers, so that
their locations are chosen quite naturally in close proximity to customers.
Consider now a typical scenario, in which multiple facilities of one type exist.
In such a case, there are two issues to be addressed: on the one hand, does the firm
or the customer choose the facility he is dealing with? The second issue concerns
the facility customer interaction. In particular, we distinguish the case in which the
firm is responsible for the interaction (e.g., the transport), and the case in which this
interaction is done by the customer. This leaves us with four cases. In the first class
of models, the firm chooses the facility the customer is served from and the firm also
deals with the interaction. Typical examples of this case are trucking, waste collec-
tion, ambulances, and many others. Models with this feature are often referred to
as allocation (as the firms allocate facilities to their customers) or shipping models.
On the other hand, consider models in which customers choose the facility they are
dealing with and are also responsible for the interaction with the chosen facility.
Models in this category include retail stores, libraries, ATM machines, and other,
service-related facilities. Models of this type are frequently called customer choice
models or shopping models. Much less are “mixed models”, in which one agent is
choosing the facility, while the other is responsible for the interaction. Examples
for the firm (in the widest sense, typically a government agency) chooses the facil-
ity customers are to patronize, while customers are responsible for the interaction.
Scenarios of this type include polling stations, public schools, and similar facilities.
Finally, there are rather few cases, in which customers choose the facility, while the
firm takes care of the interaction. At first glance, it appears that a customer’s choice
of the branch of a chain of retail, e.g., furniture, stores falls into that category, as the
store may then ultimately take care of the interaction with the customer by deliver-
ing the goods. While this is true, the customer cannot choose (and has no interest in
choosing) the warehouse, from which the furniture is delivered to him. This is then
again a standard allocation/shipping model.
Another possible way to distinguish between location models was suggested by
ReVelle et al. (1970), see also ReVelle and Eiselt (2005), where the authors dis-
tinguish between the location of public and private facilities. In some cases, such
a distinction will not reveal any insight. Consider, for instance, the location of a
library, typically a public facility. The city or whichever government department
is responsible for the planning, will try to make the library accessible to as many
people as possible, so that the average facility—customer (or, in this case, poten-
tial patron) is as short as possible. On the other hand, a firm that plans the location
4 H. A. Eiselt et al.

of a regional distribution center will locate the center in such a way that the sum
of distances from the center to the customers is minimized, as this can be seen as
a proxy for cost. (Note that this argument is not as straightforward as it appears:
first, it assumes that special trips are made to the customers rather than milk runs
to serve multiple customers on a single trip, which would result in much more dif-
ficult location-routing problems, and secondly that there is an established cost per
ton-mile for trucking, such as $ 1.38 as illustrated in The Trucker’s Report 2015). In
other words, the decision maker’s behavior will be the same (locating the facility in
close proximity to the customers), while the underlying motives differ: accessibility
for the library planner, and cost minimization for the trucking firm. On the other
hand, the public vs. private distinction may be quire relevant in other aspects: typi-
cally, public decision makers face many stakeholders in their decision making and,
as a result, many conflicting objectives and criteria are to be included. On the other
hand, private decision makers can much more easily agree on (profit maximizing or
cost minimizing) objectives, which are much easier to incorporate in optimization
models given the relative ease with which they can be quantified.
As outlined above, location models deal with the spatial separation and the inter-
action of customers and facilities. The separation of facilities and customers is
expressed in terms of their distance (shortest path in networks and some Minkowski
distance in the plane are the most frequently used measures). These distance can
then appear in the objective function of a mathematical optimization problem as is
the case in typical median or center problems, or they appear in the constraints, as
is the case in covering models. Naturally, mixed forms are possible and have been
used. In addition to the usual models in which proximity of facilities to customers is
desired (they minimize a function of the distances, where customers “pull” a facil-
ity towards themselves, see, e.g., Eiselt and Laporte 1995) or undesirable (where
the customers “push” the facility away from their location, in case the facility is
deemed undesirable), there are models in which only the distances between facil-
ities are relevant. Typical examples are dispersion and defender models; see, e.g.,
Daskin (2008).

1.2 A Locator’s Toolkit

This section will delineate some tools locators have at their disposal. We will first
survey generic formulations of a number of the standard location models. We then
discuss a few of the approaches to multiobjective optimization, and finally, we will
review some of the main techniques in multicriteria decision making.
In order to facilitate our discussion, we first distinguish between two main cate-
gories of location problems: discrete problems and continuous problems. Whereas
there exists only a finite number of potential locations in discrete problems, there is
an infinite number of location to choose from in continuous problems. If a discrete
problem is formulated on a network, the finite set of potential locations is often equal
to (or a subset of) the set of nodes. This is, however, not a necessary requirement.
1 Location Analysis in Practice 5

The location variables, i.e., the variables that deal with where to locate a facility in
the given space, in discrete spaces are simply defined as zero-one variables: for each
potential facility location, a variable is created that assumes the value of one, if we
decide to locate at that site, and zero otherwise. Such an approach is obviously not
feasible in continuous problems, in which an infinite number of potential locations
exists. In such cases, in which a facility can either be located in the plane or, in more
general cases, in any d-dimensional space, or at any point on a network, the facility
location can then be determined by the coordinates of the facility location in space,
or the distance of the location from any existing point. To simplify our exposition,
we will describe the main location models by using discrete location problems.
In order to do so, first define a vector of location variables y = (y1 , y2 , . . . , yn )
that indicate whether or not a facility will be located at the potential locations. In
addition to the location variables yj , we need to refer to customers i = 1, . . . , m
and, at least whenever multiple facilities are to be located, allocation variables X =
(xij ), where xij = 1, if customer i is served by facility j (or, more precisely, if the
customers at site i are served by a facility at site j). Furthermore, define w = (wi ) as
a vector of weights (which, for instance, denote the number of customers at customer
site i), D = (dij ) as the aligned of distances between customers i and potential
facility sites j, and f = (fj ) the fixed costs of a facility, if it were to be constructed
at site j. We also need g(X, y) as a set of constraints and S as the set of feasible
solutions. We can then easily describe some of the main types of location problems.
First, consider the simple plant location problem (SPLP) and its close cousin,
the capacitated plant location problem (CPLP). The main idea is to locate a num-
ber of facilities (its number is determined endogenously), so as to minimize the
sum of facility location and allocation, i.e., transportation, costs. Typically, the con-
straints will ensure that all customers receive service, and that customers can only be
served from existing facilities. The closely related capacitated version of the prob-
lem adds realism to the problem by associating a capacity to each facility. (Actually,
it would also be possible to consider a finite number of possible capacities to dif-
ferent versions of facilities, each of which would then have its own cost). While the
introduction of capacities appears to be a minor step in the sense that it just adds a
few constraints to the problem, the consequences are much more severe. While the
optimizer will automatically assign each customer to its closest facility (by virtue
of the minimization objective function), this may no longer be possible if capaci-
ties are introduced. This also gives rise to two different versions of the capacitated
problem: the single-source assumption, where all customers at one site are served
by the same facility, and the multiple source version, in which this is not the case. In
some applications, the single-source problem is more realistic, for instance in waste
collection, in which the collected garbage anywhere in one town is hauled to the
same landfill or transfer station.
The simple plant location problem and capacitated plant location problems can
be formulated as follows:
6 H. A. Eiselt et al.

SPLP/CPLP: Min wDX + fy

s.t. g(X, y) ∈ S
X ∈ {0, 1}n×n
y ∈ {0, 1}n .

This problem is actually a general version of a number of other well-known prob-

lems. The best-known such problem is the p-median problem (p-MP). Loosely
speaking, it requires us to locate a given number of p facilities, so that, given that
each customer is served by its closest facility, the total transportation costs are min-
imized. As the SPLP and CPLP before it, this model assumes that each service
requires a special trip. It can be formulated as

p − MP: Min wDX

s.t. ey = p
ĝ(X, y) ∈ S
X ∈ {0, 1}n×n
y ∈ {0, 1}n ,

where e = (1, 1, . . . , 1) is the vector of ones and ĝ refers to the appropriate con-
straints. As a matter of fact, the p-MP can be derived from the SPLP by setting f = 0
and introducing the additional constraint ey = p. In other words, while the number
of facilities in the SPLP/CPLP is endogenous, it is exogenous in the p-MP.
The anti p-median (also referred to as maxian) is a problem very similar to the
p-median, except that it attempts to maximize the average distance between facili-
ties and customers. It was designed to be used for hazardous or noxious facilities,
which customers would like to push away as much as possible from their own
location. However, formulating p-MP as an anti-median problem takes more than
simply replacing the “Min” by a “Max” function. The reason is that the “Min” func-
tion automatically allocated each customer to its closest facility. Replacing it by a
“Max” objective will result in customers automatically being allocated to their far-
thest facilities. However, what the planner is trying to do is to push the closest, not
the farthest, facility as far away from himself as it is the closest facility that is pol-
luting his location most. In order to formulate this, we will need additional O(n2 )
constraints that will guarantee the appropriate allocation.
Consider now the location set covering problem (LSCP). Its purpose is to mini-
mize the sum of facilities that are to be located while ensuring that each customer
has a facility within a predetermined service distance from himself. Many such cov-
ering problems are found in the context of the location of emergency facilities, such
as fire halls, hospitals, police stations, and others. The LSCP can be seen as a special
case of the SPLP by deleting the allocation variables X and setting the fixed setup
costs of all potential locations equal to each other, i.e., f = e The problem can then
be written as
1 Location Analysis in Practice 7

LSCP: Min ey
s.t. g (y) ∈ S
y ∈ {0, 1}n .

In many applications, covering each customer is not feasible, especially in situa-

tions, in which customers are very much spread out in the given space. In such case,
which gave rise to the max cover problem (MCP); see, e.g., Church and ReVelle
(1974). In this problem, the decision maker specifies p, the number of facilities to
be located and, given that each customer node has a weight associated with it, the
objective is then to maximize the total weight that can be covered within a given
distance from the facilities. The problem can be formulated as follows:
MCP: Max wz
s.t. ey = p
g(y, z) ∈ S
y, z ∈ {0, 1}n ,

where z is a vector of zero-one variables whose i-th component equals1, if customer

i is within a prescribed distance from its closest facility, and 0 if it is not.
There are other problems such as center problems (in which the objective is to
locate facilities, so that the longest customer—facility distance is as short as pos-
sible) and anti-center problems (in which the shortest customer—facility distance
is as long as possible), but very few of them have been applied in practice. One
of the problems associated with these problems is their unique focus on the worst
case. Potentially, objectives of this type can be applied as secondary objectives in
cases in which the worst case is to be avoided if at all possible, such as in hazmat
transportation or similar instances.
Many practical location problems have more than a single objective. In gen-
eral, we will distinguish between multiobjective (linear) optimization problems
(MOLP), which are typically mixed-integer optimization problems, and multi-
criteria decision-making problems (MCDM), which feature a small number of
decision alternatives, whose consequences are measured on a number of differ-
ent criteria. Among multiobjective optimization problems, vector optimization is
a prominent approach. First described by Zeleny (1974), vector optimization prob-
lems are designed to generate nondominated solutions, i.e., solutions which cannot
be improved upon on all objectives while remaining in the feasible set. Since the
generation of all nondominated solutions is not practical, techniques have been
devised to approximate the set of nondominated solutions. Cohon (1978) described
a number of techniques for that purpose. Prominent among them are the weight-
ing method and the constraint method. The weighting method associates positive
weights with each of the objectives (typically explained as measuring tradeoffs
between units of the objectives) and then constructs a composite objective as the
sum of the weighted individual objectives. The resulting single-objective problem
can then be solved by the pertinent methods. The constraint method chooses one
8 H. A. Eiselt et al.

objective to remain part of the model as an objective, while all other objectives are
transformed into constraints by using upper or lower limits on them. These limits
can, and usually are, parametrically modified in a sequence of sensitivity analy-
ses. While the constraint method will always find an extreme point of the original
feasible set, the weighting method may not.
Another approach to multiobjective optimization problems is goal programming.
Based on Charnes et al. (1955) and Charnes and Cooper (1961), Ignizio (1982)
was one of the main proponents of the approach. Its basic premise is as follows.
Constraints are rarely as rigid in real life as they are in a mathematical optimization
problem. For instance, a budget typically requires that the amount of money spent
does not exceed the amount available. Inserted in an optimization problem, such a
constraint is absolute and cannot be violated. In practice, though, it is, of course,
possible to borrow money, albeit at a cost (interest). This feature is captured in
goal programming by using so-called deviational variables dk+ and dk− for positive
and negative deviations from a set target value of tk , respectively. The variables dk+
and dk− are commonly interpreted as “overachievement” and “underachievement,”
respectively. As an illustration, consider a covering problem, in which it is desired
to cover each customer at least once. Each customer, who is not covered, causes a
penalty of, say, 5, while each customer, who is covered more than once receives a
penalty of 1 for each unit of excess coverage. The reason for the latter are “equity”
considerations, which, at least in public decision making, are often considered rather
important. As above in the MCP, define a zero-one variable yi , which is equal to 1,
if a customer at site i is covered and 0, if this is not the case. In addition, let the
vector ai• = (aij ) equal 1 in its j-th component, if customers at site i can be covered
by a facility at site j. Then ai• y equals the number of times customers at site i are
covered, and we can formulate the goal constraint

ai• y + dk− − dk+ = 1,

where the subscript of the deviational variables simply indicates that this is the k-th
goal. As usual, we assume that the deviational variables have to satisfy the non-
negativity constraints. Hence, if customer i is not covered, d − = 1 and d + = 0,
indicating that there is a coverage deficit of 1. On the other hand, if customer i is
covered twice, d + = 1 and d − = 0, indicating that there is an excess coverage
of 1. In case customer i is covered exactly once, both deviational variables will
assume the same value. Consider the penalties of 5 and 1 for overachievement and
underachievement of the target value, respectively, the contribution to the objective
function by this goal constraint is then

Min . . . + 5wi d − + 1wi d + . . .

This part of the objective will penalize each solution that does not cover site i and
its wi customers by 5 for each customer, while each customer generates one unit of
penalty for each time he is covered more than once. Furthermore, in case the target
of single coverage is satisfied, both deviational variables will not only be equal, but
1 Location Analysis in Practice 9

will equal 0, as any other solution with d + = d − will have a higher value of the
objective function in the minimization function.
Finally, goal programming problems come in a variety of versions. A popular,
albeit somewhat problematic, version features preemptive priorities, where a higher-
ranking goal is considered “infinitely more important” than a goal on the next lower
level, so that no finite tradeoff exists. Another important issue concerns the com-
mensurabilities, as deviational variables from different constraints are added in the
objective. Care must be taken that they are transformed to similar units.
The last multiobjective optimization technique reviewed here is fuzzy program-
ming. Fuzzy logic was pioneered by Bellman and Zadeh (1970). It deals with the
vagueness of an achievement, such as “this machine will cost a lot more than $
250,000.” In order to deal with the problem, a fuzzy membership function is set up
that assigns a degree of membership to each achievement or outcome, for which a
lower and an upper bound are specified by the decision maker. The value of this
membership function is a number between 0 and 1; it assumes a value of 1, if the
(maximization) objective achieves the upper bound or is even higher, it assumes a
value of 0, if the objective falls short of the lower bound, and is a value in between
for achievements between lower and upper bounds. A maximin function with the
membership function values as achievements is then the single objective, which can
then be solved with the pertinent methods.
Consider now multicriteria decision making problem, in which a decision maker
faces a “payoff” matrix C = (cij ), so that cij denotes the achievement of solution i on
criterion j. For simplicity, it is often assumed that all criteria are of the “maximiza-
tion” variety, i.e., higher values indicate better solutions. In the context of location
models, the decisions (rows) refer to the potential locations for the (single) facility.
An interesting reference that illustrates the use of multicriteria decision analysis to
location problems is Larichev and Olson (2001). The book includes models that site
facilities to dispose of hazmat, pipeline locations, and the selection of municipal
solid waste systems.
A simple generic method for multicriteria decision making problems is as fol-
lows. Suppose that the decision maker were able to specify weights λ for all criteria
, then aggregate achievements could be determined by calculating weighted aver-
age achievements for all decision, which then allows the decision maker to choose
the decision with the highest average achievement. Some refinements are easily
incorporated: a decision maker could delete decisions whose achievements on spe-
cific crucial criteria fall short of a prescribed threshold, or achievements beyond
some upper bound will no longer considered to be relevant as contributing towards
the aggregated average. Clearly, any version of this technique will have to be
followed by extensive sensitivity analyses.
The next class of techniques to deal with multicriteria decision problems are so-
called reference point methods. One technique in this class is TOPSIS, and it was
first described by Hwang and Yoon (1981). The idea of these techniques is to specify
a norm or target value and then measure how far the individual decisions are from
this absolute value. In addition to a metric to measure the distance, this requires the
specification of tradeoffs between criteria. Note the similarity of this approach and
10 H. A. Eiselt et al.

goal programming and its target values. A number of interesting approaches in this
class are described by Eiselt and Marianov (2014).
In contrast to reference point methods, which compare decisions and their out-
comes to exogenously given standards, data envelopment analysis (DEA) compares
decisions with each other and determines whether or not they are efficient. In order
to do so, analysts first have to group the input factors (such as costs, required
manpower, and others) as well as the output factors of the locations (employment,
pollution, and others) together. Then a set of simple linear programming problems
will determine whether or not a decision (we will consider one decision at a time) is
efficient compared to the other decisions. Note that this approach only demonstrates
relative efficiency, as a decision may be the best of the lot, while it may, in the grand
scheme of things, still be quite terrible and unacceptable.
Preference cones are an interesting and relatively easy to incorporate concept in
multicriteria decision making. The idea is that a decision i dominates a decision ,
if the aggregated outcome of the i-th decision with any set of nonnegative, finite
weights (as in the generic method described above) is no less than the aggregated
outcome of the· -th decision for the same weights. The advantage of this method
is that it requires no numerical input from the decision maker. However, it may
very well turn out that no dominances exist, so that no reduction of the system is
possible. We are not aware of any applications in location analysis that have applied
this concept.
Another approach is taken by outranking methods. In essence, they apply a weak
dominance concept and enable a persuasive visualization. Outranking methods were
first suggested by Benayoun et al. (1966), with Roy’s (1971) ELECTRE methods
(many other versions exist) providing the first workable concept. The basic tech-
nique requires that the decision maker specify weights for each of the criteria that
indicate its relative importance. The technique then compares each pair of decisions
and determines a concordance aligned and a discordance aligned on the basis of the
weights. More specifically, the concordance of one decision over another equals the
sum of all weights, in which the former decision is preferred over the latter. The dis-
cordance aligned can be constructed according to similar lines. A graph, in which
each node represents a decision then has an arc leading from a node i to a node  if
the concordance of decision i over decision  exceeds a prespecified threshold and
the discordance of that relation does not exceed a prespecified threshold.
Based on earlier work by Brans in the early 1980s, Brans and Vincke’s (1985)
PROMETHEE method takes a different route. On each criterion separately, the
method determines the differences of the achievements for each pair of decisions.
This difference is then translated into a preference by way of a preference function.
These preference functions could be step functions, piecewise linear, or nonlin-
ear functions, not unlike decay functions in covering models (see, e.g., Church
and Roberts 1983, or Berman and Krass 2002). Given the weights of the crite-
ria specified by the decision maker, these preferences are then aggregated. The
aligned of these aggregated preferences is then the basis, on which the graphs for
1 Location Analysis in Practice 11

the different versions of PROMETHEE are constructed. Karagiannidis and Mous-

siopoulos (1997) and Hokkanen and Salminen (1997) are among the authors who
have suggested to apply outranking methods to location problems.
Finally, techniques that allow inconsistent evaluations by decision makers have
become quite popular. Most prominent among them is the analytic hierarchy pro-
cess (AHP), or it successor, the analytic network process ANP). These techniques are
based on the seminal work by Saaty (1980). There is a large number of studies that
have suggested using the method for actual applications; see, e.g., Aragonés-Beltrán
et al. (2010) and Tavares et al. (2011). The technique requires decision makers to
specify their degree of preference between each pair of decisions on all criteria
separately. While reciprocity must be satisfied asserting that, if decision i is, say,
considered three times as good as decision , this necessitates that decision  is con-
sidered only 1/3 as good as decision i, transitivity does not (i.e., the decision maker
may determine that decision i is twice as good as decision  which, in turn, is three
times as good as decision r, while claiming that decision i is five times as good as
decision r on the same criterion). After this substantial input, the method will pro-
vide the decision maker with two outcomes: first, it will determine the degree of
consistency of the decision maker’s input (which means, that if it falls short of an
agreed-upon threshold, the decision maker will have to rethink his evaluations), send
secondly, the technique will provide a final ranking of the decisions under consid-
eration. There has been much discussion concerning the theoretical underpinnings
concerning the method, and some alternatives have been suggested, such as the geo-
metric mean method. Surveys of this and many other methods can be found in Olson
(1996) and in Eiselt and Sandblom (2004).

1.3 Location Applications in the Literature

This section will summarize applications of location analysis that have been
reported in the literature. While there are many contributions that use real data to
test their models and methods, not all of them are actual applications. For the pur-
pose of this paper, we would like to distinguish between realistic applications and
real applications. Realistic applications are those papers, whose main focus is on
the description of a new model or method, which is subsequently tested on data that
have been gleaned from real life. On the other hand, a real application’s main focus
is on the actual application. Clearly, the distinction between realistic and real appli-
cations is fuzzy, which may explain why we may have not included some papers
others may deem true applications.
In our discussion, it will be beneficial to distinguish between the firm’s view and
the (collective) customers’ view. More specifically, while in both cases it is the firm
(or government agency or whoever the planner may be) that locates the facility in
question, the two distinct angles of view emphasize who the ultimately beneficiary
will be. Models that espouse the firm’s view typically include a simple cost mini-
mization or profit maximization objective. This does, of course, not mean that the
12 H. A. Eiselt et al.

models themselves are necessarily simple: they may (and often do) include complex
functions that attempt to model customer behavior. Modeling the customers’ view
is much more contentious, as a single model will have to somehow express the col-
lective view of customers. As an example, consider the (re-) location of a hospital.
Here, a planner (typically not a representative of the customers themselves) will set
up objectives for the customers’ benefit. In doing so, he will usually consider aver-
age travel times of the patients to the hospital, and try to locate that facility, so that
as many potential patients are within a specific travel time from the hospital. Such
a planner may also wish to include in the model some measure of “equity,” which
is almost always ill-defined and, in order to find a quantifiable proxy, replaced by a
measure of equality.
We would like to point out that there is no clear relation between the firm’s and
the customers’ view on the one hand and the shipping and shopping models on
the other: in case of models that take the firm’s view, firms can locate trucking
terminals in the context of a delivery (i.e., shipping) model, or may locate retail
outlets (a shopping model). Similarly, in case of the customers’ view, planners may
locate ambulances (a shipping model, as dispatchers will decide which ambulance
to use and they will be responsible for the transport), or they may locate libraries (a
shopping model).
Before we go into any further detail, we would like to offer a few thoughts on the
scale of the decision-making process. Two major tools from the vast toolkit of loca-
tion planners are mathematical optimization problems and methods in multicriteria
decision making with multiobjective models occupying a middle ground between
the two. More often than not, simple, single-objective optimization models are used
on the macro scale in the first stage of a decision-making process. The solution may
identify a specific point, which decision makers will usually take as an indication of
the general area, in which the facility in question should be located. Subsequently,
the process zooms in on this area and at this point, more things are asked of the
location, which is accomplished by considering only a smaller number of available
locations, and additional criteria. This is what multicriteria approaches do.
Models that use the firm’s view may, in addition to the usual cost minimization
or profit maximization objective, include features that relate to customer behavior.
For instance, when locating a gas station, planners often look not at customer loca-
tion, but at typical trips made by customers, as it has been observed that customers
usually fill up their cars on their way to work or while they are on trips for other
purposes. Similar behavior has been observed for other small purchases and the use
of child care. Once such a criterion has been determined, planners may include a
flow capturing (or flow interception) feature in their model. For details on these
approaches, see, e.g., Hodgson (1981).
Another feature often included in models that use the firm’s view is routing. The
main reason is that the usual median models assume that each facility—customer
interaction requires a separate trip. Once this is no longer true, routing features may
be incorporated in the model. Clearly, this is a serious complicating factor: p-median
problems as well as even (for practical purposes) simple routing problems such as
1 Location Analysis in Practice 13

standard vehicle routing are NP-hard by themselves, so that in most cases a heuris-
tic method for the solution will be required. For a survey, readers are referred to
Nagy and Salhi (2007). In addition to the many existing heuristics and metaheuris-
tics, ranging from simulated annealing to tabu search, neighborhood search, genetic
algorithms, and many others (for a survey, see, e.g., Richards 2000) Nagy and Salhi
(2007) provide a number of heuristic methods, which are tailor-made for location-
routing problems. They distinguish between sequential methods (which solve the
location problem first, followed by a routing heuristic), iterative techniques (which
shuttle between location and routing components—a sequential method with a feed-
back loop), clustering-based methods (which subdivide the set of customers into
clusters, find a route within each cluster, and then determine the facility locations),
and hierarchical techniques.
The most frequently used feature included in models that use the customers’
point of view is accessibility and “equity.” A good example is the work by Burkey
et al. (2012), whose hospital model uses the average time for a customer to reach
a hospital as one criterion, the second is the coverage of as many potential patients
within a half-hour travel time (this can be considered providing equitable service),
and also the Gini index (Gini 1912 as well as Ceriani and Verme 2012), which uses
the Lorenz curve to measure the degree of equality of a solution.
In addition to these typical concerns included in individual models, there are also
industry-specific features that may be included. Typical examples include the access
to sources of water in case of fire stations, energy and materials recovery in case of
the location of facilities in solid waste management, penalties for overcrowding in
the planning of jails, proximity to highways, ports or airports in case distribution
centers or manufacturing companies are to be located, to sources of pure water for
beer breweries.
Table 1.1 below summarizes some of the applications of location analysis that
were published in the last 15 years. The table concentrates on the main features of
the model, the objectives, the solution approach, the type of facility, the country it
is located in, and the journal in which the piece was published. The exact references
are found at the end of this section. To save space in the table, we are using some
abbreviations: EJOR refers to the European Journal of Operational Research, C &
OR stands for Computers & Operations Research, JORS denotes the Journal of the
Operational Research Society and ITOR is the International Journal of Operational
Research.
Table 1.1 Summary of location applications 1990–2015
14

Year of Country of Facilities to be located Methodology Objectives Authors of study Journal

application application
1990 U.S. Telemarketing centers Integer programming Min communication, labor, & Spencer III et al. Interfaces
real estate costs
1990 Arizona Emergency medical Nonlinear integer Max the expected number of Goldberg et al. EJOR
vehicles programming model calls that can dealt with within 8
min
1990 U.S. Service terminal Analytic hierarchy Operating costs, availability of Hegde, Tadikamalla EJOR
process staff, ease of service,
convenience for customers, etc.
1992 Australia Maintenance deports Max cover Distance Rose et al. EJOR
(roads)
1993 Connecticut Closure of a fire station Spatial waiting line model Retain first due travel time, Swersey et al. Interfaces
service
1994 Kentucky Ambulances Expected covering, Max coverage within 10 min Repede, Bernardo EJOR
simulation
1994 Northern Ireland Ambulances p-median Distances/response time McAleer, Naqvi EJOR
1995 Turkey Brewery Integer programming Min transport costs & inventory Köksalan et al. EJOR
costs
1995 Connecticut Vehicle emissions Set covering, queuing, Stay within service constraints Swersey, Thakur Management
testing stations simulation (distance & waiting time) Science
1996 U.S. Multimodal rail plants Integer programming Min logistics costs Miller et al. Location
Science
1996 U.S. Blood collection & Integer programming with Min transportation costs Jacobs et al. Interfaces
distribution facilities different scenarios
1996 Israel Hospital Nonlinear & integer Min transportation costs, Mehrez, JORS
programming, analytic penalty for overcrowding, AHP Sinuany-Stern,
hierarchy process with population diversity, Arad-Geva,
H. A. Eiselt et al.

employment & service Binyamin

Table 1.1 (continued)
1

Year of Country of Facilities to be located Methodology Objectives Authors of study Journal

application application
1997 India Propane bottling plants Integer programing Min the sum of facility costs & Sankaran, Raghavan Interfaces
transportation costs
1998 Dubai Fire stations Set covering & Min cost, time, min service Badri et al. EJOR
11-criterion integer goal overlap, min water problems,
programming &c.
1998 Worldwide Army installations Dynamic integer Min allocation, closing, & other Dell Interfaces
programming costs
Location Analysis in Practice

1999 Portugal Landfills & transfer Integer programming Min sum of transportation costs Antunes Interfaces
stations & location setup costs
1999 Turkey Malt processing plants Integer programming Min discounted transportation Köksalan, Süral Interfaces
& facility opening costs
2000 U.S. Manufacturing/distributionVector optimization Max profit, min access time, Melachrinoudis, EJOR
plant max local incentives Min
2001 Hong Kong Satellite depots Integer programming Min installation & Cheung et al. Interfaces
transportation costs
2002 Georgia & Preventive health care Integer programming Maximize the number of Verter, Lapierre Annals of OR
Quebec facilities participants in preventive health
care
2002 Colorado Fire station Graphical aggregation Proximity, soft constraints Hewitt Interfaces
2002 Finland Landfill MCDM (stochastic Environmental, cost, property Lahdelma et al. EJOR
acceptability analysis) values, land use, etc.
2002 Worldwide Repair-parts warehouse Analytic network process Costs, taxation, risk, labor, Sarkis, Sundarraj EJOR
delivery time, etc.
2004 Illinois Automatic meter Integer programming Location set covering model Gavirneni et al. Interfaces
reading machines
2004 North Carolina Ambulances Integer programming Max expected covering Tavakoli, Lightner C & OR
15

(2004)
 16

Table 1.1 (continued)

Year of Country of Facilities to be located Methodology Objectives Authors of study Journal
application application
2005 Florida Disaster recovery Heuristics min avg distance, min max Dekle et al. Interfaces
centers distance, min number of centers
to be opened, max the prob that
a center is available after
disaster
2005 Chile Jails Integer programming Location costs, expansion costs, Marianov, Fresard JORS
penalties for overpopulation,
transportation for shuttling
inmates between jails
2006 Italy Organ transplant Integer programming Transportation costs plus Bruni et al. Health Care
centers penalty term for inequality Mg’mt Sci
between regions
2007 Portugal Disposal facilty for Integer programming p-median Gomes et al. Envir Mod &
treated wood Software
2007 Spain Incinerators Tabu search Location-routing Caballero et al. EJOR
2008 U.S. Military units to be Integer programming Min net present value of Dell et al. Interfaces
closed (various types of) costs
2008 Greece Different solid waste Lexico minimax Min greenhouse effect, min Erkut et al. EJOR
facilities disposal to landfill, max energy
recovery, max material
recovery, min total cost.
2008 Ohio Park & ride facilities Noninferior solutions & Max demand covered, min avg Farhan, Murray C & OR
tradeoff curves. travel time, max the number of
existing facilities that are used.
H. A. Eiselt et al.
 1

Table 1.1 (continued)

Year of Country of Facilities to be located Methodology Objectives Authors of study Journal
application application
2009 Florida Hydrogen refueling Integer programming with p-facility flow capturing model Kuby et al. Int J of
stations different scenarios Hydrogen
Energy
2009 British Columbia Location of clinical Dynamic integer Min linear convex combination Santibáñez et al. Interfaces
services programming of patient travel distances & &
deviation of solution from
existing solution
Location Analysis in Practice

2009 ./. Diopter strength Integer programming p-median, p-center, covering Francis Interfaces
2009 Texas Priests Integer programming Min traveling, max coverage Butler et al. Interfaces
2010 Greece Landfills for electronic ELECTRE III Cost, accessibility, public Achillas et al. Waste
waste acceptance Management
2011 Portugal Charging stations for Integer programming Max demand Frade et al. Transp. Res.
electric vehicles coverage—number of stations Record
located
2011 Worldwide CARE warehouses Integer programming Min average response time Duran et al. Interfaces
2011 Turkey Fire stations Multiperiod integer Max cover problem Çatay OR Insight
programming
2012 China Bank branches Integer programming Max cover Wang et al. Interfaces
2012 Spain Resort MCDM Borda’s Water quality, existing facilities, Crecente et al. Landscape
voting/consent method environment, etc. and Urban
Planning
17
 18

Table 1.1 (continued)

Year of Country of Facilities to be located Methodology Objectives Authors of study Journal
application application
2012 Chile Schools Integer programming Opening closure, operating, Araya et al. ITOR
transportation, & tuition costs
2013 Portugal Recycling plant Cost-benefit analysis, no Costs Coelho, deBrito J of Cleaner
optimization Production
2013 Turkey Fire stations Integer programming Covering problems Aktas et al. Interfaces
2013 Serbia Landfill Fuzzy analytic hierarchy Hydrogeological, Milosevic, Waste
process, Vikor meteorological, spatial, Naunovic Management
socio-political, legal, economic & Research
criteria
2013 Sierra Leone Landfill Analytic hierarchy Water, topography, slope, Gbanie et al. Applied
process, weighted linear access, cost of land, Geography
combination transportation costs
2014 Columbia Bioethanol processing Integer programming Max total benefits Duarte, Sarache, Energy
plant Costa
2014 Pakistan Warehouses Integer programming min (all sorts of) costs Brahimi, Khan 4OR
2014 Italy Schools Integer programming p-median Bruno, Genovese, Procedia–Soc
Piccolo, Sterle and Behav Sci
2014 Brazil Storage hubs for Integer programming min costs Milanez, deAzevedo Simp Brasil de
soybeans et al. Pesquisa Oper
H. A. Eiselt et al.
 1

Table 1.1 (continued)

Year of Country of Facilities to be located Methodology Objectives Authors of study Journal
application application
2014 South Korea Battery exchange Integer programming Modified p-median with max Ko, Shim Int J of
stations distance constraints Sustainable
Transporta-
tion
2015 Turkey Freight villages Integer programming Min transport cost Aksoy, Özyörük Appl. Math
Modeling
2015 China Watchtowers for forest Integer programming Covering problems Bao et al. Fire Safety
Location Analysis in Practice

fire monitoring Journal

19
20 H. A. Eiselt et al.

1.4 Summary & Outlook

This paper has described some of the main approaches to location analysis under
special consideration of different types of applications, followed by a survey of
descriptions of actual applications of location analysis. As expected, the main
tools are various types of integer programming and techniques from multicriteria
decision-making. The summary also reveals that location models are applied all
across the globe. In addition to the classics: finding optimal locations for landfills,
fire stations, ambulances, and bank branches, there are a number of new trends. It
is hardly surprising that many of these are fueled by technology: (mobile) vehicle
inspection stations (see the piece by Geetla et al. in this volume), charging stations
for electrical vehicles (or those that use hydrogen fuel), cell phone towers, wifi hot
spots, toll stations for electronic road pricing, and similar facilities. What is entirely
missing are applications for non-physical location problems, such as models that
position employees in positions in a firm (finding the best fit), locating brands that
locate close to existing customer tastes, and other “facilities.” The groundwork for
such location models has been made a long time ago (Niemi and Weisberg 1976 for
political models, and Eiselt and Marianov 2008, for employee positioning), but the
calibration of data is probably one of the key difficulties for these models. One of the
very positive approaches that exemplify “thinking outside of the box” is presented
by Swersey et al. (1993), who, in their quest to keep public services at an acceptable
level, while saving money, considered dual training of first responders as firefighters
and paramedics. This integration of services appears very promising.
Still, the survey also demonstrates that modeling is pretty much still at the “ad
hoc” level. In other words, there is not generally accepted “fire station location
model” that functions as transferable technology and is, with appropriate modifi-
cations for the local situation, usable in many different places. It would not free
modelers from the task of putting together a problem description for their specific
situation, but it would allow them to skip the basic steps, which are already done,
and skip to the fine tuning right away. This has not yet happened, but, as the saying
goes, hope springs eternal.

References

Aragonés-Beltrán P, Pastor-Ferrando JP, García-García F, Pascual-Agulló A (2010) An analytic

network process approach for siting a municipal solid waste plant in the metropolitan area of
Valencia (Spain). J Environ Manage 91:1071–1086
Başar A, Çatay B, Ünlüyurt T (2012) A taxonomy for emergency service station location problem.
Optim Lett 6:1147–1160
Bellman RE, Zadeh LA (1970) Decision making in a fuzzy environment. Manage Sci 17:141–164
Benayoun R, Roy B, Sussman B (1966) ELECTRE: Une méthode pour guider le choix en presence
de points de vue multiples. SEMA Note 49
Berman O, Krass D (2002) The generalized maximal covering location problem. Comput Oper
Res 29:563–581
1 Location Analysis in Practice 21

Bernasco W, Block R, Ruiter S (2013) Go where the money is: modeling street robbers’ location
choices. J Econ Geogr 13:119–143
Brans JP, Vincke Ph (1985) A preference ranking organization method. Manage Sci 31:647–656
Burkey ML, Bhadury J, Eiselt HA (2012) A location-based comparison of health care services in
four U.S. states with efficiency and equity. Socio-Econ Plan Sci 46:157–163
Carroll L (1880–5) A tangled tale. Knot two: eligible apartments. https://ebooks.adelaide.edu.au/
c/carroll/lewis/tangled/knot2.html. Accessed 29 March 2015
Ceriani L, Verme P (2012) The origins of the Gini index: extracts from Variabilità e Mutuabilità
(1912) by Corrado Gini. J Income Inequal 10(3):421–443
Charnes A, Cooper WW (1961) Management models and industrial applications of linear
programming. Wiley, New York
Charnes A, Cooper WW, Ferguson RO (1955) Optimal estimation of executive compensation by
linear programming. Manage Sci 1:138–151
Church RL, ReVelle CS (1974) The maximal covering location problem. Pap Reg Sci Assoc
32:101–118
Church R, Roberts K (1983) Generalized coverage models and public facility location. Pap Reg
Sci Assoc 53:117–35
Cohon J (1978) Multiobjective programming and language. Academic Press, New York. Reissued
in 2004 by Dover Publications, Mineola, NY
Daskin MS (2008) What you should know about location modeling. Nav Res Logist 55(4):283–294
Eiselt HA (1992) Location modeling in practice. Am J Math Manag Sci 12(1):3–18
Eiselt HA, Laporte G (1995) Objectives in location problems. In: Drezner Z (ed) Facility location:
a survey of applications and methods. Springer-Verlag, New York, pp 151–180
Eiselt HA, Marianov V (2008) Employee positioning and workload allocation. Comput Oper Res
35(2): 513–524
Eiselt HA, Marianov V (2014) Multicriteria decision making under uncertainty: a visual approach.
Int Trans Oper Res 21:525–540
Eiselt HA, Sandblom C-L (2004) Decision analysis, location models, and scheduling problems.
Springer-Verlag, Berlin
Gini C (1912) Variabilità e mutuabilità. Contributo allo studio delle distribuzioni e delle relazioni
statistiche. C. Cuppini, Bologna
Hodgson MJ (1981) The location of public facilities intermediate to the journey to work. Eur J
Oper Res 6:199–204
Hokkanen J, Salminen P (1997) Locating a waste treatment facility by multicriteria analysis. J
Multi-Criteria Decis Anal 6:175–184
Hwang CL, Yoon K (1981) Multiple attribute decision making: methods and applications.
Springer-Verlag, New York
Ignizio JP (1982) Linear programming in single- & multiple-objective systems. Prentice-Hall,
Englewood Cliffs
Karagiannidis A, Moussiopoulos N (1997) Application of ELECTRE III for the integrated
management of municipal solid wastes in the Greater Athens Area. Eur J Oper Res 97:439–449
Larichev OI, Olson DL (2001) Multiple criteria analysis in strategic siting problems. Kluwer
Academic Publishers, Boston
Lucas MT, Chhajed D (2004) Applications of location analysis in agriculture: a survey. J Oper Res
Soc 55(6):561–578
Nagy G, Salhi S (2007) Location-routing: issues, models, and methods. Eur J Oper Res
177(2):649–672
Niemi RG, Weisberg HF (eds) (1976) Controversies in American voting behavior. WH Freeman
and Co., San Francisco
Olson DL (1996) Decision aids for selection problems. Springer-Verlag, New York
ReVelle CS, Eiselt HA (2005) Location analysis: a synthesis and survey. Eur J Oper Res 165(1):
1–19
ReVelle C, Marks D, Liebman JC (1970) An analysis of private and public sector location models.
Manage Sci 16(11):692–707
22 H. A. Eiselt et al.

Richards E (2000) Heuristic algorithms. In: Eiselt HA, Sandblom C-L (eds) Integer programming
and network models. Springer-Verlag, Berlin, pp 229–258
Rosing KE, ReVelle CS (2000) Defendens imperium romanum: a classical problem in military
strategy. Am Math Mon 107(7):585–594
Roy B (1971) Problems and methods with multiple objective functions. Math Progr 1:280–283
Saaty TL (1980) The analytic hierarchy process. Mc Graw-Hill, New York
Swersey AJ, Goldring L, Geyer ED (1993) Improving fire department productivity: merging fire
and emergency medical units in new haven. Interfaces 23(1):109–129
Tavakoli A, Lightner C (2004) Implementing a mathematical model for locating EMS vehicles in
Fayetteville, NC. Comput Oper Res 31:1549–1563
Tavares G, Zsigraiová Z, Semiao V (2011) Multi-criteria GIS-based siting of an incineration plant
for municipal solid waste. Waste Manage (Oxford) 31:1960–1972
The Trucker’s Report (2015) The real cost of trucking in the United States. http://www.
thetruckersreport.com/infographics/cost-of-trucking/. Accessed 26 March 2015
Tryby ME, Boccelli DL, Uber JB Rossman LA (2002) Facility location model for booster
disinfection of water supply networks. J Water Resour Plann Manage 128(5):322–333
Weber A (1909) Über den standort der industrien. Mohr Verlag, Tübingen (Theory of the location
of industries). The University of Chicago Press, Chicago
Weiszfeld E (1937) Sur le point pour lequel la somme des distances de n points donnés est
minimum. Tohoku Math J 43:355–386
Wilamowsky Y, Epstein S, Dickman B (1995) How the oldest recorded multiple facility location
problem was solved. Locat Sci 3(1):55–60
Zeleny M (1974) Linear multiobjective programming. Lecture notes in economics and mathemat-
ical systems, vol 95. Springer-Verlag, Berlin
 Part I
Business
Chapter 2
Location Analysis in Banking: A New
Methodology and Application For a Turkish
Bank

Ayfer Başar, Özgür Kabak, Y. İlker Topçu and Burçin Bozkaya

2.1 Introduction

In recent years, technology has improved and distribution channels such as credit
cards, telephone-internet banking, Automated Teller Machines (ATMs) etc. have
become alternative opportunities for reaching services of banks. However, banks
generally gain new customers and develop customers’ loyalty at their branches.
Since branches are the indispensable contact points between the banks and their
customers, no bank can easily avoid opening new branches or reorganizing the lo-
cations of current ones. According to the current statistics of The Banks Association
of Turkey, the number of total bank branches has increased by 5.35 % from 10,450
to 11,009 in the last year. (The Banks Association of Turkey 2014). According to
the statistics of Retail Banker International, JPMorgan & Chase opened 89 new
branches in June 2013, which increased number of its branches from 5,608 to 5,697.
In the same period, BB&T increased the number of its branches from 1,775 to 1,851
(Retail Banker International 2013). This shows that, due to the effects of increase in
total population, population per bank branch and individual earnings, banks try to
increase the number of their branches by locating them in the right places. There-
fore, the branch location problem is a fundamental topic for banks in reaching their
strategic goals.

B. Bozkaya ()
Sabanci School of Management, Sabanci University, Tuzla, 34956 İstanbul, Turkey
e-mail: [email protected]
A. Başar · Ö. Kabak · Y. İlker Topçu
Industrial Engineering Department, Istanbul Technical University,
Maçka Campus, 34357 İstanbul, Turkey
e-mail: [email protected]
Ö. Kabak
e-mail: [email protected]
Y. İlker Topçu
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 25

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_2
26 A. Başar et al.

Although location problems for different types of places such as fire stations,
schools, post offices, etc. are frequently studied in the literature, the problem of
selecting the best places for new bank branches is much less often visited. Also,
it is seen that the bank branch location problem is studied mostly with a multi-
ple criteria decision making (MCDM) approach. The mathematical programming
models in the literature, on the other hand, concentrate on modeling and solving a
pre-defined problem. They do not provide any information on how to specify the
decision criteria nor do they consider the multiple criteria nature of the problem.
Zhao et al. (2004) state that the criteria that are taken into consideration while
opening a new branch may change according to the banks’ strategies. Banks can
focus on the places where they or their competitors have no branches, often areas
involving industrial and commercial activities, organized industrial zones, shopping
centers, collective housing areas, touristic regions, universities etc. in order to open
new branches. Economic development level, population and demographic charac-
teristics, distribution network, latent customers and their proximity to the potential
markets, physical location, credit and deposit potential of the candidate regions may
be important indicators for the branch location problem depending on the bank’s
strategies. Meanwhile, banks prefer to open various types of branches (individual,
commercial, corporate, private, entrepreneur, etc.) due to the different characteris-
tics of their target customers to minimize their costs and maximize their business
process efficiency. Thus, the importance and the effect of the characteristics of po-
tential locations can differ depending on the branch type. For example, in regions
where commercial and industrial activities are high and big investments are made,
commercial and corporate branches are generally located, while places where popu-
lation and/or level of collective housing is high are selected for individual banking.
Also, branches providing private banking services are mostly located where average
household income is high.
This chapter presents an integrated decision support methodology for bank
branch location problems in order to find relevant criteria and their importance
weights to be used in a new mathematical programming modeling framework. In
our methodology, we first select a set of criteria with the help of a detailed litera-
ture review and expert judgments. The criteria affecting the selection of optimum
location for bank branches may differ due to the banks’ strategies, customer profile
and characteristics of the region. According to a survey of the literature, transaction
volume of potential sites is one of the most important performance criteria of bank
branches (Manandhar and Tang 2002; Cook et al. 2004; Camanho and Dyson 2005;
Portela and Thanassoulis 2007). Therefore in this study we intend to discover the
criteria affecting transaction volume based on a detailed literature survey and ex-
perts’ opinions. In addition to the transaction volume, penalty of opening branches
close to each other, cost of opening a new and closing an existing one are also con-
sidered as the other main criteria depending on the experts’ judgments. Secondly, we
identify importance of these criteria for different types of bank branches based again
on expert opinions using pairwise comparisons. Furthermore, we develop a novel
mathematical programming model, which helps banks to plan the location of their
branches for different types of branches. Since the model is NP-hard and an optimal
2 Location Analysis in Banking 27

solution cannot be found for large problems, we propose a tabu search approach and
compare our results to those of CPLEX 12.2. We find that tabu search gives better
and faster solutions than CPLEX (based on limited run times). Finally, we apply the
proposed methodology in locating branches of a Turkish bank in Istanbul.
The remainder of this chapter is organized as follows: in Sect. 2.2, we present
the studies in the literature about general location planning and bank branch loca-
tion; in Sect. 2.3, we provide the details of our proposed methodology; in Sect. 2.4,
we describe its application in Turkey. Finally, we provide conclusions and further
research directions in Sect. 2.5.

2.2 Literature Review

Problems about finding the most appropriate place are important especially for
strategic planning in both public and private institutions. In the literature, there
are lots of studies involving location problems of different types such as for fire
brigades, emergency medical services, schools, post offices, police patrols, military
services, etc. Başar et al. (2012) classified the site selection problems in different
ways: Being deterministic or stochastic according to the model structure; being
static or dynamic, etc. Arabani and Farahani (2012) presented a review of stud-
ies in different location problems; essentially, coverage based and p-median models
are the most common and well-known location models in the literature.

2.2.1 Coverage Based and p-median Models

The very basic deterministic model is the set covering problem (SCP) described
by Toregas et al. (1971). The objective of SCP is to find the minimum number of
potential sites covering all demand points. SCP tries to cover all demand points
by at least one site without taking into account the population of demand points.
Hwang (2002) described the stochastic set covering model and Baron et al. (2009)
proposed SCP with stochastic demand. Murray et al. (2010) proposed a location
set covering model, assuming that each demand area can be covered not only by
one facility but also by two or more so that each facility covers a percentage of
demand. Another common coverage based model is maximal coverage location
problem (MCLP) proposed by Church and ReVelle (1974). The aim of MCLP is to
maximize the population or the number of the demand points covered by the poten-
tial sites. Daskin (1983) proposed the maximal expected covering location problem
with an equal busy probability assigned to all potential sites. Berman and Krass
(2002) proposed a generalized MCLP by using the distance to the nearest potential
site as a non-increasing function. Both SCP and MCLP depend on single coverage
of demand points. This means that, if a server is busy due to serving a demand
28 A. Başar et al.

point, other demand points covered by this server will no longer be covered. There-
fore, backup coverage models have been proposed to avoid this situation. Daskin
and Stern (1981) proposed the modified maximal covering location problem (MM-
CLP) with a second objective of maximizing the demand points covered multiple
times. Curtin et al. (2010) and Erdemir et al. (2010) have also studied backup cov-
erage problems. Since backup coverage problems are based on multiple coverage of
demand points in a single travel time or distance restriction, models using double
coverage with different time or distance standards were required in the literature.
Therefore, Gendreau et al. (2000) introduced the double standard problem (DSP)
maximizing the demand covered multiple times using two different travel time re-
strictions. The objective of DSP is to maximize the demand covered at least twice in
the shorter travel time limit. The constraints include a set covering requirement of
all demand points within the longer travel time and a given proportion of the popu-
lation to be covered within the shorter travel time limit. Başar et al. (2011) extended
the double coverage problem.
Another common type of model used for location problems is p-median studied
by Kuehn and Hamburger (1960), Hakimi (1964) and Manne (1964). In p-median
problems, the average total weighted access time is minimized while p service
facilities are located. ReVelle and Swain (1970) solved the p-median model by
using linear programming and branch-bound algorithm. Since p-median problems
are NP-Hard, they cannot be solved in polynomial time. Therefore, many heuris-
tics and exact methods have been proposed to solve the problem such as variable
neighborhood search (Hansen et al. 2009) and simulated annealing (Brimberg and
Drezner 2013), etc.
The p-median model and its extended versions in the literature all have an
assumption that customers make dedicated visits to the bank to and from their
home/work location. While this assumption may not hold for many types of loca-
tion problems, it holds to a reasonable extent in the case of locating bank branches.
Based on experience and observations, we find that many customers of certain type
or age (e.g. housewives or home-office working people, retirees, part-time work-
ers) who live in residential as well as commercial zones make special trips to a bank
branch. Such a trip is usually unrelated to any other tasks that may precede or follow
the bank visit, as the main purpose of the bank visit is to perform a task (other than
money withdrawal) that cannot be completed online or via other channels (phone,
ATM, mobile, etc.). There is, of course, still a possibility that such customers may
combine the visit with other stops nearby the branch location.

2.2.2 Bank Branch Location Literature

The literature on finding the best place for bank branches can be classified into three
groups: Statistical models, MCDM models and mathematical programming models.
2 Location Analysis in Banking 29

2.2.2.1 Statistical Models

In most studies, different statistical techniques and criteria are discussed for the
decision-making problem of finding the best places for bank branches. Claw-
son (1974) proposed stepwise linear regression to solve the bank branch location
problem, setting realistic performance standards for different potential sites, and
specifying remedial actions for poorly performing branches from a sample of 26
branches. Boufounou (1995) used regression analysis and conducted some statis-
tical tests to determine the importance and significance level of related criteria
for a Greek bank to evaluate existing branches’ performance, assess performance
of potential sites and place branches in more efficient locations. Ravallion and
Wodon (2000) also used regression analysis to explain the relationship between
demographic, economic, employment indicators and the location of Bangladesh’s
Grameen Bank’s branches. They found that the bank branch location problem is
mainly affected by the economic indicators. On the other hand, industrial charac-
teristics and banking products served to these industries by the branches are also
important factors. They discussed that the bank branch location decision was influ-
enced by the potential gain from switching to more profitable non-farm activities in
rural areas. All these factors and their relative influences on the location decision
process, resulting from the regression modeling approach, could allow an analyst to
pin down the most relevant set of criteria, which can subsequently be used to locate
new facilities in future analyses.

2.2.2.2 MCDM Models

Since they consider multiple criteria, MCDM models are very common for bank
branch location problems. As stated by Carlsson and Fuller (1996), MCDM is a
fundamental field of study dealing with decision making including multiple crite-
ria, attributes and objectives. The main objective of the MCDM methods is to help
decision makers solve complicated decision making problems systematically. There
are four major groups of methods in MCDM: (i) the outranking, (ii) the value and
utility theory based, (iii) the multiple objective programming, (iv) group decision
and negotiation theory based methods. As a consequence of the developments in
fuzzy set theory, fuzzy MDCM are also commonly discussed in the literature. Chen
and Hwang (1993) made distinctions between fuzzy ranking methods and multiple
attribute decision making methods all included in the groups (i)-(iv).
Fuzzy set theory is used very frequently for the bank branch location problem,
because of the uncertainties in the comparisons of the criteria and the alternatives.
For instance, Min (1989) proposed fuzzy goal programming method to find the most
appropriate places for commercial bank branches in Ohio. In order to select one
among six cities in South Eastern Anatolia for opening a new branch, Cinar (2009)
used fuzzy analytical hierarchy process (AHP) to find the importance of the related
criteria and TOPSIS to rank the cities. Rahgan and Mirzazadeh (2012) used fuzzy
30 A. Başar et al.

AHP to specify the importance of criteria and evidential reasoning to order the al-
ternatives. As a result of the improvements in information technologies, Morrison
and O’Brien (2001) used geographical information systems (GIS) and a spatial in-
teraction model (Huff 1963) in four stages: The probability of customers visiting a
branch is estimated, the expected distribution of customers is determined, the ex-
pected number of transactions at a given branch is computed and the impact of
removing one or more branches from the network is analyzed.

2.2.2.3 Mathematical Programming Models

Although the mathematical programming literature related to facility location prob-

lems is rich, studies specific to bank branch location problems are scarce. Min
and Melachrinoudis (2001) proposed a three-level hierarchical location-allocation
model for bank branches by considering risk and uncertainty, where a chance con-
strained goal programming model was developed using forecasting techniques. The
methodology was applied in a state in the USA. Miliotis et al. (2002) introduced a
two-step methodology in which firstly the minimum number of branches to meet the
coverage requirement was found and then locations of branches to maximize cov-
erage was determined. They used GIS and applied their methodology to a bank in
Greece. Wang et al. (2002) considered the problem of locating ATMs, internet mir-
ror sites, or other immobile (permanent location) service systems of limited service
capacity. They model these service facilities as simple M/M/1 queuing systems and
solve the model using three different heuristic approaches. Wang et al. (2003) de-
veloped a mathematical model for the branch location in Amherst, New York. They
improved greedy interchange, tabu search, and Lagrangian Relaxation to apply this
NP-Hard model for the branch location in New York by using 270 generated prob-
lems. Unlike the p-median model, the model consists of a budget constraint related
to opening and closing branches. Zhang and Rushton (2008) proposed a model max-
imizing total benefit with the budget constraints related to opening branches and the
waiting time of the customers. They used a genetic algorithm to solve their pro-
posed problem. Alexandris and Giannikos (2010) proposed a new model for the
MCLP and illustrated the applicability of the proposed model by means of a case
study concerning the location of bank branches. The aim of the model is to maxi-
mize the total population covered by the selected branches. Xia (2010) formulated a
mathematical programming model considering operations and rental costs, demand
and distance between branches by proposing a hybrid nested partitioning algorithm
to solve the large-scale problem.
As a result of the literature review, there is no integrated methodology that
combines problem structuring phase (i.e., as in MCDM models) and mathematical
programming models for the bank branch location problem. Therefore this study
proposes a novel integrated methodology for not only identifying the criteria and
their importance but also finding the exact locations of bank branches through a
mathematical programming model.
2 Location Analysis in Banking 31

2.3 Proposed Methodology

The basic aim of this study is to solve the bank branch location problem using an
integrated methodology that combines the problem structuring phase and mathe-
matical programming model. To this aim, firstly related criteria are listed based on
a literature survey and expert opinions. Secondly, the importance of each criterion
is identified depending on the pairwise comparison method conducted with experts
working in the banking sector. Furthermore, a new mathematical model is developed
to determine the specific locations of branches. Since the model is NP-Hard, a tabu
search approach is developed to find (near-) optimal solutions for the big problems
and the results are compared to those of the CPLEX 12.2. The methodology is ap-
plied to the real life practice of a Turkish national bank branches’ location problem
in Istanbul.

2.3.1 Identifying the Criteria

Banks prefer providing service with different types of branches having different
customers such as individual, commercial, private, corporate, entrepreneur etc. to
decrease their costs and increase the efficiency of business processes. Therefore,
they need to open different types of branches. According to expert opinion, finding
the best locations especially for individual, entrepreneur, corporate and commercial
bank branches is an important topic for banks.
Individual branches can serve busy and standard transactions to every kind of
customer. Entrepreneur branches are generally located in the industrial zones, trad-
ing estates and agricultural spheres. Corporate branches make high contribution to
the gross national product with their predominantly agricultural customers and are
located in the regions where there is no need to open a commercial branch. Com-
mercial branches’ customers generally deal with industrial, business and agricultural
activities and also make the highest contribution to the gross national product. En-
trepreneur, corporate and commercial branches generally do not offer busy and
standard service to their customers. On the other hand, customer segmentation rules
of the banks may affect customer profile and the location regions’ specification for
these branches.
Firstly, we analyze the criteria used in the literature to identify the criteria for
bank branch location problem. We see that there are lots of criteria taken into
consideration while opening or closing a bank branch. Clawson (1974) found that
population of middle age, average per capita domestic income and home ownership
rate are the most important factors. Meidan (1983) pointed out that commercial po-
tential, population, state of employment in the candidate regions and the location
of competitor bank branches were fundamental aspects to be chosen. Boufounou
(1995) indicated that total population in terms of gender and age, average house-
hold size, population growth rate, domestic per capita income, number of firms
32 A. Başar et al.

in terms of sector and location of competitor banks were important criteria for
bank branch performance in Greece. Min (1989) proposed demographic charac-
teristics, ease of access and business operations of candidate regions as important
criteria for bank branch location problems in Ohio. Pastor (1994), Kaufman and
Mote (1994) stated that economic and demographic variables had an important ef-
fect on location of bank branches in Chicago. Ravallion and Wodon (2000) showed
that demographic characteristics, employment and economic indicators of the candi-
date locations were the most important factors for bank branch location problems in
Bangladesh. Abbasi (2003) developed a decision support system for bank branch
location problems in Jordan and used population, income level, cultural charac-
teristics, number of firms, capital, growth potential and number of competitors as
variables in their proposed model. Zhao et al. (2004) emphasized that financial
indicators, demographic characteristics, customer segmentation, position of com-
petitor banks and means of access of the potential points had to be considered for
bank branch location. They indicated the main and sub criteria for location of bank
branches in Sydney as follows:
• Service variables (number of people older than 18, population growth rate, aver-
age annual household income, employee rate, number of small firms, number of
competitor branches),
• Location variables (number of small firms, competitor banks, and employed pop-
ulation within 200 m distance to the potential location; number of bank branches
within the context of application; shopping centers within 500 m distance to the
potential location; transportation index; closeness to the public transportation).
Cinar (2009) identified 5 main criteria, including 21 sub criteria, affecting a bank’s
mission and strategy as follows:
• Demographic (total population, annual population growth and urbanization rate),
• Socioeconomic (average household size, literacy rate, rate of higher education
graduate, average per capita domestic income, employee and employer rate),
• Banking indicators (number of banks and branches, deposit and credits per bank
branch-person),
• Employment rate in accordance with the sectors (agriculture, building, manufac-
turing, and service sectors),
• Trade potential (number of firms and industrial zones).
Rahgan and Mirzazadeh (2012) introduced a hierarchical model for bank branch
location problems using cost, demography, banking, geographical conditions and
accessibility as main criteria. Criteria used in the literature for bank branch location
problems are given in Table 2.1.
As one of the aims of our study is to find the related criteria for the banking
sector, expert opinion is also considered. By the help of one-on-one interviews,
necessary information is gathered through an unstructured questionnaire. Experts
are asked to indicate their opinion on both how banks decide where to locate new
branches in current conditions and the effective strategies they employ. Interviews
2 Location Analysis in Banking 33

Table 2.1 Criteria used in the literature

Criteria Paper
Population Clawson 1974; Olsen and Lord 1979; Doyle
et al. 1981; Meidan 1983; Boufounou 1995;
Abbasi 2003; Zhao et al. 2004; Cinar 2009
Average per capita domestic income Clawson 1974; Boufounou 1995; Cinar 2009
Home ownership rate Clawson 1974; Olsen and Lord 1979
Commercial potential Meidan 1983; Cinar 2009
Location of competitor bank branches Doyle et al. 1981; Meidan 1983; Boufounou 1995,
Abbasi 2003; Zhao et al. 2004; Cinar 2009
State of employment Olsen and Lord 1979; Doyle et al. 1981;
Meidan 1983; Zhao et al. 2004; Cinar 2009
Demographic characteristics Olsen and Lord 1979; Doyle et al. 1981; Min 1989;
Kaufman and Mote 1994; Ravallion and Wodon 2000;
Rahgan and Mirzazadeh 2012
Ease of access Doyle et al. 1981; Min 1989; Zhao et al. 2004;
Rahgan and Mirzazadeh 2012
Business operations Min 1989
Average household size Boufounou 1995; Cinar 2009
Population growth rate Boufounou 1995; Abbasi 2003; Zhao et al. 2004;
Cinar 2009
Number of firms Doyle et al. 1981; Boufounou 1995; Abbasi 2003;
Zhao et al. 2004; Cinar 2009
Income level Abbasi 2003; Zhao et al. 2004
Total deposit Abbasi 2003
Cultural characteristics Abbasi 2003
Literacy rate Cinar 2009
Deposit-credits per bank branch-person Cinar 2009

are conducted with five experts who work in the Turkish banking and information
technology sector as senior executives.
All the experts indicate that average transaction volume is the most effective
factor for Turkish bank branch profit. This result is also supported by the literature
(Jablonsky et al. 2004; Camanho and Dyson 2005; Portela and Thanassoulis 2007).
There are different parameters affecting banks’ daily transaction volume such as
payment of wages, education fees, bills etc., banking transactions of which have to
be done at a specified date. Similarly, annual transaction volume of the sites can be
misleading due to the long periods involved. Thus, based on the experts’ judgment,
average monthly transaction volume of potential sites is identified as one of the main
criteria for deciding the location of bank branches in Turkey. Transaction volume is
defined so as to include all types of banking operations such as credits, deposits,
bonds, cheques, money transfers etc.
34 A. Başar et al.

Table 2.2 Criteria used for the estimation of average monthly transaction volume
Criteria Proposed sub criteria Related criteria in the literature
Number of potential Daytime population Population (Clawson 1974;
customers Olsen and Lord 1979; Doyle
et al. 1981; Meidan 1983;
Boufounou 1995; Abbasi 2003;
Zhao et al. 2004; Cinar and
Ahiska 2010)
Socioeconomic situation Education level, number of Literacy rate and education level
houses and summer houses (Cinar 2009)
Social potential Number of education and Social potential (Abbasi 2003)
entertainment places, parks,
hospitals
Commercial potential Number of private institutions, Commercial potential
shopping centers, car parks, (Meidan 1983; Cinar 2009)
financial institutions, restaurants
and car services
Growth potential Population growth rate Population growth rate
(Boufounou 1995; Abbasi 2003;
Zhao et al. 2004; Cinar 2009)
Ease of access Ease of travel, accessibility and Ease of access (Doyle et al. 1981;
proximity to public transport Min 1989; Zhao et al. 2004;
Rahgan and Mirzazadeh 2012)
Competition Number of competitors’ bank Number of banks and branches,
branches position of competitor banks
(Meidan 1983; Boufounou 1995;
Abbasi 2003; Zhao et al. 2004;
Cinar 2009)
Financial situation Average household income Home ownership rate
(Clawson 1974); income level
(Abbasi 2003; Zhao et al. 2004);
employee and employer rate
(Meidan 1983; Zhao et al. 2004;
Cinar 2009)

Since forecasting the average monthly transaction volume accurately is almost

impossible, different criteria and sub criteria are identified depending on both the
literature survey and expert opinion. These are listed in Table 2.2 along with the
related sub-criteria. According to the literature review and experts’ judgment, “pop-
ulation growth rate,” “number of competitor bank branches” and “average household
income” are determined as measures of growth potential, competition and financial
situation, respectively.
Since the selected criteria cover them, a number of important criteria mentioned
in the literature survey are not directly used in this study. For instance, “employee-
employer rate” mentioned in Zhao et al. (2004) and “average household income”
represented by Cinar (2009). Credit/deposit per bank branch/person are the banks’
2 Location Analysis in Banking 35

financial data, therefore it is not possible to directly use these data because of se-
curity and confidentiality rules of banks; thus, “average household income” is used
instead.
Since data related to daytime population are not available, number of potential
customers for all candidate places is estimated by different criteria affecting day-
time crowdedness of the regions which are total population, number of work and
education places, financial institutions, and hospitals. Similarly, proximity to public
transport is not readily available; hence total distance to the neighboring districts,
calculated on the transportation network, is used as a cost measure for ease of ac-
cess. Also, after analyzing correlation between criteria values among 763 districts in
Istanbul; number of houses, parks and restaurants are eliminated since there exists
high correlation between these criteria and total population, number of car parks and
financial institutions, respectively.
Moreover, according to the experts’ judgments, the distance between all the
potential points is another main criterion, especially to avoid opening multiple
branches close to each other. Thus, we determine that the opening of new branches
in close proximity should be penalized. Moreover, experts indicate that costs of
opening a new branch as well as closing an existing one are other important crite-
ria for bank branch location problems. Finally, the selected criteria of the proposed
model are presented in Fig. 2.1.

2.3.2 Specifying the Importance of the Criteria

As analyzed in Sect. 2.3.1, bank branch location problems can be modeled as a hi-
erarchy containing the decision goal, which is selecting the best place, and related
criteria affected by different factors. Since it is difficult to use an objective weight-
ing method such as entropy method, regression analysis and the CRITIC (Criteria
Importance Through Intercriteria Correlation) method, because of data confiden-
tiality in the banking sector, a subjective weighting technique is required to find
the importance of the criteria. Miller (1956) argues that the brain of an average
human can simultaneously process, differentiate and deal with at most seven fac-
tors, with this limit decreasing to five for some, while increasing to nine for others,
and people are generally more consistent when they do pairwise comparisons than
when they just assign relative importance to the criteria. Thus when different ap-
proaches such as rating, point allocation, ratio, raking, pairwise comparison and
tradeoff are analyzed, pairwise comparison is found to be the most efficient ap-
proach since it focuses on only two alternatives at each time (Malczewski 1999).
Also since decision makers focus on finding the relative importance of two criteria
without being affected by external factors and generally have deep knowledge with
which to compare the criteria, pairwise comparison generally gives more accurate
results compared to the other weighting methods (Badri 2001). Pairwise compari-
son is also the main approach of AHP which is one of the most common techniques
in decision making literature as originally proposed by Saaty (1980). The AHP has
 36

SelecƟng the best

locaƟon for bank
branches

Distance between Cost of opening a Cost of closing a

TransacƟon volume
branches branch branch

Number of
Socioeconomic Commercial Growth Ease of Financial
potenƟal Social potential CompeƟƟon
status potenƟal potenƟal access situaƟon
customers

Distance to the Number of Average

PopulaƟon
Total Private Financial EducaƟon EducaƟon Summer EducaƟon Entertainment Private Shopping Car Financial Car neighbour compeƟtors’ household
Hospitals Hospitals growth
populaƟon insƟtuƟons insƟtuƟons places level houses places places insƟtuƟons centers parks insƟtuƟons services potenƟal bank income
rate
branches branches

Fig. 2.1 Selected criteria of the proposed model

A. Başar et al.
2 Location Analysis in Banking 37

Goal

Criteria

Sub
Criteria

Fig. 2.2 A typical decision hierarchy

Table 2.3 Pairwise comparison scale used in the study

Intensity of importance Definition
1 Equal importance
3 Moderate importance of one over another
5 Strong or essential importance
7 Very strong or demonstrated importance
9 Extreme importance
2, 4, 6, 8 Intermediate values

been applied in a variety of studies including bank branch location problems (e.g.
Chen et al. 2001; Sato 2004; Ishizaka and Lusti 2004; Ahsan and Bartlema 2004;
Tzeng et al. 2002; Aras et al. 2004; Wu et al. 2007; Fernandez and Ruiz 2009).
Thus, pairwise comparison is determined as an effective method for establishing the
importance of main and sub criteria by making a series of expert judgments. The
method is applied by following these steps:
Step 1 The decision hierarchy, including a goal, criteria and sub criteria (if exist), is
prepared to make judgments on the elements of the hierarchy in pairs with respect
to the parent element. A typical decision hierarchy is illustrated in Fig. 2.2.
Step 2 A pairwise comparison matrix A is built to compute the importance of dif-
ferent criteria and sub criteria. Each entry (aij ) of the matrix A shows the importance
of the criterion i compared to the criterion j.
The relative weights of the criteria are determined according to the scale of 1–
9 corresponding to the linguistic comparisons, which are described in Table 2.3
(Saaty 1990).
Step 3 After building matrix A, relative importance weights of the criteria w = (w1 ,
w2 ,. . .,wn ) are found by solving Aw = nw, which is called the principal right
eigenvector of A (Saaty 1990).
38 A. Başar et al.

2.3.3 The Proposed Mathematical Programming Model

As stated in the literature review, there are only a few mathematical models proposed
for bank branch location problem. Experts indicate that a maximization model cal-
culated by the help of criteria in parallel with banks’ mission and strategy is more
useful than a model using to minimize distance or time between demand points
and branches. The number of each type of branch to open is definite due to the
budget constraint, which is generally planned annually. Moreover, opening a new
branch and closing an existing one both have associated costs and some of existing
branches cannot be closed due to company strategy, according to experts’ opinion.
Thus, a new mathematical programming model to locate bank branches in Turkey
using criteria defined in Sect. 2.3.1. is defined as follows. (Table 2.4)
The average monthly transaction volume hij can be forecasted according to
the criteria identified in Sect. 2.3.1 (See Fig. 2.1 and Table 2.2), and the weights
specified in Sect. 2.3.2 (Table 2.5) as follows:

hij = w1j NPCi + w2j SSi + w3j SPi + w4j CMPi + w5j
GPi + w6j EAi w7j Ci + w8j FSi (2.1)

where wyj (y = 1,. . ., 8, j ∈ J) is the weight of criterion y for location type j as defined
in Table 2.5.
In Eq. 1, NPCi , SSi , SPi , and CMPi have sub-criteria according to the criteria
hierarchy given in Fig. 2.1. Therefore these are calculated as follows (see Table 2.11
for the notation):

NPCi = w9j TPi + w10j PI i + w11j FI i + w12j Hi + w13j EP (2.2)

SSi = w14j ELi + w15j NSi , (2.3)
SPi = w16j EPi + w17j Hi + w18j ENPi , (2.4)
CMPi = w19j PI i + w20j SCi + w21j CPi + w22j FI i + w23j CSi (2.5)

where wyj (y = 9,. . ., 23, j ∈ J) is the weight of criterion y for location type j as
defined in Tables 2.6, 2.8 and 2.9.
      
Z (max) c1j hij xij − c2j λimj − c3j xij − c4j (1 − xij )
i∈I j ∈J i∈I m∈I,i=m j ∈J i ∈K
/ j ∈J j ∈J i∈K

s.t. xij = Pj ∀j ∈ J (2.6)
i∈I

xlj = 1 ∀  ∈ L, j ∈ J (2.7)
λimj ≥ kim (xij + xmj − 1) ∀i, m ∈ I, i  = m, j ∈ J (2.8)
xij ∈ {0, 1} ∀i ∈ I, j ∈ J (2.9)
λimj ≥ 0 ∀i, m ∈ I, j ∈ J (2.10)
2 Location Analysis in Banking 39

Table 2.4 Notation used in the model

Notation Definition
Sets
I Set of potential sites to open bank branches
J Set of types of branches (individual, entrepreneur, corporate and commercial)
K Set of sites of current branches
L Set of sites of the branches that cannot be closed by strategy
Indices
i Potential site to open bank branch
j Type of branch (individual, entrepreneur, corporate and commercial)
Parameters
hij Average monthly transaction volume of the potential site i, when branch type j is
located at site i
NPCi Number of potential customers at the potential site i
SSi Socio-economic situation of the potential site i
SPi Social potential of the potential site i
CMPi Commercial potential of the potential site i
GPi Growth potential of the potential site i
EAi Ease of access in the potential site i
Ci Competition of the potential site i
FSi Financial situation of the potential site i
TPi Total population of the potential site i
PI i Private institutions of the potential site i
FI i Institutions of the potential site i
Hi Hospitals of the potential site i
ELi Education level of the potential site i
NSi Number of summer houses of the potential site i
EPi Education places of the potential site i
ENPi Entertainment places of the potential site i
SCi Shopping centers of the potential site i
CPi Car parks of the potential site i
CSi Car services of the potential site i
dim Distance between the potential sites of i and m (in meters)
kim Parameter depending on distance between i and m (takes value between 0 and 1)
Pj Maximum number of branches of different types
S Threshold (in meters) to consider the penalty for opening branches near each other
c1j Benefit coefficient (as percentage) of transaction volume for opening branch type j
c2j Penalty coefficient (as percentage) for opening branches type j near each other
c3j Coefficient (as percentage) for opening a branch type j
c4j Penalty coefficient (as percentage) for closing a branch type j
40 A. Başar et al.

Table 2.4 (continued)

Notation Definition
Decision Variables
λimj Penalty of opening j type branches in potential nearby sites i and m
xij Binary variable which is 1 if a type j branch is located at site i, and 0 otherwise

Fig. 2.3 The relationship be- k im

tween distance and parameter
based on the threshold value
1

d im
S

The objective function maximizes total net profit as the difference between total
benefits of expected average monthly transaction volume minus the penalty of open-
ing branches nearby to each other and cost of opening new and closing existing
branches. Constraint (2.6) shows how many branches are to be opened for each type
given the budget. Constraint (2.7) avoids closing branches that cannot be closed due
to company strategy (in our case, branches opened in the last 3 years, according
to the expert opinion). Constraint (2.8) indicates penalty of opening branches near
each other according to the threshold values and distances. λimj is the penalty of
opening type j branches in potential nearby sites i and m, taking a value between 0
and 1. It is designed to be 1 if the distance between branches i and m is zero, and
0 if the distance is equal to or more than a threshold value S, linearly decreasing in
between. Obviously, if a branch is not opened in either site i or site m, λimj will be
zero. To guarantee λimj gets values as designed, a parameter namely kim is defined.
kim is formulated based on distance dim between potential sites i and m as given in
Fig. 2.3. λimj is equal to kim if branches are opened in both sites i and m, and 0
otherwise.
Constraints (2.9) and (2.10) show that decision variables of bank branch loca-
tion are binary and penalties of opening branches near each other are nonnegative,
respectively. The model permits both opening new and closing existent branches at
the same time. Also, while multiple branches of the same type cannot be opened in
the same locations, different types of branches are allowed to be opened in the same
candidate location.
In order to solve any bank branch location problem using the proposed model,
the related data components have to be compatible with one another. Therefore, we
convert the data on education level of people living around the potential site to nu-
merical data using a four-point scale, one for the people who are unable to read and
write, two for primary school, three for high school, four for university graduates.
Furthermore, we normalize the values for all the criteria. To this aim, we first spec-
ify S based on the experts’ judgment and the geographical structure of the potential
sites. Thus, after collecting the distances dim , using the transportation network and
2 Location Analysis in Banking 41

the help of a GIS, between the best-served potential sites, we calculate the kim val-
ues. Since dim is based on the transportation network instead of the straight-line
distance between i and m, dim may differ from dmi for all i  = m. Also, for the dis-
tricts having an existing branch, the place where the branch is located is considered
as the reference point. Thus, the location of the branches that will continue to serve
will remain unchanged. In order to provide concurrency, these data have to be nor-
malized as benefit attributes, since the bigger kim values are, the more penalty will
be imposed to the potential site. Furthermore, the average monthly transaction vol-
ume of the potential site is forecasted in accordance with the hierarchy defined in
Sect. 2.3.1. This means that, after normalizing all the data and using the importance
weights found by pairwise comparisons, the value of the related criteria is specified
beginning from the bottommost and moving up through the hierarchy. At this point,
the direction of the effect must also be considered. Therefore, while the “number
of competitors’ bank branches” and “distance to the neighbor potential branches”
have to be normalized as a cost attribute, all the other criteria are evaluated as ben-
efit attributes. Thus, each term in the objective function can take a value between
0 and 1.
It can be proven that proposed mathematical model is NP-hard by considering
a special case of the problem where L = Ø and λimj = 0 such that there are no
branches that cannot be closed due to strategy and the distance between the all po-
tential locations is bigger than S. In this case, the problem reduces to MCLM, which
is known to be NP-hard (Marianov and ReVelle 1995; Berman and Krass 2002).
Therefore, optimal solution of the problem cannot be easily found for large problem
sizes.

2.3.4 Proposed Tabu Search Algorithm

Banks generally plan to open branches at the city or metropolitan level and prefer
big cities to locate new branches. Therefore, it is not possible to solve the location
problem of bank branches in large cities having lots of candidate points. Accord-
ingly, efficient heuristic and meta heuristic approaches are required. Youssef et al.
(2001) and Arostegui et al. (2006) experimentally pointed out that tabu search gives
better results than genetic algorithms or simulated annealing for location problems.
Also, Başar et al. (2011) discussed that tabu search is a very efficient approach for
a mathematical model similar to the proposed one in our context. Therefore, we
propose a tabu search approach to solve the mathematical model in Sect. 2.3.3.
Tabu search is a metaheuristic originally developed by Glover (1977), which uses
an initial solution, searches the neighbors of the incumbent solution at each iteration
to reach better solutions and improve the best solution obtained. It keeps a tabu list
to avoid the repetition of the same solutions, by forbidding a certain set of moves.
In principle, a tabu move can still be accepted if it gives a better best known solu-
tion. Due to its efficiency and effectiveness, tabu search is commonly used in the
42 A. Başar et al.

literature for different optimization problems (Osman and Kelly 1996; Pardalos and
Resende 2002; Ribeiro and Hansen 2002).

2.3.4.1 Initialization

We use three initialization methods in our implementation of tabu search to ob-

serve their effects on the overall solution quality: (i) random method, (ii) linear
programming, relaxation-based method and (iii) criteria-based method. Since both
feasibility conditions (constraints 2.6 and 2.7) are satisfied by opening Pj branches
for each type and not closing the branches in the set of L, all initialization methods
are feasible.
In the random method, after satisfying the feasibility conditions, locations of re-
maining branches for each type are determined randomly. The initial solution thus
found is expected to be of low quality since the method is not rule-based. In the lin-
ear programming relaxation method, xij decision variables, which cause the model
to be NP-hard are relaxed and the resulting linear program is solved. The candidate
point as many as Pj for each branch type with the biggest xij values are selected
to locate the branches. This initial solution is expected to be of high quality since
the xij values are considered. In the criteria-based method, after satisfying feasibil-
ity conditions, locations of remaining branches for each branch type are determined
according to the best value of the main criterion with the highest importance. In Ta-
ble 2.4, we find that transaction volume is the most important main criterion for all
types of branches. Therefore, tabu search selects candidate points with the highest
transaction volume to open the remaining branches.

2.3.4.2 Tabu search procedure

We propose a neighborhood search structure in which one branch is closed without

violating the feasibility (which is not in the set L) and a new one is opened (without
causing the same location to have multiple branches of the same type) simulta-
neously for each type. Since initial solutions used are feasible, the proposed tabu
search also maintain the feasibility with respect to constraint (2.6) and (2.7). The
tabu search algorithm searches all such possible moves for four types of branches
and determines the candidate solution as the one providing the largest increase in
the objective function value at each iteration. An alternative approach would be to
open multiple branches and close the same number of branches without violating
feasibility. However, it is clear that this will increase the computational time signif-
icantly because of searching high number of neighbors. Also, it is not ensured that
such a move structure will lead to better results, so we choose not to implement it.
The tabu search algorithm keeps track of a tabu list to avoid cycling of solutions.
It does this by forbidding the same moves for a number of iterations called the
tabu list size. Since tabu list size affects solution quality meaningfully, it has to be
determined properly according to the problem structure and size. Although, there are
2 Location Analysis in Banking 43

studies that use static tabu list size (Malek et al. 1989; Glover 1989), there are also
works including examples of dynamic tabu list size (Zhang et al. 2003; Grabowski
and Wodecki 2004). The tabu search procedure identifies a move and then it checks
to see if the move is in the tabu list or not. If the move is tabu and the aspiration
criterion is satisfied by leading to a solution better than best obtained so far, the
corresponding solution is accepted for the next iteration; otherwise more neighbors
are reinvestigated.
Our definition of a tabu move is such that we keep the closed and opened branch
in the tabu list for four types of branches separately. This means that the exact station
pair to be opened and closed simultaneously is forbidden for a specific number of
iterations. It is clear that using two separate tabu lists for closed and opened branches
for each type will cause the algorithm to be stuck when the number of candidate
regions is not high. So we choose not to implement this alternative, either.
Generally, tabu search may get trapped at a local optimum solution after a while
when it can no longer find a better solution than the best-so-far solution. Therefore,
we propose using a diversification strategy to overcome this problem. If tabu search
cannot improve the best-so-far solution after k2 consecutive iterations, it closes a
branch (which is not in the set L) and opens another branch (without causing the
same location to have multiple branches of the same type) randomly for four types
without violating feasibility. Başar et al. (2011) experimented also with randomly
closing and opening multiple stations for the location problem of emergency ser-
vice stations, which, however, has led to lower performance compared to randomly
closing and opening a single station.
Also, the current solution may not improve for k1 consecutive iterations. In such
a situation, another solution providing the least decrease in the current objective
function value is selected instead of the one giving the highest increase. Lastly, a
stopping criterion is determined to end the algorithm at the end of k3 iterations.

2.3.4.3 Experimental Study

Forty problem instances are generated randomly in order to calibrate the parame-
ters, 5 each with 50, 100, 200, 300, 400, 500, 750 and 1,000 candidate locations.
The number of branches to open (Pj ) is assumed to be 10 %, 4 %, 2 %, and 1 % of
the total number of candidate locations for individual, entrepreneur, corporate and
commercial branches, respectively, depending on the relation between the number
of sub-districts and branches in the cities of Turkey. Also, the number of branches
to open is rounded down in case it is found fractional (e.g. since the number of com-
mercial branches for a town with 50 potential locations will be 50(0.01) = 0.5, it will
be used as 0, thus no commercial branch will be opened in this area. On the other
hand, we assume that banks do not prefer to decrease the number of their current
branches. Therefore, Pj is set equal to at least the number of current branches for
each type. Moreover, the locations of current branches and the ones which cannot be
closed due to company strategy (that were opened in the last 3 year) are generated
randomly depending on the problem size. The proposed tabu search procedure is
44 A. Başar et al.

coded in Microsoft Visual C++ and executed on 2.3 GHz Intel Pentium with 4 GB
of RAM. All problem instances are solved using OPL Studio 6.3 with ILOG CPLEX
12.2 on the same processor.
We conducted an extensive experimental study to set the parameters of tabu
search, thus we determined the best value of k1 as 8 and k2 as 12. Also, we observe
that bigger tabu list sizes give better results for the instances with more candidate
locations. Therefore, tabu list size is decided to be dynamic by using 5, 7, 8, 10, 13,
15, 16 and 19 for the problems with 50, 100, 200, 300, 400, 500, 750 and 1,000,
respectively. On the other hand, 5,000 and 10,000 iterations are used as k3 values to
assess the performance of tabu search with respect to the high number of iterations.
The experimental study shows that tabu search gives very good results compared
to the solutions found by CPLEX, independent from the initialization approach and
it outperforms CPLEX in terms of both the objective function value and computation
time for most problems.

2.4 Locating Bank Branches of a Turkish Bank in Istanbul

The proposed tabu search algorithm is applied to branch location problem of a

Turkish National Bank (for confidentiality reasons, the name of the bank cannot
be provided) in Istanbul based on the real data collected by the help of Turkish Sta-
tistical Institute and The Banks Association of Turkey. GIS is used to determine the
candidate location set for the branches. There are 763 sub-districts in Istanbul, which
are all potential sites for a new branch location. The location of the bank’s current
branches, their types and opening dates are obtained for use in the methodology.
The coordinates of bank’s current branches are considered in order not to change
the location of the working branch if it is in the solution. Otherwise, coordinates of
the central points of the districts are taken into account due to the assumption that
these places are on the most visited roads. Distances of candidate branches to each
other are calculated based on the transportation network using the GIS. Thus, the
penalty of opening the same type of branches near each other and ease of access of
all candidate locations are determined.
Currently, the bank has 152 individual, 18 entrepreneur, 12 corporate and 3 com-
mercial branches in Istanbul. Also, 16 individual, 18 entrepreneur, 10 corporate and
1 commercial branches cannot be closed due to company strategy since they were
opened within the last 3 years. According to the experts’ opinion and structure of
the sub-districts in Istanbul, S is specified as 1,000 m.
Through one-on-one interviews that involve a series of pairwise comparisons for
four different types of branch, we sought the help of 10 experts working as senior
executives in the Turkish banking sector in exploring the hierarchy and evaluating
all the criteria and sub-criteria shown in Fig. 2.1. As a result, six pairwise compari-
son matrices are prepared for each branch type: One for four main criteria in the first
level of the hierarchy, one for eight sub-criteria in the second level of the hierarchy
to find their importance on transaction volume, and four for the lowest level of the
2 Location Analysis in Banking 45

hierarchy to establish the importance of the sub-criteria for the number of potential
customers, socio economic situation, social and commercial potential. After collect-
ing the evaluations of each expert based on pairwise comparisons, the importance
of the main and sub-criteria are identified using Eq. (2). As a result, the importance
weights given in Tables 2.5–2.10 are calculated.
As seen in Table 2.5, transaction volume is the most important main criterion for
all branch types. According to the results summarized in Table 2.6, the most im-
portant criterion affecting transaction volume is the number of potential customers

Table 2.5 Weights of main criteria

Criteria Notation Individual Entrepreneur Corporate Commercial
Transaction volume c1j 0.51 0.56 0.51 0.52
Distance between branches c2j 0.32 0.27 0.30 0.30
Cost of opening a new branch c3j 0.10 0.09 0.12 0.12
Cost of closing a branch c4j 0.07 0.08 0.07 0.06

Table 2.6 Weights of criteria for the transaction volume

Criteria Notation Individual Entrepreneur Corporate Commercial
Number of potential customers w1j 0.24 0.14 0.07 0.05
Socioeconomic situation w2j 0.06 0.06 0.06 0.07
Social potential w3j 0.08 0.08 0.07 0.06
Commercial potential w4j 0.16 0.18 0.39 0.44
Competition w5j 0.12 0.14 0.12 0.10
Financial situation w6j 0.16 0.22 0.18 0.15
Ease of access w7j 0.10 0.09 0.05 0.05
Growth potential w8j 0.08 0.09 0.06 0.08

Table 2.7 Weights of criteria for the number of potential customers

Criteria Notation Individual Entrepreneur Corporate Commercial
Total population w9j 0.45 0.36 0.15 0.12
Private institutions w10j 0.24 0.29 0.33 0.42
Financial institutions w11j 0.19 0.23 0.36 0.35
Education places w12j 0.07 0.07 0.09 0.08
Hospitals w13j 0.05 0.05 0.07 0.03

Table 2.8 Weights of subcriteria for socio-economic status

Sub criteria Notation Individual Entrepreneur Corporate Commercial
Education level w14j 0.68 0.71 0.68 0.62
Summer houses w15j 0.32 0.29 0.32 0.38
46 A. Başar et al.

Table 2.9 Weights of subcriteria for social potential

Sub criteria Notation Individual Entrepreneur Corporate Commercial
Education places w16j 0.51 0.51 0.33 0.30
Hospitals w17j 0.31 0.31 0.39 0.38
Entertainment w18j 0.18 0.18 0.28 0.32

Table 2.10 Weights of subcriteria for commercial potential

Sub criteria Notation Individual Entrepreneur Corporate Commercial
Private institutions w19j 0.35 0.37 0.41 0.41
Shopping centers w20j 0.25 0.17 0.16 0.11
Car parks w21j 0.08 0.06 0.06 0.05
Financial w22j 0.27 0.34 0.33 0.39
institutions
Car services w23j 0.05 0.06 0.04 0.04

for individual, financial situation for entrepreneur; but commercial potential for cor-
porate and commercial branches. From the perspective of socioeconomic status,
education level is the most fundamental sub-criterion for all types of branches. As
a social potential indicator; number of education places in the potential sites has
the highest weight for individual and entrepreneur branches, while the number of
hospitals is the most important criterion for corporate and commercial branches.
Moreover, the number of private institutions has the highest importance for all types
of branches as a commercial potential indicator. Table 2.11 summarizes final im-
portance weights of all sub-criteria depending on the hierarchical structure and by
branch type.
According to the weights in Table 2.5, the total distance from a place chosen for
a branch to the other potential locations has almost equal importance as transaction
volume. This shows that banks would be ill-advised to open branches very near
to each other. Since the distance between potential locations cannot be merged to
the decision model, a mathematical programming model is proposed to solve the
problem.
Next, the pairwise comparison results found in above are used for the benefit-
cost parameters in the objective function and for the determination of transaction
volume. As it is proven in Sect. 2.3.3 and shown in Sect. 2.3.4.3, the proposed
mathematical model is NP-Hard. Therefore, an optimal solution cannot be found
for aforementioned bank branch location problem in Istanbul, which has 763 sub-
districts and hence the proposed tabu search is applied.
The bank management wants to increase the number of individual branches in
Istanbul to 161, entrepreneur branches to 20 and corporate branches to 13 by open-
ing 9 new individual, 2 new entrepreneur branches and 1 new corporate branch.
The best values of the tabu search parameters obtained in our experimental study
are used as k1 = 8, k2 = 12 and tabu list size = 16 (as with the random instance of
2 Location Analysis in Banking 47

Table 2.11 Final weights of criteria for transaction volume

Sub Criterion Individual Entrepreneur Corporate Commercial
Total population 0.11 0.05 0.01 0.01
Education level 0.04 0.05 0.04 0.04
Number of summer houses 0.02 0.02 0.02 0.03
Education places 0.05 0.05 0.03 0.02
Hospitals 0.04 0.04 0.03 0.03
Entertainment places 0.01 0.01 0.02 0.02
Private institutions 0.11 0.11 0.18 0.21
Shopping centers 0.04 0.03 0.06 0.05
Car park 0.01 0.01 0.02 0.02
Financial institutions 0.09 0.09 0.15 0.19
Car services 0.01 0.01 0.02 0.02
Number of competitor bank branches 0.12 0.14 0.12 0.10
Average household income 0.16 0.22 0.18 0.15
Ease of access 0.10 0.08 0.05 0.05
Growth potential 0.09 0.09 0.07 0.06

750 potential locations). The linear programming relaxation method is selected as

the initialization approach due to its high performance in comparison to the random
and criteria-based methods. Finally, the results of the tabu search algorithm with
k3 = 5,000 and 10,000 are shown in Table 2.12 (where TS refers to tabu search).
The solution found with maximum run time set to 10 h by CPLEX is used as
the benchmark for the solution obtained by tabu search. The deviation (gap) in Ta-
ble 2.12 is calculated as (TS/CPLEX solution)—1. Since we try to maximize the
objective function value, the positive gap values mean that tabu search gives better
results than CPLEX. As seen in Table 2.12, the tabu search solution is 0.032 %
higher than CPLEX solution and it stays the same after k3 = 5,000 iterations
until we reach k3 = 10,000 iterations. Also, it is observed that current 12 individ-
ual branches are closed and 21 new branches are opened, current 5 entrepreneur
branches are closed and 7 new branches are opened, current 2 corporate branches
are closed and 3 new branches are opened instead at the end of k3 = 10,000 itera-
tions. The number of commercial branches in Istanbul (3) stays the same while the
location of a branch is changed. The results of the location problem of a Turkish
National Bank’s branches are summarized in Table 2.13.
The changes in the locations of this Turkish National Bank’s branches in Istanbul
given in Table 2.13 are illustrated in Fig. 2.4.
In Fig. 2.4, the dark and light grey colored boxes labeled 1, 2, 3 and 4 sym-
bolize the locations of individual, entrepreneur, corporate and commercial branches
that are closed and opened, respectively. Also, labels (1, 2, 3, 4) of closed (opened)
branches are written with white (black) font inside the dark (light) grey boxes. The
dark grey colored box labeled (1, 2) illustrates both individual and entrepreneur
48 A. Başar et al.

Table 2.12 Results for the location problem of a Turkish National Bank in Istanbul
Number of CPLEX CPLEX Initial Initial Initial TS TS Gap TS Time TS Gap
current (OFV) Time Solu- Solu- Solu- (Time) (%) (s.) (%)
branches (s.) tion 2 tion 2 tion 2 5000 it- 5000 it- 10,000 10,000
according (OFV) Time Gap erations erations itera- iterations
to their (s.) (%) tions
types
respectively
161,20,13,3 38.871 36,000 32.836 327.55 − 15.53 3,968.13 0.032 7,786.96 0.032

Table 2.13 Change in the locations of a Turkish National Bank’s branches in Istanbul
Branch type Tabu search (District—Sub-district)
Closed Branches Individual Bahçelievler-Siyavuşpaşa; Bahçelievler-Çobançeşme;
Bakırköy-Zeytinlik; Beylikdüzü-Barış; Beyoğlu-Tomtom;
Eyüp-Karadeniz; Fatih-Saraç İshak; Güngören-Güven;
Kadıköy-Eğitim; Kağıthane-Ortabayır; Küçükçekmece-Fatih;
Küçükçekmece-Tevfikbey
Entrepreneur Bağcılar-Evren; Başakşehir-Ziya Gökalp; Beylikdüzü-Barış;
Şişli-Kaptanpaşa; Tuzla-Aydınlı
Corporate Bayrampaşa-Muratpaşa; Güngören-Mehmet Nesih Özmen
Commercial Bağcılar-Evren
Opened Branches Individual Küçükçekmece-Halkalı Merkez; Üsküdar-Kandilli;
Üsküdar-Salacak; Maltepe-Zümrütevler; Şişli-Esentepe;
Ataşehir-Barbaros; Bakırköy-Zuhuratbaba; Şişli-Fulya;
Ümraniye-İstiklal; Pendik-Yenişehir; Bahçelievler-Hürriyet;
Esenler-Kazım Karabekir; Bakırköy-Sakız Ağacı;
Ataşehir-Fetih; Sarıyer-Bahçeköy Yeni;
Sultanbeyli-Hasanpaşa; Çekmeköy-Aydınlar; Kartal-Yalı;
Gaziosmanpaşa-Sultançiftliği; Ataşehir-Mimar Sinan;
Zeytinburnu-Telsiz
Entrepreneur Beşiktaş-Akat; Sarıyer-İstinye; Şişli-Teşvikiye,
Kadıköy-Suadiye; Avcılar-Denizköşkler; Pendik-Yenişehir;
Kartal-Cevizli
Corporate Şişli-Esentepe; Beşiktaş-Bebek; Kadıköy-Caddebostan
Commercial Ümraniye-Yukarı Dudullu

branches’ closure in the sub-district Beylikdüzü-Barış. Similarly, the dark grey

colored box labeled (2, 4) illustrates both entrepreneur and commercial branches’
closure in the sub-district Bağcılar-Evren. As it is stated in Sect. 2.3.3, while mul-
tiple branches of the same type cannot be opened in the same locations; different
types of branches are allowed to be opened in the same candidate location. Thus,
the light grey colored box labeled (1, 2) illustrates both individual and entrepreneur
branches’ opening in the sub-district Pendik-Yenişehir while (1, 3) illustrates in-
dividual and corporate branches’ opening in Şişli-Esentepe. As it can be seen in
Fig. 2.4, closed branches are mostly located in the western (European) side of the
2 Location Analysis in Banking 49

Fig. 2.4 Solution for the branch location problem of a Turkish National Bank’s in Istanbu

city, since a considerable part of the current branches are located here. It is ob-
served that closed branches are the ones located in the sub-districts having multiple
branches or near to the current branches of same type. On the other hand, locations
with high transaction volume or those with no surrounding nearby branches are se-
lected as new branch locations. Moreover, it is easily seen that the service network
is enlarged by closing branches which are very close to each other and opening
new ones in the easily accessible locations instead. The results are also approved by
the experts of the related bank and other experts who contributed in identifying the
criteria and specifying the importance of the criteria.
In order to test the validity of the results and evaluate the robustness of the so-
lution, a sensitivity analysis is conducted by selecting four main criteria, namely
transaction volume, distance between branches, cost of opening a new branch and
cost of closing a branch. Their importance weights are varied by ± 10 % and ± 20 %.
The results of the criteria sensitivity obtained at the end of k3 = 10,000 iterations
are summarized in Table 2.14.
As seen in Table 2.14, the solution does not alter for the changes in four main
criteria in the interval of + 10 % and − 10 %. If the weights obtained via expert
judgments are changed by 20 % for the most important criteria, which are trans-
action volume and distance between branches, only the results for a few scenarios
change slightly. Thus, we find the results of the method to be robust.
50 A. Başar et al.

Table 2.14 Sensitivity analysis of the solution in Istanbul obtained by tabu search
Change ratio (%) Change
Original solution – Table 2.13
Transaction volume + 20 Do not close individual branches in
Bakırköy-Zeytinlik and Güngören-Güven
sub-districts; do not open individual branches in
Pendik-Yenişehir and Sultanbeyli-Hasanpaşa
sub-districts. Do not close the entrepreneur branch
in Şişli-Kaptanpaşa sub-district and do not open the
entrepreneur branch in Avcılar-Denizköşkler
sub-district
+ 10 Same as the original solution
− 10 Same as the original solution
− 20 Do not close individual branches in
Beylikdüzü-Barış, Beyoğlu-Tomtom and
Kağıthane-Ortabayır sub-districts; do not open
individual branches in Şişli-Esentepe,
Bakırköy-Zuhuratbaba, Esenler-Kazım Karabekir
and Kartal-Yalı sub-districts, select
Bağcılar-Demirkapı sub-district to open a new
individual branch. Do not close the entrepreneur
branch in Beylikdüzü-Barış sub-district and do not
open the entrepreneur branch in
Avcılar-Denizköşkler sub-district
Distance between + 20 Do not close individual branches in
branches Beylikdüzü-Barış and Kağıthane-Ortabayır
sub-districts. Do not open individual branches in
Şişli-Esentepe and Esenler-Kazım Karabekir
sub-districts
+ 10 Same as the original solution
− 10 Same as the original solution
− 20 Do not close the individual branch in
Bakırköy-Zeytinlik sub-district and do not open the
individual branch in Sultanbeyli-Hasanpaşa
sub-district
Cost of opening a + 20 Same as the original solution
new branch + 10 Same as the original solution
− 10 Same as the original solution
− 20 Same as the original solution
Cost of closing a + 20 Same as the original solution
branch + 10 Same as the original solution
− 10 Same as the original solution
− 20 Same as the original solution
2 Location Analysis in Banking 51

An important limitation of our study lies with the trouble of finding real data
for some criteria such as daytime population for number of potential customers.
Moreover, since we do not have real transaction volume data or banking outputs
such as volume of deposits or credits of the mentioned bank, we were unable to use
an objective weighting method and confirm the accuracy of the importance weights
obtained via pairwise comparison. On the other hand, we can see that the results
of the model can easily be applied in real life directly, since we have considered
the most important criteria for the branch location problem. Also, we can use the
proposed mathematical model for any bank since it considers both opening new and
closing inefficient branches at the same time. Furthermore, opening a new branch in
the most exact place is suggested by penalizing the opening of same-type branches
near one another, taking ease of access into consideration and selecting the most
frequently visited places as candidate locations by the help of GIS.

2.5 Conclusion and Future Research

In this study, we presented an integrated methodology to make decisions for the lo-
cation of bank branches. As we stated in the literature review, MCDM is a common
technique employed, and there are not many studies using mathematical models to
solve this problem. However, both MCDM and mathematical modeling approaches
have some shortcomings. MCDM deals mainly with analyzing decision problems
and evaluating the alternatives/criteria based on decision makers’ preferences. It
is clear that MCDM lacks the capability of an optimization aspect. On the other
hand, mathematical models have limited capability of considering expert opinion.
By combining these two techniques, we benefit from the strengths and discard the
shortcomings of both methods. To this aim, we determine the most relevant criteria
and obtain their importance weights via extensive literature review and expert opin-
ion. We also develop a new mathematical model to determine the exact locations of
the branches. Our model permits both the opening and closing of branches at the
same time, which shows that it can be used not only for deciding the best locations
for branches, but also for measuring the performance of existing branches. Since the
model is NP-Hard and an optimal solution for the problems with large number of
candidate locations might not be so easily found using an exact method, we develop
and propose a tabu search algorithm. We experimentally show that tabu search gives
better solutions than CPLEX in terms of both the objective function value and com-
putational time. We apply our proposed methodology to the branch location problem
in Istanbul for one of the biggest banks in Turkey. We find the results to be robust
and validate them by both sensitivity analysis and expert opinion.
Future studies may include the presentation of a multi-period mathematical
model to support banks’ long-term strategies, application of a different meta-
heuristic to compare with the tabu search results and researching the applicability
of the proposed methodology in similar location problems (e.g., supermarkets,
restaurants etc.)
52 A. Başar et al.

References

Abbasi GY (2003) A decision support system for bank location selection. Int J Comput Appl
Technol 16:202–210
Ahsan MK, Bartlema J (2004) Monitoring healthcare performance by analytic hierarchy process:
a developing-country perspective. Int Trans Oper Res 11:465–478
Alexandris G, Giannikos I (2010) A new model for maximal coverage exploiting GIS capabilities.
Eur J Oper Res 202:328–338
Arabani AB, Farahani RZ (2012) Facility location dynamics: an overview of classifications and
applications. Comput Ind Eng 62:408–420
Aras H, Erdogmuş S, Koç E (2004) Multi-criteria selection for a wind observation station location
using analytic hierarchy process. Renewable Energy 29(8):1383–1392
Arostegui MA, Kadipasaoglu SN, Khumawala BM (2006) An empirical comparison of tabu search,
simulated annealing, and genetic algorithms for facilities location problems. Int J Product Econ
103(2):742–754
Badri MA (2001) A combined AHP—GP model for quality control systems. Int J Product Econ
72:27–40
Banking and sector information (2014) http://www.tbb.org.tr/tr/banka-ve-sektor-bilgileri/banka-
bilgileri/subeler/65. Accessed 23 June 2014
Baron O, Berman O, Kim S, Krass D (2009) Ensuring feasibility in location problems with
stochastic demands and congestion. IIE Trans 41:467–481
Başar A, Çatay B, Ünlüyurt T (2011) A multi-period double coverage approach for locating the
emergency medical service stations in Istanbul. J Oper Res Soc 62(4):627–637
Başar A, Çatay B, Ünlüyurt T (2012) A taxonomy for emergency service station location problem.
Optim Lett 6(6):1147–1160
Berman O, Krass D (2002) The generalized maximal covering location problem. Comp Oper Res
29:563–581
Boufounou PV (1995) Evaluating bank branch location and performance: a case study. Eur J Oper
Res 87:389–402
Brimberg J, Drezner Z (2013) A new heuristic for solving the p-median problem in the plane.
Comp Oper Res 40:427–437
Camanho AS, Dyson RG (2005) Cost efficiency measurement with price uncertainty: a DEA
application to bank branch assessments. Eur J Oper Res 161:432–446
Carlsson C, Fuller R (1996) Fuzzy multiple criteria decision making: recent developments. Fuzzy
Sets Syst 78:139–153
Chen SJ, Hwang CL (1993) Fuzzy multiple attribute decision-making, methods and applications.
Lecture notes in Economics and Mathematical systems 375. Springer, Heidelberg
Chen H, Barna BA, Rogers TN, Shonnard DR (2001) A screening methodology for improved
solvent selection using economic and environmental assessments. Clean Prod Processes 3:
290–302
Church R, ReVelle C (1974) The maximal covering location problem. Pap Reg Sci Assoc 32:
101–118
Cinar N (2009) A decision support model for bank branch location selection. World Acad Sci Eng
Technol 60:126–131
Cinar N, Ahiska SS (2010) A decision support model for bank branch location selection. In
YAEM 2010. Proceedings of the 2010 International Conference on Industrial Engineering and
Operations Management, Sabancı University, Istanbul, June 30-July 2, (CD-ROM)
Clawson CJ (1974) Fitting branch locations, performance standards, and marketing strategies to
local conditions. J Marketing 38:8–14
Cook WD, Seiford LM, Zhu J (2004) Models for performance benchmarking: measuring the effect
of e-business activities on banking performance. Omega 32:313–322
Curtin KM, Hayslett-McCall K, Qiu F (2010) Determining optimal police patrol areas with
maximal covering and backup covering location models. Netw Spat Econ 10(1):125–145
2 Location Analysis in Banking 53

Daskin M (1983) A maximum expected covering location model: formulation, properties and
heuristic solution. Transp Sci 17:48–70
Daskin MS, Stern EH (1981) A hierarchical objective set covering model for emergency medical
service vehicle deployment. Transp Sci 15:137–152
Doyle P, Fenwick I, Savage GP (1981) A model for evaluating branch location and performance. J
Bank Res 12:90–95
Erdemir ET, Batta R, Spielman S, Rogerson PA, Blatt A, Flanigan M (2010) Joint ground and air
emergency medical services coverage models: a greedy heuristic solution approach. Eur J Oper
Res 207:736–749
Fernandez I, Ruiz MC (2009) Descriptive model and evaluation system to locate sustainable
industrial areas. J Cleaner Prod 17(1):87–100
Gendreau M, Laporte G, Semet F (2000) A dynamic model and parallel tabu search heuristic for
real time ambulance relocation. Parallel Comput 27:1641–1653
Glover F (1977) Heuristics for integer programming using surrogate constraints. Decis Sci 8:
156–166
Glover F (1989) Tabu search—part I. ORSA J Comput 1(3):190–206
Grabowski J, Wodecki M (2004) A very fast tabu search algorithm for the permutation flow shop
problem with makespan criterion. Comput Oper Res 31:1891–1909
Hakimi S (1964) Optimum locations of switching centres and the absolute centres and medians of
a graph. Oper Res 12:450–459
Hansen P, Brimberg J, Urosevic D, Mladenovic N (2009) Solving large p-median clustering
problems by primal-dual variable neighborhood search. Data Min Knowl Disc 19:351–375
Huff D (1963) A probabilistic analysis of shopping center trade areas. Land Econ 39:81–90
Hwang H (2002) Design of supply-chain logistics system considering service level. Comput
Industrial Eng 43:283–297
Ishizaka A, Lusti M (2004) An expert module to improve the consistency of AHP matrices. Int
Trans Oper Res 11:97–105
Jablonsky J, Fiala P, Smirlis Y, Despotis DK (2004) DEA with interval data: an illustration using
the evaluation of branches of a Czech bank. Central Eur J Oper Res 12:323–337
Kaufman G, Mote R (1994) A review from the Federal Reserve Bank of Chicago. Federal Reserve
Bank of Chicago, Chicago
Kuehn A, Hamburger M (1960) A heuristic program for locating warehouses. Manage Sci 9:
643–666
Malczewski J (1999) GIS and multicriteria decision making. Wiley, New York
Malek M, Guruswamy M, Pandya M, Owens H (1989) Serial and parallel simulated annealing and
tabu search algorithms for the traveling salesman problem. Ann Oper Res 21:59–84
Manandhar R, Tang JCS (2002) The evaluation of bank branch performance using data envelop-
ment analysis: a framework. J High Technol Manage Res 13:1–17
Manne A (1964) Plant location under economies of scale. decentralization and computation.
Manage Sci 11:213–235
Marianov V, ReVelle CS (1995) Facility location. Springer, Berlin
Meidan A (1983) Distribution of bank services and branch location. Int J Phys Distrib Managerial
Manage 13(3):5–18
Miliotis P, Dimopoulou M, Giannikos I (2002) A hierarchial location model for locating bank
branches in a competitive environment. Int Trans Oper Res 9:549–565
Miller GA (1956) The magical number seven, plus or minus two: some limits on our capacity for
processing information. Psychol Rev 63:81–97
Min H (1989) A model based decision support system for locating banks. Inform Manag 17:
207–215
Min H, Melachrinoudis E (2001) The three-hierarchical location-allocation of banking facilities
with risk and uncertainty. Int Trans Oper Res 8:381–401
Morrison PS, O’Brien R (2001) Bank branch closures in New Zealand: the application of a spacial
interaction model. Appl Geogr 21:301–330
54 A. Başar et al.

Murray AT, Tong D, Kim K (2010) Enhancing classic coverage location models. Int Reg Sci Rev
33(2):115–133
Olsen LM, Lord JD (1979) Market area characteristics and branch bank performance. J Bank Res
Summer 10:102–110
Osman IH, Kelly JP (1996) Meta-heuristics: theory and applications. Kluwer Academic Publishers,
Boston
Pardalos PM, Resende MGC (2002) Handbook of applied optimization. Oxford University Press,
New York
Pastor JT (1994) Bicriterion programs and managerial location decisions: application to the
banking sector. J Oper Res Soc 45(12):1351–1362
Portela MCAS, Thanassoulis E (2007) Comparative efficiency analysis of Portuguese bank
branches. Eur J Oper Res 177:1275–1288
Rahgan SH, Mirzazadeh A (2012) A new method in the location problem using fuzzy evidential
reasoning. Eng Techno 4(22):4636–4645
Ravallion M, Wodon Q (2000) Banking on the poor? Branch location and nonfarm rural
development in Bangladesh. Rev Devel Econ 4:121–139
Retail Banker International (2013) US Branch numbers fall for fourth year running (2014).
https://dscqm8cqg6d5o.cloudfront.net/uploads/articles/pdfs/mnetisgnefhmblsrdclkablzye_
rbioct13issue694usbranches.pdf. Accessed 23 June 2014
ReVelle CS, Swain R (1970) Central facilities location. Geogra Anal 2:30–42
Ribeiro CC, Hansen P (2002) Essays and surveys in metaheuristics. Kluwer Academic Publishers,
Norwell
Saaty TL (1980) The analytic hierarchy process. McGraw-Hill Inc, New York
Saaty TL (1990) How to make a decision: the analytic hierarchy process. Eur J Oper Res 48:9–26
Sato Y (2004) Comparison between multiple-choice and analytic hierarchy process: measuring
human perception. Int Trans in Ope Res 11:77–86
The Banks Association of Turkey (2014) Available at http://www.tbb.org.tr/tr/banka-ve-sektor-
bilgileri/banka-bilgileri/subeler/65. Accessed 12 Apr 2015
Toregas CR, Swain R, ReVelle CS, Bergman L (1971) The location of emergency service facilities.
Oper Res 19:1363–1373
Tzeng GH, Teng MH, Chen JJ, Opricovic S (2002) Multi criteria selection for a restaurant location
in Taipei. Int J Hosp Manage 21(2):171–187
Wang Q, Batta R, Rump CM (2002) Algorithms for a facility location problem with stochastic
customer demand and immobile servers. Ann Oper Res 111:17–34
Wang Q, Batta R, Bhadury J, Rump CM (2003) Budget constrained location problem with opening
and closing of facilities. Comput Oper Res 30:2047–2069
Wu CR, Lin CT, Chen HC (2007) Optimal selection of location for Taiwanese hospitals to ensure
a competitive advantage by using the analytic hierarchy process and sensitivity analysis. Build
Environ 42(3):1431–1444
Xia L, Yin W, Dong J, Wu T, Xie M, Zhao Y (2010) Hybrid nested partitions algorithm for banking
facility location problems. IEEE Transautomation Sci Eng 7(3):654–658
Youssef H, Sait SM, Adiche H (2001) Evolutionary algorithms, simulated annealing and tabu
search: a comparative study. Eng Appl Artif Intell 14:167–181
Zhang L, Rushton G (2008) Optimizing the size and locations of facilities in competitive multi-site
service systems. Comput Oper Res 35:327–338
Zhang G, Habenicht W, Spieß WEL (2003) Improving the structure of deep frozen and chilled
food chain with tabu search procedure. J Food Eng 60(1):67–79
Zhao L, Garner B, Parolin B (2004) Branch bank closures in Sydney: a geographical perspective
and analysis. Proceedings of the 12th International Conference on Geoformatics, Sweden. June
7–9:541–548
Chapter 3
Location Modeling for Logistics Parks

Joyendu Bhadury, Mark L. Burkey and Samuel P Troy

3.1 Background

Location theory is an extensive field with hundreds of published papers drawing

upon disciplines as diverse as operations research, operations management, regional
science, geography, economics, computer science and mathematics. Some recent
and well-known references that survey the field are Hamacher and Drezner (2002),
Current et al. (2004), Revelle and Eiselt (2005) and Eiselt and Marianov (2011).
Yet, the discipline suffers from a paucity of publications that involve real-life appli-
cations, especially case studies. This paucity is noted in Current et al. (2004), which
begins with the statement “Much of the literature on facility location modeling
has not been directed to specific applications (that is, case studies)” and pointedly
reinforced in Revelle and Eiselt (2005, p. 15) who state “. . .when it comes to ap-
plications, there appears to be a significant deficit, at least as compared to other,
similar, fields.” This is of particular concern since substantial inefficiencies can oc-
cur if facilities are located sub-optimally—for example, Rushton (1988, p. 101),
illustrates this fact in an application involving the location of health clinics in In-
dia. Despite this paucity however, there are a few applications of facilities location
models documented in the published literature from a variety of different disciplines
and countries, as first noted in Eiselt (1992). For example, Rushton (1988) contains
references on the use of location-allocation models in siting public-service facilities
such as health clinics and schools, in the developing world.

J. Bhadury () · S. P Troy

Bryan School of Business and Economics, University of North Carolina at Greensboro,
Greensboro, NC 27402, USA
e-mail: [email protected]
S. P Troy
e-mail: [email protected]

M. L. Burkey
School of Business and Economics, North Carolina A & T State University,
Greensboro, NC 27411, USA
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 55

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_3
56 J. Bhadury et al.

This work extends the extant literature on location related case studies based on a
project related to the location of a Logistics park in the southeastern region of North
Carolina, USA. While a portion of this work draws upon that project, the primary
objective of this case study is more than to simply to describe the location model-
ing process used in that application. Instead, the main objective is to distill lessons
learned that can help guide future academic research in location theory. To that end,
this case study begins with solving the problem at hand using a theoretical approach
so that a contrast may be drawn with how the modeling was done in practice. In turn,
that contrast forms the basis of the final section on conclusions and lessons learned.
The remaining case study is organized as follows. Sect. 3.2 gives the necessary
background to the case study by providing the contextual framework of the project
and an introduction to NCSE, detailing the socio-economic and infrastructural infor-
mation that informed the research. The next two sections are devoted to modeling
the location of a Logistics Park in NCSE. While Sect. 3.3 details how a “theoreti-
cal” modeling approach would have worked, Sect. 3.4 describes the actual approach
used when conducting the study within a structured framework, referred to by its
acronym, SIRC. Sect. 3.5 ends this work by describing the follow-up to the Seven
Portals Study, lessons learned and recommendations for location scientists involved
in modelling real-life location/site selection problems.

3.1.1 Background: Logistics Parks, Seven Portals Study

and NC Southeast

A “Logistics Park” is defined as a “development concept in which distribution cen-

ters typically seen in a suburban area are built in a park like setting, created by
landscaping” (Gardner, Kansas Official Website 2015). It is usually populated by
warehouses, distribution centers and logistics-related companies/offices. In almost
all cases, it is also an intermodal facility where truck trailers and containers are
transferred between trucks and the railroad. The concept of Logistics Park has been
applied and used in various international settings and may be referred to by other
names such as Freight Village, Güterverkehrszentrum, Interporto etc. Many such fa-
cilities exist throughout the world and a few examples include Burlington Northern
Logistics Park (Illinois, USA), large cargo airports such as Hong Kong and Mem-
phis, Pinghu Logistics Park in Shenzen, China and Schipol Logistics Park in the
Netherlands. In the last decade, academic research has begun to emerge in loca-
tion theory around logistics parks; for instance, Lee et al. (2001) and Lee and Yang
(2003) discuss models and strategies for the location of such facilities, whereas El
Amrani (2007) investigates the impact of international Logistics Parks on supply
chains of multinationals. To the best of our knowledge, however, this case study is
the first that is based on an actual application related to Logistics Park location.
As to the history of why the state of North Carolina undertook planning for lo-
cating Logistics Parks, one needs to begin with the state-wide strategic plan for
logistics (List et al. 2008). In order to implement the same, the North Carolina State
3 Location Modeling for Logistics Parks 57

Legislature established the Governor’s Logistics Task Force (GLTF) on December

7, 2009 and assigned this group seven tasks (see GLTF-Final Report 2012) which
included the following:
• “Conduct a thorough inventory and evaluation of existing public and private
transportation and commerce assets, including ports, inland ports, airports,
highways, railroads, major distribution centers, and business and industrial parks.
• Report on the current system for moving goods and people, including the
condition of the system, its overall performance, and its safety.
• Project future needs for the state’s multi-modal transportation system and explore
challenges and opportunities in meeting those needs.
• Explore innovative ideas in transportation and economic development that can
help support the state’s logistics capacity, including public private partnerships.”
The final report from GLTF (GLTF-Final Report 2012) details the accomplishment
of all its objectives, including the four listed above. In order to implement its charge,
GLTF undertook some key studies, one of them being the “Seven Portals Study”
(List and Foyle 2011), which was administered by the North Carolina Department of
Transportation (NCDOT). The purpose of that study was to “describe ways in which
North Carolina’s transportation infrastructure investments can help with economic
development and the creation of jobs” (List and Foyle 2011). The overall goal was
investigating potential Logistics Parks throughout the state and identifying what
infrastructure improvements are needed to support such a facility at that location. It
was made clear that the study would not recommend specific sites above others but
merely present facts about each to assist the North Carolina State Legislature make
the final decisions. The final report of the Seven Portals Study by List and Foyle
(2011) was based on smaller regional reports each of which focused on one specific
region of North Carolina. This case study focuses on one such region, namely, North
Carolina’s Southeastern Region (NCSE) (see Fig. 3.1) and is drawn, in part, from
the final report submitted for the same (Bhadury and Troy 2011).
For purposes of regional economic development, the North Carolina Depart-
ment of Commerce (http://www.nccommerce.com/) has divided the state into seven
distinct geographical regions, each with an established “Regional Partnership”
agency to foster a collective approach to regional economic development; see
http://www.thrivenc.com/ for a complete listing. North Carolina’s Southeast Re-
gion, abbreviated as NCSE, (see Figs. 3.1 and 3.2) is one of these seven. The region
is composed of 14 counties stretching from the Atlantic Ocean to Sandhills re-
gion of North Carolina. The region’s 500,000 strong workforce, is employed in a
diverse cross-section of workers engaged in agriculture, wood products, manufac-
turing, wholesale trades (i.e., distribution), construction, healthcare, government and
the professions. Economically, NCSE is substantially diverse. Along with a strong
agricultural, wood products and food processing sector supported by companies
such as Smithfield Foods, International Paper Company and Campbell Soup, it is
also home to high tech companies like GE Nuclear Energy, Pharmaceutical Product
Development Company (PPD) and Corning.
58 J. Bhadury et al.

Fig. 3.1 Relative position of NCSE in Eastern USA

Fig. 3.2 North Carolina’s south east region. (Source: http://www.ncse.org/)

Two sectors are of outstanding importance to overall economy of NCSE and de-
serve a special mention: military (Fort Bragg) and shipping (Port of Wilmington).
As for the military sector, it is centered on Fort Bragg near Fayetteville. Established
in 1918, Fort Bragg is now the largest US Army base by population in USA and
3 Location Modeling for Logistics Parks 59

includes 10 % of the US Army’s active component forces. The base covers 161,047
acres and the military population comprises almost 60,000 officers and enlisted men
and women with an additional 21,000 civilian workers employed directly by the
army base. The BRAC (Base Realignment and Closure) process undertaken since
2005 by the US Department of Defense has contributed significantly to the growth of
Fort Bragg. For example, European-based forces have been relocated to Fort Bragg.
Additionally, several aircrafts are now located at Fort Bragg to form an Air Force
Reserve/active duty associate unit as is Air Force Reserve Command operations and
maintenance. Finally, BRAC has resulted in two major headquarters, Army Forces
Command and Army Reserve Command to be located at Fort Bragg. With such a
significant presence, it is no surprise that Fort Bragg and thus, the military sector,
is a major economic driver in NCSE. The current annual payroll of Fort Bragg is
estimated at $ 3.5 billion which generates a direct and indirect annual impact of
approximately $ 10.9 billion in the immediate region (NCSE Regional Economic
Profile 2013).
As for Port of Wilmington, it is North Carolina’s sole container shipping port.
The port is owned and operated by the North Carolina State Ports Authority
(http://www.ncports.com) and offers terminal facilities serving container, bulk and
break-bulk operations. The 42 ft navigational channel is complemented on the land
side by available modern transit and warehouse facilities, state-of-the-art Panamax
container cranes and support equipment, nine berths with 6768 ft. of wharf frontage
and the latest in cargo management technology (North Carolina State Ports Au-
thority 2015). Railroad service is provided by CSX Transportation which has daily
service for boxcar, tanker and general cargo services. In order to facilitate the po-
sition of the port as a hub of international trade, it is a designated Foreign Trade
Zone that is administered by North Carolina Department of Commerce. The port
itself hosts almost 1 million sq. ft. of storage. The port has an annual traffic of over
2.1 million containers, which makes it smaller compared to the neighboring East
Coast ports of Charleston, Savanah and Norfolk. Nonetheless, it provides container
and bulk shipping to/from most world markets. Ocean carriers that call on the Port
of Wilmington include major international shipping lines such as “K” Line Amer-
ica, Cosco Container Lines, Hanjin Shipping Company, Maersk Line and National
Shipping Company of Saudi Arabia. As for the economic impact of the port and the
shipping industry overall, Findley et al. (2014) estimate that this port has an annual
economic impact of $ 12.9 billion and along with a smaller bulk port in Morehead
City, supports 76,000 direct and indirect jobs.
Notwithstanding the above, the NCSE is a predominantly rural part of North Car-
olina, which suffers from significant unemployment and poverty issues. Figure 3.3
reveals that the region has few towns/counties that have a population above 1000.
Also, as evident from Table 3.1, the region had an overall unemployment rate of
6.46 % as of December 2014 (above the state rate of 5.5 %) with 11 out of the 14
counties having a worse than state average unemployment rate. In particular, Scot-
land County, with a 10 % unemployment rate, was the 2nd highest in the state and
has historically been in a similar position for a long while. With regards to family
income, the picture is better balanced. As of end of fiscal year 2013-14, the average
60 J. Bhadury et al.

Fig. 3.3 NCSE—towns over 1000, over 10,000 labeled

Table 3.1 NCSE counties at a glance. (Source: http://quickfacts.census.gov/, http://www.ncse.org/

and https://desncc.com/deshome)
County Population Median family income Unemployment in
(2014) 2013-14 (US $) Dec 2014 (%)
Anson county 26,948 34,659 6.6
Bladen county 35,190 42,473 8.1
Brunswick county 113,235 54,172 6.1
Columbus county 54,899 41,957 7.3
Cumberland county 344,167 55,675 6.3
Duplin county 58,710 33,172 6
Hoke county 46,265 49,400 5.7
Montgomery county 27,798 32,946 5.8
New Hanover county 196,320 65,219 5
Pender county 54,546 53,225 5.5
Richmond county 46,893 44,189 7
Robeson county 132,092 41,304 7.7
Sampson county 65,513 46,947 5.1
Scotland county 37,059 49,852 10.1
NCSE region (weighted 51,091 6.46
average of 14 counties)
NC (overall) 9,848,060 46,334 5.5
3 Location Modeling for Logistics Parks 61

Table 3.2 NCSE’s employment profile (3rd quarter, 2012). (Source: http://www.ncse.org/)
Industry Number of people Average weekly wages
employed in NCSE by in the industry (in $)
the industry
Agriculture, forestry, fishing & hunting 6075 477.74
Mining 232 843.94
Utilities 2192 1576.91
Construction 15,764 713.37
Manufacturing 37,270 868.16
Wholesale trade 10,029 898.68
Retail trade 49,036 453.30
Transportation & warehousing 9910 786.35
Information 5068 841.25
Finance & insurance 7633 913.92
Real estate & rental & leasing 5210 648.31
Professional & technical services 14,642 1139.03
Management of companies & enterprises 2463 830.73
Administrative & waste services 19,126 513.31
Educational services 33,587 727.48
Health care & social assistance 61,993 760.62
Arts, entertainment & recreation 5557 333.21
Accommodation & food services 41,116 270.90
Other services except public administration 8360 483.12
Public administration 31,981 866.52

median family income was $ 51,091, which is higher than the state-wide average
of $ 46,334. The $ 51,091 figure was buoyed primarily by the two urban centers
in the region, Wilmington and Fayetteville, and the economic sectors around them.
Overall, half of the region had a median family income below the state average. The
distribution of the employment in this region across the various industry sectors and
the median weekly wages in each are displayed in Table 3.2.
The region also has a strong logistical infrastructure for all modes of transporta-
tion. With regards to roadways, the bulk of this region is contained within a triangle
of three major US interstates—Interstate 95, Interstate 40 and Interstate 73/74 and
more importantly, the NCDOT has targeted these highways for improvement in the
next decade. These three collectively make the region accessible to over half of the
population of USA within one day’s driving. Wilmington hosts North Carolina’s
only container shipping port, namely, the Port of Wilmington. The railroads in the
region include both of USA’s major freight carriers: CSX and Norfolk-Southern
as well as several short-haul local lines such as Aberdeen & Rockfish, Carolina-
Southern and Clinton Terminal Railroad. Finally, air service is provided at two
primary airports: Wilmington and Fayetteville that offer connections via major
American carriers such as Delta, US Airways, and American Eagle.
62 J. Bhadury et al.

The regional economic development agency, known as North Carolina’s South-

east (http://www.ncse.org/) has identified the following nine as the targeted industry
clusters for future economic development of the region: advanced textiles, aviation
and aerospace, biotechnology, building products, defense, distribution and logistics,
energy, food processing and agri-industry, metalworking. As evident, critical to each
of these is the presence of an efficient logistical infrastructure in the region including
state of the art multi-modal Logistics Park with expeditious connection to the Port
of Wilmington. This formed the basis for their desire to locate one in their region.

3.2 A Theoretical Approach to Determine the “Optimal”

Location

In order to provide a contrast with how location decisions for Logistics Parks are
made in practice as opposed to theory, we will begin with modeling the problem
of locating a logistics park in NCSE as a traditional 1-median problem in location
theory and analyze the results.
As is typical for any location model that seeks to determine the optimal location
of a service facility, we have to begin by modeling the “consumers” of the ser-
vices that would be provided by the logistics park. For that purpose, we represent
NCSE as a collection of demand points on the two-dimensional Euclidean plane.
With that, assume that there are N demand points located in NCSE, each of which
is referenced as demand point i, where i = 1,2,. . .,N. We assume here that each de-
mand point represents an appropriately defined cluster of businesses or citizens of
NCSE that are potential users of the services provided by a logistics park. It must be
noted here that such aggregations essentially convert what is a location problem with
continuous demand into one with discrete demand. In turn, that generates agglom-
eration errors that are well studied in location theory, see Francis et al. (2009) for a
survey. Nonetheless, for purposes of illustration we will proceed with this demand
agglomeration.
Further, for each demand point i above, let (ai ,bi ) represent the abscissa and the
ordinate respectively and let wi represent the total demand for the services provided
by the logistics park. Additionally, let X= (aX ,bX ) represent any arbitrary point on
this same plane. Then, the lp distance between X and demand point is i, denoted by
dp (X, i), is given by the following formula:
 1
dp (X, i) = |ax − ai |p + |bX − bi |p p (3.1)
Perhaps the most commonly used version of (1) above is when p = 2, i.e., L2
distance, given by

d2 (X, i) = (aX − ai )2 + (bX − bi )2 (3.2)
As is well known, the l2 distance formula (2) shown above is also known as the
Euclidean, or straight-line distance, which measures the length of the line segment
3 Location Modeling for Logistics Parks 63

connecting X and demand point i. This metric assumes that travel takes place along
straight lines, and without barriers.
Another distance function that is used, albeit less commonly is the squared
Euclidean distance given by

d 2 (X, i) = (aX − ai )2 + (bX − bi )2 (3.3)

The squared Euclidean distance metric (3) assumes that the disutility of travel in-
creases quadratically. For example, the disutility of commuting is more costly for
the 2nd mile than for the 1st mile, and more costly for the 10th mile than for the 9th.
There are two reasons behind this assumption. First, people normally get increasing
disutility from higher levels of a bad thing, such as work or pollution. Secondly,
there is an increasing opportunity cost of an individual’s time as more and more
time is spent driving. It must also be mentioned that this metric is widely used by
practitioners in selecting locations for distribution centers and warehouses.
Another commonly used distance metric is the l1 metric when p = 1, also known
as the Manhattan metric or the rectilinear distance metric. In this case, the distance
between X and demand point i is given by:

d1 (X, i) = |aX − ai | + |bX − bi | (3.4)

Two good sources of information for different kinds of lp metrics and their im-
plications in practice are Brimberg and Love (1995) and Burkey et al. (2011). As
shown in Brimberg et al. (1994) and Fernández et al. (2002), variants of the lp norm
with intermediate values of p between 1 and 2 best predict road distances in real-
world applications. This is why the analytical results in this section include those
for p = 1.5.
Notwithstanding the definition of distance, the objective of locating a Logistics
Park would obviously be to be as proximate to the users as possible. To that end, for
any given lp -metric, the 1-median of these of demand points is given by a point X*,
with the property that

N 
N
wi dp (X∗ , i) ≤ wi dp (X, i) for all points X in NCSE (3.5)
i=1 i=1

In other words, a logistics park located at X* would be guaranteed to have the min-
imum possible total weighted average distance to all the customer demand points
i = 1, 2,. . .,N, where the weight of each point i is wi and represents the total demand
at that point. It is useful to note that when p = 2, i.e., in the case of the Euclidean met-
ric, the 1-median is also referred to as the Fermat-Weber point, a classical problem
in geometry going back to the seventeenth century.
As for computing the 1-median for a given set of points, the best known algo-
rithm is given in Weiszfeld (1937), which has recently been translated into English
(Weiszfeld and Plastria 2009). Numerous improvements to the original Weiszfeld
algorithm have been proposed over time, see Plastria (2011) for a comprehensive
update on them. Finally, we note that in the special case where the distance metric
64 J. Bhadury et al.

used is the Squared Euclidean Metric given in (3), the 1-median is referred to as the
center of mass or center of gravity—we will use the former term. Determining the
center of mass is straightforward, as it involves taking the weighted average of the
coordinates of the demand points. However, it is important that one uses a rectangu-
lar projection of abscissa and ordinate coordinates rather than longitude and latitude
(which are spherical coordinates).
With the backdrop above, it is clear that an analytical approach to locating a
logistics park in NCSE would essentially determine the 1-median as the optimal
location. The next questions then are: how do we define the appropriate clusters
to aggregate demand and what distance metric (i.e., the value of p) do we use in
the distance calculation formula in determination of the 1-median. Given that many
such answers are possible, we provide below an analysis of how different solutions
for the 1-median problem (5) may be obtained based on different answers to each
of the two questions above. We refrain from prescribing one solution over another,
but do present a rationale for why they differ from each other.
The first issue that needs to be resolved is a determination of where people live
or work in NCSE, since that information is the foundation of all location modelling.
There are many choices for the demand points. Two broad categories of choices are
(i) choice of demander, and (ii) choice of scale. The simplest choice to use for the
demanders would be to use several specific points if there are only a few relevant de-
mand points, e.g., specific distribution centers or factories. Another common option
(for small area studies) is to use the number of people at the location of their resi-
dences as the intensity and location of demand points. However, in the current case,
it might make more sense to use the location (and size) of businesses in formulating
the demand points.
Whether one uses locations of people or businesses, the coarsest practical scale
for NCSE would be to use the 14 counties as the demand points. The drawback
to doing this is that one must assume a single point to represent many demand
points that are spread out over a large area, probably non-uniformly. When using
people to represent the demanders, the finest practical scale is to use census blocks
or block groups1 . The block is the smallest unit of data released by the US Census
Bureau. Typically, only population counts are released at this level, and at roughly
29,000 blocks per county in North Carolina2 , might be unnecessarily fine. Table 3.3
gives a breakdown of several intermediate scales that might be chosen, using North
Carolina’s numbers for reference.
When using businesses as the demand points, the finest practical scale for most
analyses is to use business data aggregated at the zip code level. A convenient source
of such data in the US is the economic census from the BLS. For each zip code, data
are available on how many firms there are, how many total employees work in these
firms, and the total payroll.

1 Though some work on a small scale, such as for a single city or county could use individual

addresses as demand nodes e.g., Qabaja and Bikdash (2014).

2 https://www.census.gov/geo/maps-data/data/tallies/tractblock.html.
3 Location Modeling for Logistics Parks 65

Table 3.3 Various scales for Type Number Average size (mi2 )
data for North Carolina
County 100 486.19
Zip code (or ZCTA) 1048 46.39
Tract 2195 22.15
Block group 6155 7.90
Block 288,987 0.17

In what follows, we will compare several different solutions for finding the 1-
median of NCSE. As mentioned above, the key difference in each solution is a
different answer to how the demand is being calculated and what distance metric
is being used. As for distance metrics, we will show the solutions for four different
distance measures: p = 1 (l1 or Manhattan metric, given in (4)), p = 1.5, p = 2 (l2 or
Euclidean metric given in (2)) and the squared Euclidean metric described in (3).
As for estimating the demand points i = 1,2,. . ., N, we will also compare using two
different weights: the number of people living in each zip code3 in the 2010 census
of the population and how many employees were recorded working at businesses
in each zip code in the 2012 economic census. Because codes often cross county
boundaries, zip codes were selected if their centroids fell within the set of counties
under study (Fig. 3.4).

Fig. 3.4 Different solutions to the 1-median problem for NCSE

3 Or ZCTA (Zip Code Tabulation Area), an approximation of zip code boundaries using census
data.
66 J. Bhadury et al.

The results from the computation of the 1-median as given in (5) are shown
in Fig. 3.4 and the endings EMP and POP are for points using employment and
population as the demand weights, respectively. Here, l1 , l1.5 , and l2 represent the
distance metric used, and MEAN denotes the center of mass.
These results lend themselves to some interesting observations. First, note that
all of the solutions result in locations in northern Bladen County, whose county seat
is Elizabethtown with a population of around 4000 people.
Second, we see that all of the measures using employees as the weights are to
the southeast of their counterparts weighted on population, which may be explained
as follows. There are only two major cities in this region of 14 counties: Fayet-
teville/Fort Bragg, and Wilmington. Of the two, Fayetteville/Fort Bragg has a much
higher population represented in its zip codes (256,233 vs. 186,140), but the num-
ber of employees counted in the economic census is about the same (78,155 vs.
77,164). Despite this, the observed shift to the southeast is likely to be due to the
fact that while Fort Bragg and surrounding ZIP codes have a large population, rel-
atively few of those employees, who are primarily employed in the military sector,
would be measured by the economic census4 . Therefore, one can expect areas with
large military bases, universities, or other categories of employer not considered to
be “industries” in the economic census are likely to see sizeable variation in results
based on whether population or employees are chosen as weights.
A third interesting observation is that the centers of mass (labeled MEANPOP
and MEANEMP in Fig. 3.4) are to the southeast of the 1-median obtained using l1 ,
l1.5 , and l2 distance measures. The difference comes from the underlying objective
of the two types of measures. While with the l1 , l1.5 , and l2 we minimize the sum of
the weighted distances, the center of mass minimizes the sum of squared distance.
This means that points farther away will have a larger influence on the center of
mass, because the square of a large distance becomes much larger than the sum of
many medium-sized distances. Thus the distant lying sparsely populated regions of
NCSE have tended to “pull” the center of mass in their direction.
Finally, note from Fig. 3.4 that all the solutions to the 1-median problem for l1 ,
l1.5 , and l2 metrics in NCSE occur in an area that is largely agricultural land, and
not in the middle of cities or towns. Also, several state highways (seen labeled with
numbered ovals in Fig. 3.4) run nearby many of these locations, also with fairly con-
venient access to federal and interstate highways. This makes these sites well-suited
for the location of a logistics park. Despite the same, Sect. 3.4 will make evident
that none of the initial candidate sites selected in the actual implementation of the
project were even proximate to the location indicated in Fig. 3.4. This discrepancy,
in turn, is behind one of the key conclusion of this chapter that analytical results are,
at best, only a small part of location modelling in practice. Other factors, be they
social, economic or political, are usually much more important in influencing final
outcomes.

4 The economic census covers businesses of all sizes, but excludes most government-owned

industries. Also excluded are schools (primary through university, including private
schools), agriculture, and religious organizations (see http://www.census.gov/econ/census/help/
naics_other_classification_systems/codes_not_covered.html for a complete list).
3 Location Modeling for Logistics Parks 67

3.3 Location in Practice: Site Selection Process and Analysis

This section presents a synopsis of the actual approach taken in modelling the site
selection process for a Logistics Park in NCSE and is based in large part on Bhadury
and Troy (2011). The process followed may be structured formally within a frame-
work that we introduce here and refer to by its acronym SIRC, which stands for the
following steps that need to be executed in their prescribed order.
Step 1 Situational analysis of the region where the logistics park will be located.
Step 2 Initial selection of candidate sites for locating the logistics park in the region.

Step 3 Readiness assessment for each of the candidate sites selected in Step 2 above.
Such assessment should include current infrastructure, desired infrastructure for
peak performance of the logistics park and the gap between the two.

Step 4 Competitive summary of the candidate sites, stating strengths and weak-
nesses of each. This step is obviously based on the data collected in the prior three
steps.

While the SIRC framework is general enough to be used for any real-life location
modeling problem, the remaining portion of this section is devoted to illustrating its
implementation in the context of the current application.

3.3.1 Step 1 of SIRC: Situational Analysis

The application of this first step of SIRC comprised of collecting social and eco-
nomic information about NCSE that is relevant to a public works project such as
the location of a logistics park. Such information comprised of population, em-
ployment, labor wages and their trends. Information was also gleaned about the
industry clusters that were targeted for development by the regional economic de-
velopment agency. Besides studying published reports, 14 interviews were held with
34 selected industry leaders, economic development officials, public officials and se-
nior management from the existing logistical facilities such as Port of Wilmington,
Wilmington International Airport, Fayetteville Airport, Laurinburg-Maxton Airport
and International Logistics Park to name a few. The objective of these interviews
was to collect background information and assess the current and anticipated lo-
gistical needs of the region and their views on the Logistics Park in NCSE. The
results from this first step of Situational Assessment forms the basis of most of the
information presented in Sect. 3.2 above.
68 J. Bhadury et al.

Fig. 3.5 Initial candidate sites

3.3.2 Step 2 of SIRC: Initial Selection of Candidate Sites

The next step in the location modelling process is to select initial candidate sites
for location. In this case, this was done on the basis of the data collected in the first
step and an analysis of all available sites across the NCSE. The criteria used for the
selection were sites proximate to highway, rail, and air access and having at least
400–500 acres of developable land to accommodate the needs of the companies
that were expected to locate in the logistics park. Additionally, the opinions and
ideas of the local economic and political leaders, as gleaned from the interviews,
weighed heavily on the decision making process. Based on all the above, four sites
were finally selected as candidate sites for locating a logistics park5 and as depicted
in Fig. 3.5, they were as follows: Laurinburg-Maxton Airport (MEB); Fayetteville
Regional Airport (FAY); International Logistics Park of NC (ILP) and Wilmington
International Airport (ILM). Additionally, as a result of numerous interviews and
research conducted to evaluate the four designated sites, the research team added
a fifth possible site, a “virtual logistics park” for the Greater Wilmington Area. A
brief overview of each site is presented below.

5 An analytical approach to the modelling would have selected all sites in NCSE that matched the

criteria being used, consider this entire set as initial candidates for location and determine the opti-
mal location of the logistics park by solving the for the 1-median of this resulting discrete location
problem. Nonetheless, other factors (preferences of interviewees, need for economic development
in economically depressed counties, need to promote sites that were already developed etc.) pre-
vailed and contributed to only the above four sites from this set being selected as candidate sites in
Step 2 of the SIRC process.
3 Location Modeling for Logistics Parks 69

1. Laurinburg-Maxton Airport (MEB): The Laurinburg-Maxton Airport (see

http://www.lmairport.com) is a small airport that is located in a rural region
that borders Scotland and Robeson counties and is operated by the Laurinburg-
Maxton Airport Commission that reports to the governing bodies of the cities
of Laurinburg and Maxton. No commercial flights operate at MEB and its chief
current use is for military training. Among the features of this small airport are a
6500-foot lighted runway, equipped with high-intensity lights, new LED taxiway
edge lights, signage and pavement marking and a new and full ILS (instrument
landing system). It also serves as the last landing site for many planes that are
targeted to be decommissioned. MEB offers significant opportunity for develop-
ment as a result of the availability of developable land, adequate and improving
roads, and ready access to most infrastructural needs. In addition, as a direct re-
sult of its location, MEB can serve as a link to the Port of Charleston making it an
attractive site as a distribution center for both North and South Carolina as well.
However, the airport currently lacks a strategic plan and facilities need major
costly improvement and while there are immediate plans and funds available to
improve two of the three runways, the improvements do not fill all the needs and
requirements to make MEB truly competitive. In addition, the area has access to
national railroads operated by CSX but to take full advantage of these resources
costly rail connections are needed. Finally, while most infrastructure is available
at the MEB site, much of it needs substantial upgrades and retrofitting.
2. Fayetteville Regional Airport (FAY): Fayetteville Regional Airport (http://www.
flyfay.com/) is located in Cumberland County and is operated by an airport direc-
tor for the City of Fayetteville. The primary advantage of this as a possible site
for a logistics park is proximity to potential demand given the industrial presence
around Fayetteville and the growing military sector in Fort Bragg. The site is ge-
ographically well located with regards to US interstates. It has access to CSX,
Norfolk Southern and short line railroads; while these connections are not on the
airport site, they are all located within the City of Fayetteville. The area is also in
close proximity to major interstate highways, the Port of Wilmington, Research
Triangle Park, recreational facilities and developable land including shovel ready
sites. In addition, and perhaps most significantly, it is located near an expanding
military facility of Fort Bragg.
3. International Logistics Park (ILP): The International Logistics Park of North
Carolina (http://brunswickedc.com/sites-buildings/available-sites/international-
logistics-park-of-nc) is built on the Columbus/Brunswick County line on US
74/76 just 15 miles from the Port of Wilmington. This is important because the
Port of Wilmington promotes the “At Port” site location model whereby facilities
(e.g., warehouses, distribution centers) located within 20 miles of the port enjoy
the same state, regional, local and port tax incentives as those located within the
port premises and a lower rate from trucking companies. ILP is a joint venture
between Brunswick and Columbus Counties, and has an undeveloped 1100 acre
park that hopes to capitalize upon the “at port” site location promoted by the
Port of Wilmington. The most significant attribute of ILP is the vast amount
of developable land that includes shovel ready sites where utilities, gas, wa-
ter electricity and sewer are readily available. Further, the existing roads and
70 J. Bhadury et al.

planned road projects make ILP accessible, especially to the Port of Wilming-
ton and Wilmington International Airport. Although ILP lacks a rail connection,
located directly across the street is the site of another large industrial park, the
Mid-Atlantic Logistics Center, that has CSX rail access.
4. Wilmington International Airport (ILM): The Wilmington International Airport
(http://www.flyilm.com/) is located in New Hanover County and serves southeast
North Carolina. The airport, located off I-40 and I-140, is operated by the New
Hanover County Airport Authority. Although underutilized, it is a full-service
airport that offers commercial, cargo and general aviation facilities and a state
of the art Federal Inspection Services (FIS) facility. Runways are adequate and
expandable to meet demand. Shovel ready industrial sites that meet the Port of
Wilmington’s “at port” location criteria are available at the airport and larger
tracks of developable land are located nearby in Pender, Brunswick and Colum-
bus Counties. The existing basic infrastructure (communications, water, sewer
and power) at ILM may be classified as being of average quality. The airport also
has at-site rail service with an indirect link to the Port of Wilmington. In addi-
tion, the high quality of life in the area provides the ability to attract highly skilled
workers. However, a weakness for ILM, like the nearby International Logistics
Park, is its risk exposure to the future viability of the Port of Wilmington.
5. A “Virtual Logistics Park” for the Greater Wilmington Area: The final assess-
ment of the research team was an innovative idea that, to the best of knowledge
of the authors, has never before been mentioned in the location modeling liter-
ature. This was to create a “virtual logistics park” from the Greater Wilmington
area that comprises of all logistics assets of the region and is coordinated by a
central organization to work synergistically as a unified and coherent institution.
These assets include: Port of Wilmington, ILM, ILP discussed above but also two
additional industrial parks in this region, namely, Mid-Atlantic Logistics Center
and the Pender Commerce Park all of which are within 30 miles of each other.
There are numerous reasons in support for this idea. First, geographical proximity
puts all of these existing industrial parks within the requirement to be considered
“at port” with regards to the Port of Wilmington. Second, the “virtual logistics
park” collectively has all modes of transportation—shipping (Port of Wilming-
ton); air freight (ILM), rail (CSX rail access at ILM, Port of Wilmington as well as
Mid-Atlantic Logistics Center) and abundant trucking services. Additionally, de-
velopable land parcels of all sizes, many shovel-ready, are found in the metro area
and/or ILP and the Greater Wilmington area has excellent utilities infrastructure
(communications, water, sewer and power) to accommodate growth. Therefore,
all that would be needed to create a “virtual logistics park” would be some minor
infrastructure needs and for the area’s political and economic leaders to work
together and cooperatively in creating an umbrella organization to administer
this park as one coherent entity. And it is this last condition that engenders the
most important challenge involved with the “virtual logistics park,” namely, the
practical difficulties involved in initiating such an entity that brings all these dis-
parate organizations under one umbrella and thereafter, getting them to work
cooperatively towards common regional goals rather than individual ones.
3 Location Modeling for Logistics Parks 71

3.3.3 Step 3 of SIRC: Readiness Assessment of Candidate Sites

This step is the heart of the location modeling process espoused by the SIRC frame-
work. As implied in the title, this step was devoted to conducting an in-depth
assessment of each of the five candidate sites mentioned in Step 2 above and on
the basis of the same, develop a comprehensive summary of the site’s logistical as-
sets, shortcomings and an estimate of what improvements would be needed to locate
a regional logistics park at that site and time and cost estimates of the same. After
reviewing existing practitioner literature and speaking to local economic develop-
ers, it became clear that for each site, the research team would have to focus on all
infrastructure within a 10 and 60 miles radius since those would be the most heavily
impacted areas if a logistics park was located at the site. In order to assess the local
workforce available to work at companies expected to operate within the logistics
park, existing commuter data for the state indicated that examining the workforce
qualifications within a 30 min drive time from each site would provide the best in-
formation. Finally, it was important to study and assess the facilities available at the
site itself. As an example of this portion of the location modeling, see Fig. 3.6 for an
example of the 60 miles radius area examined for the FAY site. Also, Fig. 3.7 shows
the 30 mile radius that comprises the “virtual logistics park.”
After examining the areas mentioned above, three tables were produced for each
site. The first table, labeled the “assessment matrix” presented an overall summary
of the readiness of each site. The second table, labeled “site summary” presented
detailed information on existing infrastructure such as roads, airports etc. at each
site and the improvements to them that would be needed in order to locate a logistics
park at the site. The third and final table was based on the first two and summarized
the strengths, weaknesses and recommendations for each site. See Tables 3.4, 3.5
and 3.6 for examples of what was produced for the FAY site.
This was followed by a narrative explaining the basis for the information pre-
sented in the three tables. For example, the narrative accompanying the assessment
of the FAY site was as follows (see Bhadury and Troy 2011):
“The Fayetteville Regional Airport (FAY) is located in Cumberland County and is
operated by an Airport Director for the City of Fayetteville. Cumberland County is a
Tier 2 county and has an unemployment rate of 9.2 % (November 2010) and is home
to an increasingly important military base, Ft. Bragg. . . . “Proximity” is the key
word needed to understand the Fayetteville Regional Airport area and its potential
as a ‘logistics village.’ Fayetteville is located half-way between Miami and Maine.
It is on I-95 and 30 min away from I-40. Although not at the airport, the area has
rail access on CSX and NS as well as a local short line railway. It is near the Port of
Wilmington, Research Triangle Park and recreational facilities at the ocean and near
the Pinehurst resorts. In addition, and perhaps most significantly, it is in very near the
expanding Ft. Bragg. Airport facilities are in good condition with a 6500 ft. runway
that could be expanded to 8500 ft. The airport has several parcels of land available
for development and adjoining land in private hands could be developed as well.
Furthermore, additional industrial land is available in the immediate area. Some of
72 J. Bhadury et al.

Fig. 3.6 60 mile radius for the FAY site. (Source: http://www.ncdot.gov/doh/preconstruct/tpb/
research/download/2010-34-3southeastregionreport.pdf)

the land is ‘shovel ready.’ Access from the airport to I-95 is excellent. To reach its
potential as a ‘logistics village,’ FAY would have to develop cargo facilities. It is
currently considered a ‘truck’ market. Water and sewer issues at the airport would
need to be addressed and should include a pumping station for sewer service. To help
make some of this happen, the Airport Plan and the Fayetteville Area Plan should
be coordinated to maximize the economic development benefits of the airport.”
Given the novelty of the idea of creating a “virtual logistics park”, see Table 3.7
and Table 3.8 for the assessment matrix and for the strengths, weaknesses and
recommendations regarding the same.
Based on Table 3.7 and Table 3.8, the narrative assessment for the “virtual
logistics park” was presented as below (Bhadury and Troy 2011).
3 Location Modeling for Logistics Parks 73

Fig. 3.7 30 mile radius for the “virtual logistics park”. (Source: http://www.ncdot.gov/doh/
preconstruct/tpb/research/download/2010-34-3southeastregionreport.pdf)

“. . .the Greater Wilmington Area is strategically positioned to capitalize upon

the area’s logistics assets and establish a ‘virtual logistics park.’ A ‘virtual logistics
park’ for the Greater Wilmington Area would not only incorporate two sites under
consideration in this study, the International Logistics Park and Wilmington Inter-
national Airport but also the Port of Wilmington, Pender Commerce Park and the
Mid-Atlantic Logistics Center. In fact, the ‘virtual logistics park’ would include all
of the logistics assets within a 30 mile radius of Wilmington International Airport.
While all components of a ‘virtual logistics park’ are not located in one spot in the
metro area, they are all in close enough proximity to each other to satisfy the ‘at
port’ logistics model criteria for the Port of Wilmington. This is a significant factor
that will serve as a strong foundation for developing logistics assets in the Greater
Wilmington Area.
74 J. Bhadury et al.

Table 3.4 Assessment matrix for FAY Site. (Source: http://www.ncdot.gov/doh/preconstruct/tpb/

research/download/2010-34-3southeastregionreport.pdf)
Measure Fayetteville regional airport (FAY) area
Facility typea Commercial and general aviation
Target industries Defense/homeland security; distribution; aerospace; general
manufacturing
Speed of implementation 5–10 years
Infrastructure—transport
Highway Very good
Rail Not at site but good connections in metro area
Air
Existing runways Two-only one FAA supported Lengths: 7712’ & 4800’ ILS
Category 1 capability
Can runways be extended? Yes, at a cost
Maritime (ship/barge) 92 miles to port of wilmington; fayetteville is accessible by
barge via cape fear but not currently operational
Infrastructure—other
communications Very good
Power Adequate
Water Improvements needed
Sewer Improvements needed-pumping station
Land available (acreage) Some existing sites available at airport, improvements needed
for other sites
Labor force
Availability Yes, including skilled labor
Education programs to support Yes
local industry
Specialty criteriab Fayetteville economic development seeking FTZ; no cargo
facilities
a Facility type major business practice at this location (intermodal facility, warehouse, distribution

center, light manufacturing, and so forth)

b Specialty criteria foreign trade zone, customs, high-security, and so forth

The ‘virtual logistics park’ should be established, developed and coordinated by

area leaders through the creation of a central ‘facilitation’ board. . . . The combined
logistics assets of a ‘virtual logistics park’ far outweigh the value of Wilmington In-
ternational Airport and the International Logistics Park when evaluated separately.
. . . Most significantly, for the short-term, the ‘virtual logistics park’ is essentially
ready to begin operation without additional major investments as the infrastructure
already exists to support this effort. In fact, only two things are immediately needed
to support the creation of a ‘virtual logistics park,’ the establishment of a coordi-
nating board and a lighted intersection at the entrance to ILP and the Mid-Atlantic
Logistics Park.”
3 Location Modeling for Logistics Parks 75

Table 3.5 Site summary for FAY. (Source: http://www.ncdot.gov/doh/preconstruct/tpb/research/

download/2010-34-3southeastregionreport.pdf)
Facility Current status Improvements needed
Roads Adequate plus See roads recommendations
Airports Central location, halfway Runway/taxiway upgrades
between Maine and Miami needed (see water/sewer)
Moderate commercial and Cargo facility required
general aviation facility (no Additional sources for capital
cargo facility) improvements
One FAA supported runway
which can only handle narrow
body aircraft. Cash flow positive
but limits growth (debt free)
Railroads Inadequate at FAY but good Spur required if cargo issue is
metro access; FAY 5–8 miles resolved; costly
from main line
Ports 92 miles to port of Wilmington
with adequate highway
Shovel ready sites Moderate airport sites available Water and sewer improvements
in NW corner with additional needed for development of
private tracts available existing property
Workforce Better than adequate skilled Needs to be promoted for
labor & technical labor available recruitment of industry to area
(military discharges around 6000
per year)
Transport servicing Trucking facilities are plentiful
Communication Good to very good
Power Adequate, owned by city
Water/sewer Adequate, presuming “no Pumping station at FAY for
Water/sewer (continued) growth” additional sewer capacity
No sewer facilities on Southside
of airport
Taxiways need remediation for
water problems Overall
improvements needed for future
growth
Specialty criteria Applying for FTZ, location
undetermined

3.3.4 Step 4 of SIRC: Comparative Summary of Candidate Sites

This is the final step in the SIRC framework and is made necessary by the fact
that in most real-life location problems, the decision makers do not ask for a single
prescriptive recommendation, opting instead to be presented with alternate solu-
tions with rationale for each so that they can make the final decision. Therefore,
after completion of the individual assessments for each of the five candidate sites
76 J. Bhadury et al.

Table 3.6 Strengths, weaknesses, and Recommendations for FAY Site. (Source: http://www.ncdot.
gov/doh/preconstruct/tpb/research/download/2010-34-3southeastregionreport.pdf)
Strengths Weaknesses Recommendations
Proximity (highways, Ft. No cargo facility See that city and economic
Bragg, recreation, port, RTP, No at site rail connection development proactively
etc.) Port 92 miles away Only one recruit a champion, military is
Ft. Bragg expansion will FAA supported runway that a good candidate to spur
stimulate airport growth can handle narrow body growth and provide dollars for
(commercial/industrial, aircraft logistics village at airport.
general aviation, and military Additional capital sources are Update Fayetteville area plan
related traffic) not currently available and coordinate with airport
Military outsourcing leads to No updated fayetteville area plan in order to promote
airport growth potential plan and airport plan not tied economic development at FAY
(hangers, runways, land sites) to economic development Correct water and sewer issues
Better than adequate skilled driven issues to maximize growth potential
and technical labor available Water and sewer issues need to
(around 6000 military be addressed
discharged per year) Limited number of large
Roads are better than adequate parcels of land for growth
in all directions Expansion of runways will be
expensive

in Sect. 3.4.3 above, a comparative summary was drawn up for all the five sites.
This summary ensured that it did not recommend one site over another. Instead,
it attempted to compare the relative strengths and weaknesses of each site so that
the ultimate decisions makers (NCDOT and North Carolina State Legislature) could
choose the final location for the logistics park.
The first part of the comparative summary was to enlist the common charac-
teristics of all five sites studied. Some examples were as follows (Bhadury and
Troy 2011):
• “The area shares an excellent network of existing roads including interstates, lim-
ited access highways and four-lane highways. Also, completion of planned road
construction for improvement of the regional highway infrastructure will further
enhance area transportation and provide a stronger foundation for a distribution
and logistics based cluster for NCSE.
• While not excellent, the region generally has good access to railroads, CSX, Nor-
folk Southern and numerous short line connections. Improvements such as the
establishment of an intermodal rail/truck facility in or near the NCSE region that
has a direct rail link from the Port of Wilmington and the reestablishment of some
preserved rail corridors could improve the attractiveness of the area to industry
especially for supply chain based companies.
• The region has a strong agricultural and food processing base. Additionally, the
growing military presence in the Fayetteville area and the Port of Wilmington are
regional assets and are critical components for economic development in NCSE.”
Thereafter, an added narrative was devoted to comparing the site specific resources
and infrastructure needs. Specific topics addressed were: highway infrastructure,
3 Location Modeling for Logistics Parks 77

Table 3.7 Assessment matrix for the “virtual logistics park” for the greater Wilmington
area. (Source: http://www.ncdot.gov/doh/preconstruct/tpb/research/download/2010-34-3south-
eastregionreport.pdf)
Measure Virtual logistics park
Facility typea Airport, trucking, rail, port, logistics and distribution, warehousing
center, and manufacturing
Target industries See ILM, ILP and NCSEs targeted clusters
Speed of implementation Immediate
Infrastructure—transport
Highway Good to very good and improving (see “cape fear commutes 2035:
transportation plan”)
Rail Available with limitations
Air ILM
Existing runways Two at ILM Lengths: 8000’ & 7004’ ILS category 1 capability
Can runways be Yes, at a cost
extended?
Maritime (ship/barge) Port of Wilmington
Infrastructure—other
Communications Good to excellent
Power Adequate
Water Adequate, but secondary water source needed. Overall area has
future water concerns
Sewer Septic and traditional available
Land available (acreage) Excellent- numerous shovel ready of all sizes
Labor force
Availability Overall excellent- some skills training may be needed
Education programs to Yes
support local industry
Specialty criteriab ILM has full-service customs FIS facility; area meets “at port”
logistics model; inactive FTZ at port could be reactivated and
sub-zones established as needed at other “virtual logistics village”
sites
a Facility type major business practice at this location (intermodal facility, warehouse, distribution
center, light manufacturing, and so forth)
b Specialty Criteria foreign trade zone, customs, high-security, and so forth

railroads, air connections, utilities (including broadband), availability of devel-

opable land, labor force availability and specialty criteria (such as the need to
establish free trade zones at sites other than the Port of Wilmington).
Two examples of such narratives from the report (Bhadury and Troy 2011) are
presented below. The first is for the highway infrastructure at these sites and the
second focusses on the amount of developable land available.
78 J. Bhadury et al.

Table 3.8 Strengths, weaknesses, and recommendations: “virtual logistics park” for the greater
Wilmington area. (Source: http://www.ncdot.gov/doh/preconstruct/tpb/research/download/2010-
34-3southeastregionreport.pdf)
Strengths Weaknesses Recommendations
“Virtual logistics All transportation Absence of strong Enhance coordination
village” for the greater modes coordination with a “facilitation”
Wilmington area Adequate Highway and rail group (e.g., similar to
infrastructure access (“the last mile”) Aerotropolis
Shovel ready sites Risk exposure to future Leadership Board for
International trade viability of the Port of PTI)
(Port of Wilmington) Wilmington Capitalize on “At Port”
Major industrial parks logistics model
including: international Improve highway and
logistics park, rail access
Mid-Atlantic logistics Continue to invest in
center, pender maritime activities
commerce park

“Highways: In general, all five sites have good road access and planned improve-
ments will only make access easier. When I-74 is complete to the Piedmont Triad
and beyond, NCSE access to the Midwest will improve as truck traffic will be able to
bypass the congested Raleigh area. This is especially true for the International Lo-
gistics Park (ILP), the Wilmington International Airport (ILM), the ‘virtual logistics
village’ and the Laurinburg-Maxton Airport (MEB) sites. For these same sites, The
Monroe Bypass, along with the completion of I-74 to Rockingham, and completion
of I-140 in the Wilmington area will provide better access to the Charlotte market
and beyond. Key to the near term success of the ILP is a full lighted intersection
at the entrance to the Mid-Atlantic Logistics Center and ILP. In the future the state
should also consider building a full interchange at the intersection of Highway 87
and US 74 that would provide links via access roads to not only ILP but the Mid-
Atlantic Logistics Center and allow the removal of the interim lighted intersection.
While the Port of Wilmington by itself is not a targeted site, it is part of the ‘vir-
tual logistics village’ and this site will benefit significantly from all road projects
and will help the Port to implement its ‘at-port’ logistics model and become more
globally competitive. Planned road projects for the Fayetteville area will slightly
enhance traffic around the Fayetteville Regional Airport (FAY); however, overall,
FAY already has excellent immediate access to I-95 and access to I-40 is just 30 min
away via I-95.”
“Developable land: each of the five possible sites studied for a ‘logistics village’
has land available for development. Similar to the Global TransPark, MEB has land
available to meet almost any need. Some of that land is already shovel ready and
located in a certified industrial site fronting the future I-74. The ILP has 1100 acres
available and, immediately across the street from ILP, the Mid-Atlantic Logistics
Center has an additional 1100 acres with rail access. FAY has some parcels immedi-
ately available but infrastructural improvements would be needed to build on other
existing parcels. Additional land is also available near the airport. ILM has around
150 acres available at site and other possible sites adjoin the airport. If the ‘virtual
3 Location Modeling for Logistics Parks 79

Table 3.9 Comparative summary for all candidate sites

Site Strengths Weaknesses Needs
Laurinburg-Maxton Significant capacity for Facilities require Needs champion
airport (MEB) expansion significant costly (private/government)
Proximity to Ft. bragg improvements
and camp McCall Lack of strategic plan
for airport
Fayetteville airport “Proximity” Lack of capital sources Proactively recruit a
(FAY) Ft. bragg expansion for infrastructure champion (military?)
Adequate labor improvement Update Fayetteville
available Disconnect between area plan and
airport plan and coordinate with airport
economic development plan
Correct water and
sewer issues
International Logistics Abundance of shovel Risk exposure to future Planned road projects
Park (ILP) ready sites viability of the Port of (full intersection)
Meets “at port” criteria Wilmington Cooperate with
Tier 1 (economically Mid-Atlantic logistics
disadvantaged) county center
status makes it eligible Leverage regional
for tax incentive logistics assets
support from state and
local governments
Wilmington Modern full-service Risk exposure to future Leverage regional
international airport (FIS) airport viability of Port of logistics assets
(ILM) Shovel ready industrial Wilmington Utilize professional
sites skills of local retirees
Quality of life
“Virtual logistics park” All the strengths of Risk exposure to future Form a coordinating
in the greater regional logistics viability of Port of board that can leverage
Wilmington area assets including ILP Wilmington regional logistics
and ILM above assets in creating this
“virtual” organization

logistics village’ is considered significant, developable land is available at ILP, the

Mid-Atlantic Logistics Center, and Pender Commerce Park and at or adjacent to
ILM. A large percentage of this land is shovel ready and meets the criteria for the
Port of Wilmington’s ‘at port’ logistics model.”
The complete comparative summary is available in Bhadury and Troy (2011);
key points from that are captured in Table 3.9.

3.4 Epilogue and Conclusions

This final and most important section of the case study will first describe the
epilogue of events after the completion of the Seven Portals Study. Thereafter,
conclusions are presented on the lessons learned from the study as well as its
implications for future research in location theory.
80 J. Bhadury et al.

After the completion of the research, the key findings were presented in early
2011 to Governor’s Logistics Task Force (GLTF) in the form of the final report
Bhadury and Troy (2011). That report became a part of the final report submitted by
the lead research team of the Seven Portals Study project, see List and Foyle (2011).
In turn, List and Foyle (2011) became an integral part of the report submitted by
GLTF, namely, the GLTF Final Report (2012). However, in 2012, all recommenda-
tions from prior studies undertaken by NCDOT, including the Seven Portals Study,
were placed on hold, and NCDOT announced that it would substantially change
the way transportation infrastructure in North Carolina was financed. Subsequently,
NCDOT announced a new method for evaluating infrastructure projects such as the
ones recommended in the Seven Portals Study. This new method is referred to as
Strategic Transportation Investments and a description of the same is available at
http://www.ncdot.gov/strategictransportationinvestments/. As of the end of 2014, no
logistics park had been located by NCDOT anywhere in the state directly as a result
of the Seven Portals Study. Nonetheless, smaller recommendations from the study
were implemented; for example, cold storage facilities were added at the Port of
Wilmington in order to facilitate exports of agricultural products.
There are numerous lessons to be learned by location theorists from an actual
application such as the one described in this chapter. Three key lessons are as
follows:
1. Most Problems Involve Location of a Single Facility: In most cases in real life,
site location problems consider the location of only one facility. In case multiple
facilities are involved, the decision-makers usually divide the area to be served
into smaller regions and focus primarily on locating one facility in each of them,
much like the original Seven Portals Study was broken up into smaller, regional
ones, each involving the location of one logistics park in a specific region of the
state.
2. Most Problems Have Multiple, often Conflicting, Objectives: In real life location
problems, decision-makers have multiple objectives, some of which are not even
quantifiable. This makes it almost impossible to determine an optimal location. In
fact, the fuzziness of some of the objectives, as for example, the objective to stim-
ulate economic development through location of logistics parks in the case of the
Seven Portals Study, make it questionable if an optimal location can even be de-
fined that is acceptable to all decision makers. For example, whereas minimizing
travel distance might attract the location of a facility towards the demand centers
(usually high population urban centers), governments locating public facilities
also have a tendency to consider sites in high unemployment areas (usually
sparsely populated rural areas) so as to stimulate economic development. There-
fore it is important that in the analysis, researchers stick to evaluating strengths
and weaknesses of candidate sites with regards to the various objectives of the
decision-makers rather than being prescriptive and recommending one particular
site.
3 Location Modeling for Logistics Parks 81

3. Rigorous Scientific Analysis is a Small Part of The Modeling Process6 : When

locating public facilities, especially large and expensive facilities such as logis-
tics parks, theoretical analysis such as the one presented in Sect. 3.3, is at best a
small part of the actual location modeling process. In fact, as is evident from this
chapter, the 1-median locations determined by the analytical process of Sect. 3.3
and as presented in Fig. 3.4 are not even proximate to any of the initial candi-
date sites selected by the research team (Fig. 3.5) on the basis of the Situational
Analysis step of SIRC. This exemplifies that site selection of public facilities in
practice mostly involves factors other than analytical ones. Public facilities such
as logistics parks are viewed by the citizenry as job creators as well as a nuisance
(causing congestion, pollution etc.). As a result, public perception, as reflected
directly by the citizenry, as well as through their elected leaders and/or special in-
terest groups and the clout that these have in the decision-making process have a
far larger bearing on the modeling process in practice than does the mathematical
analysis presented in Sect. 3.3.
The import of lessons 2 and 3 is an important guideline for location theorists who are
called upon to perform location modelling of public facilities. The guideline is that
in Step 1 of the SIRC framework (namely, situational analysis), researchers must
first identify all stakeholders involved and make sure that all necessary background
information is collected about them. Such background research should involve a
study of the socio-economic profile of the region as well as their future trends, as
we presented in Sect. 3.2. In addition to this in-depth study of the region being con-
sidered for location, it is also important for researchers to get to know the important
social, economic and political figures in the region and interview them to find out
their expectations from the facility being located as well as the process that ought
to be used in determining the location. That is why the research team conducted
14 interviews with 34 different people in NCSE as a part of completing the Situa-
tional Analysis for this project. Not only does this create the information that forms
the bedrock of the subsequent steps (Steps 2–4 of the SIRC framework), but also
lends credibility to the work being done which, in turn, lays the groundwork for the
eventual acceptance by stakeholders of the findings and conclusions of the research
project.
These lessons presented above also lead to two important recommendations
about future research in location theory and we conclude this case study with the
same. The first of these points to the importance of single-facility location models.
While multi-facility location models are theoretically much richer than single facil-
ity location models, more research needs to be focused on the latter. The primary
source of complexity in real life location models comes not from having to locate
multiple facilities but from having to locate one with numerous, often conflicting
and non-quantifiable, objectives. Thus, it is our recommendation that more work

6A member of the research team succinctly stated this principle as “ Location in practice is 80 %
politics and 20 % science”.
82 J. Bhadury et al.

needs to be done on multi-objective, single facility location models where some of

the objectives are allowed to be fuzzy and/or qualitative. By nature, such models
will need to combine quantitative analysis with qualitative research and should add
to the richness of extant literature in location theory.
The second observation about future academic research stresses the importance
of exploring alternate optimal and near-optimal solutions in location models. As
mentioned above under lessons learned, it is difficult to define “optimality” with a
high degree of precision for most real life location problems. As a result, decision-
makers look not for a single prescriptive recommendation but a range of alternate
recommendations that do “well enough” on most objectives. In the language of op-
timization, this implies that research in location theory should focus on efficient
ways to generate alternate optimal locations as well as near-optimal solutions to
complex, multi-objective problems. The goal of location modeling in practice is to
then present the best of these alternate and near-optimal solutions to the decision-
makers along with strengths and weaknesses of each to enable them to make final
decisions.

Acknowledgment This work draws, in part, upon the experiences of the first and the third author
in the implementation of the Seven Portals Study project. That project was supported by NCDOT,
which is gratefully acknowledged.

References

Bhadury J, Troy SP (2011) Seven portals study—an investigation of economic development

in North Carolina through logistics villages (SouthEast Region Report). http://www.ncdot.
gov/doh/preconstruct/tpb/research/download/2010-34-3southeastregionreport.pdf. Accessed
10 April 2015
Brimberg J, Dowling PD, Love RF (1994) The weighted one-two norm distance model: empirical
validation and confidence interval estimation. Location Science, 2:91–100
Brimberg J, Love RF (1995) Estimating distances. In: Drezner Z (ed) Facility location. Springer,
New York, pp 9–31
Burkey ML, Bhadury J, Eiselt HA (2011) Voronoi diagrams and their uses. In: Eiselt HA, Marianov
MV (eds) Foundations of location analysis. Springer, New York, pp 445–470
Current J, Daskin M, Schilling D (2004) Discrete network location models. In: Drezner Z,
Hamacher HW (eds) Facility locations: applications and theory. Springer, Berlin, pp 81–118
Eiselt HA (1992) Location modeling in practice. Am J Math Manage Sci 12(1):3–18
Eiselt HA, Marianov V (2011) Foundations of location analysis. (Vol. 155). Springer Science &
Business Media, Berlin
El Amrani A (2007) The impact of international logistics parks on global supply chains. (Doctoral
dissertation, Massachusetts Institute of Technology)
Fernández J, Fernández P, Pelegrin B (2002) Estimating actual distances by norm functions: com-
parison between the k, p, θ-norm and the ib1, b2, θ-norm and a study about the selection of the
data set. Comput Oper Res 29:609–623
Findley DJ, Small JD, Tran W, Heller A, Bert SA, Searcy SE, Hall WW (2014) Economic con-
tribution of the North Carolina ports. http://www.ncports.com/elements/media/files/economic-
contribution-north-carolina-ports.pdf. Accessed 10 April 2015
3 Location Modeling for Logistics Parks 83

Francis RL, Lowe TJ, Rayco MB, Tamir A (2009) Aggregation error for location models: survey
and analysis. Ann Oper Res 167(1):171–208
Gardner, Kansas Official Website (2015) What is a logistic park. http://www.gardnerkansas.gov/
images/uploads/Administration/Intermodal/WhatisaLogisticsPark.pdf. Accessed 10 April 2015
Governor’s Logistics Task Force, Final Report (2012) http://www.ncdot.gov/download/business/
committees/logistics/GovernorsReport_Jun2012.pdf. Accessed 10 April 2015
Hamacher HW, Drezner Z (eds) (2002) Facility location: applications and theory. Springer Science
& Business Media, Berlin
Lee H, Yang HM (2003) Strategies for a global logistics and economic hub: Incheon international
airport. J Air Transp Manage 9(2):113–121
Lee KL, Huang WC, Kuo MS, Lin SC (2001) Competitiveness model of international distri-park
using the virtual value chain analysis. J East Asia Soc for Transp Stud 4(4):313–325
List GF, Foyle RS (2011) Seven portals study: an investigation of how economic develop-
ment can be encouraged in North Carolina through infrastructure investment. https://apps.dot.
state.nc.us/Projects/Research/ProjectInfo.aspx?ID=2761 Accessed 10 April 2015
List GF, Foyle RS, Canipe H, Cameron J, Stromberg E (2008) Statewide logis-
tics plan for North Carolina. http://www.ncdot.gov/download/business/committees/logistics/
StatewideLogisticsPlan_080513.pdf. Accessed 10 April 2015
North Carolina’s Southeast regional economic development partnership (2013). www.ncse.org.
Accessed 10 April 2015
North Carolina State Ports Authority (2015) Port of Wilmington. http://www.ncports.com/
elements/media/files/port-wilmington-fact-sheet.pdf. Accessed 10 April 2015
Plastria F (2011) The Weiszfeld algorithm: proof, amendments, and extensions. In: Eiselt HA,
Marianov V (eds) Foundations of location analysis. Springer, New York, pp 357–389
Qabaja H, Bikdash M (2014) Identification of closest safe places and exit routes during evac-
uation from GIS Data. ASE@360 Open Scientific Digital Library. http://ase360.org/handle/
123456789/141. Accessed 10 April 2015
ReVelle CS, Eiselt HA (2005) Location analysis: a synthesis and survey. Eur J Oper Res 165(1):
1–19
Rushton G (1988) The roepke lecture in economic geography location theory, location-allocation
models, and service development planning in the third world. Econ Geogr 64(2):97–120
Weiszfeld E (1937) Sur le point pour lequel la somme des distances de n points donnés est
minimum. Tohoku Mathematical J 43:355–386
Weiszfeld E, Plastria F (2009) On the point for which the sum of the distances to n given points is
minimum. Ann Oper Res 167(1):7–41
Chapter 4
An Introduction to Industrial Forestry from a
Location Perspective

Eldon Gunn

4.1 Introduction

Forestry is the natural resource industry most dominated by spatial considerations.

Forests are an extensive resource typically spread out over millions of hectares.
In this paper we focus on the industrial aspects of forestry, and some location issues
that flow from these. This is in contrast to a focus on location analysis and then
asking about its applications in forestry.
We are conscious of the distinction made by Revelle and Eiselt (2005) that loca-
tion analysis involves the siting of facilities in a given space with the facilities being
“small relative to the space in which they are sited and interaction between the facil-
ities may or may not occur,” while in the case of layout problems, the facilities
are “fairly large relative to the space in which they are positioned, and interaction
between facilities is the norm.” In a strict sense, some readers may feel that this
paper does not deal with location problems since forestry problems that we discuss
do not necessarily fit neatly into this framework. In some of the cases discussed,
such as the location of processing facilities, the fit within this framework is obvi-
ous, but in other cases, the facilities are not always obvious. For example, forested
stands where harvest occur or harvest regions which serve as a source of wood flow,
can be seen as facilities with the siting seen as the choice of stands to harvest in
a given period. These are location problems in the sense that the facilities are a
small proportion of the total area in any one period and the facilities do not interact
in terms of the flow of material from one to another. The flows from the “harvest
facilities” occur to the various customer plants that use the forest products produced
from harvesting. Typically the metrics are total profits for delivering wood products.

E. Gunn ()
Department of Industrial Engineering, Dalhousie University,
Halifax, B3H 4R2, NS, Canada
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 85

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_4
86 E. Gunn

In this review, we propose to touch on at six interrelated areas where location

issues. These include:
1. Long term location of harvest by species, log type and age class in response to
strategic decisions on ecosystem management
2. Location of processing facilities relative to harvest
3. Joint location of harvest and facilities
4. Shorter term location of harvest w.r.t adjacency and road building
5. Shorter term location of facilities (camps)
6. Operational location of harvest equipment
Our goal is to illustrate the hierarchy of inter-related location problems and indicate
some of the accomplishments in these areas. These problems can be seen can be
seen as arising from the long-term, medium term, and short-term, design of the
forest products supply chain. However, it is worth emphasizing that this is more
than a supply chain, it is what many of us referred to as a value chain, in that, many
different value streams emanate from the forest.

4.2 Some Terms

Since some of the terms may not be obvious to the non-specialist, we begin with a
few terms that make up the spatial issues of the problem.
Stand A relatively small area over which the forest can be considered as homoge-
nous with respect to features such as species distribution, age, site quality, and costs
of harvest . In a GIS, a stand is often equated to a polygon.
Ecodistrict Part of the larger ecosystem classification of the landscape into eco-
zones, ecoregions, ecodistricts and ecosites. An ecodistrict is a connected landbase
that has similar landform, soil and vegetation. In my home province of NS, the 4.5
million ha of forested land are divided into 38 ecodistricts.
Watersheds An area of land that drains to a common river mouth. Here we are
thinking of fairly significant river systems. Again in Nova Scotia, the land base can
be partitioned into 46 watersheds1 .
Timbershed A region from which logs are harvested. The purpose of a timbershed
is to provide a location that can be used as a point to estimate transportation costs
(see Gunn 2009).
Sawmills Mills that process round logs into semi-finished lumber. The main fea-
ture here is that the process of sawing produces multiple lumber products as well
as chips, sawdust, shavings and bark. Typically the lumber will be less than 45 %

1 https://www.novascotia.ca/nse/water.strategy/docs/WaterStrategy_NSWatershedMap.pdf.
4 An Introduction to Industrial Forestry from a Location Perspective 87

of the input volume, with chips as high as 40 %. The percentages are dependent
on the lumber products being produced, and on the size of logs being processed.
Products are distinguished by species (fir, spruce, pine, birch, maple, oak). Mills
are often classified as softwood and hardwood mills depending on the species most
commonly sawn.

Pulpmills/biorefineries Here we mean a more general concept of mills that are

aimed at processing the fibres in the wood as opposed to solid wood. Some aim
at thermo-mechanical processing where heat and grinding reduced the wood to a
pulp. Others deal with the wood chemistry and attempt to break out the cellulose,
hemi-cellulose and lignin components of the wood. Typical is the kraft process with
a primary aim at producing cellulose but more recently new processes are aimed at
high end chemical processing of the lignin and hemi-cellulose components. These
have come to be referred to as biorefineries. Pulpmills/biorefineries are usually
based on wood chips from sawmills and other sources. If they receive round wood,
the bark byproduct typically is for energy generation either onsite or elsewhere.

Energy Facilities Energy facilities may be relatively small scale, typically heat
or combined heat and power. Energy facilities may be located concurrently with
sawmills, pulpmills, or they may be located independently in large consumers, such
as hospitals or universities. Energy facilities consume woodchips, bark, shavings.
In some cases, of standalone large-scale energy facilities, the inputs may need to be
preprocessed in the form of pellets, either green or torrified.

4.3 The Sustainable Forest Management Setting

Although this paper is focused on industrial force management, one also has to be
cognizant of the fact that all forestry operates under commitments to sustainable
forest management. Internationally, one has the Montréal Process (2014), which
has been signed by more than 20 countries. In Canada, we operate under Cana-
dian Council of Forest Ministers (2003) Criteria and Indicators (CCFM C&I) for
sustainable forest management. The main criteria under the CCFM are the follow-
ing: (1) Biodiversity, (2) Ecosystem condition and productivity, (3) Soil and water,
(4) Global cycles, (5) Multiple economic benefits, and (6) Societies responsibility.
Biodiversity is often discussed in terms of preservation of habitats for endangered
species, location of old forest conditions, but also, importantly, in terms of dis-
tributions of representative habitats across the landscape. Because ecosystems are
represented as distinct spatial entities, the representativeness of species and age
distributions on an ecodistrict is one important measure of ecosystem condition.
Watersheds, by their very nature, depend of the location of major lakes and rivers.
The last two criteria (5–6) involve supply to industries and affects on communi-
ties, both aboriginal communities and small rural communities and again require
88 E. Gunn

explicit recognition of relative location of the forest activities and the location of
both industrial sites and communities. Perhaps only the global cycles criterion (4),
with its main focus on the carbon cycle can be thought as aspatial. Yet, even here
the location of fire, insect out-breaks, and forest harvesting play significant roles
and the net contribution of the green house gases will depend on transportation and
the relative location of harvesting processes, and wood consuming facilities.
In most of the situations discussed this paper, we will focus on the economic
issues, but, the criteria for sustainable forest management, will always be lurking in
the background. It is true that there are opportunities for direct locational analysis in
sustainable forest management. This often arises in issues of location of protected
and special management areas and are dealt with elsewhere in this book (Chap. 6).
In looking at industrial forest management, we will often take the position that some
agency has adopted certain strategies to deal with the criteria and indicators of sus-
tainable forest management and that these strategies impose locational constraints
on what we can do logistically.

4.4 Strategic, Tactical and Operational Planning

Within forestry there has long been a focus on distinguishing to strategic, tacti-
cal, and operational management problems. We discuss these briefly below. For a
good example of the forestry planning hierarchy from an information perspective
see Marques (2011).
Strategic level decisions are those that affect the resources available to the
organization (Anthony 1965). Here we will discuss three closely related decision
problems. The first is the problem of how much to harvest, where, and when, to
make the most effective use of the forest’s growth capabilities. Often, such decisions
have been made with no regard to the location to which the harvest of material will
be sent. More recently, models which take account the location and capacity of the
wood using facilities are being discussed. The second problem involves the design of
the facilities making up the overall supply chain. This problem has usually focused
on the location of facilities, particularly intermediate processing facilities, and the
flow of materials between them. What has been missing is a focus on the appro-
priate capacity and location of the larger scale facilities. This is important because
large-scale facilities such as pulpmills are subject to enormous economies of scale
in terms of both capital and labor. However, there are also diseconomies in terms
of transportation and materials acquisition cost both from the forest directly and
from intermediate facilities. The third problem involves the joint design of both for-
est harvesting and the facilities making up the forest supply chain. This recognizes
the dual facts that appropriate location of the right types of facilities can facilitate
much improved forest harvesting strategies, and that the appropriate forest harvest-
ing strategies can significantly improve the economics of investment in new capital
facilities.
4 An Introduction to Industrial Forestry from a Location Perspective 89

Tactical level problems involve how to best exploit the resources provided
through strategic decisions. In most organizations, tactical level problems typically
have something like 1–5 year time horizons. Forestry has been unusual in that its
tactical problems have often been posed over time horizon’s ranging up to 20 years.
Typically these problems involve location of harvest and the decisions on road build-
ing to provide access to the harvest. The dominant feature in these decision problems
involves issues of harvest of stands and their relative locations to each other and to
the road network. One of the problems in this area that has attracted significant
research attention is the adjacency problem in which we try to prevent harvesting
too many adjacent stands that will create overly large openings in the forest.
Operational problems are those that occur at a relatively short time frame and
typically involve minimization. We will focus only on the operational problems that
occur within the forest since these tend to have the most location oriented aspects.
These include problems of location of landings and equipment, problems of location
of skid roads, and problems of crew scheduling and movements between harvest
locations. A problem that has recently come up that involves the location and sizing
of camps for the forestry work crews. These problems arise in the northern regions
of Canada where sparse settlement and long distances may make staying in regional
towns prohibitive.

4.5 Previous Reviews

This review is not meant to be comprehensive. Thus the reader should refer to
Church et al. (1998) for additional material particularly from a location perspective.
Bare et al. (1984) is an early review focussing on location is Gibson and Roden-
burg (1975). Martell et al. 1998 updates this work with an emphasis on the complex
hierarchy of forestry decisions. Baskent and Keles (2005) gives a focus on spatial
forest planning. Melo et al. (2009) reviews facility location issues associated with
supply chains. Carlsson et al. (2009) review the issues of of supply chain plan-
ning in the pulp and paper industries. Chauhan et al. (2009) look at the planning
of the multi-commodity supply network where many products emanate from the
forest. D’Amours et al. (2008) reviews operational research applications to supply
chain planning in the forest industry. For the serious reader, Davis et al. (2001) is a
fundamental work on forest management issues.

4.6 Strategic Harvest Planning

The first of our location models we will examine is the strategic harvest planning
model (see Gunn 2007, 2009 for a longer discussion). This is the problem of decid-
ing where and how much to harvest from various forest types over time subject to
90 E. Gunn

requirements of a reasonably constant supply of wood. It arose from an initial con-

cern (Ware and Clutter 1971) that if forest stands were harvested at their time of peak
output (economic or biological) then this would result in a very irregular supply that
could not be used by the industry. This is the longest standing forest location prob-
lem, but it had its origin largely in an aspatial context. As will be evident below, this
class of models are not strategic in the sense of computing strategy. The sustainable
forest management strategic decisions define both the constraints and variables of
the problem and are given a priori. Thus these are models for assessment of strategy,
not choosing strategy.
The early paper by Ware and Clutter (1971) did focus on “cutting units” with rev-
enues based upon “net stumpage values.” This does imply a location, but the main
emphasis has been on defining cutting units as homogenous units in terms of future
growth capabilities. When these types of models were used early on, there was a ten-
dency to combine stands into “timber classes” (Navon 1971) or “macro-stands” with
a corresponding loss of location identity. The initial modelling (Model I) was based
around the concept of assigning a cutting unit to a prescription where a prescription
was an entire plan for harvesting, planting, thinning, harvesting, etc over the entire
planning horizon. With many cutting units and many possible prescriptions, mod-
els rapidly grow. This was one reason for the amalgamation of stands with similar
growth characteristics into a much smaller number of cutting units with a subse-
quent loss of identity. Commercial software that came into widespread usage in the
1990s was based on the Model II ideas of Johnson and Scheurman (1977) which
automatically generates prescriptions through an acyclic network model. The nodes
of the network correspond to points at which the forest is regenerated and the arcs
correspond to decisions of what is to be done between regeneration points. Stand
identity is lost as all stands that regenerate at a point in time merge. These mod-
els (both Model I and Model II) came to be widely used to develop the concept of
“annual allowable cut” without much thought about where this cut was coming from
or where it was going (Church 2007; Gunn 2007, 2009). These models have been
widely used is a variety of government and industrial analyses.
In Gunn (2009) we formulated a model that has a more spatial orientation than
the conventional. This model was developed into a case study in Martin (2013) Thus
this model is not typical of those in common use in government and industry. To see
the type of model we have in mind, we use the following notation, taken from Gunn
(2009):
Sets:
I: the set of stands
Pi : the set of allowable prescriptions for stand i
Y: the set of yields under strategic consideration
Yw: the set of wood products (Y w ⊂ Y)
R(): the set of regions applicable for yield type· . If  ∈ Y w , then R() is a
timbershed for wood product 
M: the set of mills (forest products demand centres)
S(m): the set of demand segments for mill m
4 An Introduction to Industrial Forestry from a Location Perspective 91

Data:
Ai : area of stand i
LYrt : lower limit on yield  in region r in period t
U Yrt : upper limit on yield  in region r in period t
LDmt : lower limit on demand in mill m in period t
max :
dsmt upper limit on demand segment s for mill m in period t
yiklrt : yield of type  produced in region r in period t if prescription k is used for
stand i
∂m : conversion coefficient of log type  to meeting the demands at mill m
psmt : demand curve price for segment s at mill m in period t (discounted to
time 0)
Cik : silvicultural costs for managing stand i with prescription k (all costs
discounted back to time 0)
τrmt : unit transportation costs for logs of type  from region r to mill m in
period t (discounted to time 0)
Variables:
Xik : area of stand i managed using prescription k
Yrt : yield of type  produced in region r in period t
zrmt :: transportation of logs of type  from region r to mill m in period t
Dsmt : amount of demand segment s supplied to mill m in period t
Then Dsmt allow us to define demand curves at each mill m using the price endoge-
nous ideas described in Lebow et al. (2003). Since the model is aimed at strategic
assessment, it is worth noting that these demand curves here also need to be defined
in strategic terms. The main objective is not to predict pricing, but rather to link
supply chain mill capacity to the forest strategy. This requires an ability to repre-
sent the fact that some forest products may command premium prices psmt and the
extent of the market assumed to exist for these products. There may be limited abil-
ity to use certain log types (species, size) at high prices and the representation of the
low price section of the demand curve can useful to represent disposal. The mill sub-
script m allows representation of both product type and location. In a strategic model
we might expect these to be broadly defined, with several facilities that require the
same type of wood product and located in the same general region aggregated into
a single mill type.

  
I 
Pi 
Max/Min psmt Dsmt − cik Xik − τrmt zrmt
t m s∈S(m) i=1 k=1 t m  r
(4.1)


Pi
s.t. Xik = Ai , i ∈ I (4.2)
k=1
92 E. Gunn


Pi
yikrt Xik = Yrt ,  ∈ L, r ∈ R(), t ∈ T (4.3)
k=1

LYrt ≤ Yrt ≤ U Yrt ,  ∈ L, r ∈ R(), t ∈ T (4.4)


zrmt = Yrt ,  ∈ Lw , r ∈ R(), t ∈ T (4.5)
m∈M
 
∂m zrmt = Dsmt , m ∈ M, t ∈ T (4.6)
 r s∈S(m)

Dsmt ≥ LDmt , m ∈ M, t ∈ T (4.7)
s∈S(m)

0 ≤ Dsmt ≤ dsmt
max
, m ∈ M, s ∈ S(m), t ∈ T (4.8)

The objective (4.1) is to maximize the net present value of all cash flows. These
include the discounted revenues associated with the delivery of logs to markets, the
discounted harvesting and silvicultural costs associated that result from applying
silvicultural prescriptions to stands and the discounted transportation costs of deliv-
ering harvested log types from regions to mills. Equation (4.2) requires that the areas
treated by all prescriptions equal the stand area. Equation (4.3) computes the yield
of type l in each period in each region r where l is relevant. Equation (4.4) constrains
this yield to lie within bounds chosen by the user over time. For those yields that
correspond to wood products, Eq. (4.5) requires that the sum of the amount trans-
ported from the region to mill centres equals the amount produced in each period.
Lumber products are converted to market demands via conversion coefficients in
(4.6) and the total of those products delivered to market m must equal the amount
sold in that market summed over the price ranges in that market. Market minimum
demand is defined by the user in (4.7) and the amount that can be sold at each price
is limited in (4.8).
Let us point out a few differences between the model (4.1–4.8) and conven-
tional practice. Equations (4.5–4.8) and the corresponding terms in the objective
function are not conventional. Most uses of the strategic harvest modelling focus
on the land base and the production of timber. Constraints (4.2–4.3) are conven-
tional although the most popular commercial software uses a Model II formulation
instead of the Model I formulation in (4.2). The yield steering constraints (4.4)
is one of the main differences. What is common instead of (4.4) is some sort of
flow constraints for timber yields. Non-declining yield asserts that Yrt ≤ Yr(t+1) ;
level flow Yrt = Yr(t+1) and many practioners allow some percentage variation
from either level flow or non-decline. Although the mathematical difference may
be minor, philosophically there is quite a difference between non-declining yield
and yield steering (4.4). The first is a normative model, the second is a strategic
choice of the decision makers. This becomes more pronounced when we start to
consider more than just timber yields over one or a small number of regions. The
indicators to model model biodiversity (e.g., habitat), ecosystem condition (e.g., age
4 An Introduction to Industrial Forestry from a Location Perspective 93

class distribution), watershed condition (percentage forest cover) are very complex
(Gunn 2005) and the idea of non-decline or constant makes little sense. Note also
the definition of yields in terms of regions for that yield. This is necessary to dis-
cuss the SFM criteria. This regional representation of yields implies a partitioning
of the landscape where each element of the partition represents a unique combina-
tion of the underlying spatial regions (eg a stand is in a particular combination of
GIS layers-ecodistrict, watershed, species habitat, economic region, etc.). One con-
sequence of this is that Model II formulations rapidly grow in size as we attempt to
account for this spatial complexity because we cannot logically merge stands that
lie in different elements of the partition.
As indicated above, a second aspect to note is the explicit representation of where
harvests occur in terms of the transportation of the harvested material to the mills
where it can be used (Eqs. 4.5–4.8). More conventional models value the forest prod-
ucts produced by harvest at the stand and do not take into account that the demands
for these products occur at specific locations with limited capacity to absorb these
products. This can lead to an unrealistic assessment of overall system economics.
The third aspect is the objective function (4.1). The use of net present value is
itself problematic for strategic planning. It is quite possible for an NPV objective to
have positive cash flows in early years, but reducing, possibly to negative cash flows
in subsequent years. This is obviously a problem for any real economic systems.
Church and Daugherty (1999) has pointed out that other objectives may address the
issues of intergenerational equity and may produce more sustainable economics.
However, a somewhat surprising phenomenon found in many jurisdictions, is that
the objective is purely maximization of volume harvested. In this viewpoint, the
role of the forest is to produce wood, possibly subject to ecological constraints and
it does not really matter where or what type of wood is produced. It is up to the
industry to decide how to use the wood.
Martin (2013) and Martin et al. (2014a) shows that this form of model is capable
of dealing with an extensive spatial representations of the indicators of sustainable
forest management without blowing up as is the case when more conventional model
II representations are used. In Martin (2013) and Martin et al. (2014b), a case study
on Crown Lands (provincial lands) in central Nova Scotia shows that taking into
account the location of facilities and the transportation issues can lead to substantial
changes in the level and timing of harvest of various types of timber.

4.7 Strategic Location of Mills, Processing and Transport

Facilities

If the role of the strategic harvesting model is to decide how much wood is to be
produced, of what type and where, a natural converse is where should this wood
go to make the maximum economic contribution. This raises the question of what
types of mills we should have, how large should they be and where should they be
located? There are several factors that make these questions very important. The first
94 E. Gunn

involves economies of scale. Sawmills and pulpmills, like most process industries,
are subject to substantial economies of scale in both capital cost and labor cost. This
suggests a few big mills. A second is transportation cost. Wood coming from the
forest is about 50 % moisture; moisture that will eventually have to be removed in
processing. Moving wet wood long distances is expensive suggesting the need for
many small mills. A third is integration. Forest harvesting produces wood with a
broad variety of species, dimensions and quality attributes. Small material is either
sent directly to pulpmills or to some intermediate facility where it is debarked and
chipped. Larger material goes to sawmills where, depending on species and log
quality, it can be debarked, and sawn into a variety of types of lumber. However
the sawing process produces a substantial percentage of chips, sawdust, planer mill
shavings and bark. A volume breakdown of 40 % lumber, 45 % chips, 8 % sawdust,
7 % shavings and 10 % bark would not be unrealistic2 . The chips, sawdust, shavings
and bark will then go to pulpmills and/or energy facilities.
Thus, at the overall forest system level, there is an overall capacitated flow net-
work design problem to be solved. In Gunn (2012) and Gunn and Martell (2015) a
mathematical model like the following is presented for this design problem.
Sets:
T: set of time periods (1 . . . T)
D: the set of districts
L: the set of facility locations
C: the set of capacity types
F: the set of wood flow types
M: the set of markets
P(m): the set of price ranges for market m.
Data:
ARes
fdt : resource availability of wood type f ∈ F in district d ∈ D in time period
t ∈ T.
ρmpt : unit revenue at price level p in market m in period t
oct : operating cost per unit production location , capacity type c in period t
τdfW ct : transportation cost from district d of type f wood to location  which
has capacity type c in period t
τF1 fct : transportation cost from location 1 , of type f wood to location  which
has capacity type c in period t
M :
τmfct transportation cost to market m, of type f wood from location  which
has capacity type c in period t
+ −
γfc, γfc : maximum, minimum percentage of wood flow type f in operating levels
of capacity type c
βfc : conversion coefficients for input wood flow type f to operating of
capacity type c

2 The reason for these percentages summing to 110 % is that, in Canada, tree stem volumes are
usually measured inside bark.
4 An Introduction to Industrial Forestry from a Location Perspective 95

αfc : amount of output wood flow type f for unit operating level of capacity
type c
MA+
mpt : maximum demand at price level p for market m in period t
MA−
mpt : minimum demand at price level p for market m in period t
Variables:
MREV t : total market revenues in period t
OpCostt : total operating cost in period t
TranCostt : total transportation cost in period t
Cap
Xct : capacity of type c installed in location  at beginning of period t
CAPct : capacity of type c at location  available throughout period t
WF fdct : wood flow type f (typically logs) from district d to location  which
has capacity type c in period t
CF f 1 ct : wood flow type f (typically intermediate products) from location 1 to
location  which has capacity type c in period t
MF mf t : wood flow type f (typically market products) from location  to market
m in period t
OLct : operating level of capacity type c at location  in period t
Mmpt : market sales at market m and price level p in period t
The model can then be written as:
  
Max (i + 1)−t MREVt − CapCostt − OpCostt − TranCostt (4.9)
t
  
Cap
s.t.CapCostt = C Xct t = 1, . . . , T (4.10)
,f,t

MREV t = ρmpt Mmpt t = 1, . . . , T (4.11)
m,p

OpCostt = oct OLct t = 1, . . . , T (4.12)
,c,t
 
TranCostt = τfWdct WF fdct + τfF1 ct CF f 1 ct
f,d ,c,t f,1 ,c,t

+ M
τmf ct MF mf t t = 1, . . . , T (4.13)
m,f ,t

Cap
Capct = Xct + Capc,t−1 ,  ∈ L, c ∈ C, t = 1, . . . , T (capital additions)
(4.14)

WF fdct ≤ ARe
fdt , f ∈ F, d ∈ D, t = 1, . . . , T (wood availability)
s
(4.15)
,c

OLct ≤ Capct ,  ∈ L, c ∈ C, t = 1, . . . , T (capacity limitation) (4.16)

96 E. Gunn
⎧ ⎫
−
⎪
⎨ γf c OLct ≤ WF f dct ≤ γf+c OLct ⎪
⎬
d
−  ∈ L, c ∈ C, t = 1, . . . ,T
⎪
⎩γf c OLct ≤ CF f 1 ct ≤ γf+c OLct ⎪
⎭
1 receipe inputs (4.17)

⎧ ⎫
⎪
⎨ βf c WF fdct + βf c CF f 1 ct = OLct ⎪
⎬
d 1
 ∈ L, c ∈ C, t = 1, . . . ,T
⎪
⎩ αfc OLct = CF f 2 c2 t + MF mf t ⎪
⎭
c 2 ,c2 m input-output balances (4.18)
⎧ ⎫
⎨ MF mf t = Mmpt , m ∈ M, t = 1, . . . , T ⎬
f, p
⎩ ⎭
MA− +
mpt ≤ Mmpt ≤ MAmpt m ∈ M, p ∈ P (m), t = 1, . . . , T market demand
(4.19)
Equations (4.10–4.13) are accounting equations for capital cost, market revenues,
operating costs and transportation costs with the ρ, o, τ in Eqs. 4.3, 4.4 and 4.5
being the appropriate revenue/cost coefficients. There are several different ways
of modelling capacity costs ranging from distinct choices, fixed costs plus linear
or piecewise linear. We have left these unspecified in Eq. (4.10). Equation (4.11),
together with Eq. (4.19), models market m as having piecewise constant demand
curves (Lebow et al. 2003; Gilless and Buongiorno 1985). Equation (4.14) keeps
track of capacity. The initial capacities Capc0 are given data and subsequent capac-
ities computed from additions. Equation (4.15) limits the wood flows of every flow
type from each district to an assumed amount given in each period. This model
has a simplified notion of processing. Each facility is assumed to have a recipe that
dictates the inputs and outputs of the facility. Equation (4.16) limits the amount of
processing to the available capacity. Equation (4.17) limit the relative proportions
of each type of forest wood flow (WF) and intermediate product wood flow (CF) to
be within certain proportions of recipe inputs. The two types of equations given by
(4.18) indicate that the inputs, which may be in different units (tonnes, m3 , cords),
adjusted using conversion factors βfc , have to add up to the total recipe amount in
operating level units. The outputs of the recipe are in fixed proportions αfc . The
outputs of the recipe can be intermediate products that go to other facilities (CF)
and/or products that go to final markets (MF). Equation (4.19) specify that flows to
a market m are sold at one of a variety of prices with the amount that can be sold at
a given price range p is limited by lower and upper bounds. This enables the use of
price endogenous demand models (Lebow et al. 2003).
For simplicity, capacity in the model described appears to last forever. Frieden-
felds (1981) discusses a variety of ways of modelling capacity which is either time
limited, diminishes over time, or is retired. However, this raises the issue of time
horizon? Typically, time scales for industrial investment are 20 years or less, but the
value of end of horizon capacity is probably not zero. How to value that capacity is
an important question. These problems are closer to classic location problems, par-
ticularly in the realm of supply chain design (Melo et al. 2009) in that they involve
the direct location of facilities, such as various types of sawmills, pulpmills and
energy facilities.
4 An Introduction to Industrial Forestry from a Location Perspective 97

In these problems there is little direct interaction between facilities of the same
type but there are strong capacity influences in terms of a total capacity to process
forest products and the number, size and location of mills to provide this capacity
with substantial economies of scale in these capacity decisions. There are strong
interactions with the forest regions and the flows of wood that emanate from these
regions. There are also strong interactions with mills of other types. For example,
sawmills depend on pulpmills, and energy facilities to make use of the substantial
(more than 50 %) byproducts of the sawing process in terms of chips, sawdust, shav-
ings and bark. The relative location, and hence the transportation cost to these other
facilities matters.
In much of the literature on capacity planning for the forest products industry,
there has actually been quite limited discussion on the location of the main pro-
cessing facilities (see the review in D’Amours et al. 2008). Much of the design has
focused on the flow of materials through the supply chain with some emphasis on
the location of material handling/transformation facilities to facilitate this flow and
on the choice of processing technologies at the various facilities. (e.g., Gunnarsson
et al. (2004, (2006; Carlsson and Ronnqvist 2005; Carlsson et al. 2009 and Chauhan
et al. 2009). To some extent this has been caused by the depressed state of the indus-
try which has led to greatly reduced production and the closure of many mills (see
Canada 2009). Few new mills have been built in the past 30 years. However, going
forward, as mills continue to age the need to replace obsolete capacity will create
demand for new capacity even if inherent demand does not increase. Importantly,
as we move to the new biorefineries, the question will arise where to put them and
how big they should be.

4.8 Dealing with the Dicotomy: Strategic Harvest; Strategic

Capacity

As discussed above, if we want to know how to deal with the locational issues of
harvest with some reasonable approach to the economics of the forest, we need to
know something about the capacity to use the wood produced and where this capac-
ity is located. This can be about more than economics in some jurisdictions. For
example, if a certain species, present in a large number of stands, which can only be
used by a few (or zero) mills of limited capacity for that species, then what we see is
a total harvest of other species that is much less than the total capacity of the mills.
This implies we need a possibility of disposal at low, (possibly negative) value so
as to make possible the harvest of the more desireable species. This is the extreme
case but in general we can see that the value of the delivered wood, the capacity
to use the wood and the cost of transport from the harvest region to the location of
the capacity will determine the amount of harvest. Moreover the amounts actually
harvested need to correspond in some reasonable way to the amounts planned to be
harvested. This is because of the regeneration and growth issues of the forest. If a
stand is not harvested, it cannot be regenerated and hence cannot grow as rapidly.
98 E. Gunn

This is the fundamental problem with the more conventional practice of using mod-
els without the economics in Eqs. (4.6–4.8). It has been observed in several cases
that the aggregate harvests carried out in early periods may be approximately those
in the plan for several important species but well below plan for some less desireable
species. Some will equate this under usage of certain species as overall sustainabil-
ity but this is not the case. The failure to harvest consistent with the plan will lead
to less wood being available over the long term. (see Paradis et al. 2013). It is a case
of use it or lose it! In summary, as shown in the recent work in Martin (2013) and
Martin et al. (2014b) the harvest plan depends on the capacity to actually use the
wood produced.
The mirror image is in the capacity model (4.9–4.19). The key feature is in the
Eq. (4.15) where the amount of wood of a given type consumed from a given region
in a particular time period must be less than or equal than the amount available. If
these availabilities come from a plan developed using the strategic harvesting model
and the capacity model solution satisfies (4.15) as equality, then the capacity plan
will be consistent with the forest capacity as computed. There are two main issues
here. If the amounts of a wood type used from given locations and time periods in
the capacity plan is inconsistent (i.e., <) with the amounts provided in the strate-
gic harvest plan, then the strategic harvest plan will not come true not just for the
given wood type but possibly for all wood types. The second problem is that if the
original harvest plan was developed using estimates of profitability that discouraged
the production of a certain type of wood, then this wood will not be available to the
capacity planning model and hence the capacity that might profitably make use of it
cannot be installed. Martin (2013) showed that by installing a mill of the right type
in the right place, it is possible to raise the economical harvest to be close to that
achieved under volume maximization.
One answer is to combine (4.1–4.8) and (4.9–4.19) into a single model with
the availability ARes
f,d,t in Eq. (4.15) being the yields Yrt of Eq. (4.3) where the
yield type l corresponds to the wood flow type f and the region r corresponds to
harvesting district d of Eq. (4.5). The obviously difficulty with this approach is the
complexity of solving the mathematical model. Although the model (4.1–4.8) is just
a linear program, with a substantial number of SFM indicators and a large number of
stands, these models become very large. Combing these with the integer program-
ming model (4.9–4.19) is likely to be unsolvable. A current Ph.D. project (Gunn and
Shereshti 2012) will look at iterative strategies for joint analysis.

4.9 Understanding Harvest and Access Costs: Relative

Location of Stands and Roads

The Strategic Harvesting model (1)–(8) ignores the question of can we actually
access the stands to be harvested at the time we want to harvest them within a given
region. This region could be an ecodistrict, a watershed or a timbershed. These are
4 An Introduction to Industrial Forestry from a Location Perspective 99

often referred to in forestry as tactical problems. There are at least two main issues
here. One is adjacency of stands, the other is roads to access the stands.
If we harvest a set of adjacent stands simultaneously, this creates an opening in
the forest canopy. Many jurisdictions have constraints on the area of the opening
that can be created either within one period or in terms several periods required
for harvested stands to “green up” to a forested condition. The question is can we
meet a certain target trajectory of harvests of certain wood types from a region
over time without violating these adjacency constraints. Initial approaches to this
problem emphasized heuristics such as simulated annealing (e.g., Lockwood and
Moore 1993). Snyder and Revelle (1996) was an early mathematical programming
approach. More recent approaches, Weintraub and Murray (2006) have emphasized
exact integer programming approaches. These can be simple models to formulate
but quite complex in their data creation process requiring an enumeration of either
all clusters (possible sets of allowable openings, see Goycoolea et al. 2005), all paths
(possible sets of forbidden openings, see Tóth et al. 2013). Alternatively, one can
use somewhat more complex models (Constantino et al. 2008). All three approaches
become challenging as one gets to problems of realistic scale. For example a timber-
shed can easily contain 10,000 stands. Gunn and Richards (2005) pointed out that if
a simple “stand centred” adjacency constraint were used, large scale problems could
be solved relatively easily. The constraint states that if a particular stand k is har-
vested then the harvest from all stands immediately adjacent to it must be less than
the maximum allowable opening minus the area of the stand under consideration. It
takes the form

ai xi ≤ Mk (1 − xk ) + (Amax − ak )
i∈A(k)

where k indexes the stand under consideration, A(k) is the set of stands adjacent to
k, ai is the area of stand i and Amax is the maximum allowable opening size. The
xi are binary variables with xi = 1 if stand i is harvested. The value Mk is set to a
“sufficiently large” number. Gunn and Richards show how these constraints can be
improved through solving strengthening and lifting subproblems.
This constraint only applies to those stands whose area plus that of its adjacent
stands exceeds the allowable opening. Thus the number of adjacency constraints is
less than or equal to the total number of stands in the timbershed. The method is
not rigorous in the sense that it does not prevent “long stringy openings,” where
stands i1 , i2 , i3 , . . ., ik-1 , ik are harvested and stand i1 is adjacent to i2 , . . ., ik-1 is
adjacent to ik but stand i1 , and ik are not adjacent. Gunn and Richards point out that
in the optimal solution to practical situations the number and total area of such long
stringy openings is very small relative to the number of stands and the total area of
the region.
The road building problem is quite closely related to the adjacency problem. If
we plan to harvest certain stands over time, then when a particular stand is harvested,
there must be a road available to allow equipment to access the stand and to remove
the wood produced. Modeling the connection between the dynamics of stand har-
vesting and road building are demonstrated in Andalaft et al. (2003) although this
paper does not deal with adjacency issues. Nelson and Finn (1991) formulated prob-
lems of road building with adjacency and solve these with random search heuristics.
100 E. Gunn

However they dealt with cut blocks, not stands. Stands had been pre-assigned to cut
blocks of such a large size that cutting any two adjacent cut blocks would violate
adjacency. Richards and Gunn’s (2000) tabu search approach included solving this
problem with individual stand adjacency. In this case the problem of building roads
to access the stands scheduled for cutting is a dynamic Steiner tree problem (Imase
and Waxman 1991). Richards (1997) developed an efficient heuristic for this Steiner
problem, which served as an evaluator for any particular choice of harvest decisions
during the tabu search. Richards and Gunn (2006) demonstrated that the same large
scale problems solved in Richards and Gunn (2000) can be solved as integer pro-
gramming problems by using the stand-centred adjacency constraints together with
tight linkages to the road network. Sample problems with 1039 stands, 135 road
segments and 5 planning periods (5 year) produce optimality gaps of less than 1 %
within 300 s in 24 of 30 problems. Gaps less than 5 % are produced within 500 s
in all cases. The road building problem in Richards and Gunn deals with a higher
level with fairly long potential road segments. Real road building problems are more
complex. Meignan et al. (2012) discuss more detailed road location problems that
take the stand harvests as given.
In terms of the longer term strategic forest harvesting and capacity models, these
tactical models serve two roles. One is to estimate how much wood can actually be
produced once adjacency is accounted for (Eq. 4.15). The second is to estimate the
cost of accessing this wood. Per unit road costs are a key part of the cost coefficients
in (4.1) and in (4.13) and the particular road segments constructed also influence the
hauling distance to access the main highway network. Again we have a dichotomy.
We need solutions to the strategic harvesting model to estimate the amount of wood
we want to harvest from a particular region and we need the solution to these adja-
cency/road building problems to estimate access costs and assess realistic harvest
levels.

4.10 Understanding Materials Handling and Logistics

The costs of moving material from a harvest to the particular facilities depend on
the logistics of the movement. In both the forest harvesting strategy models (4.1–
4.8) and the capacity models (4.9–4.18), one needs the cost of harvesting wood
and delivering it to some facility for processing. However the supply chain is more
complicated than this. We have choices of how to harvest, with many different types
of machines and systems. These harvesting systems may produce either full trees,
or cut-to-length wood assortments or wood chips or combinations of all three. Full
trees may be the cheapest harvest in the woods but the trucks that can access the
woods operations are inefficient for long hauls. Also the full tree can be broken
down into different products (sawlogs, studwood, pulpwood, biomass chips) that
might need to go to different locations. The contribution by Epstein et al. (2007a)
reviews some of the issues in designing the forest transportation system. D’Amours
et al. (2008) treat this as the tactical problem of designing the procurement for the
4 An Introduction to Industrial Forestry from a Location Perspective 101

forest products supply chain. From a location theory point of view, perhaps the
main interest here involves the decisions to create processing terminals where wood
from the forest, perhaps tree length, perhaps otherwise, can be sorted and processed
(cut to length, chipped), and transferred to more efficient trucks or possibly rail for
movement to other facilities. How many such terminals, the types of processing
available, the transport modes and their location become the design problem. One
could look at such terminals as just additional processing facilities with products
that go to other facilities in the context of the capacity model (4.9–4.19). However,
as can be seen in Gunnarsson et al. (2004), the modelling of these logistics facilities
rapidly produces very large models, even in an environment where energy facilities
are given.
Transportation can be between 30 and 50 % of the costs of delivering wood to
mills from the forest. Thus the logistics system can have significant effects on where
we harvest and what capacity we build. Again we see the situation where in order to
properly to develop forest harvest strategy and capital investment strategy, we need
good estimates of transportation costs and we need supply chain logistics design
models to estimate these costs. However the supply chain logistics models need to
know where the wood is coming from and the facilities to which it can go. Thus we
need solutions to the forest harvest strategy models and capacity investment models
to develop good logistics systems.

4.11 Operational Location of Camps

In some settings in Canada (the boreal forest), harvesting occurs at considerable

distance from communities with housing and other services. Companies who wish
to harvest have a choice of dealing with the cost and inefficiencies of transporting
workers long distances daily, or establishing camps with accommodation, recreation
and other facilities where crews can be accommodated near the workplace. In the
forest harvesting strategic model we have a single period consisting of 5 years of
harvests in a region. However in order to understand the harvesting costs associated
with this plan we need to know how the workers will be accommodated and trans-
ported. The problem is dynamic in two senses, first the harvest moves around the
region over the 5 years. However, there are also strong seasonal influences. Many of
the stands in the region can only be harvested in the winter when the ground is frozen
sufficiently to permit both harvesting and transport. Jena et al. (2012) have shown
that the problem of deciding the location, number and capacity of these camps is
a very interesting dynamic location problem. The modular aspect where capacity
is provided in terms of fairly standard trailers which can be moved from one loca-
tion to another if necessary, and the dynamic aspect, where camps can be open and
closed at various times during the year, make the solution of this dynamic location
problem challenging. They have also been able to show quite substantial savings in
total costs (transportation costs plus camp location and relocation costs) through the
application of their approach.
102 E. Gunn

Again we have the situation that in order to understand the costs of harvesting
for the forest harvesting model, we need to know how much it will cost to accom-
modate and transport the workers. However, solving this problem of where and how
to accommodate the workers depends on the original 5 year plan of total harvest in
the region.

4.12 Operational Problems; Machine and Crew Scheduling;

Product Transport

As we work down the hierarchal planning system, it will not be surprising to find
that there are a host of problems in which location plays a major role Epstein et al.
(2007b) characterizes these into seven main areas. We mention three of these here.
These include the machinery location problem, the related crew scheduling problem
and the detailed transportation planning problems that examine truck routing and
load assignment.
Epstein et al. (2006) looks at the complexity of machinery location and road
design. These problems are applied to somewhat small landbases (1000 ha or less)
and are often known in the industry as the landing location problem. This was one
of the early applications of location analysis in forestry (Dykstra and Riggs 1977).
They can be very challenging (Legues et al. 2007). A landing is a point on the road
network to which logs are delivered by the harvesting machinery so that they can
then be sorted and trucked onwards to the logistics system. These problems typically
deal with time periods of days and weeks and quite fine land divisions (10 m2 ) for
sources of timber. The problem is how many landings to have, where they should
be located and what connecting roads to build to remove the harvested wood. These
problems are particularly important in mountainous areas where road building is
expensive and the landing may correspond to the location of logging towers with
aerial cables to deliver the wood from where it is cut to the landing. On more level
ground, the landing is the site from which mobile skidders/tractors travel to the cut
trees, pick them up and deliver them back to the landing. The more landings that you
have, the easier it is to remove the wood from the forest to the landing. However,
all landings require a fixed cost to establish. Epstein et al. (2007a) provide a history
of the problem, a mixed integer programming formulation and a heuristic algorithm
for solving these complex problems.
The crew scheduling problem is the dynamic problem of assigning a set of cut-
ting crews to harvest various cut locations. There are three main issues governing
the costs. First, different crews may have different types of equipment and, depend-
ing on topography, each type of equipment may be more or less productive on a
given cut location. This affects both the time and the cost to harvest the location.
Second, each crew may have a home base and there is a cost of daily travel from
the home base to the cutting location. Third, once the crew has finished harvesting
a given location, their equipment must be moved to the next location. This cost of
relocation can be substantial and depends on the distance between the cutting loca-
tions. Complicating this is an ongoing demand for various types of logs from a set of
4 An Introduction to Industrial Forestry from a Location Perspective 103

customer mills. Each of the cutting locations may differ in terms of the species and
quality of logs produced by harvesting. Murphy (1998) deals with a static problem
of assigning crews to stands to maximize equipment productivity. Mitchell (2004)
deals with the complete dynamic problem. Mitchell’s work utilizes David Ryan’s
(1992) set covering and constraint branching ideas, originally applied in air crew
scheduling, to solve these complex problems quite efficiently. In some cases it is
hard to quantify the ability of certain crews to work in a given location’s topogra-
phy and soil conditions. Gallant and O’Brien (2004) developed an approach where
the company scheduler can decide a sequence of harvest locations for each crew.
A VBA application within MS Access computes the timing and costs for each crew
and a series of simple network flow problems solves the allocation to customers. The
user can assess whether or not harvest location assignments and sequence should be
changed.
Product transport is an enormous area; one that we can’t deal with in any detail
in this setting. The short message is that location implies the need for transportation.
Thus transportation systems are challenged by the location of harvest activities and
the location of the facilities that make up the supply/value chain. Conversely, effi-
cient transportation systems can be the means of overcoming the location challenges
in forestry. Two of the most noteworthy groups in this area have been the Chileans
and the Swedes although some considerable progress is being made by the FORAC
group in Quebec as well as the more broadly based VCO Network. The paper by
Epstein et al. (1999) is illustrative of the long standing work in Chile. The ASICAM
system, described further in Weintraub et al. (1996), deals with daily harvesting and
truck movements with substantial savings in costs. One of the interesting location
issues discussed in Epstein et al. (1999) is that the way trees are optimally bucked
(cut up into logs of various sizes) depends on the location of the stand relative to
the various mills that can use these logs. The Swedish work is illustrated in Carl-
gren et al. (2006), Forsberg et al. (2005), Karlsson et al. (2003) and Carlsson and
Ronnqvist (2005). The latter of these is notable for the inclusions of the possibility
of backhauls into the transportation calculations. Frisk et al. (2010) point out that in
multi-company situations where companies have their own harvesting locations and
facility locations, that an overall global solution that ignores company issues can
produce substantial transportation savings over and above the sum of the individual
company optimal transport costs. This raises questions of how these companies can
collaborate, sharing benefits appropriately, to capture these potential savings while
making all companies better off.

4.13 Uncertainty

Uncertainty is fundamental to forestry (see Martell et al. 1998). This includes uncer-
tainty in products prices/demands, fuel costs, labour costs, and productivity that
are typical of all production planning problems. However, there are fundamental
forestry uncertainties. These include uncertainty in the amount and quality of trees
104 E. Gunn

growing on a piece of land. Although new technologies are now enabling quite
precise inventories (Pitt and Pineau 2009; Næsset 2007), some uncertainty in the
amount and product breakdown of harvest will always remain. Similarly uncertainty
in harvesting machine productivity, moisture levels and process yields will always
affect the overall cost structure. These are natural products and come with all the
variability of nature. This suggests that a well structured hierarchical planning pro-
cess is a necessity (Gunn 1996). Within this hierarchical process, location issues
play a large role in decision making.
Natural disturbance is a fundamental characteristic of forests. Fires, insect and
disease can quite rapidly fundamentally alter the availability of forests for harvesting
and the location of this availability. It can also change the various spatial indicators
of biodiversity, ecosystem condition, and watershed condition. (James et al. 2007)
This means that following a major disturbance that substantial replanning of forest
harvest strategy and possibly industrial location strategy may be necessary.
There is a location issue that comes up in dealing with natural disturbance;
namely how to allocate the prevention effort. Martell (1994) has indicated ways
of valuing the forest fighting effort and indicating the return for this effort to pre-
vent fires. This is part of a large body of work stemming from Martell’s group at the
University of Toronto (see Martell 2015). At a level of more spatial detail, Minas
et al. (2014) have pointed out that appropriate cutting patterns can reduce the spread
of fire and limit adverse impacts. Insects, such as the spruce budworm have also
resulted in considerable spatial analysis in developing decision support tools (see
MacLean et al. 2001, for example). Daugherty and Fried (2007) link fuel treatments
to location of biomass facilities.
At an operational level the detection and the fighting of disturbance also poses
locational challenges. MacLellan and Martell (1996), Islam and Martell (1998) and
Islam et al. (2009) study how the location of bases for forest fire airtankers, the
allocation of the airtankers to these bases and the strategy of deployment of the air-
tankers to the fire locations affects both cost and effectiveness of the fire suppression
effort.

4.14 Summary

We have initially focused on the models for strategic forest management and indus-
trial capacity location. The forest management models start with the requirement
to deal with the spatial entities required to characterize biodiversity, ecosystem
condition and watershed condition, key criteria of sustainable forest management.
Strategic decisions on the indicator levels for these spatial entities will drive or con-
strain harvest strategies. Other criteria of sustainable forest management, namely
multiple economic benefits require us to deal with location of the forest resource
relative to the processing plants and markets and relative to the resource dependent
communities, including First Nations communities.
4 An Introduction to Industrial Forestry from a Location Perspective 105

We have seen that for harvests to be economical we need to plan these harvests
with regard to the location of the sawmills, pulpmills and energy facilities that will
be the destination for the wood produced. Harvest costs at the forest stands, trans-
portation costs to the facilities and between facilities and the prices at the facilities
will determine what harvest plans make sense. Put simply in order to plan harvests
we need to know where the wood will go. However, we have also seen that deter-
mining what facilities make sense in terms of number, type, size, and location that
we need to know how the amount of wood harvested varies over time by species
type, quality/size, location and harvest cost. Again put simply, in order to decide on
facility capacity and location, we need to know where the wood will come from.
This dichotomy probably cannot be resolved completely algorithmically. There are
issues both of model size and solvability but there are also issues of how system
level benefits can be shared among the parties.
However it is also important to be aware that both of these strategic problems
depend on more detailed models. The assumed availability and cost of wood in
a region depends on the ability to access that wood, both in terms of ecological
acceptability as portrayed through the adjacency models and in terms of road build-
ing. Similarly the ability to transport wood efficiently from the forests may well
depend on investments in the logistics system in terms of investments in terminals to
facilitate intermediate processing and transfer to more efficient long distance trans-
port, including multimodal. These are often seen as tactical issues, but in reality they
are also quite strategic. Investments here change costs, which permit revisiting the
harvest and capacity strategies.
This story is repeated as we become more operational. Better logistics processes
again change costs and again we should revisit out strategies. Even without uncer-
tainty, understanding the complexity of forestry location related decisions requires
a hierarchical family of models used not to get ultimate answers but to continually
explore for opportunities of integration, collaboration and continuous improvement.
Once uncertainty raises its head, a hierarchical approach is inevitable.

References

Andalaft N, Andalaft P, Guignard M, Magendzo A, Wainer A, Weintraub A (2003) A problem

of forest harvesting and road building solved through model strengthening and Lagrangean
relaxation. Oper Res 51(4):613–628
Anthony RN (1965) Planning and control systems: a framework for analysis. Harvard University,
Graduate School of Business Administration, Boston
Bare BB, Briggs DG, Roise JP, Schreuder GF (1984) A survey of systems analysis models in
forestry and the forest products industries. Eur J Oper Res 18(1):1–18
Baskent EZ, Keles S (2005) Spatial forest planning: a review. Ecol Model 188(2):145–173
Canada Parliament Senate. Standing Senate Committee On Agriculture And Forestry (2009) The
Canadian forest sector: past, present, future, interim report of the Standing Senate Committee
on Agriculture And Forestry, Dec 2009. http://www.parl.gc.ca/Content/SEN/Committee/402/
agri/rep/repintdec09-e.pdf. Accessed 13 Apr 2015
106 E. Gunn

Canadian Council of Forest Ministers. (2003). Defining Sustainable Forest Management in Canada
2003. http://ccfm.org/pdf/CI_Booklet_e.pdf. Accessed 13 Apr 2015
Carlgren CG, Carlsson D, Rönnqvist M (2006) Log sorting in forest harvest areas integrated with
transportation planning using backhauling. Scand J For Res 21(3):260–271
Carlsson D, Rönnqvist M (2005) Supply chain management in forestry—-case studies at Södra
Cell AB. Eur J Oper Res 163(3):589–616
Carlsson D, D’Amours S, Martel A, Rönnqvist M (2009) Supply chain planning models in the pulp
and paper industry. INFOR: Inf Syst Oper Res 47(3):167–183
Chauhan SS, Frayret JM, LeBel L (2009) Multi-commodity supply network planning in the forest
supply chain. Eur J Oper Res 196(2):688–696
Church RL (2007) Tactical-level forest management models. In: Weintraub A, Romero C, Bjørndal
T, Epstein R (eds) Handbook of operations research in natural resources, Springer, New York,
pp 317–341
Church RL, Murray AT, Weintraub A (1998) Locational issues in forest management. Locat Sci
6(1):137–153
Church R, Daugherty PJ (1999) Considering intergenerational equity in linear programming-based
forest planning models with MAXMIN objective functions. Forest Sci 45(3):366–373
Constantino M, Martins I, Borges JG (2008) A new mixed-integer programming model for harvest
scheduling subject to maximum area restrictions. Oper Res 56(3):542–551
D’Amours S, Rönnqvist M, Weintraub A (2008) Using operational research for supply chain
planning in the forest products industry. INFOR: Inf Syst Oper Res 46(4):265–281
Daugherty PJ, Fried JS (2007) Jointly optimizing selection of fuel treatments and siting of forest
biomass-based energy production facilities for landscape-scale fire hazard reduction. INFOR:
Inf Syst Oper Res 45(1):17–30
Davis LS, Johnson KN, Bettinger PS, Howard TE (2001) Forest management, 4th edn. McGraw-
Hill, New York
Dykstra DP, Riggs JL (1977) An application of facilities location theory to the design of forest
harvesting areas. AIIE Trans 9(3):270–277
Epstein R, Morales R, Seron J, Weintraub A (1999) Use of OR systems in the Chilean forest
industries. Interfaces 29(1):7–29
Epstein R, Weintraub A, Sapunar P, Nieto E, Sessions JB, Sessions J, Bustamente F, Musante H
(2006) A combinatorial heuristic approach for solving real-size machinery location and road
design problems in forestry planning. Oper Res 54(6):1017–1027
Epstein R, Rönnqvist M, Weintraub A (2007a) Forest transportation. In: Weintraub A, Romero C,
Bjørndal T, Epstein R (eds) Handbook of operations research in natural resources, Springer,
New York, pp 317–341
Epstein R, Karlsson J, Rönnqvist M, Weintraub A (2007b) Harvest operational models in forestry.
In: Handbook of operations research in natural resources. Springer, New York, pp 365–377
Forsberg M, Frisk M, Rönnqvist M (2005) FlowOpt–a decision support tool for strategic and
tactical transportation planning in forestry. Int J For Eng 16(2):101–114
Freidenfelds J (1981) Capacity expansion: analysis of simple models with applications. North-
Holland
Gallant R, O’Brien S (2004) A decision support system for Bowater’s Mersey woodland opera-
tions, senior year project. Department of Industrial Engineering. Dalhousie University, Halifax
Gibson DF, Rodenberg J (1975) Location models for the forest products industry and other
applications. AIIE Trans 7(2):143–152
Gilless JK, Buongiorno J (1985) PELPS: Price-endogenous linear programming system for eco-
nomic modeling. Research division of the college of agricultural and life sciences. University
of Wisconsin, Madison
Goycoolea M, Murray AT, Barahona F, Epstein R, Weintraub A (2005) Harvest scheduling subject
to maximum area restrictions: exploring exact approaches. Oper Res 53(3):490–500
Gunn EA (1996) Hierarchical planning in forestry: a stochastic programming, decision analytic
perspective. In: Martell DL, Davis LS, Weintraub A (eds) Proceedings of the workshop on
hierarchical approaches to forest management in public and private organizations. University of
4 An Introduction to Industrial Forestry from a Location Perspective 107

Toronto, Petawawa National Forestry Institute, Information Report PI-X-124, Toronto, Ontario,
Canada, 25–29 May, Canadian Forest Service (1996), pp 85–97
Gunn EA (2005) Sustainable forest management: control, adaptive management, hierarchical
planning. In: Bevers M, Barrett M, tech. comps. 2005. System analysis in forest resources:
proceedings of the 2003 symposium. Gen. Tech. Rep. PNW-GTR-656. U.S. Department of
Agriculture, Portland, OR, Forest Service, Pacific Northwest Research Station, pp 7–14
Gunn EA (2007) Models for strategic forest management. In:Weintraub A, Romero C, Bjørndal T,
Epstein R (eds) Handbook of operations research in natural resources, Springer, New York,
pp 317–341
Gunn, Martell (2015) Decision support needs for strategic planning of Canadian forest value
chains, chapter 4. In D’Amours S., Ouhimmou M., Audy J.-F., Feng Y (eds) Forest value chain
optimization and sustainability. CRC Press/Taylor & Francis.
Gunn E (2009) Some perspectives on strategic forest management models and the forest products
supply chain. INFOR: Inf Syst Oper Res 47(3):261–272
Gunn EA, Richards EW (2005) Solving the adjacency problem with stand-centred constraints. Can
J For Res 35(4):832–842
Gunn E, Shereshti N (2012) What are the benefits and risks associated with tightly integrated indus-
try structure. What can one do in the design of the industry to mitigate the risk of uncorrelated
final product markets? Project 1.17, NSERC/FPInnovations Strategic Research Network, Value
Chain Optimization in the Forest Industry (2012–2015)
Gunnarsson H, Rönnqvist M, Lundgren JT (2004) Supply chain modelling of forest fuel. Eur J
Oper Res 158(1):103–123
Gunnarsson H, Rönnqvist M, Carlsson D (2006) A combined terminal location and ship routing
problem. J Oper Res Soc 57(8):928–938
Imase M, Waxman BM (1991) Dynamic Steiner tree problem. SIAM J Discret Math 4(3):369–384
Islam KM, Martell DL (1998) Performance of initial attack airtanker systems with interacting bases
and variable initial attack ranges. Can J For Res 28(10):1448–1455
Islam KS, Martell DL, Posner MJ (2009) A time-dependent spatial queueing model for the daily
deployment of airtankers for forest fire control. INFOR: Inf Syst Oper Res 47(4):319–333
James PM, Fortin MJ, Fall A, Kneeshaw D, Messier C (2007) The effects of spatial legacies
following shifting management practices and fire on boreal forest age structure. Ecosystems
10(8):1261–1277
Jena SD, Cordeau JF, Gendron B (2012) Modeling and solving a logging camp location problem.
Ann Oper Res 1–27 doi:10.1007/s10479-012-1278-z
Johnson KN, Scheurman HL (1977) Techniques for prescribing optimal timber harvest and invest-
ment under different objectives—discussion and synthesis. For Sci 23 (Suppl. 18):a0001–z0001
Karlsson J, Rönnqvist M, Bergström J (2003) Short-term harvest planning including scheduling of
harvest crews. Int Trans Oper Res 10(5) 413–431
Lebow PK, Spelter H, Ince PJ (2003) FPL-PELPS, a price endogenous linear programing system
for economic modelling; supplement to PELPS III, V 1.1, USDA forest service, forest products
laboratory research paper FPL-RP-614, Madiaon, WI. 32 p
Legües AD, Ferland JA, Ribeiro CC, Vera JR, Weintraub A (2007) A tabu search approach for
solving a difficult forest harvesting machine location problem. Eur J Oper Res 179(3):788–805
Lockwood C, Moore T (1993) Harvest scheduling with spatial constraints: a simulated annealing
approach. Can J For Res 23(3):468–478
MacLean DA, Erdle TA, MacKinnon WE, Porter KB, Beaton KP, Cormier G, Morehouse S, Budd
M (2001) The spruce budworm decision support system: forest protection planning to sustain
long-term wood supply. Can J For Res 31(10):1742–1757
MacLellan JI, Martell DL (1996) Basing airtankers for forest fire control in Ontario. Oper Res
44(5):677–686
Marques AF, Borges JG, Sousa P, Pinho AM (2011) An enterprise architecture approach to for-
est management support systems design: an application to pulpwood supply management in
Portugal. Eur J For Res 130(6):935–948
Martell DL (1994) The impact of fire on timber supply in Ontario. For Chron 70(2):164–173
108 E. Gunn

Martell DL (2015) Fire management systems laboratory. http://www.firelab.utoronto.ca/

publications/dlm_pubs.html. Accessed 9 Apr 2015
Martell DL, Gunn EA, Weintraub A (1998) Forest management challenges for operational
researchers. Eur J Oper Res 104(1):1–17
Martin AB (2013) A linear programming framework for models of forest management
strategy. MASc Thesis, Department of Industrial Engineering, Dalhousie University.
http://dalspace.library.dal.ca/handle/10222/37840. Accessed 9 Apr 2015
Martin AB, Richards EW, Gunn EA (2014a) A comparison of model one and model two for mod-
eling of forest management strategy. Working paper (unpublished). Department of Industrial
Engineering, Dalhousie University
Martin AB, Gunn EA, Richards EW (2014b) Modelling forest management strategy with industry
representation. Working paper (unpublished), Department of Industrial Engineering, Dalhousie
University
Meignan D, Frayret JM, Pesant G, Blouin M (2012) A heuristic approach to automated forest road
location. Can J For Res 42(12):2130–2141
Melo MT, Nickel S, Saldanha-da-Gama F (2009) Facility location and supply chain management—
a review. Eur J Oper Res 196(2):401–412
Minas JP, Hearne JW, Martell DL (2014) A spatial optimisation model for multi-period landscape
level fuel management to mitigate wildfire impacts. Eur J Oper Res 232(2):412–422
Mitchell SA (2004)Operational forest harvest scheduling optimisation: A mathematical model and
solution strategy. (Doctoral dissertation, ResearchSpace@ Auckland)
Montreal Process (2014) Montréal process criteria and indicators for the conservation and
sustainable management of temperate and boreal forests. Technical Notes on Imple-
mentation of the Montréal Process Criteria And Indicators Criteria 1–7, 3rd edn. June
2009 (Rev. July 2014). http://www.montrealprocess.org/documents/publications/techreports/
MontrealProcessTechnicalNotes3rdEditionRevisedJuly2014.pdf. Accessed 9 Apr 2015
Murphy G (1998) Allocation of stands and cutting patterns to logging crews using a tabu search
heuristic. J For Eng 9(1):31–37
Næsset E (2007) Airborne laser scanning as a method in operational forest inventory: status of
accuracy assessments accomplished in Scandinavia. Scand J For Res 22(5):433–442
Nelson JD, Finn ST (1991) The influence of cut-block size and adjacency rules on harvest levels
and road networks. Can J For Res 21(5):595–600
Paradis G, LeBel L, D’Amours S, Bouchard M (2013) On the risk of systematic drift under
incoherent hierarchical forest management planning. Can J For Res 43(5):480–492
Pitt D, Pineau J (2009) Forest inventory research at the Canadian Wood Fibre Centre: notes from a
research coordination workshop, June 3–4, 2009, Pointe Claire, QC. For Chron 85(6):859–869
ReVelle CS, Eiselt HA (2005) Location analysis: a synthesis and survey. Eur J Oper Res
165(1):1–19
Richards EW (1997) A tabu search method for a tactical forest planning problem, Ph. D. Thesis,
Dept. of Industrial Engineering. Dalhousie University, Halifax, N.S. Canada, p 247
Richards EW, Gunn A (2000) A model and tabu search method to optimize stand harvest and road
construction schedules. For Sci 46(2):188–203
Richards EW, Gunn EA (2006) Integer programming models for tactical harvest and access
planning. INFORMS Conference, Pittsburg, PA. (unpublished)
Ryan D (1992) The solution of massive generalised set partitioning problems in aircrew rostering.
J Oper Res Soc 43(5):459–467
Snyder S, ReVelle C (1996) The grid packing problem: selecting a harvesting pattern in an area
with forbidden regions. For Sci 42(1):27–34
Tóth SF, McDill ME, Könnyü N, George S (2013) Testing the use of lazy constraints in solving
area-based adjacency formulations of harvest scheduling models. For Sci 59(2):157–176
Ware GO, Clutter JL (1971) A mathematical programming system for the management of industrial
forests. For Sci 17(4):428–445
Weintraub A, Murray AT (2006) Review of combinatorial problems induced by spatial forest
harvesting planning. Discret Appl Math 154(5):867–879
Weintraub A, Epstein R, Morales R, Seron J, Traverso P (1996) A truck scheduling system
improves efficiency in the forest industries. Interfaces 26(4):1–12
Chapter 5
Layout Planning Problems in Health Care

Ines Arnolds and Stefan Nickel

5.1 Introduction

In general, layout planning problems can be classified as in-house location prob-

lems where the aim is to minimize traveling or material handling costs based on
distances by deciding on the relative positions of any kind of organizational units
inside a building. This class of operations research problems originates from indus-
trial applications, for example, planning the location of different machines of an
assembly line needed to manufacture a product or the arrangement of racks and
shelves within a warehouse.
The special case of layout planning problems in health care has been first intro-
duced by Elshafei in 1977 (Elshafei 1977). He modeled a hospital layout problem
as a quadratic assignment problem (QAP) and developed heuristics to solve it. In
the framework for hospital planning and control the hospital layout planning prob-
lem is classified as a resource capacity planning problem on a strategic level (Hans
et al. 2012). Although it is a long-term decision, the spatial organization within hos-
pitals also directly influences the quality and efficiency of health care and secondary
services of the daily routine (Choudhary et al. 2010; Hignett and Lu 2010) as well
as patient satisfaction (Chaudhury et al. 2005). In practice, hospital buildings are
commonly planned by architects based on experience, design aspects and legal reg-
ulations. Instead of that, it is important to develop and follow a holistic approach
in order to combine the architectural and legal aspects with logistics, i.e., patient,
personnel and material flows inside the future hospital building. In this context, the
established operations research methodologies, especially optimization and simu-
lation techniques, can be applied in order to support finding an optimal or robust
hospital layout. On the one hand, optimal can mean to minimize traveling costs for

I. Arnolds () · S. Nickel

Discrete Optimization and Logistics at the IOR, Karlsruhe Institute of Technology,
Englerstr. 11, Gebäude 11.40, 2. OG, 76131 Karlsruhe, Germany
e-mail: [email protected]
S. Nickel
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 109

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_5
110 I. Arnolds and S. Nickel

personnel or traveling distances and/or times for patients and/or material. Although
these objectives might be conflicting, they not only result in more efficient work-
flows and, thus, in patient and personnel satisfaction but also in economic efficiency.
On the other hand, robustness implies that a layout plan has a good performance
for different scenarios with uncertain input data, for example, uncertain clinical
pathways depending on the patients’ recovery.
A hospital layout planning problem where all functional departments, wards,
surgery rooms and other necessary and supporting areas have to be assigned to loca-
tions inside the hospital building is referred to as a layout planning problem on the
macro level. In contrast, when only planning the layout of a single functional depart-
ment, ward, etc. in the building it is called a hospital layout planning problem on
the micro level. In the next section an overview of the literature on hospital layout
planning problems on both levels is given. Nevertheless, the focus of the applica-
tions detailed in Sect. 5.3 as well as the framework presented in Sect. 5.4 lies on
the macro level. In Sect. 5.5 a summary is given and some practical challenges are
discussed.

5.2 Literature Review

A survey on layout planning problems was conducted by Drira et al. (2007). Further-
more, Tompkins et al. (2010), Heragu (2008), and Francis et al. (1992) published
textbooks presenting different modeling and solution techniques for layout planning
problems in general. In most applications the layout is considered as a long-term
decision (Drira et al. 2007). In such static layout problems all relevant parameters
are assumed to remain constant during the entire planning horizon. Nevertheless,
there may also come up issues that make it necessary to rearrange a given layout,
for example, the development of new products with different production processes
or new treatment procedures that change the clinical pathways of patients with a spe-
cific disease. Thus, dynamic layout problems were developed in order to consider
varying input data during the planning horizon. Two approaches exist to reflect this
variability (Drira et al. 2007; Moslemipour et al. 2012): developing a robust layout
that is best in sum over all periods during the planning horizon (see, for exam-
ple, Kouvelis 1992; Benjaafar and Sheikhzadeh 2000; Azadivar and Wang 2000;
Aiello 2001; Kulturel-Konak et al. 2004; Enea et al. 2005; Braglia et al. 2005; Nor-
man and Smith 2006; Pillai et al. 2011; Arnolds and Nickel 2013b), or developing
a layout plan for multiple periods where layout adaptations are allowed for while
incurring rearrangement costs (see, for example, Lacksonen 1994; Urban 1998;
Yang and Peters 1998; Kochhar and Heragu 1999; Balakrishnan and Cheng 2000;
Chang et al. 2002; Krishnan et al. 2006, 2008; Kulturel-Konak 2007a; Ulutas
and Islier 2009; Bashiri and Dehghan 2010). Reviews on dynamic layout prob-
lems and solution approaches are given in Balakrishnan and Cheng (1998), and
Kulturel-Konak et al. (2007).
5 Layout Planning Problems in Health Care 111

Regarding hospital layout planning problems, an exhaustive search for relevant

literature follows. The objective of this literature review is threefold: First, it is
designed to give an overview on recent advances in layout planning problems in
health care with a focus on hospitals both on the macro and micro level. Second,
a taxonomy, i.e., a structured way of classifying the reviewed papers, is developed
to support discovering linkages between various publications as well as compar-
ing them. Third, the extent to which diverse issues on layout planning in health care
have already been covered in the literature is identified. Thus, existing research gaps
can be revealed.
The following search string was used in the search engine Scopus:
TITLE-ABS-KEY ((“layout” OR “facility planning” OR “facilities planning”
OR “facility design” OR “facilities design”) AND (“hospital” OR “clinic”) AND
(“heuristic*” OR “optimization” OR “mixed integer program*” OR “mathematical
program*” OR “integer program*” OR “linear program*” OR “binary program*”
OR “quadratic program*” OR “dynamic program*” OR “goal program*” OR
“discrete event simulation” OR “discrete-event simulation” OR “discrete-event-
simulation”)).
The asterisk (*) may be replaced by different character combinations. For
example, searching for “mathematical program*” can result in, for example,
“mathematical program” or “mathematical programming”.
Using the presented search string, 59 papers were retrieved. The abstracts of
these papers were screened in order to identify irrelevant articles and sort them
out. After that, 22 relevant papers remained. Based on these, both a forward and
a backward search were conducted. As an indicator for relevant papers the titles,
abstracts and keywords of the forward and backward search results were scanned
for layout and health care relevant applications. Furthermore, the papers cited by
the literature reviews of Jun et al. (1999), Günal and Pidd (2010), Forsberg et al.
(2011) were examined such that an additional 33 papers were retrieved.
In order to categorize the total of 55 papers, the following taxonomy was devel-
oped where most of the papers can be categorized into at least one topic of each
category:
• Scope: This category differentiates between hospital layout planning problems
on the macro and micro level. The micro level is further broken down to differ-
ent organizational units such as operating theater, ward, radiology, emergency
department or other patient service centers.
• Modeling technique: This category refers to the modeling approach such
as quadratic assignment problems (QAP), mixed integer programs (MIP) or
discrete-event simulation models (DES).
• Solution technique: In this category it is differentiated, for example, between
optimal solution approaches, heuristics or process analysis.
• Objectives: This category distinguishes between general facility design aspects,
patient or resource centered objectives, amongst others.
For each of the categories of the developed taxonomy a table is built which gives an
overview on the retrieved research papers (see Tables 5.1–5.4).
Table 5.1 Scope of the reviewed papers
112

Review Micro level Macro level

General Corridor Emergency Gynecology/ Laboratory Wards/ Operating Patient service General Hospital/ Other
allocation department Obstetric Inpatient theater center/ health care settings
problem clinics units Ambulatory setting
care/Outpatient
clinics
Acar et al. (2009) x
Amaral (2012) x
Amladi (1984) x
Arnolds and Nickel x
(2013)
Arnolds and Nickel x
(2013a)
Arnolds and Nickel x
(2013b)
Ashby et al. (2008) x
Assem et al. (2012) x
Baumgart et al. x
(2009)
Becker and Parsons x
(2007)
Borzo (1992) x
Boucherie et al. x
(2012)
Bromley (2012) x
Burn (1982) x
I. Arnolds and S. Nickel
 5

Table 5.1 (continued)

Review Micro level Macro level
General Corridor Emergency Gynecology/ Laboratory Wards/ Operating Patient service General Hospital/ Other
allocation department Obstetric Inpatient theater center/ health care settings
problem clinics units Ambulatory setting
care/Outpatient
clinics
Butler et al. (1992) x
Butler et al. (1992a) x
Ceglowski et al. x
(2005)
Choudhary et al. x
Layout Planning Problems in Health Care

(2010)
Elshafei (1977) x
Fetter and x
Thompson (1965)
Forsberg et al. x x x
(2011)
Freeman et al. x x x
(1974)
Gibson (2007) x
Grigg et al. (2009) x x
Günal and Pidd x x
(2010)
Hahn and Krarup x
(2001)
Halpern et al. (2012) x
113
Table 5.1 (continued)
114

Review Micro level Macro level

General Corridor Emergency Gynecology/ Laboratory Wards/ Operating Patient service General Hospital/ Other
allocation department Obstetric Inpatient theater center/ health care settings
problem clinics units Ambulatory setting
care/Outpatient
clinics
Hamacher et al. x
(2002)
Hancock et al. x
(1978)
Hassan and Tucker x
(2010)
Hassan and Tucker x
(2010a)
Ierardo et al. (2008) x
Iskander and Carter x
(1991)
Jiang and Hu (2012) x
Jun et al. (1999) x x
Kiechle et al. (2005) x
Lee et al. (2012) x
Leung (1992) x
Levy et al. (1989)
Lin et al. (2015) x
Mahachek and x x
Knabe (1984)
I. Arnolds and S. Nickel
 5

Table 5.1 (continued)

Review Micro level Macro level
General Corridor Emergency Gynecology/ Laboratory Wards/ Operating Patient service General Hospital/ Other
allocation department Obstetric Inpatient theater center/ health care settings
problem clinics units Ambulatory setting
care/Outpatient
clinics
Molyneux (2010) x x
Murali (1988) x
Nassar (2010) x
Nickel et al. (2000) x
noAuthor (1975) x
Layout Planning Problems in Health Care

Pagell and Melnyk x

(2004)
Seelye (1982) x x
Sepulveda et al. x
(1999)
Swisher et al. (2000) x x
Thompson and x x
Goldin (1975)
Thorwarth and x
Arisha (2012)
Vos et al. (2007) x
Walus et al. (1997) x x
Weiss et al. (2002) x
Yeh (2006) x
115
 116

Table. 5.2 Modeling techniques of the reviewed papers

Simulation Discrete event Quadratic Multicriteria Stochastic Subset QAP Pedestrian flow Process Mixed Binary Integer Linear Graph- Hopfield Quadratic
simulation assign- QAP QAP modeling modeling integer linear linear model theoretical neural integer goal
ment program program program model network program
program (MIP) (BLP) (ILP)
Acar et al. (2009) x x
Amaral (2012) x
Amladi (1984) x
Arnolds and Nickel x x
(2013)
Arnolds and Nickel x x
(2013a)
Arnolds and Nickel x x
(2013b)
Ashby et al. (2008) x
Assem et al. (2012) x
Baumgart et al. x
(2009)
Boucherie et al. x
(2012)
Butler et al. (1992) x x
Butler et al. (1992a) x x
Choudhary et al. x
(2010)
Elshafei (1977) x
Fetter and x
Thompson (1965)
Forsberg (2011) x
Gibson (2007) x
Günal and Pidd x
(2010)
Hahn and Krarup x
(2001)
I. Arnolds and S. Nickel
 5

Table. 5.2 (continued)

Simulation Discrete event Quadratic Multicriteria Stochastic Subset QAP Pedestrian flow Process Mixed Binary Integer Linear Graph- Hopfield Quadratic
simulation assign- QAP QAP modeling modeling integer linear linear model theoretical neural integer goal
ment program program program model network program
program (MIP) (BLP) (ILP)
Hamacher et al. x x
(2002)
Hancock et al. x
(1978)
Hassan and Tucker x x
(2010)
Hassan and Tucker x
(2010a)
Iskander and Carter x
(1991)
Jiang and Hu (2012) x
Layout Planning Problems in Health Care

Jun et al. (1999) x

Lee et al. (2012) x
Leung (1992) x
Levy et al. (1989) x
Mahachek and x
Knabe (1984)
Nassar (2010) x
Nickel et al. (2000) x
Pagell and Melnyk x
(2004)
Sepulveda et al. x
(1999)
Swisher et al. (2000) x
Thorwarth and x x
Arisha (2012)
Vos et al. (2007) x
Yeh (2006) x
117
 118

Table. 5.3 Solution techniques of the reviewed papers

Simulation Optimization/ Simulation- General Simulated Hill Genetic Evidence Process Questionnaire/ Graph- Systematic Fuzzy Spiral Social Fractional
Exact optimization annealing climbing algorithm based analysis Survey/Interview theoretical layout constraint tech- net- facto-
algorithm design/ approach planning theory nique work rial
Experience (SLP) analy- design
sis
Acar et al. x
(2009)
Amaral (2012) x x
Amladi (1984) x
Arnolds and x
Nickel (2013)
Arnolds and x
Nickel (2013a)
Arnolds and x
Nickel (2013b)
Ashby et al. x
(2008)
Assem et al. x x
(2012)
Baumgart et al. x x
(2009)
Becker and x
Parsons (2007)
Borzo (1992) x
Boucherie et al. x x
(2012)
I. Arnolds and S. Nickel
 5

Table. 5.3 (continued)

Simulation Optimization/ Simulation- General Simulated Hill Genetic Evidence Process Questionnaire/ Graph- Systematic Fuzzy Spiral Social Fractional
Exact optimization annealing climbing algorithm based analysis Survey/Interview theoretical layout constraint tech- net- facto-
algorithm design/ approach planning theory nique work rial
Experience (SLP) analy- design
sis
Bromley x x
(2012)
Burn (1982) x
Butler et al. x
(1992)
Butler et al. x
(1992a)
Ceglowski x
et al. (2005)
Elshafei (1977) x
Layout Planning Problems in Health Care

Fetter and x
Thompson
(1965)
Forsberg x
(2011)
Francis et al. x
(1992)
Gibson (2007) x
Grigg et al. x
(2009)
Hahn and x
Krarup (2001)
Halpern et al. x
(2012)
119
 120

Table. 5.3 (continued)

Simulation Optimization/ Simulation- General Simulated Hill Genetic Evidence Process Questionnaire/ Graph- Systematic Fuzzy Spiral Social Fractional
Exact optimization annealing climbing algorithm based analysis Survey/Interview theoretical layout constraint tech- net- facto-
algorithm design/ approach planning theory nique work rial
Experience (SLP) analy- design
sis
Hamacher et al. x
(2002)
Hancock et al. x
(1978)
Hassan and x x x x
Tucker (2010)
Hassan and x x x x
Tucker (2010a)
Ierardo et al. x
(2008)
Iskander and x
Carter (1991)
Jiang and Hu x
(2012)
Jun et al. x x
(1999)
Kiechle et al.
(2005)
Lee et al. x x
(2012)
Leung (1992) x x
I. Arnolds and S. Nickel
 5

Table. 5.3 (continued)

Simulation Optimization/ Simulation- General Simulated Hill Genetic Evidence Process Questionnaire/ Graph- Systematic Fuzzy Spiral Social Fractional
Exact optimization annealing climbing algorithm based analysis Survey/Interview theoretical layout constraint tech- net- facto-
algorithm design/ approach planning theory nique work rial
Experience (SLP) analy- design
sis
Levy et al. x
(1989)
Lin et al. x x
(2015)
Mahachek and x
Knabe (1984)
Nassar (2010) x
Layout Planning Problems in Health Care

Nickel et al. x
(2000)
Pagell and x x
Melnyk (2004)
Sepulveda et al. x
(1999)
Swisher et al. x x
(2000)
Thompson and x
Goldin (1975)
Thorwarth and x
Arisha (2012)
Vos et al. x
(2007)
Weiss et al. x
(2002)
Yeh (2006) x x
121
Table. 5.4 Objectives of the reviewed papers
122

Patient Personnel Resource Occupancy Facility Distance Workflow Pedestrian/ Cost Closeness
centered centered utilization rate design minimiza- optimiza- Occupant flows minimization rating maxi-
tion tion/Efficiency mization
Acar et al. (2009) x
Amaral (2012) x
Amladi (1984) x x
Arnolds and Nickel (2013) x x
Arnolds and Nickel (2013a) x x x x
Arnolds and Nickel (2013b) x x x
Ashby et al. (2008) x x
Assem et al. (2012) x
Becker and Parsons (2007) x x
Borzo (1992) x x
Boucherie et al. (2012) x x x
Bromley (2012) x
Burn (1982) x
Butler et al. (1992) x x
Butler et al. (1992a) x x
Ceglowski et al. (2005) x x
Choudhary et al. (2010) x
Elshafei (1977) x x
I. Arnolds and S. Nickel
 5

Table. 5.4 (continued)

Patient Personnel Resource Occupancy Facility Distance Workflow Pedestrian/ Cost Closeness
centered centered utilization rate design minimiza- optimiza- Occupant flows minimization rating maxi-
tion tion/Efficiency mization
Fetter and Thompson (1965) x x
Gibson (2007) x x x
Grigg et al. (2009) x x
Hahn and Krarup (2001) x
Halpern et al. (2012) x x x
Hamacher et al. (2002) x x x
Hancock et al. (1978) x
Layout Planning Problems in Health Care

Hassan and Tucker (2010) x

Hassan and Tucker (2010a) x
Ierardo et al. (2008) x
Iskander and Carter (1991) x
Jiang and Hu (2012) x
Jun et al. (1999) x x
Kiechle et al. (2005) x
Lee et al. (2012) x
Leung (1992) x
Levy et al. (1989) x x x
Lin et al. (2015) x
Mahachek and Knabe (1984) x x x
Molyneux (2010) x
123
Table. 5.4 (continued)
124

Patient Personnel Resource Occupancy Facility Distance Workflow Pedestrian/ Cost Closeness
centered centered utilization rate design minimiza- optimiza- Occupant flows minimization rating maxi-
tion tion/Efficiency mization
Murali (1988) x x x
Nassar (2010) x
Nickel et al. (2000) x x x
Pagell and Melnyk (2004) x
Sepulveda et al. (1999) x x
Swisher et al. (2000) x
Thompson and Goldin x x
(1975)
Thorwarth and Arisha (2012) x x
Vos et al. (2007) x x x x
Weiss et al. (2002) x
Yeh (2006) x
I. Arnolds and S. Nickel
5 Layout Planning Problems in Health Care 125

<1974
25 21 1974-1978
20 1979-1983
1984-1988
15
1989-1993
9
10 6 7 1994-1998
5
5 3 1999-2003
1 2 1 2004-2008
0 2009-2013

Fig. 5.1 Number of published papers per 5-year interval

Figure 5.1 shows the statistics with respect to the number of published papers per
5-year interval from 1965 to 2013. Obviously, the topic gets more and more attention
within the scientific community. This could particularly be observed during the last
5 years where the number of published papers more than doubled.

5.3 A Graph-Theoretical Layout Planning Approach Applied

to a Major German Hospital

As already mentioned, hospital buildings still are mainly planned by architects.

They are experts in the field of design, usually have experience from other hospital
planning projects and know the relevant legal regulations with respect to hospital
buildings. Furthermore, there exist first example projects where architects take into
account the processes, i.e., patient, personnel and material flows, when planning a
hospital. But, lack of knowledge can still be detected in some cases and, conse-
quently, no application of decision supporting operations research methodologies
such as optimization or simulation techniques. Nevertheless, in the last years the
authors have had the experience that responsible hospital planners are becoming
more and more interested and open minded towards such kind of decision support-
ing operations research methodologies. Last but not least, this could be a result of
the increasing financial pressure on hospitals.
In what follows, a project is detailed where the authors applied a graph-
theoretical heuristic to support developing a layout plan for a new building of a
major German hospital.

5.3.1 The Cooperating Hospital

The project was initiated by the Institute of Operations Research of the Karlsruhe
Institute of Technology (http://dol.ior.kit.edu) and a major German community hos-
pital with more than 1500 beds. Over the years, the hospital has grown on an area
of 155,758 m2 and currently comprises 23 buildings. The historical development of
126 I. Arnolds and S. Nickel

the different buildings results in long travel distances and times for both patients
and personnel. This impedes process efficiency and, thus, treatment quality as well
as economic efficiency. At the same time, the technical infrastructure of the existing
buildings is not expandable such that, for example, the requirements for the instal-
lation of medical technology cannot be covered anymore. Summarizing, there is a
need for a new building which is also manifested in the target planning program
of the hospital for the upcoming years. According to the hospital’s plan, the new
building shall comprise the following organizational units:
• Wards of different intensities of care.
• Interdisciplinary inpatient and outpatient operating theaters.
• Walk-in clinics for different disciplines.

5.3.2 Goal and Methodology

For the hospital, the objective of the layout planning project was to reduce the long
travel distances and times for patients and personnel which can be observed in the
current setting. This can be achieved by an efficient planning of the location of
the organizational units inside the new building according to the patients’ clinical
pathways and logistic processes that will take place during daily routine.
Three main characteristics of the given layout planning problem have to be con-
sidered when deciding for an appropriate modeling and solution technique: Firstly,
the high number of wards and functional departments to be located; secondly,
the different sizes of the organizational units, and, thirdly, the necessity to plan a
multi-floor building. These characteristics make the problem too complex to find an
optimal solution by formulating and solving a mathematical model, as, for exam-
ple, a quadratic assignment or a mixed integer program. Moreover, these models are
rather appropriate if organizational units have to be (re-)located within a given struc-
ture of an existing building with defined dimensions (length and width of each level
of the building). Contrarily, in this project there still were some degrees of freedom
with respect to the architectural dimensions since a completely new building had to
be developed where only the dimensions of the ground level were fixed beforehand.
To solve this kind of layout problems, a graph theoretical approach was developed
(see, for example, Francis et al. 1992). Particularly, an advantage of this procedure
is that it is illustrative and visually presentable which makes it easier to convince
the responsible hospital managers to apply the approach.

5.3.3 A Graph-Theoretical Approach for Layout Planning

The graph-theoretical approach for layout planning and its theoretical background
is thoroughly described by Francis et al. (1992). In this section a short overview of
5 Layout Planning Problems in Health Care 127

Fig. 5.2 Planar graph with five inner and one outer facet

the procedure and the underlying theory is given. For more detailed information the
interested reader is referred to Francis et al. (1992). As a basis for the application
an interaction or flow matrix is needed, where each entry represents the number
of interactions between two organizational units, for example, during 1 year. This
number can also be interpreted as the importance of locating two organizational
units close to each other. The higher the interaction between two organizational
units, the higher is the importance of a direct adjacency of the corresponding units.
Organizational units can be, for example, machines in a manufacturing environment,
offices in any service setting or wards and functional departments in a hospital. The
idea of the approach is to construct a graph in the first step and then to derive a block
layout from its dual graph in the second step. The objective is to maximize the sum
of the interactions of organizational units that are adjacent in the resulting layout
plan.
In the graph to be constructed each node represents one organizational unit. An
edge between two nodes means that the represented units are adjacent in the result-
ing block layout. In order to derive a layout from the graph the latter has to be planar.
A graph is called planar if it can be drawn on a plane without any crossover of edges
(see Fig. 5.2). Furthermore, a planar graph is called a maximally planar graph if
and only if the characteristic of planarity gets lost when adding a further edge (see
Fig. 5.3).
By depicting a graph, the plane is subdivided in facets γ , i.e., a set of one
or more inner facets and one outer facet. An inner facet is an area in the plane
bounded by nodes and edges where, first, the bounding nodes and edges form an
elementary circle (consisting only of different nodes and edges), second, a pair-
wise intersection of two facets is empty, and, third, no subset of a facet features
the former two characteristics. In Fig. 5.2 the following areas are inner facets:
γ1,2,3,4 , γ1,4,6,7 , γ2,3,5,8 , γ3,4,5,6 and γ5,6,7,8 . An outer facet is the area in
the plane which is not covered by inner facets. In Fig. 5.2 the outer facet is γ1,2,7,8 .
A facet which is bounded by the three nodes i, j and k and the three edges [i, j],
[j, k], and [k, i] is called a triangle δi,j,k (see Fig. 5.3). If all facets of the graph are
128 I. Arnolds and S. Nickel

Fig. 5.3 Maximally planar

graph (deltahedron) consist-
ing of eight triangles (seven
inner and one outer facet)

triangles it is called a deltahedron. A planar graph is maximal if and only if it is a

deltahedron.
A graph is connected if there is a walk between every pair of vertices. A walk in
a graph is an alternating sequence of vertices and edges W = v0 , e1 , v1 , ..., en , vn
such that for j = 1, . . ., n the vertices vj −1 and vj are the endpoints of edge ej . A
simple graph is a graph that has no self-loops or multi-edges. A simple graph is a
complete graph if every pair of vertices is joined by an edge.
If a planar graph representing the adjacency requirements of several organiza-
tional units can be constructed, then it is possible to derive a compatible block layout
using its dual graph. Given a connected, undirected and planar graph, then the dual
graph is built as follows:
• The dual graph includes exactly one node for each facet of the primal graph.
• The dual graph includes exactly one edge for each edge in the primal graph that
separates two facets. In the dual graph, this edge connects the two nodes that
represent these two facets in the primal graph.
To solve the layout problem by applying the graph-theoretical approach, first, a
planar graph has to be determined in which the sum of the edge weights, which rep-
resent the corresponding entries of the interaction matrix, is maximized. Thus, the
search for such a graph can be limited to maximally planar graphs and, consequently,
to deltahedrons.
If a graph is a deltahedron with n nodes and m edges, then the following relation
holds: m = 3n − 6 (see Francis et al. 1992). Using this relation, an upper bound for
the sum of the edge weights of a layout problem can be derived by simply summing
up the 3n − 6 highest values of the interaction matrix.
Given an interaction matrix U = (uij ) and a simple, complete and undirected
graph G = [V , E, U] with nodes v ∈ V and edges e ∈ E which are weighted
with the values
 in U. The problem to identify a maximally planar subgraph G =
V , E , U of G with E  ⊂ E, which has the highest sum of edge weights, can be


formulated as follows:
Decision variables:
⎧  
⎨ 1 if edge i, j ∈ E 
xij = ∀i, j with i < j
⎩
0 else
5 Layout Planning Problems in Health Care 129

Model:

n 
n
Max uij xij
i=1 j =1
s.t. G = [V , E  ] is a maximal planar graph


xij ∈ {0, 1]∀i, j with i < j .

Since the problem is NP-hard, large instances cannot be solved to optimality. Thus,
the following heuristic procedure was developed to construct a maximally planar
graph (see Leung 1992).
• Prerequisite: Calculate the row sums of the interaction matrix and sort them
according to monotonic decreasing values.
• Initialization: Build an initial deltahedron using the four nodes with the highest
row sums.
• Iterations: Integrate the remaining nodes into the deltahedron in the order of
decreasing row sums. Always include them into the triangle where the objec-
tive function value is increased the most. Here, including means connecting the
new node with the three existing nodes of the triangle. Thus, the new node is
connected with the nodes of that triangle with which it has the most interactions.
Regarding the quality of this heuristic, Leung (1992) showed that the sum of the
edge weights of the obtained graphs using this heuristic lie between 92.4 and 99.8 %
of the upper bound on the optimal value generated by summing up the weight of the
(3n − 6) edges of maximum weight.
Example 5.1. Given the following interaction matrix U = (uij ) of a complete graph
with six nodes v ∈ V and row sums ui :
⎛ ⎞ ⎛ ⎞
− 4 7 1 7 10 29
⎜ ⎟ ⎜ ⎟
⎜ 4 − 2 5 2 4⎟ ⎜17⎟
⎜ ⎟ ⎜ ⎟
⎜ 7 2 − 8 8 1⎟ ⎜26⎟
⎜ ⎟ ⎜ ⎟
U=⎜ ⎟, ui = ⎜ ⎟
⎜ 1 5 8 − 1 3⎟ ⎜18⎟
⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎝ 7 2 8 1 − 5⎠ ⎝23⎠
10 4 1 3 5 − 23

Initialization
• Nodes with highest row sums: V  = {1,3, 5,6}
• Corresponding edges: E  = {[1,3], [1,5], [1,6], [3,5], [3,6], [5,6]}
• Corresponding triangles: Δ = {δ135 , δ136 , δ156 , δ356 }
• Objective function value: Z = 7 + 7 + 10 + 8 + 1 + 5 = 38
• Resulting graph (Note: In order to construct the initial deltahedron, one node has
to be chosen to be put in the center, here node 3 is chosen.):
130 I. Arnolds and S. Nickel

Iteration it = 1
• Choose v = 4, because u4 = max{u2 = 17, u4 = 18}

it 1
v 4
δ135 1 + 8 + 1 = 10
δ136 1 + 8 + 3 = 12∗
δ156 1 + 1 + 3 = 5
δ356 8 + 1 + 3 = 12
Note: The table shows the values representing the increase of the objective func-
tion value when integrating node 4 in the existing triangles. The highest increase is
marked with an asterisk, ties are broken arbitrarily.
• Resulting graph:
5 Layout Planning Problems in Health Care 131

Iteration it = 2
• Choose the last remaining node v = 2

it 1 2
v 4 2
δ135 10 8
δ136 12∗ −
δ156 5 10
δ356 12 8
δ134 11
δ146 13∗
δ346 11
• Resulting graph:

If all nodes are inserted the objective function value can be calculated by summing
up the edge weights (see Fig. 5.4): Z = 63
Before constructing the dual graph and deriving a block layout, an additional
node has to be integrated into the maximally planar graph. This node represents the
environment of the new building and, thus, the origin of all flows before entering the
building. Constructing the block layout from the dual graph without this additional
node would result in one organizational unit located outside the building. In Fig. 5.5,
node (7) represents the environment. This new node has to be connected to all the
nodes that formerly established the outer facet of the maximally planar graph.
When constructing the dual graph it has to be ensured that the node representing
the environment lies outside the dual graph (see Fig. 5.6).
One possible block layout to be derived from the dual graph is shown in Fig. 5.7.
Note that no area restrictions are given for the organizational units.
132 I. Arnolds and S. Nickel

Fig. 5.4 Constructed graph

with edge weights

Fig. 5.5 Additional node (7)

for the environment

Fig. 5.6 Construction of the

dual graph (red)

Fig. 5.7 Block layout derived

from the dual graph
5 Layout Planning Problems in Health Care 133

5.3.4 Application of the Graph-Theoretical Heuristic to Construct

a Hospital Layout

The hospital layout planning problem instance addressed in this chapter consists of
25 organizational units to be located (see Table 5.5). Each organizational unit com-
prises a set of rooms. In accordance with the hospital management, the rooms were
grouped to organizational units since for the total of more than 700 rooms it would
not have been possible to reliably derive an interaction matrix from the available
data sources. By applying the graph-theoretical heuristic a two-dimensional block
layout in the plane, i.e., representing one level, can be constructed. Thus, in the last
step the organizational units have to be assigned manually to the different floors of
the new building.

5.3.4.1 Data

In order to derive a layout by applying the graph-theoretical heuristic, the data

described in the following sections has to be obtained.

5.3.4.1.1 Area of each Organizational Unit

For planning purposes the hospital developed a so-called space allocation plan for
the new building. This plan details the kind and area of each organizational unit.
Table 5.5 gives an overview of the space allocation plan.

5.3.4.1.2 Further Restrictions Regarding the Layout

Before applying the described approach, the hospital management already had
defined the following basics for the new building: The building shall have six levels
where with each level the floor area decreases. The ground level shall have a dimen-
sion of 214 m of length and 54 m of width, including aisles, waiting areas, technical
equipment areas, elevators, stairs etc. For the remaining levels of the building the
area had not been fixed yet.

5.3.4.1.3 Interaction Matrices

The data that was delivered from the hospital to infer the interaction matrix for the
organizational units was the hospital’s specific case mix:
• Number of cases per discipline, for example, eye clinic, dermatology or women’s
clinic.
• Number of surgery cases per discipline, for example, trauma surgery or children’s
surgery.
134 I. Arnolds and S. Nickel

Table 5.5 Space allocation plan

Organizational unit Area [m2 ]
1 Emergency department (ED) 312
2 Medical services (MS) 336
3 Outpatient clinic for anesthesiology (A) 304
4 Outpatient clinic for eyes (OC1) 304
5 Outpatient clinic for abdominal surgery (OC2) 224
6 Outpatient clinic for vascular surgery (OC3) 280
7 Outpatient clinic for otolaryngology (OC4) 680
8 Outpatient clinic for oral and maxillofacial surgery (OC5) 352
9 Outpatient clinic for urology (OC6) 272
10 Outpatient cancer center (OC7) 152
11 Inpatient surgery unit (IS) 2720
12 Outpatient surgery unit (OS) 272
13 Standby rooms (SR) 192
14 General care unit 1 (GCU1) 1064
15 General care unit 2 (GCU2) 1064
16 General care unit 3 (GCU3) 1064
17 General care unit 4 (GCU4) 1064
18 Intermediate care unit 1 (IMC1) 1064
19 Intermediate care unit 2 (IMC2) 1064
20 Intensive care unit 1 (ICU1) 794
21 Intensive care unit 2 (ICU2) 794
22 Intensive care unit 3 (ICU3) 794
23 Intensive care unit 4 (ICU4) 794
24 Sterile goods supply (SG) 1000
25 Medical device supply (MD) 96

The assumptions which had to be made in order to derive the quantitative interac-
tions between all pairs of organizational units from the case mix data are specified
in the following section. Since no detailed information on quantitative flows was
available directly, it had to be derived by making these assumptions. Furthermore,
patient, personnel and material flows were not weighted differently such that the
entries in the interaction matrix equal the sums of the different kinds of flows.
5 Layout Planning Problems in Health Care 135

5.3.4.2 Assumptions

As it is usual in practical applications, not all data which is needed to apply the
approach was available directly. This section deals with the assumptions that had
to be made in order to derive quantitative information regarding the general, sur-
gical and emergency patient flows as well as the sterile goods and medical device
supply and the flows to and from the standby rooms for physicians. Based on these
assumptions and (derived) data the interaction matrix was then set up. Since this is
sensitive data for the hospital management, the absolute values, i.e., the entries of
the interactions matrix, cannot be shown here.
The section on general patient flow refers to assumptions relevant for all in-
and outpatients. Also, the surgical and emergency patient flows are not necessar-
ily excluding each other, that means, an emergency patient may, for example, have
to undergo a surgery. In most of the cases only patient flows were considered as
these could be derived more reliably from the available data than the usually highly
variable personnel and material flows. Nevertheless, for some organizational units,
where patients do not have access to, for example, the central sterilization unit or
standby rooms for physicians, relevant personnel and material flows were inferred
from expert interviews.
Furthermore, there is assumed a 15 % increase on the number of patients per year
based on the current numbers. This rise was stipulated by the hospital management.

5.3.4.2.1 General Patient Flow

Regarding the general patient flow, the first contact point inside the hospital for
all outpatients is the organizational unit called “medical services”. In contrast, all
inpatients are directly guided to their assigned ward. Outpatients leave the hospital
from the corresponding outpatient clinics and inpatients from the discharging ward.
From the case mix data an average length of stay of 6 days from admission
to discharge was calculated and taken as a basis for all patients and intensities of
care. During their treatment process, patients recover and are transferred to less
intense care units. It is assumed that due to reconvalescence 50 % of the intensive
care patients are directly transferred to a general care unit, whereas the remaining
50 % have to be transferred to an intermediate care unit first before being moved to
a general care unit later. The transfers take place in equal shares to each of the four
general care and two intermediate care units, respectively. Furthermore, patients are
only discharged from general care units.

5.3.4.2.2 Surgical Patient Flow

Regarding the surgical patient flow, it is divided between inpatient (59 %) and out-
patient (41 %) surgeries, with a current total of about 21,600 surgeries per year. It
is assumed that each outpatient is operated once. Contrarily, inpatients are at least
136 I. Arnolds and S. Nickel

operated once during their hospital stay, and 10 % have to undergo a second surgery.
Patients may come from each of the care units or from the emergency department.
After surgery, the patients are transferred back to the care unit where they came
from or, in case of the emergency patients, are assigned to one of the different care
units in equal shares.
Before surgery, each outpatient has to sign a consent form and is given a surgical
clearance in the clinic for anesthesiology. Since the clearance has to be completed
at least 1 day before surgery the patients leave the hospital afterwards and return on
the day of their surgery.

5.3.4.2.3 Emergency Patient Flow

About 40 % of all arriving patients are emergency patients. Less than 1 % of those
emergency patients have to be transferred to one of the intensive care units imme-
diately. From the remaining emergency patients 20 % are assigned to one of the
outpatient clinics and 80 % have to be operated immediately.

5.3.4.2.4 Sterile Goods Supply

A sterile good unit is a defined unit of volume used for sterilization of materials in
an autoclave. Sterile good units are needed for each surgery. For the transportation
of sterile goods between the operating theaters and the unit for sterile goods supply
the personnel uses trolleys with a capacity of 12 sterile good units. From the case
mix data it is known that 59 % of the trolleys have to be transported to/from the
inpatient operating theater and 41 % to/from the outpatient operating theater.

5.3.4.2.5 Medical Device Supply

The unit for medical device supply prepares all devices which are needed in the
operating theaters. These are, for example, respiration machines or medical appli-
ances needed for treatment or examination during a surgery. Again, 59 % are
transported to/from the inpatient operating theater and 41 % to/from the outpatient
operating theater.
Furthermore, in the intensive care units artificial respiration devices are used that
have to be cleaned and sterilized for new patients. These devices have also to be
transported between the intensive care units and the unit for medical device supply.
It is assumed, that each intensive care unit needs the same amount of respiration
devices.
5 Layout Planning Problems in Health Care 137

Fig. 5.8 Primal graph

5.3.4.2.6 Standby Rooms for Physicians

According to the medical personnel’s information, it can be assumed that each night
each of the 13 physicians on standby is called twice on average. The physician then
either has to go to the emergency department or to one of the care units. After the
patient’s treatment the physician usually returns to the standby room.

5.3.4.3 Results

The solution of the implemented graph-theoretical heuristic procedure shows how

to construct the resulting maximally planar graph. This primal graph is shown in
Fig. 5.8 with a color code that highlights the different groups of organizational units.
It can be observed that the inpatient surgical unit (11) is connected to the other
organizational units with the highest number of edges or, mathematically spoken, it
is the node with the highest degree.
The environment of the building is represented by node (26). The heuristic has to
be slightly adapted in order to guarantee that this node is part of the outer facet in the
resulting graph. Since all patients enter the hospital building from the environment
this node is one of those chosen in the initialization step of the heuristic. Regarding
138 I. Arnolds and S. Nickel

Fig. 5.9 Construction of the dual graph

the further steps, it then has to be ensured that new nodes are only inserted into one
of the existing inner facets.
Figure 5.9 shows how to construct the dual graph that is needed to derive the
single-floor block layout. Within each facet of the primal graph a new node is located
which is colored red. The red nodes are connected such that each new (red) edge
cuts one edge of the primal graph. Node (26) which represents the environment is
not included in any of the inner facets because, obviously, it has to be outside the
building.
After removing the primal graph from Fig. 5.9 a cluster of outpatient clinics can
be recognized as well as two care clusters, each consisting of one intensive care unit,
two intermediate care units and two general care units (see Fig. 5.10). The inpatient
surgical unit (11) has a quite central position and the medical service unit (02) which
is the first contact point for all outpatients is adjacent to the environment where all
the patients come from.
In the next step, the edges of the dual graph are rectified. The result is a single-
floor block layout as depicted in Fig. 5.11. Again, both the outpatient cluster as well
as the two care clusters are clearly recognizable. Up to now, the real areas of the
organizational units have not been taken into account yet. The sterile goods supply
(24) and the medical device supply (25) units are close to the inpatient (11) and
outpatient (12) surgical units. The emergency department (1) is located at a central
5 Layout Planning Problems in Health Care 139

Fig. 5.10 Dual graph

Fig. 5.11 Single-floor block layout without area restrictions

position next to the medical service for admission (2), the cluster of outpatient clin-
ics (3)-(10) and the outpatient surgery unit (12). The standby rooms for physicians
(13) are placed between the outpatient clinics for anesthesiology (3) and general and
abdominal surgery (5) as well as the medical service (2).
140 I. Arnolds and S. Nickel

Fig. 5.12 Single-floor block layout with area restrictions

Considering the given areas for each organizational unit, the layout plan shown
in Fig. 5.12 can be constructed. Still, this is a single-floor block layout, but now
it considers the dimensions of each organizational unit. In the center, the inpatient
surgical unit (11) is located. To its left and right the two care clusters are placed.
The sterile goods (24) and medical device (25) supply units are adjacent to both the
inpatient (11) and outpatient (12) surgical units. The emergency department (1) and
the outpatient clinics (3)–(10) are located in the south-western part of the floor plan.
The standby rooms are placed between the outpatient clinics for anesthesiology (3),
abdominal surgery (5), vascular surgery (6), urology (9), the cancer center (10) and
one of the intensive care units (14). As all outpatients first have to contact the med-
ical service unit (2), from there they can easily reach the outpatient clinics via the
aisle that has been incorporated additionally.
It can be observed that due to the given area restrictions of each organizational
unit, some of the required adjacencies cannot be preserved anymore. But in accor-
dance with the hospital management, the identified outpatient and care clusters are
kept. Furthermore, the hospital management had already decided for a building with
six levels such that the single-floor block layout had to be adapted accordingly. Dur-
ing this manual adaptation step, the desired width-length dimensions of the ground
floor as well as the requirement that with an increasing level the floor areas shall
decrease (see Sect. 5.3.4.1) have to be kept.
Figure 5.13 depicts the manually generated multi-floor layout plan where the
outpatient and care clusters can still be identified. Level-01 contains the inpatient
surgery unit (11), the sterile goods (24) and medical device (25) supply units as well
as the emergency department (1).
5 Layout Planning Problems in Health Care 141

Fig. 5.13 Multi-floor layout plan

All outpatient clinics (3)–(10) as well as the outpatient surgical unit (12), the
doctor’s standby rooms (13) and the first contact point for all elective patients, i.e.,
the medical services unit (2) are located on level 00.
Through the assignment of the emergency department (1) and the medical ser-
vices unit (2) to different levels, the emergency and elective patient arrivals can
be separated from each other so that overcrowding is prevented. Although the two
surgery units for inpatients (11) and outpatients (12) are located on different levels,
they could easily be connected via “sterile stairs.” These are connecting stairs which
only the personnel may use in order to move from one unit to the other without
the necessity to, for example, change clothes due to hygienic requirements. Nev-
ertheless, the assignment of personnel to either the inpatient or outpatient surgery
142 I. Arnolds and S. Nickel

unit during one shift is usually fixed. Consequently, the usage of the “sterile stairs”
would in most cases only be necessary at the beginning or end of a shift, during the
rest period or due to personnel shortage.
Levels 01 and 02 each contain one care cluster consisting of one general, one
intermediate, and one intensive care unit. Also, levels 03 and 04 each comprise one
general and one intensive care unit. The arrangement in care clusters is advanta-
geous regarding, on the one hand, the possibility of a transfer from one care level
to another within the same specialty located on the same floor due to an improve-
ment or deterioration of the patient’s medical condition. On the other hand, transfers
between the same care levels of different specialties which are located on different
floors are very unusual such that vertical transports can be avoided. But, by locating
the intensive care units on levels 03 and 04 directly above the intermediate care units
on levels 01 and 02, the transfer of patients via an elevator between the intensive and
the intermediate care unit is as easy and convenient as possible since it includes as
less as possible horizontal movements. Furthermore, each care group can be directly
connected to the inpatient surgical unit via elevators what makes the transport to and
from the operating theater very comfortable.

5.3.4.4 Discussion

In this section the results of applying the graph-theoretical heuristic to the real-world
hospital layout problem will be discussed thoroughly with respect to the assump-
tions, the multi-floor aspect, possible layout distortions, and the development of the
layout plan. Although all the potential drawbacks of the presented approach have
been addressed during the application process with common sense solutions there
is still some potential for future research on a more methodological level.

5.3.4.4.1 Assumptions

As already detailed in Sect. 5.3.4.2, some assumptions had to be made to derive the
necessary data for the graph-theoretical heuristic. As far as possible, the available
data was used to set up assumptions, especially regarding realistic patient, person-
nel and material flows through the hospital. Here, the focus was laid on the patient
flow because the available target planning program of the hospital, which served as
data basis to plan the layout, contained no information on the personnel and mate-
rial flows but only on the patients’ case mix related data such as the number of
cases or surgery numbers per discipline. Furthermore, on the one hand, the person-
nel usually belongs to and, consequently, moves within an organizational unit and,
on the other hand, the material flow is highly variable and, therefore, hardly pre-
dictable. Obviously, this assumption regarding the input data influences the result
of the graph-theoretical heuristic. Nevertheless, according to the hospital manage-
ment the constructed layout shows a reasonable arrangement of organizational units
which justifies the data-based assumptions. The results have not been compared
5 Layout Planning Problems in Health Care 143

to another hospital due to the difficulty of finding a similar hospital with respect
to patient flows which are the essential and critical input data for the proposed
approach.
Furthermore, only data relevant for interactions between organizational units
within the new building but not to and from other buildings on the hospital’s cam-
pus were taken into account. Nevertheless, some organizational units which will
not be moved into the new building as, for example, the radiology or endoscopy
departments will have interactions with organizational units which will be located
in the new building. These interactions are not taken into account since the patient
and personnel flows within the building are assumed to be more important than the
flows between buildings. The only effect which the flows between buildings could
have on the solution would be a shift of organizational units inside the new building
nearer to the entrance if they have high interaction rates with organizational units
in the other buildings. But the distance within the new building has a much smaller
share on the total distance than the distance to the corresponding organizational unit
in another building anywhere on the campus. Obviously, the latter part of the dis-
tance cannot be influenced by solving the layout problem because the locations of
all organizational units in any of the other buildings are fixed. This means that the
total distances between organizational units in the existing and new buildings would
only be influenced marginally by different layouts. Furthermore, the hospital man-
agement’s focus was laid on the operational processes inside the new building and
not between buildings.

5.3.4.4.2 Multi-Floor Layout

The layout generated by the graph-theoretical heuristic only comprises one level but
due to the limited space available, the new building has to be a multi-floor building.
Consequently, a manual adaptation is necessary which was justified by identifying
different clusters in the single-floor layout. Nevertheless, some of the adjacencies
had to be neglected when deriving the multi-floor layout plan from the single-floor
layout plan. Obviously, in this step, other layout plans could also be identified but
since expert opinions were included when deriving the clusters and the final layout
the solution could be proven to be adequate. Nevertheless, future research aims at
measuring the potential error of the graph-theoretical approach and developing a
procedure to avoid high errors.

5.3.4.4.3 Layout Distortion

The graph-theoretical heuristic does not consider any area restrictions of the orga-
nizational units. In the presented application the areas of the organizational units
differ with a factor of up to 1:10. Consequently, it has to be ensured manually that
small organizational units are not stretched too much in order to be adjacent to other
organizational units. This problem can be solved quite easily by fixing the shape of
144 I. Arnolds and S. Nickel

small departments and then adjusting the shape and position of the larger depart-
ments accordingly. For example, neighboring organizational units do not have to be
adjacent along the whole width or length but only along a part of it.

5.3.4.4.4 Development of the Layout Plan

When developing the layout plan using the graph-theoretical heuristic, each orga-
nizational unit is regarded as a whole. In our application this means, for example,
that all standby rooms for medical doctors are handled as a group. The possibility
to include single standby rooms in other organizational units is not considered by
the approach. On the one hand, this shortcoming can be overcome by an appropriate
grouping of rooms to organizational units. On the other hand, care needs to be taken
to ensure that for each organizational unit the entries in the interaction matrix are
needed and that the problem becomes more complex with more organizational units
to be located.

5.4 An Iterative Simulation-Optimization Approach

for Hospital Layout Planning

The presented graph-theoretical method is only one of many possible approaches

for layout planning. As most of the procedures that can be found in the literature, it
assumes deterministic data regarding patient, personnel and material flows. In this
section an innovative framework for hospital layout planning is presented that takes
into account the impact of strategic layout decisions on the operational performance
with uncertain process flows (Arnolds and Nickel 2013b). Taking into account this
stochastic influence distinguishes the simulation-optimization approach from the
formerly presented graph-theoretical approach for layout planning where the data is
assumed to be deterministic.
In a preparatory step, the term operational performance has to be defined depend-
ing on the application. For example, if the aim is to improve patient and personnel
flows through the hospital building, the total travel times for patients and personnel
as well as the patients’ waiting times for elevators or personnel can be evaluated.
In order to incorporate process uncertainties in the layout planning phase,
optimization is combined with discrete event simulation (DES). While solving a
mathematical model results in an optimal layout under deterministic data, simula-
tion scenarios help to find a robust layout which will show a good performance even
when patient, personnel and material flows are uncertain. At a later point in time,
the DES model can further be used to test, for example, new schedules for working
hours or the influence of building modifications on workflows.
The idea of an iterative simulation-optimization approach for hospital layout
planning is adapted from Acar et al. (2009) who presented a generic approach to
combine mixed integer programming and simulation in order to solve combinatorial
5 Layout Planning Problems in Health Care 145

Fig. 5.14 Iterative simulation-optimization approach

problems under uncertainty. Regarding the application of this approach for a hos-
pital layout planning problem, an optimization model for layout planning has to be
set up in a first step. In general, any kind of problem formulation can be used. For
example, for small instances one can apply the well-known quadratic assignment
problem (QAP) which has been introduced by Koopmans and Beckmann (1957).
By solving the QAP, each organizational unit is assigned to a predefined location
inside the building. That means, the layout is not constructed from scratch but the
structure of the building with defined dimensions (length and width of each level of
the building) and locations is given. The following flow chart depicts the iterative
simulation-optimization approach adapted to the interaction of a QAP and a DES
model.
As Fig. 5.14 shows, the models interact multiple times, i.e., iteratively until the
stopping criterion is met. The first iteration consists of four basic steps: First, the
QAP is solved assuming deterministic flow data (i.e., generation of a candidate lay-
out n). Second, the DES model is applied to test the generated QAP layout with
stochastic flow data. In this step, it is important to ensure that the simulated objective
value is worse than the optimal objective value. For example, this can be guaran-
teed by assuming an increasing amount of arriving patients. Third, the difference
between the QAP and DES objective values (i.e., impact of uncertainty Nn of can-
didate layout n) is calculated. Fourth, the relevant parameters of the QAP model are
updated:
• Nn : The impact of uncertainty of candidate layout n as calculated in step three.
• Qmin : Minimum simulation result obtained so far from any simulation run.
• Tnjr : Indicator that shows if the binary variable yjr in candidate layout n is 1, i.e.,
if organizational unit j is assigned to location r in candidate layout n of the QAP
result.
• i: Current iteration − 1.
In the next iteration, solving the updated QAP either results in a new candidate
layout and the described steps are repeated or it results in a solution that has already
146 I. Arnolds and S. Nickel

been found in an earlier iteration such that the procedure stops. In the latter case,
the last found QAP solution is saved as the best robust layout. In order to realize
the feedback between the optimization and the simulation models, a generic DES
model was developed, which can be easily adapted to different hospital layout plans
according to the QAP solution.
The QAP model formulation according to Koopmans and Beckmann (1957) is as
follows:
Parameters:
fjk : Flow between organizational units j and k
dr : Distance between locations r and  ·
m: Number of organization units and locations.
The decision variables are:
⎧
⎨ 1 if organizational unit j is assigned to location r
yj r =
⎩
0 else

The model can then be written as:


m 
m 
m 
m
Min fj k drl yj r ykl (5.1)
j =1 k=1 l=1 r=1


m
s.t. yj r = 1 ∀ j ∈ {1, ..., m} (5.2)
r=1


m
yj r = 1 ∀ r ∈ {1, . . . , m} (5.3)
j =1

yj r ∈ {0,1} ∀ j, r ∈ {1, . . . , m} (5.4)

The objective function (1) minimizes the total travel distance between the organi-
zational units. Constraints (2) ensure that each organizational unit is assigned to
exactly one room whereas constraints (3) guarantee that each room is only occu-
pied by one organizational unit. Constraints (4) define the domain of the decision
variables.
In order to facilitate a feedback loop between the optimization and DES solutions
according to Acar et al. (2009), additional parameters, decision variables and con-
straints have to be introduced in the QAP. Furthermore, the objective function has
to be adapted.
The additional parameters are:
i: Current iteration − 1
M: large number
Nn : Impact of uncertainty of candidate solution n
Qmin : Minimum simulation result obtained so far from any simulation run
5 Layout Planning Problems in Health Care 147

⎧
⎨1 if binary variable yj r in candidate solution n is 1
Tnj r =
⎩
0 else

Additional decision variables are:

⎧
⎨ 1 if candidate layout n has already been suggested in a previous iteration
Zn =
⎩
0 else

The resulting model (a modified QAP) is then:


m 
m 
m 
m 
i
Min Z = fj k drl yj r ykl + Nn Zn (5.5)
j =1 k=1 l=1 r=1 n=1


m
s.t. yj r = 1 ∀ j ∈ {1, . . . , m} (5.6)
r=1


m
yj r = 1 ∀ r ∈ {1, . . . , m} (5.7)
j =1

yj r ∈ {0,1} ∀ j, r ∈ {1, . . . , m} (5.8)


m 
m
(2Tnj r yj r − yj r − Tnj r ) ≤ Zn − 1 ∀ n ∈ {1, . . . , i} (5.9)
j =1 r=1


m 
m
(2Tnj r yj r − yj r − Tnj r ) ≥ M (Zn − 1) ∀ n ∈ {1, . . . , i} (5.10)
j =1 r=1
Z ≤ Qmin (5.11)

A penalty factor is added in the objective function (5) which reflects the impact of
uncertainty Nn of a candidate layout n that has already been found and simulated
in an earlier iteration (see Fig. 5.14). Constraints (6)–(8) are the same as constraints
(2)–(4) in the original QAP. Constraints (9) and (10) ensure that the impact of
uncertainty of a previously simulated solution is only incorporated if this solution is
currently considered. Constraints (11) define an upper bound for the objective func-
tion value equal to the smallest found objective value of any previously simulated
solution.
The advantages of the presented simulation-optimization approach are manifold.
Firstly, the impact of the strategic layout decision on the operational performance
with uncertain process flows and increasing demand in the future is considered.
Secondly, the performance and robustness of hospital layouts can be compared and
148 I. Arnolds and S. Nickel

improved for various scenarios. Scenarios can be defined by changing both input
data (extrinsic configuration) and factors, which are revealed during the simulation
run (stochastic influences). The former comprises items like the control, capacity
and number of elevators, the existence of dedicated personnel elevators or personnel
shifts and schedules for breaks. The latter incorporates issues like the uncertainty
of clinical pathways depending on the state of health of patients, the used means
and speed of transportation, patient arrival rates or service durations. Thirdly, by
applying a DES model factors like the aforementioned which are hard to integrate
in mathematical models or heuristics for hospital layout planning can be studied.
Fourthly, the performance of the hospital layout can be evaluated separately for
different patient types. Thus, a fairness factor can be established. Patient types can
be defined, for example, by severity of illness or level of mobility.

5.5 Conclusion

Regarding the application of operations research methodologies to health care prob-

lems in general and to hospital layout problems in particular, it still remains a
challenge to convince the people in charge as, for example, hospital managers, med-
ical doctors or nurses of the positive and supporting effects of these methods. Until
recently, hospitals were built the way they had been built through the centuries,
i.e., hospital designers and planners used to rely mostly on experience and existing
campus outlays for inspiration (Arnolds and Nickel 2013a). Today, hospital design
is shifting towards patient logistics, opening up totally new perspectives. Hospital
budgets, quality of care and patient satisfaction will profit from this transformation.
Mostly, long-term perspectives regarding resource and capacity planning are in
the focus of hospital design. However, the emerging building will also significantly
influence short-term aspects, i.e., operational workflows. Furthermore, most archi-
tectural designers assume that the information they take into account is fixed and
deterministic. However, uncertainty can impact data, for example, on future patient
figures for certain diseases, as can processes, i.e., the flow of patients, personnel and
materials, depending on outcomes and reconvalescence. This uncertainty should be
reflected during the design process. Processes should determine how buildings are
designed, and not vice versa. Consequently, planning should integrate methods for
logistical analysis. Prior to entering the design phase for a new construction project,
an analysis of processes needs to be carried out. In particular, clinical pathways
for the patients to be cared for in the building should be investigated, providing
information on the paths of movement of patients, personnel and material.
The distances travelled can be reduced by an efficient location of organizational
units according to the processes which take place in the building, including the flows
of patients, personnel and material. Reducing distances means savings in time and,
consequently, in resources. Increased efficiency leaves more time to spend on care,
which in turn leads to improved patient and personnel satisfaction. To support these
5 Layout Planning Problems in Health Care 149

improvements in efficiency it is planned to build a layout planning data base for

benchmarking.
The challenge for researchers is now to work together with architects and hospi-
tal managers and to convince them that operations research methodologies can be
used as additional decision support tools besides expert domain knowledge for hos-
pital layout planning problems. Here, particularly discrete event simulation models
can act as a door opener for the practitioners’ acceptance of operations research
methodologies not only in hospital layout planning but also in other health care
topics.

References

Acar Y, Kadipasaoglu SN, Day JM (2009) Incorporating uncertainty in optimal decision mak-
ing: integrating mixed integer programming and simulation to solve combinatorial problems.
Comput Ind Eng 56(1):106–112
Amaral A (2012) The corridor allocation problem. Comput Oper Res 39(12):3325–3330
Amladi P (1984) Outpatient health care facility planning and sizing via computer simulation.
Winter Simulation Conference Proceedings. pp 705–711
Arnolds IV, Nickel S (2013) Multi-period layout planning for hospital wards. Socio-Econ Plan Sci
47(3):220–237
Arnolds IV, Nickel S (2013a) An Iterative Simulation-Optimization Approach for Hospital Lay-
out Planning. Working paper, Institute of Operations Research, Discrete Optimization and
Logistics, Karlsruhe Institute of Technology
Arnolds IV, Nickel S (2013b) Patient Logistics Drive Hospital Construction. http://www.european-
hospital.com/en/article/10973-Patient_Logistics_Drive_Hospital_Construction.html.
Accessed 13 April 2015
Ashby M, Ferrin D, Miller M, Shahi N (2008) Discrete event simulation: optimizing patient flow
and redesign in a replacement facility. Winter Simulation Conference Proceedings. pp 1632–
1636
Assem M, Ouda B, Wahed M (2012) Improving operating theatre design using facilities layout
planning. 2012 Cairo International Biomedical Engineering Conference, CIBEC 2012. pp 109–
113
Azadivar J, Wang F (2000) Facility layout optimization using simulation and genetic algorithms.
Int J Prod Res 38:4369–4383
Balakrishnan J, Cheng CH (1998) Dynamic layout algorithms: a state-of-the-art survey. Omega
26:507–521
Balakrishnan J, Cheng CH (2000) Genetic search and the dynamic layout problem. Comput Oper
Res 27(6):587–593
Bashiri M, Dehghan E (2010) Optimizing a multiple criteria dynamic layout problem using a
simultaneous data envelopment analysis modeling. Int J Comput Sci Eng 2(1):28–35
Baumgart A, Denz C, Bender HJ, Schleppers A (2009) How work context affects operating room
processes: using data mining and computer simulation to analyze facility and process design.
Qual Manage Health Care 18(4):305–314
Becker F, Parsons KS (2007) Hospital facilities and the role of evidence-based design. J Facil
Manage 5(4):253–274
Benjaafar S, Sheikhzadeh M (2000) Design of flexible plant layouts. IIE Trans 32:309–322
Borzo G (1992) New patient tower mixes aesthetics with practicality. Health Care Strateg Manage
10(9):18–20
Boucherie R, Hans E, Hartmann T (2012) Health care logistics and space: accounting for the
physical build environment. Winter Simulation Conference Proceedings
150 I. Arnolds and S. Nickel

Braglia M, Zanoni S, Zavanella L (2005) Layout design in dynamic environments: analytical

issues. Int Trans Op Res 12(1):1–19
Bromley EB (2012) Building patient-centeredness: hospital design as an interpretive act. Soc Sci
Med 75(6):1057–1066
Burn J (1982) Facility design for outpatient surgery and anesthesia. Int Anesthesiol Clin 20(1):135–
151
Butler TW, Karwan KR, Sweigart JR, Reeves GR (1992) An integrative model-based approach to
hospital layout. IIE Trans 24(2):144–152
Butler T, Karwan K, Sweigart J (1992a) Multi-level strategic evaluation of hospital plans and
decisions. J Oper Res Soc 43(7):665–675
Ceglowski R, Churilov L, Wassertheil J (2005) Facilitating decision support in hospital emergency
departments: a process-oriented perspective. Proceedings of the 13th European Conference on
Information Systems, Information Systems in a Rapidly Changing Economy, ECIS 2005
Chang M, Ohkura K, Ueda K, Sugiyama M (2002) A symbiotic evolutionary algorithm for
dynamic facility layout problem. Proceedings of the 2002 Congress on Evolutionary Com-
putation, vol 2. pp 1745–1750
Chaudhury H, Mahmood A, Valente M (2005) Advantages and disadvantages of single- versus
multiple-occupancy rooms in acute care environments: a review and analysis of the literature.
Environ Behav 37(6):760–768
Choudhary R, Bafna S, Heo Y, Hendrich A, Chow M (2010) A predictive model for computing
the influence of space layouts on nurses’ movement in hospital units. J Build Perform Simul
3(3):171–184
Drira A, Pierreval H, Hajri-Gabouj S (2007) Facility layout problems: a survey. Annu Rev Control
31(2):255–267
Elshafei AN (1977) Hospital layout as a quadratic assignment problem. Oper Res Q 28(1):167–179
Enea M, Galante G, Panascia E (2005) The facility layout problem approached using a fuzzy model
and a genetic search. J Intell Manuf 16:303–316
Fetter RB, Thompson JD (1965) The simulation of hospital systems. Oper Res 13(5):689–711
Forsberg H, Aronsson H, Keller CD, Lindblad SE (2011) Managing health care decisions and
improvement through simulation modeling. Qual Manage Health Care 20(1):15–29
Francis RL, McGinnis LF, White JA (1992) Facility layout and location: an analytical approach,
2nd edn. Prentice-Hall, Englewood Cliffs
Freeman J, Grimes R, Greene L (1974) Proceedings of the fifth Annual Conference of the Hos-
pital and Health Services Division A.I.I.E., Houston, Texas, Feb. 1974. Abstracts of Hospital
Management Studies 11/2, 12627MN, 268 p
Gibson I (2007) An approach to hospital planning and design using discrete event simulation.
Simulation Conference, 2007 Winter. pp 1501–1509
Grigg SJ, Garrett SK, Miller MK (2009) Helping a hospital shine. Ind Eng: IE 41(10):24–29
Günal M, Pidd M (2010) Discrete event simulation for performance modelling in health care: a
review of the literature. J Simul 4(1):42–51
Hahn P, Krarup J (2001) A hospital facility layout problem finally solved. J Intell Manuf
12(5–6):487–496
Halpern P, Goldberg S, Keng J, Koenig K (2012) Principles of Emergency Department
facility design for optimal management of mass-casualty incidents. Prehosp Disaster
Med 27(2):204–212
Hamacher H, Nickel S, Tenfelde-Podehl D, Chamoni P, Leisten R, Martin A, Minnemann J,
Stadtler H (2002) Facilities Layout for Social Institutions. Operations Research Proceedings
2001. Springer-Verlag, Berlin, pp 229–236
Hancock W, Magerlein D, Storer R, Martin J (1978) Parameters affecting hospital occupancy and
implications for facility sizing. Health Serv Res 13(3):276–289
Hans EW, van Houdenhoven M, Hulshof PJH (2012) A framework for healthcare planning and
control. In: Hall R (ed) Handbook of healthcare system scheduling. Springer, US 168:303–320
Hassan FB, Tucker A (2010) Using cellular automata pedestrian flow statistics with heuristic
search to automatically design spatial layout. Proceedings—International Conference on Tools
with Artificial Intelligence ICTAI 2. pp 32–37
5 Layout Planning Problems in Health Care 151

Hassan FB, Tucker A (2010a) Using uniform crossover to refine simulated annealing solutions for
automatic design of spatial layouts. ICEC 2010—Proceedings of the International Conference
on Evolutionary Computation. pp 373–379
Heragu SS (2008) Facilities design, 3rd edn. CRC Press, Boca Raton
Hignett S, Lu J (2010) Space to care and treat safely in acute hospitals: recommendations from
1866 to 2008. Appl Ergon 41(5):666–673
Ierardo G, Luzzi V, Vestri A, Sfasciotti G, Polimeni A (2008) Evaluation of customer satisfaction
at the Department of Paediatric Dentistry of “Sapienza” University of Rome. Eur J Paediatr
Dent 9(1):30–36
Iskander WH, Carter D (1991) A simulation model for a same day care facility at a university
hospital. Winter Simulation Conference Proceedings. pp 846–853
Jiang H, Hu Y (2012) Subset quadratic assignment problem. Lecture Notes in Computer
Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in
Bioinformatics) 7345 LNAI. pp 226–230
Jun J, Jacobson S, Swisher J (1999) Application of discrete-event simulation in health care clinics:
a survey. J Oper Res Soc 50(2):109–123
Kiechle F, Holland C, Karcher RB (2005) Laboratory design. J Clin Ligand Assay 25(4):186–197
Kochhar JS, Heragu SS (1999) Facility layout design in a changing environment. Int J Prod Res
37(11):2429–2446
Koopmans T, Beckmann M (1957) Assignment problems and the location of economic activities
econometrica. Econom Soc 25(1):53–76
Kouvelis P, Kurawarwala AA, Gutiérrez GJ (1992) Algorithms for robust single and multiple
period layout planning for manufacturing systems. Eur J Oper Res 63(2):287–303
Krishnan KK, Cheraghi SH, Nayak CN (2006) Solving dynamic facility layout problems using
dynamic from between charts. 2006 IIE Annual Conference and Exhibition
Krishnan KK, Cheraghi SH, Nayak CN (2008) Facility layout design for multiple production
scenarios in a dynamic environment. Int J Ind Syst Eng 3(2):105–133
Kulturel-Konak S (2007a) Approaches to uncertainties in facility layout problems: perspectives at
the beginning of the 21st Century. J Intell Manuf 18:273–284
Kulturel-Konak S, Smith AE, Norman BA (2004) Layout optimization considering production
uncertainty and routing flexibility. Int J Prod Res 42(21):4475–4493
Kulturel-Konak S, Smith AE, Norman BA (2007) Bi-objective facility expansion and relayout
considering monuments. IIE Trans 39(7):747–761
Lacksonen TA (1994) Static and dynamic layout problems with varying areas. J Oper Res Soc
45(1):59–69
Lee HY, Yang IT, Lin YC (2012) Laying out the occupant flows in public buildings for operating
efficiency. Build Environ 51:231–242
Leung J (1992) A new graph-theoretic heuristic for facility layout. Manage Sci 38(4):594–605
Levy J, Watford B, Owen V (1989) Simulation analysis of an outpatient services facility. J Soc
Health Syst 1(2):35–49
Lin QL, Liu HC, Wang DJ, Liu L (2015) Integrating systematic layout planning with fuzzy
constraint theory to design and optimize the facility layout for operating theatre in hospitals. J
Intell Manuf (26):87–95
Mahachek AR, Knabe TL (1984) Computer simulation of patient flow in obstetrical/gynecology
clinics. Simulation 43(2):95–101
Molyneux E (2010) Paediatric emergency care in resource-constrained health services is usually
neglected: time for change. Ann Trop Paediatr 30(1):165–176
Moslemipour G, Lee T, Rilling D (2012) A review of intelligent approaches for designing dynamic
and robust layouts in flexible manufacturing systems. Int J Adv Manuf Technol 60:11–27
Murali NM (1988) Model design layout of circular operation theatres complex and ICCU for
hospitals. IEEE/Engineering in Medicine and Biology Society Annual Conference, Publ. by
IEEE, Piscataway, NJ, United States 10/4. pp 1838–1839
Nassar K (2010) A model for assessing occupant flow in building spaces. Autom Constr
19(8):1027–1036
152 I. Arnolds and S. Nickel

Nickel S, Tenfelde D, Inderfurth K, Schwödiauer G, Domschke W, Juhnke F, Kleinschmidt P,

Wäscher G (eds) (2000) Planning and organisation in the hospital. Oper Res Proc
1999:548–553 (Springer Berlin Heidelberg)
no Author (1975) Examination of case studies in health facilities planning. Abstr Hosp Manage
Stud 11(3):13047 AR:397 p
Norman BA, Smith AE (2006) A continuous approach to considering uncertainty in facility
design. Comput Oper Res 33(6):1760–1775
Pagell M, Melnyk S (2004) Assessing the impact of alternative manufacturing layouts in a service
setting. J Oper Manage 22(4):413–429
Pillai VM, Hunagund IB, Krishnan KK (2011) Design of robust layout for dynamic plant layout
problems. Comput Ind Eng 61(3):813–823
Seelye A (1982) Hospital ward layout and nurse staffing. J Adv Nurs 7(3):195–201
Sepulveda JA, Thompson WJ, Baesler FF, Alvarez MI, Cahoon LE III (1999) Use of simulation for
process improvement in a cancer treatment center. Winter Simulation Conference Proceedings,
IEEE, Piscataway, NJ, United States vol 2. pp 1541–1548
Swisher J, Jacobson S, Jun J, Balci O (2000) Modeling and analyzing a physician clinic
environment using discrete-event (visual) simulation. Comput Oper Res 28(2):105–125
Thompson J, Goldin G (1975) The hospital: a social and architectural history. Yale University
Press, New Have
Thorwarth M, Arisha A (2012) A simulation-based decision support system to model complex
demand driven healthcare facilities. Proceedings—Winter Simulation Conference
Tompkins JA, White JA, Bozer YA, Tanchoco JMA (2010) Facilities planning, 4th edn. Wiley,
Hoboken
Ulutas BH, Islier AA (2009) A clonal selection algorithm for dynamic facility layout problems. J
Manuf Syst 28(4):123–131
Urban TL (1998) Solution procedures for the dynamic facility layout problem. Ann Oper Res
76:323–342
Vos L, Groothuis S, Van Merode G (2007) Evaluating hospital design from an operations
management perspective. Health Care Manage Sci 10(4):357–364
Walus Y, Ittmann HC, Hanmer L (1997) Decision support systems in health care. Methods Inf
Med 36(2):82–91
Weiss G, von Baer R, Riedl S (2002) Einfluss des Raumkonzepts einer Operationsabteilung auf
die Nutzungseffizienz. Der Chirurg 73(2):174–179
Yang T, Peters BA (1998) Flexible machine layout design for dynamic and uncertain production
environments. Eur J Oper Res 108(1):49–64
Yeh IC (2006) Architectural layout optimization using annealed neural network. Autom Constr
15(4):531–539
 Part II
Public Services
Chapter 6
Modeling the Potential for Critical Habitat

Richard L. Church, Matthew R. Niblett and Ross A. Gerrard

6.1 Introduction

Planners frequently have some objective or outcome that they wish to address or
understand. Location modeling is an approach that is regularly used as an aid
in understanding the effect(s) of a particular policy or plan, or to provide possi-
ble alternatives for expanding, reorganizing, or contracting a system of facilities.
For example, the Location Set Covering Problem (LSCP, Toregas et al. 1971) has
often been employed to determine the minimum number of fire stations required to
“cover” an entire population or set of demands. A demand is said to be “covered”
if it is within the limits of a specified maximum travel-time or distance standard
from a located station. The LSCP model can be used to determine the most efficient
configuration of stations such that all neighborhoods of a city can be reached within
the standard. By comparing an ideal LSCP solution to an existing set of stations,
one can estimate the efficiency with which a city has deployed their fire protection
services. For example, if a city uses 30 fire stations when it can be served by an
optimal placement of 25 stations, then it might be possible to reduce the needed
number of stations by relocation (ReVelle 1991). In fact, it is a relatively simple
task to modify a model like that of the LSCP to generate and test possible alterna-
tives of system expansion, relocation, and consolidation. The location set covering
problem is just one of many location models that have been applied in a variety
of modeling contexts. Such modeling contexts range from civil applications, such
as fire station placement (Toregas et al. 1971), to environmental applications such

R. L. Church () · M. R. Niblett

University of California, 1832 Ellison Hall Santa Barbara, Santa Barbara, CA 93106, USA
e-mail: [email protected]
M. R. Niblett
e-mail: [email protected]
R. A. Gerrard
USDA Forest Service, PSW Research Station, 1731 Research Park Dr., Davis, CA 95618, USA
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 155

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_6
156 R. L. Church et al.

as the natural reserve site selection and design problems (Underhill 1994; Church
et al. 1996; Williams et al. 2004).
Location models have been used for analysis and design in a number of
environmental problems. Select examples include: designing a groundwater mon-
itoring network (Hudak et al. 1993; Meyer and Brill Jr. 1988); natural reserve site
design and selection (Church et al. 1996; Cova and Church 2000; Fischer and
Church 2003; Malcolm and ReVelle 2005; Matisziw and Murray 2006; Williams
et al. 2004); wildlife management (Downs et al. 2008); and conservation planning
(Church 2013). Interest in environmental applications for which location model-
ing may be used to provide greater insight and planning capabilities when used in
conjunction with Geographic Information Systems (GIS) have been on the increase.
This chapter focuses on the application of a location model in habitat analy-
sis associated with the California Spotted Owl (Strix occidentalis occidentalis).
The spotted owl is a species of concern in forest planning in California, and for-
est managers and biologists have spent a great deal of effort in estimating their
habitat needs as well as planning activities (e.g., fuels removal) in patterns that
minimize the impact to the spotted owl. The application reported here involves the
use of the anti-covering location problem, which was first proposed by Moon and
Chaudhry (1984). This anti-covering location problem has also been referred to as
a packing problem (see, for example, Stephenson 2005). One of the main attributes
for the anti-covering location problem is that any two located facilities may not
be closer than a minimum separation distance standard. Although the application
doesn’t actually involve locating a configuration of facilities, it can be viewed as
locating a set of circular territories and their centers, such that all centers are located
at least a minimum separation distance from all other centers, just as in the anti-
covering location problem. One can think of each circular territory as an idealized
spotted owl territory centered about a feasible nest site or center location. Spotted
owls are quite territorial and tend to keep other owls from encroaching in their area.
In essence, maximizing the number of circular territories that can be placed within
a region is equivalent to estimating the largest number of mated pairs of spotted
owls that an area can support. This represents an estimate of the carrying capacity
of a given habitat area, which is a critical metric in habitat planning. When planned
activities (e.g. thinning of stands to reduce potential fire severity) and natural dis-
turbances (such as an invasion of the bark beetle causing high tree mortality or a
wildfire) alter both the shape and size of a habitat, the model can be used to esti-
mate whether a change in carrying capacity has occurred as well. The anti-covering
location model can be used in a similar manner for other territorial species as well.

6.2 Background

Conservationists tend to focus on protecting and sustaining what they believe to

be core habitat, hot spots of high species richness, and locations containing rare
endemic species. “Core habitat” is a concept that is often difficult to quantify and
6 Modeling the Potential for Critical Habitat 157

is frequently estimated subjectively by expert opinion (Church 2013). Often, core

habitats represent the interior of a continuous swath of good habitat that is free
of edge effects. It also must satisfy some notion of minimum size. For example,
medium sized core habitat for the black bear and moose in Vermont is an unfrag-
mented landscape of suitable habitat comprised of an area ranging between 1500
and 10,000 acres in size (Austin et al. 2006). Also, core may be a smaller area, but
be reasonably close to another area of suitable habitat. For example core habitat
for the prairie chicken is defined as a mixed grass or tall-grass prairie of at least
5000 acres or an area between 1250 and 5000 acres that is no more than 6.5 miles
from another habitat of at least 1250 acres (Hagen et al. 2004). So, size, quality, and
distance relationships may be important in defining core habitat. Modeling suitable
habitat, let alone core, can be a daunting task. The classic approach is to develop
a wildlife habitat relationship model (WHR) that can be used to predict the pres-
ence of a specific animal based upon the geographic distribution of needed habitat
elements, which may vary with season and vary when they are raising offspring.
For example, the State of California maintains a geographic database that depicts
the possible presence and suitability of 694 terrestrial vertebrates based upon the
application of WHR models to geographic data.
Our focus here is on the estimation of the carrying capacity of an area of suitable
habitat. Conservation planners often estimate the carrying capacity of a habitat for
a territorial species by first estimating the total area of suitable habitat and then
dividing that estimate by the average size of a given territory. Such estimates are
approximate at best given that suitable habitat is often distributed unevenly across a
landscape.

6.3 Determining the Carrying Capacity Using Core

Habitat Maps

To describe the basic problem of determining the carrying capacity of an area asso-
ciated with a territorial species, consider the hypothetical landscape depicted in
Fig. 1.1. In this figure each raster cell is classified in one of three ways: core habi-
tat, habitat, and matrix lands for a mated pair of fox. This type of classification
follows that used in the well-known FRAGSTATS program, which is designed to
analyze habitat patterns from the perspective of a number of metrics (McGarigal
and Marks 1995). Core habitat in this case involves high quality habitat units that
are surrounded by other core units or by units that are classified as suitable, but not
core. Those areas that are classified as habitat are either of lower quality than what
is needed to be classified as core, or they are of high quality, but are not completely
surrounded by high quality habitat. Matrix lands are classified as everything else
and may even be hostile areas for the fox. Suppose that the habitat requirements
for this territorial species has been defined as needing approximately 37 raster cells
of habitat, of which at least 30 cells should be considered core habitat. The area in
Fig. 1.1 contains 34 units of core habitat, so one might readily conclude that this
158 R. L. Church et al.

habitat could support one mated pair of this species. But, what if we consider a ter-
ritory of approximately 37 units in size? Figure 1.1 depicts two circles of such a
size. Notice that it is not possible to find a compact area containing all core habitat
or nearly so, which would meet the standards of being a feasible territory. Thus, the
territory needs to be larger than the circular territory and even somewhat oblong.
The larger a territory has to be in order to circumscribe enough core to support the
fox and the more elongated that territory is, the greater the energy requirements are
for the kit fox to hunt prey and return to their den to feed their young. The habitat
might support two mated pairs of fox if the habitat within each of the circles is of
high enough quality that each pair can survive. But, it is more likely that the core
habitat is spread out across the region to the extent that it cannot support any mated
pairs of fox, as the energetics required to find enough prey is too great a struggle. In
general, territories that are compact and almost circular are preferred to those that
are not.
Note here that the needed habitat for a mated pair is approximately 37 units
of which 30 units are core. Here there are 34 such units classified as core, so a
simple mathematical estimate of (34 units of core)/(30 units of core needed = 1.13
would suggest that the habitat could support one mated pair of fox, even though it
probably couldn’t support any mated pairs of fox. Thus, the proximity and pattern of
core habitat cannot be captured by a simple mathematical equation when calculating
whether a feasible species territory exists.
There are several ways in which one might compute a potential territorial area
that can support a demographic unit of a species (e.g., a mated pair of fox). The sim-
plest way is to assume that the best possible territory is compact and mostly circular.
Another way is to develop a model which can identify an area that is a contiguous
cluster of spatial cells that is comprised of enough good habitat to support a species,
but at the same time contains little if any hostile areas (Church et al. 2003). In this
chapter we will assume that estimated territories are compact and circular.
If we assume that territories are compact and circular, a more accurate estimate
of the carrying capacity of an area for a territorial species can be generated through
the use of a location model that involves the placement of non-overlapping circular
territories, as the limit on carrying capacity is bounded by the number of such ter-
ritories that can be placed within core habitat. Before we describe this application
and our solution approach, the next section will briefly discuss the literature on the
anti-covering location problem (Fig. 6.1)

6.4 The Anti-Covering Location Problem

The Anti-Covering Location Problem (ACLP) involves identifying a configuration

of facilities that must be separated from each other by at least a minimum speci-
fied distance. Such a problem can be defined on a network or a continuous plane.
For networks it is often assumed that each node is a potential facility site and that
on the plane all points may be considered feasible within some bounded region.
6 Modeling the Potential for Critical Habitat 159

Fig. 6.1 A hypothetical region containing habitat for the kit fox. The circles represent a compact
region of a size that if filled principally with core habitat would meet the requirements in supporting
mated pair of kit fox

The anti-covering location problem on a bounded area can be thought of as a circle

packing problem (Stephenson 2005) or as a special type of packing problem that
can include spheres, boxes and other shapes in two or more dimensions (Dowsland
and Dowsland 1992). From this point we will assume that the set of feasible loca-
tions is represented as a discrete set of points on a Euclidean plane. From among
this set, the objective is to locate as many facilities as possible while ensuring that
each located facility is at least a minimum distance, r, from its nearest neighboring
facility. For the purposes of application, we will assume that the problem is defined
on a Euclidean plane, where all feasible points are defined in advance and are rep-
resented by their x,y coordinates. The distances between any pair of facilities will
be calculated as Euclidean distance. The Anti-Covering Location Problem was first
described and formulated by Moon and Chaudhry (1984). It has also been called the
r-separation location problem (Erkut et al. 1996). Consider the following notation:
i, j indexes representing potential location sites (the entire set is denoted by I)
dij Euclidean distance from potential site i to potential site j where i, j ∈ I
160 R. L. Church et al.

r Euclidean distance separation standard (no two facilities may be closer to each
other than this distance)
Ni = {j :dij < r & i = j };
ni A sufficiently large number to impose locational restrictions (see discussion)

1, if a facility is located at site i
xj =
0, otherwise
Using the above notation we can formulate a model for the anti-covering location
problem as follows:

Max Z = xi (6.1)
i∈I

s.t.

ni xi + xj ≤ ni ∀ i ∈ I (6.2)
j ∈Ni

xi ∈ {0,1} ∀ i ∈ I (6.3)

The objective function (1) maximizes the number of facilities that are located among
the feasible set of sites. Constraints of type (2) enforce the separation requirements
among the located facilities. This constraint is written for each candidate site loca-
tion. If a facility is located at site i, the sum of the selections at all other facilities j
within r-distance of i, the neighborhood Ni , must sum to zero and are thus precluded
from being selected. In essence, any site j that is strictly within r-distance of site i
can be selected for a facility only if site i has not been selected. This type of con-
straint is often termed a neighborhood constraint. The final constraint (3) lists the
integer restrictions on the decision variables. This is a rather simple binary integer
programming problem.
The value of ni was originally defined by Moon and Chaudhry to be M, a large
number. Yoshimoto and Brodie (1994) suggested that the value of ni did not need
to be any larger than the number of members in the set Ni . This fact alone helps
to produce a tighter formulation when solving a problem with integer programming
software. Murray and Church (1996) have shown that the approach of Yoshimoto
and Brodie can be further reduced to a theoretical limit defined by a vertex packing
problem. More recently, Niblett (2014) has shown that without any loss of generality
the value of Ni for a Euclidean based ACLP problem can be set at the value of 5.
There are alternate forms for the ACLP that have been developed which involve
some form of pairwise or higher ordered restrictions. Suppose that two sites i and
j are too close to each other to be both chosen for facilities. In this case, one could
write a constraint of the form:

xi + xj ≤ 1 (6.4)

Constraint (4) is called a pairwise constraint and prevents the simultaneous selection
of both sites i and j. If three sitesi, j, k, are within r distance of each other a triplet
constraint can be written as follows.
6 Modeling the Potential for Critical Habitat 161

xi + xj + xk ≤ 1 (6.5)

This constraint will allow the selection of at most one of these three sites to be
selected, and represents pairwise restrictions for (i, j), (j, k), (i, k). Such conditions
are called clique constraints and can help reduce the computational effort when solv-
ing ACLPs by general purpose software. When solving an ACLP to optimality, the
best modeling approach reported in the literature has been found to be a combina-
tion of neighborhood constraints and clique constraints (Erkut et al. 1996); one such
form is the maximal clique model of Murray and Church (1996).
Of interest here is the potential for using the ACLP model in estimating the car-
rying capacity of a given region of habitat. Clearly, the most desirable territories are
compact and circular, which requires an individual animal to “protect” or defend the
smallest amount of boundary as well as represents a shape that would require the
least amount of energetics in foraging for food. If a territory of suitable habitat needs
to be of a certain size, then this needed size can be used to estimate the radius, s, of a
circle representing the territory. The separation distance between territory centers, r,
would be twice the needed radius of a circular territory, s. Thus, r = 2×s. Given this
needed radius and a habitat map of suitability, then one can solve an ACLP problem
to estimate the maximum carrying capacity of the landscape for such a territorial
species. What is important is that this methodology would be sensitive to the pat-
tern and undulations in the habitat within a region. Several researchers have taken
this approach including Gerrard (2006), Downs et al. (2008), and Church (2013) for
habitat analysis. Downs et al. (2008) have employed the ACLP to estimate the carry-
ing capacity for a population of greater sand-hill cranes (Grus canadensis tabida), in
critical habitat in the State of Ohio. The greater sand-hill crane is a territorial species
that rigorously defends and maintains a territory of roughly 3 km about their nesting
sites. Estimates of the maximum carrying capacity of an area were then compared
to that of the current population, so that restoration efforts could be directed towards
those areas that had not reached their maximum potential. In this chapter we report
on the work of Gerrard (2006) and Church (2013) that involved using the ACLP
model in analyzing the potential carrying capacity for the California Spotted Owl in
forests in California.
Before we describe the work on using the ACLP in modeling spotted owl habi-
tat, it is important to discuss one final element associated with habitat modeling
and planning. Although an ideal territory may be circular and compact, and many
territories may be somewhat circular is shape, they may also be somewhat oblong,
non-circular in shape. Gerrard et al. (2001) have described the features of construct-
ing a map of suitable habitat values for the San Joaquin Kit Fox based upon expert
opinions. The suitability values were based upon a ratio-scale, which means that
they could be added across a habitat patch to define a composite score for such a
patch. The suitability scores for each cell were defined in such a way that “poor”
cells detracted from the value of a patch and high quality cells contributed to the
value of the patch. Experts defined a minimum threshold for which the composite
162 R. L. Church et al.

sum of a patch needed to exceed this amount in order to be considered a viable

supporting habitat patch for some defined demographic unit (in the case of the kit
fox, it was a mated pair). Church et al. (2003) have developed a patch generation
heuristic that can generate a large sample of viable patches. The important point
here is that the ACLP model can also be used to estimate the carrying capacity of
a region by maximizing the selection of these non-circular patches, while ensuring
that no two selected patches overlap. Thus, the ACLP model can be tailored to han-
dle non-circular shapes by generalizing the forms in which neighborhood and clique
memberships are defined.

6.5 The California Spotted Owl and its Habitat

The California Spotted Owl (Strix occidentalis occidentalis) is a bird of prey that
resides primarily in the coniferous, hardwood-coniferous, and hardwood forests of
California’s Coast Range and Sierra Nevada mountain ranges. Because the spotted
owl is a protected species, forest managers must prepare management plans focused
on preserving a large portion of existing habitat with an eye towards enhancing and
increasing core habitat over time. This is also true for forest product companies that
are required to maintain spotted owl populations. Many of the areas for which the
spotted owl has been observed are in old growth conifer and mixed conifer forests
(Verner et al. 1992). Nesting sites have been associated with dense canopy cover and
a variety of tree sizes from medium to large structures (Gutiérrez et al. 1992; LaHaye
et al. 1997; Gerrard 2006). In addition to preference of larger trees and dense canopy,
the spotted owl is an intrasexual territorial species, which preferentially select nest
sites (Verner et al. 1992).
In fact, “An individual or a pair may be dominant in the core area of its home
range but not in the periphery. This tends to produce a regular dispersion [emphasis
added] by effectively excluding other individuals from breeding in the core without
necessarily excluding their presence there as subordinates engaged in other activities
(Brown and Orians 1970, p. 244).” Such nest sites are typically associated with
having high canopy closure and larger trees in the area surrounding the nest site as
well. For spotted owls, the average of the home range core surrounding a nesting
area is approximately 2.4 km2 or a circle of radius 879 m, which is equivalent to an
area of 600 acres (0.937 square miles).
California spotted owl core habitat can be mapped following the approach of
Gerrard et al. (2001). In this case, core habitat can be mapped using a vector dataset
containing the vegetation class polygons available from the US Forest Service. This
data contains estimates of the mix of trees (when present), an age/seral class, as well
as the level of canopy closure within each polygon. Often such polygons are deter-
mined through the use of air-photo interpretation or with remotely sensed satellite
data, which is usually verified through a random set of on-the-ground survey plots.
These polygons can be converted to a raster form. Figure 6.2 shows a map with the
rasterized polygons which have been classified based upon the amount of canopy
6 Modeling the Potential for Critical Habitat 163

Fig. 6.2 Rasterized map of canopy cover polygons with California spotted owl nest sites

closure. Forest biologists defined core habitat to be those areas of mature to late
seral stage forests in which a given location and the surrounding 879 m circle about
it has an average canopy closure of 60 % or greater, and habitat to be those areas in
which the canopy closure within a 879 m circle surrounding a point is greater than
50 %. It is relatively easy to generate a map depicting habitat and core habitat using
GIS (Geographic Information System) functionality. For example, the Focal Mean
Score function of Arc/GIS of ESRI Corporation can be used to calculate the average
canopy surrounding a given pixel for a circular region of 600 acres and map this
164 R. L. Church et al.

Fig. 6.3 Core habitat map derived using GIS functionality. Each raster cell is scored based upon
the average canopy values within a circle of 879 m

average value for that given pixel. Figure 6.3 shows the same area as Fig. 6.2 and
shows the results of using the focal mean function where each pixel is classified in
one of the 3 canopy cover ranges, based upon the average canopy surrounding each
pixel in an 879 m circle (600 acres). One can think of the areas that have an average
canopy > 60 % as being core habitat and potential nest site locations. This figure
also shows known nest site locations. All but three of the nest sites are in the dens-
est canopy class. None of the sites are in locations with less than 50 % canopy cover.
Those sites that are located in areas with an average of 50–59 % canopy cover are
generally very close to those locations that have dense canopy (≥ 60 %). In addition,
6 Modeling the Potential for Critical Habitat 165

one can observe that each of the spotted owl nest sites are roughly evenly distributed
from one another.

6.6 Utilizing the Anti-Covering Location Model for Estimating

Carrying Capacity for the Spotted Owl

One of the main problems in solving the anti-covering location problem in habitat
analysis is the fact that problems can be of considerable size. For the raster discussed
above for the Kings River Project Area (KRPA), nearly 75,000 cells are
possible candidates for selection as potential nest centers out of approximately
1,000,000 cells that comprise this forested area. This easily translates to a model size
that exceeds the capabilities of most commercial solvers. In fact, given that there is
one binary integer variable for each potential nest site center, a simple application
of the ACLP to the KRPA would involve nearly 75,000 binary variables and 150,000
constraints (assuming 1 neighborhood constraint for each potential nest center and
one maximal clique constraint for each potential nest center), which is clearly out of
range of a commercial solver. There are two possible tacks that one can use in solv-
ing a problem of this magnitude. The first is to develop a heuristic and the second is
to solve a set of smaller integer programming problems. The second approach can
be accomplished by dividing the region into a number of smaller areas and applying
the ACLP model to each of these subareas, or by reducing the total number of nest
sites by selecting a sample of nest sites distributed across the region and solving
the ACLP on such a sample. This second approach is also a heuristic in nature, as
the true optimum may not be identified when a problem has been divided into a set
of smaller regions or when only a sample of sites are actually used. We chose to
develop both approaches (a heuristic and an IP based heuristic), although our spon-
sor wanted a process that could quickly solve the problem to within a reasonable
estimate of the carrying capacity. Overall, we wanted to use the ACLP model to
give us a level of confidence that our heuristic could generate close to if not optimal
solutions to the ACLP.
The KRPA does exhibit several distinct areas of feasible nest centers, and it is pos-
sible that we could solve the ACLP model on each of these distinct areas. Because
such a distribution is not always so disconnected, we decided to develop a form
of the IP in which a selected sampling of sites would be used to generate a model
problem instance. We developed a visualization routine in Visual Basic to map out
a given solution to the ACLP. We added a module to this routine to produce an
MPS file that represents a given ACLP model instance and used this as input to the
XPRESS solver (ver. 25.01.05) of FICO Corporation. Our module selects an X%
sample of cells of all feasible nest center cells as potential sites. This module gener-
ated for each potential nest center, one core clique constraint (see Erkut et al. 1996)
and needed pairwise constraints to represent all separation restrictions. We found
166 R. L. Church et al.

Table 6.1 Selected sample problems involving the application of the XPRESS solver to an integer
linear optimization model of an ACLP
% of sites used Number of site Number of Objective value Solution times
selection variables constraints (secs.)
4 3406 242,434 62 230.1
5 4066 334,723 62 989.6
6 4989 492,793 63 3,899.1
7 5551 649,360 63 52,568.5

this to be the best formulation representing this problem. Table 6.1 presents data
associated with the solution of a range of problem instances based upon the per-
centages of sites that were selected for generating the problem instance. Note that
for a 5 % sample of sites, a problem instance contained 4066 binary variables and
334,723 constraints. Some of these constraints were redundant and were eliminated
by the presolve function of XPRESS. The solution time for this problem was 989.6 s
(16.5 min) and involved the selection of 62 nesting site centers and associated circu-
lar territories. Over all sample problems, the largest value for the ACLP objective for
the KRPA solved to optimality was 63. It should be noted here that each X% sample
of sites represents the entire region of feasible sites. By the very fact that it contains
fewer constraints, one might conclude that it may be possible to site more than what
would be otherwise possible, with a fully representative model. But, given that a
solution based upon an X% sample of sites is a feasible anti-covering solution for
the entire problem means that it cannot involve the positioning of any more than the
maximum possible number of circular territories. Therefore, a solution to a sample
based problem instance could be optimal to the original, fully specified problem,
but also may be less than optimal.
Figure 6.4 presents a map of the solution generated for the problem solved with
a 6 % sample of sites. This solution involved the location of 63 circular territories.
Since this is a sample based solution, it is possible that more could be located,
although it is probably unlikely.
We have also developed a heuristic solution process for the ACLP problem when
defined on a raster, where all feasible cells are considered as possible sites rather
than restricting site selection to a small sample of sites as required when solving
an integer linear program. Our heuristic design was based upon two simple obser-
vations associated with a solution of nest centers. The first observation is that for a
solution to be close to optimal, it is likely to have the circular territories packed as
closely together as possible, so that as many of these circular territories are deployed
as possible. The second is if such circular territories are densely packed and often
touch boundaries of neighbors, then local adjustments of a given solution are likely
to violate the separation requirements. We can view this property as one of brit-
tleness, that is readjusting a site center of a densely packed solution may trigger
the need to readjust many other nest centers in order to meet all of the separation
6 Modeling the Potential for Critical Habitat 167

Fig. 6.4 A map of the 63 selected nest centers and their circular territories associated with the
optimal ACLP model solution generated when the sample was set at 6 %. The darkest regions
represent those pixels that can serve as territory centers

conditions among all circular territories. That is, adjusting the position of one terri-
tory will potentially generate a domino effect in needed readjustment of a string of
nest centers. Thus, readjusting a solution to test for local improvement may take a
considerable amount of computational effort with no guarantee that it will be fruit-
ful. Because of these two observations, we elected to concentrate on developing a
heuristic that could generate a densely packed feasible solution. We could then run
the heuristic a large number of times and keep the best solution found among all of
the restarts.
The basic premise is to start the heuristic by selecting one of the feasible nest sites
as the first center for a circular territory of radius r. The second circular territory is
chosen as that feasible center which is as close as possible to the first chosen center.
If there are several possible candidates that are equally close, this second site is cho-
sen at random from among the closest candidates. The third site that is chosen, must
be a site which minimizes a combined distance to the first two chosen centers while
maintaining the desired separation among the chosen sites. The fourth site is the site
that is chosen at random from among those sites that minimize a combined distance
from the other three sites, subject to the condition that no separation conditions are
violated. The product of this first step, often results in a kernel of four clustered sites
and circular territories. Subsequent sites are chosen which minimize the combined
distance to this four site kernel, while maintaining that no site separation conditions
168 R. L. Church et al.

are violated among the chosen sites/circular territories. The process continues until
it is not possible to add any further sites without violating the separation standard.
The process is repeated for a set number of times, among which the best solution
found is reported as the final result.
There are two elements of randomness, which make this process akin to a semi-
greedy process. The first, involves the selection of the initial “seed” center, about
which other selected sites will be clustered. The second is associated with the fact
that at any stage of selection, if there are ties as to which site is closest to the growing
kernel or kernel (of the first four chosen sites), the selection is made at random with
respect to the candidates tied for best. As this is a network of cells defined along a
raster (or grid), it is common that such ties exist. Consequently, this process tends
to expand in a semi-greedy fashion about the first chosen site. As a beginning seed
locations are chosen at random from among the set of feasible nest centers, a large
number of restarts tends to sample many of the areas of a problem landscape, a form
of diversification. Finally, this heuristic by its very nature produces a densely packed
arrangement of nest centers and circular territories, a property often visualized in
optimal ACLP arrangements. We call this heuristic, Packer, as it tries to pack as
many circular territories into a given region.
The application of the Packer heuristic to the KRPA for 1000 restarts is presented
in Table 1.2. There are four columns in this table. The first column presents a given
deployment of circular territories. For example on the first line, it can be seen that 54
territories were located. The second column indicates how many times out of 1000
trials, this result was obtained. In this case it shows that 8 times out of a thousand
trials, 54 territories were located. The third column indicates the frequency with
which this particular number of territories was located, and the fourth indicates how
much such solutions depart from the best solution that was found among the 1000
restarts. For example, the highest value found was 62, which was found 21 times
out of a 1000 restarts, which was approximately 2.1 % of the time. The average over
all restarts was 57.6, and 27.5 % of the time the heuristic found a solution which
was within 5 % of the best found. Overall, the heuristic performance is close to
expectations, as it found a number of close to optimal solutions within a reasonable
amount of computation time (about 4000 s) given the enormous number of sites and
the very nature of a packed solution. It is important to note that the optimal process
was able to locate only one additional territory than the maximum of the heuristic
Table 6.2.

6.7 Final Comments and Conclusions

In this chapter we have described an application of the anti-covering location prob-

lem as a tool to estimate the possible carrying capacity of a territorial species. The
basic idea is one of locating as many compact and circular territories as possible,
while ensuring that each circular territory does not overlap with any other located
territories. Each circular territory must be centered at a location which is considered
6 Modeling the Potential for Critical Habitat 169

Table 6.2 The results of the Packer heuristic applied 1000 times to the Kings River Ranger District
Number of located Number of times Frequency (%) in which a % departure from
circular territories out of 1000 given number are located best solution found
54 8 0.8 12.9
55 54 5.4 11.3
56 181 18.1 9.7
57 282 28.2 8.1
58 200 20.0 6.5
59 163 16.3 4.8
60 32 3.2 3.2
61 59 5.9 1.6
62 21 2.1 0.0

to be ideal habitat as well as contain within the circle enough core habitat to be
considered feasible in the sense that is could support such a species. The number
of located territories, then, is an estimate of the maximum carrying capacity of the
habitat for that species. The anti-covering location problem is an exact representa-
tion of this problem when applied to a Euclidean plane, as the objective is to locate
as many facilities (territory centers) as possible such that they are all separated by at
least 2r-distance from each other (so circular territories of radius r do not overlap).
We show how the ACLP was applied to estimate the number of spotted owls
that can be supported in a given forest region. To do this requires a definition of
suitable habitat as well as what would constitute a potential nest/territorial center
with respect to the site and its surrounding circular shaped territory. We present an
application of this model to the Kings River Project Area of the Sierra National
Forest in California, which was comprised of nearly 75,000 feasible sites, among
which it was possible to locate 63 non-overlapping circular territories.
We described the solution of this problem from the perspective of two heuris-
tic approaches. The first involved solving an integer linear programming problem
where a sample of all sites is used as it was not possible to solve this problem
to optimality when using all sites as potential points for selection. The second
approach was a heuristic algorithm that was designed to generate solutions which
mimic the conditions that can be observed among optimal solutions, that is dense
tightly packed arrangements. This process was designed to be repeated a large num-
ber of times as a form of diversification where the best found is saved for display and
review. Results of both approaches applied to the Kings River Ranger District are
presented. Overall, the heuristic is able to identify close to if not optimal solutions
for such large problems.
The ACLP heuristic was developed along with a visualization routine, so that for-
est planners and biologists could apply this program to any raster data set that was
stored in an ASCII-grid format, a raster data format that can be exported from sev-
eral well-known geographical information systems (GIS) such as Arc/GIS (a product
170 R. L. Church et al.

of ESRI corporation, Redlands, CA). This stand-alone program can be used to esti-
mate the carrying capacity of other territorial species as long as one has an estimate
of the required size of circular territories, geographic information system database
on habitat, and accepted definitions for estimating whether a given territory location
can support the species under consideration. This work is now being tested on habi-
tats involving the fisher (Pekania pennanti) as well as expanded for varying radii
of separation, where larger radii are used for habitats of good quality and smaller
radii are used for habitats of high quality. Finally, feasible noncircular territories are
being modeled for the San Joaquin Kit Fox (Vulpes macrotis) and habitat plans are
being generated by optimizing the selection of these non-circular areas while each
territory is allowed to overlap by a pre-specified amount with other territories.

Acknowledgements We would like to acknowledge the research funding of the US Forest Service
which supported the development of the program Packer and the visualization routine. We also
would like to acknowledge the help and assistance from Pat Manley, Pete Stine, and John Keane
of the US Forest Service.

References

Austin J, Viani K, Hammond F (2006) Vermont wildlife linkage habitat analysis: a GIS-based,
Landscape-level identification of potentially significant wildlife habitats associated with the
State of Vermont Roadways. Vermont Agency of Transportation
Brown JL, Orians GH (1970) Spacing patterns in mobile animals. Annu Rev Ecol Syst 1:239–262
Church RL (2013) Identification and mapping of habitat cores. In: Craighead FL, Convis CL Jr
(eds) Conservation planning: shaping the future. Redlands. ESRI Press, California, (219–244)
Church RL, Stoms DM, Davis FW (1996) Reserve selection as a maximal covering location
problem. Biol Conserv 76:105–112
Church RL, Gerrard RA, Gilpin M, Stine P (2003) Constructing cell-based habitat patches useful
in conservation planning. Ann Assoc Am Geogr 93/4:814–827
Cova TJ, Church RL (2000) Contiguity constraints for single-region site search problems. Geogr
Ann 32/4:306–329
Downs JA, Gates RJ, Murray AT (2008) Estimating carrying capacity for sandhill cranes using
habitat suitability and spatial optimization models. Ecol Model 214:284–292
Dowsland KA, Dowsland WB (1992) Packing problems. Eur J Oper Res 56:2–14
Erkut E, ReVelle C, Ulkusal Y (1996) Integer-friendly formulations for the r-separation problem.
Eur J Oper Res 93:342–351
Fischer DT, Church RL (2003) Clustering and compactness in reserve site selection: an extension
of the biodiversity management area selection model. For Sci 49/4:555–565
Gerrard RA (2006) Measuring habitat quality and carrying capacity for the California spotted owl:
case study using data from Sierra national forest. Interim report compiled for Sierra Nevada
research center and Plumas-Lassen-Tahoe national forests
Gerrard R, Stine, P, Church R, Gilpin M (2001) Habitat evaluation using GIS: a case study applied
to the San Joaquin Kit Fox. Landsc Urban Plan 52:239–255
Gutiérrez R, Verner J, McKelvey KS, Noon BR, Steger GN, Call DR, Senser JS (1992) Habitat
relations of the spotted owl. In: The California spotted owl: a technical assessment of its current
status. Pacific Southwest Research Station, US Dept. of Agriculture. US Forest Service, Gen.
Tech. Rep. PSW-GTR-133. Albany, CA, p. 79–98
6 Modeling the Potential for Critical Habitat 171

Hagen CA, Jamison BE, Giesen KM, Riley TZ (2004) Guidelines for managing lessor prairie-
chicken populations and their habitats. Wildl Soc Bull 32/1:69–82
Hudak PF, Loaiciga HA, Schoolmaster FA (1993) Application of geographic information systems
to groundwater monitoring network design. J Am Water Resour Assoc 29/3:383–390
LaHaye W, Gutiérrez R, Call D (1997) Nest-site selection and reprodutive success of California
spotted owls. Wilson Bulletin 109/1:42–51
Malcolm SA, ReVelle C (2005) Representational success: a new paradigm for achieving species
protection by reserve site selection. Environ Model Assess 10:341–348
Matisziw TC, Murray AT (2006) Promoting species persistence through spatial association
optimization in nature reserve design. J Geograph Syst 8:289–305
McGarigal K, Marks B (1995) FRAGSTATS spatial analysis program for quantifying landscape
structure. General Technical Report PNW-GTR-351, USDA Forest Service, Pacific Northwest
Research Station
Meyer, Philip D, Brill Jr, E Downey (1988) “A method for locating wells in a groundwater
monitoring network under conditions of uncertainty”. Water Resour Res 24(8):1277–1282
Moon ID, Chaudhry SS (1984) An analysis of network location problems with distance constraints.
Manage Sci 30/3:290–307
Murray AT, Church RL (1996) Analyzing clique constraints for imposing adjacency restrictions in
forest models. Forest Science 42:166–175
Niblett MR (2014) The anti-covering location problem: new modeling perspectives and solution
approaches. PhD Dissertation. University of California, Santa Barbara
ReVelle, Charles (1991) “Siting ambulances and fire companies: new tools for planners”. J Am
Plan Assoc 57(4):pp 471–484
Stephenson K (2005) Introduction to circle packing: the theory of discrete analytic functions.
Cambridge University Press, New York
Toregas C, Swain R, ReVelle C, Bergman L (1971) The location of emergency service facilities.
Oper Res 19/6:1363–1373
Underhill LG (1994) Optimal and Suboptimal Reserve Selection Algorithms. Biol Conserv
70:85–87
Verner J, Gutiérrez R, Gould GI Jr (1992) The California spotted owl: general biology and eco-
logical relations. In: The California spotted owl: a technical assessment of its current status.
Pacific Southwest Research Station, US Dept. of Agriculture. US Forest Service, Gen. Tech.
Rep. PSW-GTR-133. Albany, CA, p. 55–77
Williams JC, ReVelle CS, Levin SA (2004) Using mathematical optimization models to design
nature reserves. Front Ecol Environ 2/2:98–105
Yoshimoto A, Brodie JD (1994) Comparative analysis of algorithms to generate adjacency
constraints. Can J For Res 24:1277–1288
Chapter 7
Saving the Forest by Reducing Fire Severity:
Selective Fuels Treatment Location
and Scheduling

Richard L. Church, Matthew R. Niblett, Jesse O’Hanley, Richard Middleton

and Klaus Barber

7.1 Introduction

Wildfire is a natural process which can lead to a variety of conditions in a forested

landscape, some beneficial and some not. For example, in coastal chaparral com-
munities some seeds are activated and germinate only by the heat of a fire. After
a fire, young plants can grow somewhat free of competition. As another example,
the giant sequoia (Sequoiadendron giganteum) is shade intolerant and fire is neces-
sary to clear an understory so that light can reach seedlings. Unfortunately, wildfire
may also be so destructive that sensitive habitats may be reduced in size or signifi-
cantly damaged, even to the point of reducing its viability in supporting threatened
or endangered species. Some believe that such destructive fires are the result of long
term fire suppression (Gruell 2001; Miller et al. 2009; North et al. 2007), however,
this is a hotly debated issue. Fortunately, no one questions that destructive fires often
occur when litter (woody debris from trees) and ladder fuels are abundant leading
to more destructive crown fires.
Finney (2004) has shown with a fire simulation model that fire severity and extent
may be reduced by selectively reducing forest fuels across a forest. Fuels reduction

R. L. Church () · M. R. Niblett

1832 Ellison Hall Santa Barbara, University of California, Santa Barbara, CA 93106, USA
e-mail: [email protected]
M. R. Niblett
e-mail: [email protected]
J. O’Hanley
Kent Business School, Kent University, Canterbury, UK
e-mail: [email protected]
R. Middleton
Los Alamos National Laboratory, Los Alamos, NM, USA
e-mail: [email protected]
K. Barber
US Forest Service Region 5, Vallejo, CA, USA
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 173

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_7
174 R. L. Church et al.

may involve mechanical removal of litter and ladder debris, prescriptive burns of
understory, and thinning of stands. Thinning of stands helps separate the crowns of
trees reducing the possibility of a spreading crown fire. In order to make a forest sys-
tem more resilient to catastrophic fire losses, Finney demonstrated that treatments
do not need to encompass an entire forest, but be strategically placed, scattered and
encompass approximately 25 % of the forest. In fact a set of strategically placed
fuels treatment areas can be almost as effective as a cost-prohibitive approach of
treating an entire forest. This means that there is considerable flexibility in where
treatments can be placed. Further, since fuels treatments are costly, such treatments
must be scheduled over a planning horizon.
In order to conceptualize, develop, and apply a location model for strategic and
tactical decision making, one must first understand the problem as well as the audi-
ence of users and decision makers. Within an industrial organization, a location
model formulation and application may well be done by an in-house systems mod-
eling team or by a consultant. Often these location decisions are one-of a kind,
or are accomplished periodically over a period of years in support of expansion,
consolidation, merging, and other strategic decisions (Geoffrion and Powers 1980;
Camm et al. 1997). In comparison, large governmental agencies may divide their
operations geographically into districts, in which similar problems and decisions
are made, but involve local subject matter experts. These subject matter experts in
local issues and concerns, may find it necessary to solve a variant of a given prob-
lem. Without flexibility in how data is acquired and stored and how such data might
be used to populate a model as well as having flexibility in defining the model itself,
a location model may find use in one district and not another. Thus, there are hur-
dles to overcome when a model is to be applied locally among a number of district
organizations, rather than globally by one organization. In fact it is this dimension
of modeling that was the major hurdle in developing an approach that would be
applied across ranger districts of National Forests in California.
The US Forest Service operates 17 National Forests in California, of which 15
are located in the Sierras or in the northern third of California. Altogether these
forests cover 31,264 square miles (approximately the area of the Czech Republic).
All but two of these forests have been the subject of fuels removal planning. Until
the 1980’s the US forest service managed harvesting operations in many of the west-
ern National Forests. These activities were planned at the strategic level by the use
of large scale linear programming models (Kent et al. 1991) and were assigned
to watershed units by a tactical level model, which also involved linear program-
ming (Church et al. 1998). Concern for environmental impacts to watersheds and
habitat changed the major focus of the USFS to habitat conservation and protec-
tion. Because of this, large scale harvesting virtually disappeared in public forests
in California in the 1990s, with the exception of salvage operations after major fires.
Today, stands may be thinned by harvesting selected tree sizes; such operations are
performed on a very selective basis for fire protection and stand health and must
retain any sizable trees (called green tree retention). The planning process for selec-
tive thinning and woody debris removal is described in the next section. Removal
7 Saving the Forest by Reducing Fire Severity 175

of dead woody debris and tree thinning can reduce fire severity and aid in stand sur-
vival, which is the major objective of fuels removal programs. Following that, we
describe a spatial location-scheduling model that has been developed to assist in fire
resilience planning. After the description of the model and a short description as to
how it is solved, we describe its implementation, the key element to application, as
well as give an example of the results of an application. This model system has now
been used in the majority of forests in California.

7.2 Background

The organization of the US Forest Service (USFS) is a well-defined hierarchy. The

headquarters is in Washington, DC. They direct 9 regions; each of them operate
independently of the others and are each directed by a Regional Forester. Each
region is comprised of a number of National Forests (NF). Most of these are divided
into three to five ranger districts (RD). Day-to-day operations are directed by the
staff of each ranger district. Policies and directives are developed by the Washing-
ton office and regional headquarters. Planning, environmental review, monitoring,
assessment, etc. is accomplished by NF staff and RD staff, with input and assis-
tance from the regional staff. Historically, modeling assistance has been provided
by the top analyst in a region (the regional analyst). The tools, models, and focus
of a given NF can vary considerably from other NFs in a region, based upon the
priorities, budget, and targeted conservation species.
In the last decade, operations have been directed towards the protection of critical
habitat from catastrophic, ground clearing fires as well as protecting the wildland-
urban interface (WUI). The interface between wildland and urban areas contain
cabins, homes, recreational facilities, and scattered commercial services. Protect-
ing WUI to the greatest extent possible is a political necessity and protecting habitat
is a prime objective. These two goals frequently are in conflict, as critical habitat and
WUI lands are often some distance apart and budgets for fuels removal are limited.
In order to understand how limited resources might be used to protect habitat and
WUI areas from severe fires, it is first necessary to understand the state of the art in
modeling fire. USFS and other agencies have supported fire modeling research for
quite some time. There are several fire simulation programs that have been devel-
oped over the last several decades, and the ones that have been used the most are part
of a suite of programs. These include FARSITE (Finney 2004), Behave (Andrews
et al. 2003), and FLAMMAP (Finney 2006).
Fire simulation models use data on terrain, vegetation coverage, moisture con-
ditions, etc. and simulate the spread of a fire over time from a starting location.
One can determine the potential impact of fuels removal on fire spread, by sim-
ulating a fire before and after fuels treatment. Essentially, a fuels treatment plan
will effect vegetation and woody debris conditions and changes fire spread, sever-
ity, etc. between pre and post fuels treatment can be estimated by comparing the
two simulations (Collins et al. 2010). Finney (2001) demonstrated with the use of
176 R. L. Church et al.

Fig. 7.1 Planning work flow for the Fire Cadre

simulation that a set of distributed fuels treatment areas which together comprise
approximately 25 % of the landscape can reduce fire severity as if the entire land-
scape had been treated. Imagery of recent fires, e.g., Rodeo-Chediski fire in 2002
in Arizona, showed that fire intensity was considerably lower in those areas that
had been thinned. Others have documented that fuels removal of litter and ladder
fuels also tends to lessen the intensity and speed of a fire (Collins et al. 2010). In
2001, the Regional Forester of Region 5 (Pacific Southwest) directed that the forests
of the Sierras be managed to protect the existing habitat of the spotted owl and
address the fire hazard risk to those areas by strategically locating fuels treatments
(Powell 2001). With the outbreak of several large fires (including the Rodeo-
Chediski) in 2002, the US Congress acted to improve forest health with the passage
of the “Healthy Forests Restoration Act of 2003.” In response, the Regional Forester
of Region 5 directed that a plan “aggressive enough to reduce the risk of wildfire to
communities in the wildland urban interface (WUI) while modifying fire behavior
over the broader landscape” be implemented (Blackwell 2004). This directive also
ordered that habitat of other species of concern be protected as well. This included
the Pacific fisher, the Willow flycatcher, the Great Gray Owl, as well as others.
To meet the directive, each NF was called upon to develop fuels treatment plans.
At the heart of such plans is the designation of those areas that will be treated for
fuels reduction (mechanical removal, thinning, and controlled burns), called SPLATs
(Strategically Placed Landscape Areas for fuels Treatments). To aid in this process,
a team of experts, called the Fire Cadre (FC), was assembled. The FC had experts
in fire modeling, treatment cost estimation, vegetation modeling, forest biology and
GIS. Together this team aided each NF in modeling fire with and without treatments.
The FC developed a planning process which is outlined in Fig. 7.1.
Figure 7.1 depicts the 4 major steps of the planning process (see Bahro et al. 2007
for a detailed description). The first step (Box 1) is associated with delineating those
7 Saving the Forest by Reducing Fire Severity 177

areas that are part of protected activity centers (PACs) or home range core areas
(HRCAs) for the spotted owl. Other protected habitats must also be identified. Fur-
ther, potential SPLATS are identified and a proposed fuels treatment method based
upon: location, vegetation condition, terrain, etc, is determined. The second step
(Box 2) involves selecting a set of SPLATs that are spatially distributed. Given this
set, the vegetation cover map is modified to reflect this set of treatments. Then, a
fire simulation model is run for a number of different fire scenarios (wind direction,
starting location, etc.) in order to assess the potential of the system in protecting
critical habitat, WUI, and promoting tree survival (Box 4). This is done in order to
generate a preliminary assessment as to the effectiveness in SPLAT treatments. It
may be necessary to readjust some SPLATs, generate more SPLATs, or be possible
to reduce the number of SPLATs to generate a cost effective treatment plan for fire
resilience. Once a feasible preliminary plan has been developed then a schedule of
treatments is generated by a spatial-scheduling model (FIRS) using the Initial For-
est Activities Spatial Scheduling Tool (iFASST) Decision Support System (DSS),
the main subject of this chapter. After a schedule is generated, then fire Simulations
(Box 4) are run on the impact of the first 5 years of treatments (or some other time-
frames) to estimate the impact of the schedule on overall area protection, etc. This
may mean that some readjustments in the schedule need to be made. The remainder
of this chapter describes the spatial scheduling model and its implementation.
Before scheduling, nearby SPLATs are aggregated into small project areas, called
PUCS, a short term for planning units containing SPLATs. SPLATs, themselves,
are too small to contract out for a private company. For private companies to be
interested in bidding, it is necessary to have enough SPLATS aggregated together
so that it is worth their while and will occupy a work team for several months. The
PUCS are scheduled by the iFASST model.
Our description so far has purposefully neglected to describe the spatial elements
of the scheduling problem in order to keep our description from being needlessly
complicated. But now that we have described the overall planning process, it is now
necessary to specify in greater detail the spatial elements in scheduling. As previ-
ously stated, SPLATS and therefore PUCS need to comprise approximately 25 % of
the land area. Finney (2004) demonstrated that fuels removal at this percentage of
the landscape can be as effective as reducing fuels across the entire landscape. This
has been called the Finney effect. To achieve this condition, SPLATS need to be
distributed and strategically placed in terms of potential fire spread, etc. When an
area represented by a PUCS has been treated, enough fuels have been removed so
that the Finney effect or condition has been reached. Essentially, the objective is to
achieve the Finney condition across the entire forest. But, budgets limit the extent
to which this can be achieved and to treat all PUCS may take 20 years. Besides
the budget preventing a quick resolution to fuels treatment needs, there is an impor-
tant spatial issue. If isolated PUCS are treated, then fire severity will presumably be
lessened in those PUCS areas. But, if localized critical habitat is surrounded by sev-
eral PUCS, some treated and some not, then that critical habitat is not fully protected
until all remaining neighboring PUCS have been treated. In addition to this, schedul-
ing neighboring PUCS together or in subsequent years helps to extend the Finney
178 R. L. Church et al.

effect to larger blocks of land, an added benefit in slowing the rate of fire spread. If
all landscape elements were considered equal, then the scheduling problem would
involve starting treatment in one PUC and then clustering subsequent treatments
among its neighbors, and continuing that process until all PUCS have been treated.
But, in fact, the record of decision discussed above directs the planners to protect
as a first priority, areas of critical habitat and WUI. Thus, the scheduling problem is
one in which PUCS are scheduled and clustered among WUI areas and surrounding
areas of critical habitat.
There are other constraints of a practical matter that must also be considered. If at
all possible, at least one PUC project should be scheduled each year in each ranger
district of a NF. This helps to spread the workload of those managing and monitoring
fuels treatment projects across the ranger districts. Also, many NFs wish to treat
approximately the same acreage in one year as another, as this tends to spread out
the work evenly across the planning horizon. There are other nuances of the spatial-
scheduling system that are better described in the overall model presentation, which
is the subject of the next section.

7.3 The Spatial-Scheduling Model FIRS

In order to formulate the Fire Intensity Reduction Scheduling (FIRS) model, we

begin by defining needed notation:
i, j indices used to represent a specific project areas, where j = 1, 2, 3, . . ., n
t an index used to represent planning periods, where t = 1, 2, 3 . . ., m
k an index used to represent a specific planning area, where k = 1, 2, . . ., p
cjt the cost of treating project j in time period t
aj the size of treatment area in project j in acres
fj the total acreage of project (treated and untreated)
wj the amount of wildland—urban interface acres present in project j
hj the number of acres of sensitive habitat present in project j
θj the earliest time period in which project j can be scheduled
Ht the total number of acres of habitat that can be disturbed in time period t
dt the discount factor for time t, where value is zero in time one and increases
with time
Bt the available budget for fuel removal projects in time period t
t {j: project j can be assigned for treatment in period t}
j {t: project j can be assigned for treatment in period t}
Pj {j: project j is part of planning area k}
E {(i, j): project areas i and j are adjacent where i < j}
st deviation above the average yearly level of treatment area for time t
vt deviation below the average yearly level of treatment area for time t
7 Saving the Forest by Reducing Fire Severity 179

In addition to the above notation, we will need the following decision variables:

1, if project j is scheduled for treatment in time period t
xjt =
0, otherwise

1, if planning area k is not assigned a project in period t
ukt =
0, otherwise

1, if project i and project j are both scheduled in time period t
0
zijt =
0, otherwise

1, if project i is scheduled in time t and project j is scheduled in time t + 1
1
zijt =
0, otherwise

1, if project j is scheduled in time t and project i is scheduled in time t + 1
1
zjit =
0, otherwise

Given the above notation and decision variables, the FIRS model can be formulated
as follows:

m 
Max Z1 = fj xjt (7.1)
t=1 j ∈ t


p 
m
Min Z2 = ukt (7.2)
k=1 t=1


m 
m
Min Z3 = dt wj xjt (7.3)
t=1 j ∈ t


m
Min Z4 = (st + vt ) (7.4)
t=1
⎛ ⎞
 
T −1
T
Max Z5 = ⎝ 0
zijt + 1
zijt 1⎠
+ zjit (7.5)
(i,j )∈E t=max{θi ,θj } t=max{θi ,θj }

subject to the following constraints:

If at all possible, assign at least one project to each planning unit in each period:

xjt + ukt ≥ 1 for each k = 1, 2, 3, ..., p & t = 1, 2, 3, ..., m (7.6)
j ∈Pk ∩ t

Do not exceed the allowable budget in each year:


cjt xjt ≤ Bt for each t = 1, 2, 3, ..., m (7.7)
j∈ t
180 R. L. Church et al.

Treatment within sensitive habitat in each year is limited:


hjt xjt ≤ Ht for each t = 1, 2, 3, ..., m (7.8)
j∈ t

Each project can be scheduled at most once:


m
xjt ≤ 1 for each j = 1, 2, 3, ..., n (7.9)
t=θj

Track when adjacent treatments are scheduled:

a. adjacent treatments in the same year

0
zijt ≤ xit for each (i, j ) ∈ E & max{θi , θj } ≤ t ≤ m & j > i (7.10)
0
zijt ≤ xjt for each (i, j ) ∈ E & max{θi , θj } ≤ t ≤ m & j > i (7.11)

a. adjacent treatments in previous or subsequent year

1
zijt ≤ xit for each (i, j ) ∈ E & max{θi , θj } ≤ t ≤ m − 1 & j > i (7.12)
1
zijt ≤ xjt+1 for each (i, j ) ∈ E & max{θi , θj } ≤ t ≤ m − 1 & j > i (7.13)
1
zjit ≤ xjt for each (i, j ) ∈ E & max{θi , θj } ≤ t ≤ m − 1 & j > i (7.14)
1
zjit ≤ xit+1 for each (i, j ) ∈ E & max{θi , θj } ≤ t ≤ m − 1 & j > i (7.15)

Track treatment level in each year and compute deviation from average:

m  
1
m aj xjt − aj xjt = vt − st for each t = 1, 2, 3, ..., m (7.16)
t=1 j ∈ t j∈ t

Restrictions on variables:

xjt = {0, 1} for each j = 1,2, 3, ..., n & t = θj , θj + 1, ..., m (7.17)

ukt = {0, 1} for each k = 1,2, 3, ..., p & t = 1, 2, 3, ..., m (7.18)
vt ≥ 0 for each t = 1, 2, 3, ..., m (7.19)
st ≥ 0 for each t = 1, 2, 3, ..., m (7.20)
0
zijt = {0,1} for each (i, j ) ∈ E & t = max{θi , θj }, ...., (7.21)
1
zijt = {0,1} for each (i, j ) ∈ E & t = max{θi , θj }, ...., m − 1 (7.22)
zjit = {0,1}
1
for each (i, j ) ∈ E & t = max{θi , θj }, ...., m − 1 (7.23)

The FIRS model is a multi-objective integer-linear programming model. It contains

five types of objectives. Each project, a Planning Unit Containing Splats (PUCS),
represents a set of fuel removal activities as a cluster of SPLATS within the project
7 Saving the Forest by Reducing Fire Severity 181

area. If all the activities are accomplished within the project area, then the entire area
is defined as having met the “Finney” threshold. The first objective, Z1 , involves
maximizing the total number of acres over the entire planning period that is classi-
fied as meeting the Finney threshold. The second objective, Z2 , involves minimizing
the number of time periods any planning area has not been assigned at least one
project. Planning areas are relatively large subdivisions of the forest and usually
represent ranger districts. Since each ranger district is a somewhat independent oper-
ating unit within a National Forest, it is important to keep planning and operations
staff involved in each district and time period if at all possible. The third objective,
Z3 , is designed to minimize or maximize a discounted function of treated wild-
land urban interface (WUI) lands over time. This effectively prioritizes treatments
in WUI lands in the minimization sense to be treated earlier in the planning horizon.
Prioritizing such treatment areas earlier in the planning horizon helps to protect cab-
ins, residential, and commercial areas on the edge of the urban-forest interface from
significant fire as early and quickly as possible. In addition, scheduling projects con-
taining WUI acreage as early as possible reduces the cost of fire suppression in these
areas, as fires in treated areas do not burn as intensely. The fourth objective, Z4 , min-
imizes year-to-year variation in treated acreage. Including this objective is crucial,
as many fuels treatment activities are handled by outside contractors and forest ser-
vice personnel who would like to keep the contracting and supervisory workload
consistent from year to year. The fifth objective, Z5 , maximizes project adjacency
in the same (Zij0 t ) and subsequent (Zij1 t ) time period. This objective enables a plan-
ner to cluster projects in a given year, or in a subsequent year, or to spread projects
out if the objective is minimized.
The constraints also restrict the model solution in several key ways. The first
constraint (6) assigns PUCS, xjt , to each planning unit k in each time period t, or it
forces ukt = 1. This constraint thus tracks whether or not a project has been assigned
to a planning area (ranger district) for time period t. This constraint, in conjunc-
tion with the second objective, attempts to keep a ranger district and its personnel
occupied with at least one PUCS project each year.
Constraint (7) forces the total cost of projects xjt scheduled in a time period t to
be less than or equal to the allotted budget for time period t. Constraint (8) similarly
forces every scheduled project xjt in time period t to impact no more than Ht acres
of sensitive habitat for that time period. Constraint (9) stipulates that each project
j can be scheduled at most once between the earliest possible time project j can be
scheduled and the last year of the planning horizon.
Constraints (10)-(15) track whether or not a scheduled project is adjacent to
another scheduled project. Constraints (10)-(11) track whether a project has been
scheduled adjacent to another project in the same time period t, considering the ear-
liest planning period the project may be scheduled through the end of the planning
horizon. Constraint set (12)-(15) tracks whether or not a project in time period t
has been scheduled adjacent to another project in time period t + 1, considering the
earliest planning period the project may be scheduled through the end of the plan-
ning horizon minus one. The reason for considering the earliest time a project may
182 R. L. Church et al.

be scheduled to the end of the planning horizon minus one (m − 1) is because a

project cannot be scheduled in the year past the end of the planning horizon.
Constraint (16) tracks the treatment level in each year and computes the deviation
from the average. This constraint is needed to define the level of deviation that
is used to support objective 4. Constraints (17)-(23) define the restrictions on the
variables; in this case to be binary or non-negative in value.
Although the above formulation captures the fuels treatment scheduling and loca-
tion problem, it contains several properties which make it difficult to solve optimally.
First, National Forests typically contain three or more Ranger Districts. Each of the
Ranger Districts may in themselves contain 50–100 project areas (PUCS contain-
ing SPLATS). Consequently, the number of PUCS in a forest is usually larger than
160. Since the planning period has been defined to be 20 years, and with each PUC
typically being schedulable in each of those time periods, yields a problem with the
number of integer scheduling variables exceeding 5000. Couple this large number of
integer scheduling variables with two “budget” constraints (one on costs and one on
habitat disturbance) for each time period, and it becomes a large multi-dimensional
knapsack problem. Finally, add to this a large number of “spatial-temporal clustering
constraints and variables” and the result is a very large integer linear programming
problem.
The FIRS model is embedded in a modeling and maping decision support system
called the intial Forest Activities Spatial Scheduling Tool (iFASST). This modeling
system is described in detail in the next section. This system is capable of produc-
ing a model output file that can be read by solvers like GUROBI and CPLEX. A
routine was developed to read a solver output file and produce a scheduling file
that could be mapped and analyzed by the iFASST program. A specialized heuris-
tic was also developed to solve the problem as well. The optimal approach was
used to help fine tune the heuristic. The specific details of the heuristic are given in
Niblett et al. (2014); suffice it to say that it employes elements of GRASP (Feo und
Resende 1995; Resende und Ribeiro 2005), and path relinking (Glover et al. 2000).
As most NFs do not have access to solvers like IBM CPLEX, GUROBI, or FICO
Xpress nor the time to spend seeking true optimal solutions, the heuristic was relied
on by forest analysts.

7.4 Decision Support System Implementation

As noted above, US Forest Service managers have been tasked with reducing fire
severity within their forests by using varying fuels reduction strategies with species
of concern and ecosystem restoration objectives in mind. The DSS must therefore
be flexible in its ability to handle a variety of objectives and data inputs from several
sources and users. It should also be able to: set up planning scenarios by an analyst;
display solution(s) to a scenario in mapped and data-graphic form; and to allow the
analyst to modify a scenario based upon a previously obtained solution. In addition,
the DSS should be easily deployed for use by forest managers and analysts. The
7 Saving the Forest by Reducing Fire Severity 183

Initial Forest Activities Spatial Scheduling Tool (iFASST) DSS was designed with
these capabilities in mind.
The iFASST DSS was designed to run on the Microsoft Windows operating sys-
tem, the main operating system of US Forest Service computers, with the ability
to open a variety of data types. The data types used by iFASST are either derived
from Geographic Information Systems (GIS) or from processed data exported from
software such as Microsoft Excel or Microsoft Access. In addition to its built-in
heuristic and integer linear programming model building, iFASST also has the capa-
bility to provide data-graphs and maps of solutions derived from heuristic or solver
solutions and to export solution data and statistics. The next section focuses on how
a scenario file may be generated and the data requirements.

7.4.1 Building a Scheduling Scenario

Building a scheduling scenario requires spatial and non-spatial data inputs. Building
several planning scenarios for a variety of forests and problem variations involving
several users is even harder. The iFASST DSS was designed to enable such planning
scenarios to be built easily using a Graphical User Interface (GUI). In addition, the
DSS was designed specifically to handle spatial data to manipulate and view fuels
treatment schedules and data graphs, but also to utilize data manipulated externally
by other software. It was also necessary that spatial data generated in a GIS be in
shapefile form. Shapefiles are used because they may be read in both proprietary
and open-source GIS.
A simple GUI was designed to enable analysts to quickly set up a scenario in the
iFASST DSS. Figure 7.2 shows an example of the GUI used to build the scenario
file, called a Basefile in iFASST. The Basefile stores all of the data-file linkages and
settings required to solve a fuels treatment scheduling problem, and how to display
results. In the left side of Fig. 7.2 there are several drop-boxes within the section
titled “Required Files for Basefile”. Each of these drop-down boxes contain a list
of files or options representing a data-type. For example, the first drop-down box
references the shapefile representing the spatial data of the forest. When a file is
chosen from the drop-box list, a map showing the Planning Units Containing Splats
(PUCS) of the forest is shown, and the shapefile is checked for required attributes
used for scheduling and solution output purposes. In addition, the “Shapefile Field
Selector” portion of the Basefile Creator window becomes active (not greyed out),
and the drop-box for the next required data setting becomes active if the minimum
required spatial attribute fields are present.
The shapefile field selector lets the analyst select spatial attributes that they would
like to use in modeling, as well as determining how they want an attribute displayed.
An attribute could be displayed in chart form, or as a sum total on the “Left” portion
of the iFASST DSS display.
When specifying the external data-file to use, the “Data File Field Name Set-
tings” area becomes active when an item is selected from the “BaseData File Name”
184 R. L. Church et al.

Fig. 7.2 Setting up and saving a scenario as a basefile using the basefile creator

drop-box. The selected file is checked to ensure that required attribute field headers
are present. The Data File contains data that has been processed externally from a
GIS that an analyst wishes to use in the iFASST DSS. The “Data File Field Name Set-
tings” area specifies how attributes contained in the Data File should be displayed
in iFASST.
The other “Required Files” drop-boxes contain options or links to required data.
For example, the “Adjacency File Name” drop-box contains a setting that allows the
user to specify if the iFASST DSS should create the adjacency file, or if a previously
computed one should be used. The adjacency option identifies those PUCS that are
adjacent to one another. A budget data file must also be selected which contains a
fiscal budget for the forest by year for the planning horizon, as well as the maximum
level of sensitive habitat that may be disturbed by year for the planning horizon. A
details file contains information specifying the planning horizon to be considered,
as well as the number of ranger districts within the forest to consider scheduling
projects in. The permanently fixed out (PFO) file contains a list of PUCS that should
not be considered by the iFASST DSS because they are private inholdings that are
managed outside the purview of the forest service, if present.
In addition to the “Required Files” the Basefile Creator also contains a “Con-
straints” section where constraints of the scenario may be specified. For example,
the “Earliest Time” constraint may be included by selecting the check-box. If the
constraint is selected for inclusion, an attribute field header representing the earliest
7 Saving the Forest by Reducing Fire Severity 185

Fig. 7.3 Basefile creator objectives specification

time a project could theoretically be scheduled must be chosen from the now-active
drop-box. Similarly, budget constraints may be selected. Budget constraints repre-
sent attributes used for fiscal budgeting or for environmental impact such as that
related to habitat. In addition, a constraint specifying the minimum level of treat-
able acreage to be considered as a viable project may also be specified. Once all of
the data-files and constraints that an analyst wishes to use to represent a planning
scenario are selected, the analyst may specify the objective or objectives that they
want to have considered in the iFASST DSS planning scenario.
Figure 7.3 represents the window where the objectives to be used in the planning
scenario are specified. A maximum of five objectives may be specified for consid-
eration in the iFASST DSS. In Fig. 7.3, five example objectives have been selected.
One of them is the adjacency objective. If adjacency is maximized, PUCS will tend
to be scheduled adjacent to one another (clustered) in the same time period or in
the next subsequent time period if possible. If adjacency is minimized, projects will
be spread out at the same or over time (dispersive tendency) with an attempt to not
schedule two adjacent PUCS in the same time period t or in time t + 1. The work-
flow (called “Spatial Distribution” in Fig. 7.3) objective involves projects scheduled
in ranger districts. When this objective is specified, the iFASST DSS will attempt to
schedule at least one project in each of the ranger districts of the forest in each time
period so that ranger district personnel are not idle.
The “Evenflow”, “Total Value”, and “Discount” sections of the scheduler rep-
resent an objective function that may be desirable for more than one attribute.
The numbers of attributes selected represent unique objectives. For example if two
“Evenflow” fields are selected, this would represent 2 objectives. The “Evenflow”
186 R. L. Church et al.

objective attempts to minimize the year-over-year variance of projects scheduled

over the planning horizon. For example, if amount of treated acres (TREAT) is
selected, the year-over-year variation in the number of treated acres is minimized.
Similarly, the “Total Value” objective maximizes the sum total of an attribute. For
example, if treated acres is selected, the total number of treated acreage would be
maximized. The “Discount” objective provides a “discount” to projects scheduled
toward the beginning or end of the planning horizon. For example, if an analyst
wants to preferentially treat Wildland Urban Interface (WUI) lands earlier in the
planning horizon, the analyst would maximize the WUI field as a discounted objec-
tive value. If the analyst wished to schedule WUI treatments later in the planning
horizon, the analyst would minimize the WUI field by checking the box in the
“Max/Min” list.
Once all of the objectives an analyst wishes to include in a scenario have been
set, the analyst clicks the “Set Objectives” button and the specified objectives are
saved. The analyst then saves the scenario data as a file. If the analyst wishes to
modify a scenario, a previously saved scenario may be opened, updated, and then
saved as a new scenario. The next section highlights how a scenario is solved and
how solutions may be mapped and displayed as data-graphics.

7.4.2 Using the iFASST DSS to Solve a Scenario and Display

Results

The iFASST DSS solves a planning scenario by loading the Basefile representing
it. The analyst may then weight the previously specified objectives to give added
emphasis to a particular, or set of, objectives.
Figure 7.4 shows the iFASST DSS Graphical User Interface (GUI) with the
Heuristic solution window, the Chart window, and a mapped solution. The map,
heuristic, and chart are all given in the right region of the GUI. On the left side of
the GUI, the “Left”, totals are given in this panel. Those elements appearing in this
panel were specified in the Basefile as “Left” for “Left Side Display”. The map leg-
end is given in the upper left side of the main display. Here each year is given a color,
so the thte schedule can be viewed as a transition of colors on the map itself. The
Heuristic window allows the analyst to set weights for objectives specified in the
planning scenario. In addition, it enables the analyst to manually schedule a PUCS
during a fixed time period, or to place it out of consideration as part of a solution.
The analyst can save the PUCS settings which may be re-opened for use later. The
Chart window allows an analyst to chart a particular attribute as a function of the
planning horizon. Figure 7.4 shows the treated acreage for the mapped schedule for
example. Clicking on a charted attribute during a particular year will highlight the
PUCS scheduled for that year in the map as well.
Mapping a solution interactively is also very helpful for an analyst. To achieve
this, the displayed map will update and color PUCS according to the year they were
scheduled in, if they were not able to be scheduled, or if there was not enough
7 Saving the Forest by Reducing Fire Severity 187

Fig. 7.4 Example of iFASST DSS Graphical User Interface (GUI) with Heuristic parameters, data-
graph, and mapped

acreage present for treatment based upon the minimum treatment acreage constraint.
There are also several map-query functions that have been built into the iFASST dis-
play map. For example, identifying projects scheduled for a particular year may also
be done by left clicking on the map schedule legend entry for that year. Similarly,
projects scheduled up to that particular year may also be shown by right clicking
on a year. In addition, the map may be re-colored by clicking the color next to a
year and colored using a color that an analyst chooses. Being able to change color
schemes is very important for those who have some level of color blindness. Addi-
tionally, pre-set color ramps are also included for quick re-coloring using a variety
of color schemes. Mapping functions are also included to allow the user to query
individual PUCS for attribute and schedule information to be displayed on the left
hand side information area. Also included are map re-zooming functions and the
ability to display a local mini-map showing the forest as a whole and the larger
mapped portion.
All of these DSS features enable an analyst to quickly set up, run, and tweak
a planning scenario. In addition, they enable an analyst to interact with a solution
using the map functions and chart features. This ability has enabled this DSS to be
used on a variety of forests seeking to reduce forest fuel loads while also meeting
several other varied objectives. Without this ability, forest fuels treatment scheduling
in California would not have been as effective.
188 R. L. Church et al.

7.5 Application and Implementation Issues

As we stated in the introduction, the US Forest service is divided into operational

regions. Each region contains a number of national forests, and each is directed by a
supervisor. This means that there are 17 USFS supervisors, each directing an almost
self-contained organization within region 5. The Regional Forester is the adminis-
trative head of the region. Each regional headquarters contains a staff, which helps
support different functions across the NFs as well give main directives to each super-
visor, however, there is often considerable flexibility as to how such directives are to
be met within each NF and their subunits, ranger districts. Consequently, operations
are loosely coordinated as well as differ due to local circumstances. For example,
chaparral scrub and oak woodlands, sierra mixed conifer forests and redwood forests
all differ considerably. Because of the flexibility in how the iFASST model could be
used, applications across these different domains were possible.
In those forests where the iFASST program was used extensively, results were
used to help create integrated vegetation and fuels management plans. Its use was
valuable for three principal reasons. The first reason was that the heuristic was
designed to solve a problem to near optimality in less than 60s. This meant that
analysts could test a number of different constraints and levels, and quickly gen-
erate an understanding of what could be realistically achieved. That is, the model
could be used to learn the interplay among the constraints in a timely manner, some-
thing that could not be done if an optimal solver took days to solve a given problem.
The second main reason for its use is that one could easily generate an estimate of
the cost of “promises” that had been made to special interests (or to cabin owners
for that matter) or desires of forest managers, by fixing in specific treatments in the
early years of the schedule. Then the differences in outcomes (of treated land, pro-
tected habitat, etc.) between an unfettered solution and one with such desires fixed
in would represent the cost of meeting these “deals/negotiations.” The third main
reason was that the program helped to “cap” expectations. Most planners assumed
that far more could be done in terms of treatments and meeting “Finney conditions”
than what was possible in a given time frame. Even though the program would be
used to schedule treatments out 20 years, the interest was fixed on what needed to
happen in the first 5 years, especially ensuring what was scheduled in the first 5
years did not preclude long term targets from being met. Foresters realize that con-
ditions change over time, and that virtually every NF will need to consider possible
changes in their plan 5 or 10 years from now, based upon natural disturbances, like
pests, drought, and fire. Overall, the iFASST program helped in understanding the
implications of making a near term (5-yr) schedule on a 20 year plan. One forester
characterized this as an attempt at making near term actions while ensuring long
term sustainability.
7 Saving the Forest by Reducing Fire Severity 189

7.6 Final Comments and Conclusions

Wildfire can be very destructive and, when severe, produce ground-clearing events.
Past harvesting activities have reduced the size and number of oldgrowth stands in
California, and any further losses threaten the viability of a number of threatened
species, including the California Spotted Owl. To address the concerns of biolo-
gists, the public, and special interest groups, Congress passed the Healthy Forests
Restoration Act in 2003. This along with the directive of the Regional Forester in
California outlined a policy for reducing fire severity by fuels treatments in order to
protect threatened old-growth stands as well as protect cabins, homes, recreational
facilities within the interface between urban and wildland areas. These activities
include prescribed burns during appropriate weather and plant moisture conditions,
mechanical removal of litter and downed woody debris, and selective thinning. Past
research by Finney (2004) demonstrated through the use of fire simulation models
that fire severity could be lessened by selective fuels treatments, rather than whole-
sale treatments across the entire forests. A planning process was developed by the
Region 5 analysts and researchers at UCSB to meet this challenge. At the center
of this process is a specialized fire modeling team, the Fire Cadre, who assisted
staff at National Forests in generating a plan of treatments as well as a 20-year
schedule, that address spatial, budget, and management objectives. At the heart of
this planning paradigm is a fire simulation model and a decision support system,
iFASST.
This chapter has presented the details of the spatial optimization model embed-
ded in iFASST. The model addresses the major issues in producing the “Finney
condition” across the forest as a whole with a special emphasis in clustering activ-
ities spatially around the WUI as early as possible. As each National Forest has
similar but different problems, iFASST was designed to be flexible in both con-
straints and objectives. It is because of this flexibility in model formulation, data
manipulation, and scenario generation, that the iFASST has been used in the largest
of these National Forests. Overall, the system allows planners to test different bud-
get levels, objective weights, and even fix the time in which some PUCS are treated,
aiding the planners to test a variety of strategies. At this time iFASST has now been
used in 11 of the 17 forests in California. Due to its success, work was recently com-
pleted to extend the iFASST system to maintenance treatments so that the Finney
condition is maintained over time.

Acknowledgements We would like to acknowledge the research funding of the US Forest Service
which supported the development of the FIRS model and the iFASST/mFASST program. We also
would like to acknowledge the help and assistance from Tanya Kohler and other staff members at
the Region 5 headquarters of the US Forest Service.
190 R. L. Church et al.

References

Andrews PL, Bevins CD, Robert C (2003) BehavePlus fire modeling system, version 2.0: Users
Guide. Gen. Tech. Rep. RMRS-GTR. U.S. Department of Agriculture, Forest Service, Rocky
Mountain Research Station, Ogden
Bahro B, Barber KH, Sherlock JW, Yasuda DA (2007) Stewardship and fireshed assessment: a
process for designing a landscape fuel treatment strategy. USDA Forest Service Gen Tech.
Rep. PSW-GTR-2003, 41–54, Vallejo, CA, USA
Blackwell J (2004) Sierra Nevada Forest Plan Amendment–Final Supplemental Environmen-
tal Impact Statement. Record of Decision, Department of Agriculture, Forest Service, US
Government
Camm JD, Chorman TE, Dill FA, Sweeney DJ, Wegryn GW (1997) Blending OR/MS, judgment,
and GIS: restructuring P & G’s supply chain. Interfaces 27(1):128–142
Church RL, Murray AT, Weintraub A (1998) Locational issues in forest management. Locat Sci
6:137–153
Collins BM, Stephens SL, Moghaddas JJ, Battles J (2010) Challenges and approaches in planning
fuel treatments across fire-excluded forested landscapes. J For 108(1):24–31
Feo T, Resende MGC (1995) Greedy randomized adaptive search procedures. Glob Optim
6:109–133
Finney MA (2001) Design of regular landscape fuel treatment patterns for modifying fire growth
and behavior. For Sci 47(2): 219–228
Finney MA (2004) FARSITE: fire area simulator—model development and evaluation.
U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station, Ogden
Finney MA (2006) An overview of FlamMap Fire modeling capabilities (in Fuels Management-
How to Measure Success: Conference Proceedings), USDA Forest Service Proceedings RMRS-
P-41; Fort Collins, CO: U.S. Department of Agriculture, Forest Service, Rocky Mountain
Research Station. 213–220
Geoffrion AM, Powers RF (1980) Facility location analysis is just the beginning (if you do it right).
Interfaces 10(2): 22–30
Glover F, Laguna M, Martí R (2000) Fundamentals of scatter search and path relinking. Control
Cybern 29(3):653–684
Gruell GE (2001) Fire in Sierra Nevada forests: a photographic interpretation of ecological change
since 1849. Mountain Press, Missoula, MT
Kent B, Bare BB, Field RC, Bradley GA (1991) Natural resource land management planning using
large-scale linear programs: the USDA Forest Service experience with FORPLAN. Oper Res
39(1):13–27
Miller JD, Safford HD, Crimmins M, Thode AE (2009) Quantitative evidence for increasing forest
fire severity in the Sierra Nevada and Southern Cascade Mountains, California and Nevada,
USA. Ecosystems 12(1):16–32
Niblett MR, O’Hanley JR, Church RL (2014) Scheduling fuels reduction projects using spatial
optimization. Unpublished manuscript
North M, Innes J, Zald H (2007) Comparison of thinning and prescribed fire restoration treatments
to Sierran mixed-conifer historic conditions. Can J For Res 37(2):331–342
Powell, Bradley E. (2001). Sierra Nevada Forest Plan Amendment Environmental Impact State-
ment. Forest Service, Department of Agriculture. US Government.
Resende MGC, Ribeiro CC (2005) GRASP with path-relinking: recent advances and applications.
In: Ibaraki T, Nonobe K, Yagiura M (eds) Metaheuristics: progress as real problem solvers, vol.
32. Springer, Boston, pp 29–63
Chapter 8
Locating Intelligent Sensors on a Transportation
Network to Facilitate Emergency Response
to Traffic Incidents

Tejswaroop Geetla, Rajan Batta, Alan Blatt, Marie Flanigan and Kevin Majka

8.1 Introduction and Motivation

According to the Federal Highways Administration (FHWA) study presented in Of-

fice of Highway Policy Information (2009) there are 254 million registered motor
vehicles in the U.S. Each year this number continues to grow, increasing utiliza-
tion of the road transportation network. In 2009 alone there were an estimated 2.2
million injuries related to traffic incidents and 33,808 fatalities from these injuries.
In addition, traffic incident related fatalities ranked sixth in the list of preventable
fatalities in the U.S.
Apart from being one of the major causes of fatalities, traffic incidents also
represent one of the major causes of roadway congestion. According to a survey
presented in (Cambridge Systematics and Texas Transportation Institute 2005), traf-
fic incidents account for 25 % of all non-recurring congestion reported on roads in
the US. Observing the ill effects of traffic incidents, the FHWA turned to Intelli-
gent Transportation Systems (ITS) to help increase safety and reduce congestion on
current roadways.
Nagare (2012) define Intelligent Transportation Systems (ITS) as an application
of advanced technology to solve transportation problems. ITS uses advanced com-
munication, sensor and computer technology to address transportation problems and
enhance the efficiency of the system for movement of people and goods. There are
many uses of ITS systems. These include systems that help reduce the likelihood
of a crash, usually based on sensors that issue warnings regarding approaching ve-
hicles. They also include systems that provide congestion data to users in a form
that aims to impact driver behavior on route choices. Another area of application
is in controlling entry of traffic onto highways to control congestion. There are

R. Batta () · T. Geetla

Department of Industrial and Systems Engineering, University at Buffalo (SUNY),
Buffalo, NY, USA
e-mail: [email protected]
A. Blatt · M. Flanigan · K. Majka
Center for Transportation Injury Research, CUBRC, Buffalo, NY, USA

© Springer International Publishing Switzerland 2015 191

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_8
192 T. Geetla et al.

emerging application areas of ITS, one of which is traffic-incident management.

This real-world application is the major focus of this chapter.
An effective traffic-incident management system needs the following functions:
• Rapidly detect a crash
• Characterize a crash and crash scene for incident management
• Deliver fast emergency medical services to the injured
• Provide crash related information to drivers and highway managers to reduce
congestion
At the core of any incident-detection technology is the use of sensors and wireless
communication technology. Sensors are responsible for incident detection and the
role of wireless communication channel is to transfer sensor readings to a data col-
lection node/location. A sensor is an instrument that measures a physical quantity
and reports to the user the measured quantity. In the road network systems, sensors
quantify the real road environment (e.g., congestion, weather conditions) and serve
as tools to alert and help drivers and traffic managers. For example, many cars are
now equipped with an on-board GPS with functionalities like navigation. The mod-
ern navigation design on a car can collect traffic and congestion data through radio
communication to alert a driver of potential congestions and re-route a trip to im-
prove their travel time see, e.g., BMW USA (2008). This capability arises from the
use of GPS along with data communication capabilities.
There are many sensors used by ITS systems for incident detection and man-
agement. We classify the sensors broadly into stationary and mobile. A stationary
sensor has a fixed location on the road network and monitors incidents on the road
network close to its location. Acoustic sensors and traffic cameras are some exam-
ples of stationary sensors. Mobile sensors move continuously in and out of traffic.
These mobile sensors are mounted on vehicles, which could be regular vehicles or
public utility vehicles like buses and taxis. They can serve both as sensors which
operate only in a crash situation or as sensors that relay continuous information.
They can also serve as computational engines that fuse or aggregate data from
neighboring sensors.
These mobile sensors may be operational only during some part of the day. For
example, Srinivasan et al. (1997) proposes and uses motor vehicle probes that move
along with traffic collecting vital road data like traffic flow, travel time, etc. The
mobile sensors that act as incident detection sensors observe and monitor only the
surroundings of their real-time position with certain functionality activated only if
they are involved in an incident. For example, a product offered by General Motors
(GM) called OnStar Automatic Crash Notification (ACN) will detect a crash (we
note that an earlier version of the OnStar product was a factory-installed ACN sys-
tem). Other examples of mobile sensor systems used in crash detection include the
upgraded OnStar (factory installed) system called Advanced ACN (AACN) see, e.g.,
(OnStar 2012), BMW (2010) and White et al. (2011).
Both mobile and stationary sensors with communication capabilities have an
added advantage in that they can support data fusion. Data fusion is the applica-
tion of mathematical techniques to combine data from multiple sensors. The data
8 Locating Intelligent Sensors on a Transportation Network . . . 193

in question can be of many different types, e.g. sound, change of speed, weather
conditions. The key feature of data fusion comes in when data of different types in
collected related to a common event, e.g. a crash. In this case, association of this data
towards this event needs to be made and furthermore appropriate fusion of this data
leads to a more comprehensive understanding of the event, since different sensors
collect information on different features of the event. Application of this technique
to incident detection will help combine data from multiple sensors classes to better
detect traffic incidents. In this chapter, we assume that all sensors on the road net-
work (stationary or mobile) have computational capability and are able to use data
fusion techniques.
With all the sensors available for incident detection, present day-road sensor net-
works are essentially multi-sensor networks, implying that they have a variety of
different sensors. In a multi-sensor environment, each independent sensor observa-
tion becomes a data point and all data points are collected and analyzed together to
create a situational awareness for the observer, meaning that various aspects related
to the event (situation) are developed. All sensor data needs processing to convert
the data into useful information. For example, two cars are involved in a frontal
impact motor vehicle crash. Assume that one of the cars is equipped with an AACN
and there is an acoustic sensor located very close to the crash. Both sensors generate
unique crash data. These two independent data sets have to be pooled together first
and then combined to extract crash information. Multi-sensor data fusion techniques
propose a systematic data combination to extract accurate information.
Researchers developed multisensor data fusion to combine data generated by a
large set of sensors for Department of Defense (DoD) applications in the 1980s. This
development combined ideas from various research disciplines like signal process-
ing, statistics and artificial intelligence. In recent years, data fusion has developed
into a research field of its own. Data fusion is a collection of efficient methods for
combining data from disparate sensor resources and databases to generate a unified
and in some ways better information compared to individual sensor data, see Hall
et al. (1997). This book also contains many lucid examples of how data fusion can
provide better information. In fact it provides cases where data fusion can be used to
predict future events. Our use of data fusion is at an elementary level, in that we are
trying to better characterize an event that has happened for appropriate emergency
response, not predict future occurrences of events.
One of the advantages of data fusion is it helps in authenticating a physical quan-
tity measurement observed by multiple sensors (for example, if there are multiple
sensors in a chosen area making temperature measurements, which become critical
in winter driving situations as they serve as a surrogate for slippery driving con-
ditions). Each sensor records a temperature measurement. Data fusion techniques,
using statistical estimation, help to improve confidence, reliability and reduce am-
biguity in temperature readings. The other advantage of data fusion is that once this
estimation of a physical quantity is complete, other observations use the estimate
in other data fusion processes. For example, using the temperature estimate gener-
ated by multiple sensors, the acoustic sensors estimate the speed of sound (which
varies with air temperature and use it to estimate position of the source of sound.
194 T. Geetla et al.

Data fusion provides the availability of estimated data across all sensors to generate
a unified view of the real event.
Working together using data fusion all sensors in the ITS systems used in inci-
dent detection should actively aim to provide complete information, devoid of any
ambiguity, of every crash on US roads. However, considering all the budgetary con-
straints in deploying such a massive infrastructure, operations research techniques
can help in maximizing the potential of partially deployed ITS systems. Similarly,
the use of simulation can aid in evaluation with relatively small investments.
Purchase of sensors, sensor maintenance and the support data systems may be the
three important cost drivers of any ITS project. Keeping the budgetary constraints
and to maximize the incident detection capability, the placement of the stationary
sensors becomes critical. The good news is that placement of stationary sensors can
be optimized using operations research techniques.
Many factors can lead to motor vehicle crashes. Some of the commonly known
factors are speeding, carelessness, inexperience, distraction, loss of control over the
vehicle, etc. All the conditions are mistakes originating from the driver side and
the last condition may occur due to unforeseen road weather conditions. Driver mis-
takes are the major cause of accidents, which happen more often at intersections and
entry-exit ramps, where a large number of vehicles move through and interact. An
assessment of historical crash data can provide insights into accident-prone zones
or areas that we can target for sensor installation. In this chapter, we use previous
accident data to guide sensor placement with a focus on acoustic sensors.
We proceed to state the problem and its application, which is to understand sen-
sor technology used in incident detection and use this knowledge to improve the
placement of stationary sensors. This chapter does not consider issues related to
data communication or data transmission. The assumption is that every sensor can
communicate with all the sensors and are capable of transmitting their data to a
central base location or a local data collection sensor. This assumption will be true
in the future as the DOT is working on improving communications as part of the
Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I) initiatives.
In Sect. 8.2, we present background on sensor coverage models, present the so-
lution methodology for the sensor placement problem, and develop a simulation
model used in the sensor placement model as both an evaluation tool and as an opti-
mization technique. In Sect. 8.3, to demonstrate the proposed solution methodology,
we use omnidirectional sensors as a case study and present the results. In Sect. 8.4,
we summarize the findings from our study and propose future research directions.

8.2 Sensor Location Problem

The benefits of data fusion are only realized when multiple sensors observe or de-
tect the same incident. This requires efficient placement of sensors to provide at
least double coverage (i.e., a minimum of two sensors covering the same part of
the road segment). Single coverage implies that only one sensor covers a part of the
8 Locating Intelligent Sensors on a Transportation Network . . . 195

road segment. This single coverage is a less favorable outcome for the placement of
sensors than double coverage.
The placement of sensors will greatly influence the effectiveness and utility of
data fusion technology. With a limitation on the number of sensors, a tradeoff be-
tween multiple sensors covering a smaller area versus single sensor coverage of a
larger area becomes apparent. The output from the data fusion techniques is de-
pendent on the distance between the sensors and the crash location. Data fusion of
the information from multiple sensors will take into consideration the real world
uncertainties. Since the output of data fusion is very dependent on the location of
the crash, we need a mechanism to capture the uncertainty in the location of fu-
ture crashes when assessing the value of using data fusion. One way to assess this
uncertainty in crash location is through simulation. Randomized crashes generated
using simulation help evaluate the detection capability of acoustic sensors and data
fusion opportunities. Statistical evaluation of the effectiveness of a particular sensor
placement design is necessary. Experimental design serves as a tool to achieve this
goal.
Operations research techniques have helped in innumerable sensor placement
problems and facility location problems to place sensors. The research question here
is to use the same techniques to help place incident detection sensors in a traffic en-
vironment to detect incidents and provide a data fusion-capable environment where
multiple sensors interact. Section 8.2.1 presents the literature background for the
sensor placement problem. The section also highlights the research gap this chapter
aims to fill.

8.2.1 Literature Review of Sensor Placement Problem

Sensor placement problem is a common mathematical problem arising in a large dis-

tributed sensor network. One easily relatable sensor location problem is the security
guard location problem (also called an art gallery problem) in a museum, whose
objective is to cover every display piece inside the museum by using the smallest
number of guards. The problem has to take into consideration the geometric shape
of the display area inside the room. Other research interests in the sensor placement
problem involve development of heuristics or algorithms for the sensor placement
problem. Dhillon und Chakrabarty (2003) presents two algorithms for placement of
sensors in a sensor field modeled as a grid. Nie et al. (2012) develops the problem
of locating bomb-detecting sensors in airports or shopping malls, to counter ter-
rorist attack by increasing the detection capability. Gentili und Mirchandani (2005)
presents a location of traffic flow sensors on a road environment for applications
like flow measurements and time to travel a path. One of the important measures in
traffic management is time of travel between an origin and destination pairs. Gentili
und Mirchandani (2005) proposes the use of traffic flow sensors to estimate the time
to travel between an O-D pair. A set-covering formulation is developed. Since this
is an NP-hard problem, they use a heuristic solution algorithm.
196 T. Geetla et al.

8.2.1.1 Coverage Problems

Building on the location set-covering concept, the maximal coverage location prob-
lem proposed in Church und ReVelle (1974) aims to maximize coverage using
limited resources. For example, in a cellphone base-station location model, the
demand can be customers who receive and transmit radio waves; or in a warehouse-
location-problem, the demand is from potential shops or malls. Coverage problem
formulations model demand in three ways. In the first method, the demand occurs
at specific nodes. The warehouse location problem is a good example that uses the
node demand, where malls represent the demand nodes. In the second method, de-
mand occurs on paths. For a bus-route selection problem where each trip has a
specific demand, the path based demand model is used. In the third method, de-
mand occurs at both nodes and paths. A node and path based demand model is
used to locate cellphone base stations and emergency helicopter base stations, see
Erdemir et al. (2008a, 2008b). In Erdemir et al. (2008a) cellphone base-station lo-
cation problem demand arises from customers using cellphone from nodes (housing
areas, shopping malls, etc.) and paths (roads and highways). For the sensor location
problem considered in this chapter, roadway crashes constitute demand since the ob-
jective is to maximize the incident detection. However, roadway crashes can happen
anywhere on roadways; there are no predictive mechanisms to forecast incident oc-
currence. To plan the sensor location problem we assume previous crash data serve
as a predictor for future accidents. However, a key observation from previous crash
data is that the crash demand split into crash paths and crash nodes. Crash nodes
represent high accident-prone locations (road intersections, entry and exit ramps of
a highway) and crash paths representing the rest of the road network. Crash paths
represent the other road segments that are not part of the crash nodes. The crash
paths can be parts of highways where accidents are equally likely to happen at any
part of the road segment.
Using the above observation from past crash data, where crashes happen on both
crash nodes and crash paths, the major goals of the present research are to improve
both incident detection and data fusion capabilities in this multi sensor network.
However, there is not much research on how to improve data fusion in a coverage
problem. A regular max cover problem has a single objective function that maxi-
mizes the coverage of every sensor (we note that Berman et al. 2013, have recently
considered continuous coverage). There can be extensions made to this formulation
to increase secondary coverage that might increase the data fusion capabilities.
There are many benefits provided by improving the data fusion capabilities,
which translates into requiring secondary coverage. The main drawback of data
fusion lies in the fact that it is made possible through secondary coverage, which
implies a reduction in primary coverage. While primary coverage emphasizes de-
tection capability, secondary coverage emphasizes data fusion capabilities. In this
chapter, we plan to explore the interesting research question of the trade-off between
these two objectives.
8 Locating Intelligent Sensors on a Transportation Network . . . 197

8.2.2 Sensor Classes and Solution Framework

There are many sensor classes used for incident detection. For this chapter we chose
three sensor classes:
i. Omnidirectional sensors
ii. Traffic cameras
iii. Traffic flow sensors/traffic count sensors
Acoustic sensors used in traffic incident management belong to the class of om-
nidirectional sensors (unlike microphones, which are directional acoustic sensors).
There are many ITS applications that use acoustic sensors. Traffic cameras are vis-
ible at major intersections and these sensors actively monitor road segments. A
representative traffic-count sensor (whose role, for example, is to reveal sudden
changes in traffic flow which is a typical sign that a crash has occurred) is an induc-
tive loop detector embedded inside the road segment. Although this chapter focuses
on placing these three sensor classes, extensions or modifications are possible to the
general approach presented here for placement of sensors to suit other sensor classes
as well.
Considering the three sensor classes, we propose a solution methodology that
involves three steps, which can be independently applied to all three sensor classes.
The first step formulates the problem as a coverage problem. We call this the explicit
model. The explicit model resembles a maximal coverage problem. The second step
involves a pure geometric approach to solve the sensor location problem. We call
this the implicit model. Unlike the explicit model that only considers a path covered
if and only if the entire path is covered, the implicit model allows partial cover-
age. The implicit model to solve the problem lets the decision maker/s decide the
percentage of the path that should be overlapped by the coverage of two or more
sensors. This overlap area provides a good opportunity for collecting more data and
improving the quality of information through data fusion. The implicit model can
be used as a separate method to solve the sensor location problem, independent of
the explicit model. The solution of the explicit model provides an excellent starting
point for the implicit model.
As illustrated in Fig. 8.1, a simulation based optimization procedure follows after
the implicit model’s application. This is the third step in our solution methodology.
Simulation-based optimization first evaluates all the generated feasible solutions and
then proceeds to use a local-search heuristic to generate other feasible solutions. So-
lutions that result in an improved objective function will iteratively use local search
to reach better solutions. The addition of simulation-based optimization to a cover-
age problem is unique and to the best of our knowledge never tried before. In the
next section, we present the details of the explicit and implicit models.
198 T. Geetla et al.

Fig. 8.1 Solution methodology for the sensor placement problem

8.2.3 Explicit Model

The inputs to the explicit model are:

i. Node and path demands (in our case, calculated using previous crash data)
ii. Candidate locations for sensors (in our case, a set of feasible sensor locations
on infrastructure with access to power)
iii. Coverage criteria for every sensor class
iv. Number of sensors available from each sensor class
The output of this model is the location of the available sensors.
A node is ‘covered’ if it is in the coverage zone of at least one sensor. A path
is ‘covered’ if each point on the path is in the coverage zone of at least one sensor.
The model enforces these coverage rules by the constraints. We now proceed by
introducing relevant notation, followed by our model formulation. Our model is
identical to the one found in Erdemir et al. (2008b) except that a single period case
is considered.
The input data for this model is as follows:
M set of potential sensor locations
N set of nodes
P set of paths
p1 total number of sensors to be located from omnidirectional sensors
p2 total number of traffic cameras to be located
p3 total number of traffic flow sensors to be located
8 Locating Intelligent Sensors on a Transportation Network . . . 199

Hj weight of node j
Hk weight of path k
A (Ai,j,m), where Ai,j,m = 1, if sensor at location i of type m covers node j, and
0, otherwise
B (Bi1,m1,i2,m2,k), where Bi1,m1,i2,m2,k = 1, if sensors at i1 of type m1 and
i2 of type m2 cover path k, and 0, otherwise
B’ (B’i1,m1,k), where B’i1,m1,k = 1, if sensor at i1 of type m1 covers path k,
and 0, otherwise
The outputs of the model are as follows:
xi,m 1, if a sensor of type m, is located at i, and 0, otherwise
zj 1, if node j is covered, and 0, otherwise
k,1 1, if path k is covered by a single sensors, and 0 otherwise, and
k,2 1, if path k is covered by a pair of sensors, and 0, otherwise
The formulation of the model is as follows:
 
Max Hj zj + Hk (k,1 + k,2 )∀m ∈ {1,2, 3} (8.1)
j ∈N k∈P

s.t. xi,m ≤ pm ∀m ∈ {1,2, 3} (8.2)
i∈M

Ai,m,j xi,m ≥ zj ∀j ∈ S and m ∈ {1,2, 3} (8.3)
j m

Bi1,m1,i2,m2,k xi1,m1 xi2,m2 ≥ k,1 ∀k ∈ P (8.4)
i1,m1,i2,m2

B  i1,m1,k xi1,m1 ≥ k,2 ∀k ∈ P (8.5)
i1,m1,i2,m2

k,1 + k,2 ≤ 1 (8.6)

xi,m ∈ {0,1} (8.7)
zj ∈ {0,1} (8.8)
k,t ∈ {0,1}∀t ∈ {1,2} (8.9)

In the model, the objective function (1) maximizes the total coverage, first term max-
imizing the coverage from the nodes and the second term maximizing the coverage
from the path. Constraint (2) limits the number of acoustic sensors to p1 , limits the
number of traffic cameras to p2 , and limits the number of traffic flow sensors to p3 .
Constraint (3) defines node coverage and declares that a node j is “covered” if and
only if at least one sensor covers node j. Similarly, constraint (4) defines path cover-
age and declares a path “covered” if at least one pair of acoustic sensors cover it or
a pair of traffic cameras cover it, or a pair of traffic flow sensors cover it. Constraint
5 defines path coverage and declares a path ‘covered’ if a single traffic flow sensor,
200 T. Geetla et al.

a single acoustic sensor, or a single traffic camera covers the entire path. Constraint
(6) ensures that a single sensor or a pair of sensors can “cover” a path, and only
one term can enter the objective function. Constraint (7), (8) and (9) ensure that the
decision variables are all binary.
The paper Erdemir et al. (2008a) uses a Greedy Paired Adding Heuristic (GPAH)
to solve the above quadratic maximum coverage location problem (qMCLP). GPAH
is an efficient and computationally faster algorithms compared to other algorithms.
From computational complexity theory, we know that MCLP is a hard problem
to solve, we apply the same GPAH heuristic with minor variations to solve qM-
CLP. The GPAH solution is within 6.6 % of the optimal in the computational tests
performed Erdemir et al. (2008a). Since the GPAH algorithm performs well on
large-scale problems, we use it for our situation to obtain a starting solution for
the implicit model. The GPAH is a greedy heuristic that places the sensors by first
allocating sensors to the highest weighted node or a path. At every step, the greedy
heuristic searches for the highest weighted uncovered node or path. If a node is the
highest weighted uncovered, then the greedy heuristic locates one of the sensors
near the node. If a path is the highest weighted uncovered, then the greedy heuris-
tic places a pair of available sensors at the closest potential sensor locations on the
same path.
Note that objective function or constraints in the explicit model do not include
terms dealing with the data fusion capabilities. The implicit model detailed in the
next section captures this aspect.

8.2.4 Implicit Model

The implicit model in Erdemir et al. (2008a) provides a geometric (not based on
mathematical programming) method to solve the coverage problem with demand
originating from both nodes and paths. This model allows the users the flexibility
to adjust single, double and triple coverage. We adopt this model for the sensor
location problem but only utilize single and double coverage. In the sensor location
problem, single coverage implies that a single sensor covers the entirety of a path
and double coverage implies coverage of a path by two sensors. If there is a crash in
the area covered by two sensors, each sensor makes an independent observation of
the crash. Every sensor deployed in the field has a detection range/zone. Inside this
detection zone, each sensor observes an incident with a positive probability.
For example, consider a crash scenario inside the detection range of a sensor with
true positive probability of 0.7. This means that if there is a crash in its detection
area, the sensor detects it 70 times out of hundred. However, if the crash occurs
in a region covered by two sensors (assuming the sensors have the same detec-
tion rate and are independent), the combination of data from two sensors observes
91 crashes out of 100, which is an improvement of 30 %. The advantage double
8 Locating Intelligent Sensors on a Transportation Network . . . 201

coverage provides is with data fusion techniques on data obtained through multi-
ple sensors. The advantages of ITS systems with data fusion capabilities are both
improved true positive probability and an improved crash characterization.
In this geometric approach, we measure the coverage achieved by a particular
placement of sensors on every path and every node. For example, consider a path
A with a crash demand of 0.04. Suppose that in a specific sensor placement strat-
egy 34 of the path is covered by at least one sensor, and . . . of the path is covered
by two sensors and we assign a weight (the value of the weight reflects the relative
importance given to double coverage) of 0.7α1 to single coverage and 0.3α2 to dou-
ble coverage. Then this path contributes 0.04[(3/4)0.7 + (1/4)0.3] to the objective
function.
Max f [(x1 , y1 ), (x2 , y2 ), ..., (xn−1 , yn−1 ), (xn , yn )]
= α1 s1 [(x1 , y1 ), (x2 , y2 ), ..., (xn−1 , yn−1 ), (xn , yn )] (8.10)
+ α2 s2 [(x1 , y1 ), (x2 , y2 ), ..., (xn−1 , yn−1 ), (xn , yn )]
where s1 is a function of single coverage and α1 is the weight of single coverage,
and, similarly, s2 is a function of double coverage and α2 is the weight of double
coverage.
The objective function defined in (10) maximizes the single and double cov-
erage with associated weights.The decision variables are the location of sensors
[(x1 , y1 ), (x2 , y2 ), . . ., xn , yn )]. The heuristic methodology fixes the position of n-1
sensors and then finds the optimal location for the remaining sensor using the ob-
jective function (8). Each step randomly selects a sensor location from the current
solution. Removing a sensor from the chosen location and keeping the rest of the
(n-1) sensor locations intact, a new location for the displaced sensor is needed. The
new sensor location chosen from the previously un-occupied sensor locations that
maximizes objective function in (8). For this heuristic to work effectively, the au-
thors start the heuristic with the solution from the GPAH algorithm. We adapt the
same solution methodology. There are other approaches to generate improved solu-
tions. For example, a variation to this approach is to find the best location for two
sensors in the heuristic. This heuristic can fix the location of n-2 sensors and find
the best location for the two sensors. Using the first or the second approach to solve
the implicit model does not guarantee an optimal solution.
The inputs for the implicit model are: (i) the coordinates of the two end-points
of the paths, (ii) coordinates of potential sensor locations, (iii) mean demand per
unit length of a crash path, (iv) coverage radius of the sensors, (v) weights for single
and double coverage, and (vi) the number of actual sensors available. Output for the
implicit model is the coordinate locations of sensors.

8.2.5 Simulation as an Evaluation Tool

Computer simulation is an imitation of a real-world system using abstract mathe-

matical models representing the real system using computers. Computer simulation
202 T. Geetla et al.

or simulation is used in a number of manufacturing facility systems, road network

and service industries to study and improve stochastic processes. See Kelton et al.
(1997) and Law and Kelton (1999), for general introduction on use of simulation
and general applications.
Simulation or field-testing captures this complex interaction of traffic movement,
sensor detection and sensor data accumulation. Simulation as used here is a cost-
effective evaluation tool to quantify the effectiveness of the placement strategy.
There is thus a need to construct a simulation of traffic on a road network with
ITS sensors, where crashes randomly occur and where the placement of ITS sensors
creates data fusion opportunities.
One such simulation tool currently in development is the Automated Situational
Awareness Platform “ASAP” using Arena (essentially it is a simulation capable of
modeling vehicular movement throughout a study area and a validation of the emer-
gency vehicle travel times through historical crash response data in an existing traffic
network, see Henchey et al. (2013). In designing ASAP, a major goal was to address
issues associated with the response to highway-related emergencies. In particular,
ASAP is used to:
1. Understand how data is accumulated and used in an ITS systems to create
situational awareness of an emergency event
2. Use the information generated from data fusion to generate emergency response
decisions and then evaluate these decisions.
The ASAP platform provides the necessary tools to embed sensor modules that can
interact with the road network and the traffic. These sensor modules need to reflect
the abstract detection model of each sensor. For example, the simulation models
a traffic-monitoring sensor like an inductive loop detector. Based on the type of
inductive loop detector these sensors can count the motor vehicles moving on the
road segments. In the simulation, the inductive loop sensor can be a checkpoint,
which counts the motor vehicle entities passing on the road segment. In the real
world, a sensor always has error estimating the traffic counts/traffic flow. This real
world error will also be included in the simulation. The traffic sensor handbook, see
Federal Highways Administration (2009), contains error rates.
The simulation developed in Henchey et al. (2013) consists of three modules.
The first module creates a traffic network based on the road network extracted from
the case study. More specifically, the traffic network it creates has road segments,
intersections, intersection signal lights, motor vehicles and traffic flow at intersec-
tions. The module generates traffic using real world data based on expected traffic
flow as a function of the time of day, type of vehicle and weather. This module
also simulates traffic in different weather conditions. For example during a heavy
rain condition, the simulation automatically decreases the traffic speed on the roads
to model traffic movement in severe weather conditions. We call the first module
“traffic generation module.” What we intend to do is to validate the approach on a
realistic simulation model, as a surrogate to an actual field test.
The second module simulates a random crash on the road network. This module
uses the previous crash data collected for the case study to choose a crash location.
8 Locating Intelligent Sensors on a Transportation Network . . . 203

For example, the module chooses a path to create a crash. A random point on the
path serves as the actual location of the crash for this run of the simulation. Then
the crash creation module creates all the associated parameters, e.g., the number of
vehicles involved, number of people in each vehicle, delta velocity (difference in
velocities of the objects involved in crash at impact) of the impact, number of im-
pacts, principal direction of force (PDOF), and sensor relevant data for the crash.
For example, a traffic camera needs a crash scene image for the video image pro-
cessor software and an acoustic sensor needs a crash acoustic signature for signal
processing. The crash creation module uses past accident image and acoustic sig-
nature information to create this crash related sensor data. In order to facilitate a
random accident we need a large collection of crash sensor information.
The third module is the sensor module. This part of the simulation is for the
sensor placement problem. This module imitates a real sensor behavior in a data
fusion environment. Significant functions of the sensor module are to use the crash
relevant sensor ground-truth data (here ground truth refers to the actual events and
conditions, as opposed to that observed by sensors which can have errors) gener-
ated by the crash generation module and to create crash specific data sets for every
sensor. After the sensor module receives the crash data, the sensor module eval-
uates the crash. For example if the crash generation module generates a random
crash at (43.024592, − 78.798994) with an acoustic signature, the acoustic sensor
located at (43.026178, − 78.800012) receives the acoustic signature from the crash.
The sensor module calculates the distance between the crash and the sensor loca-
tion. Through the traffic module, the sensor collects the congestion data (e.g., on
the particular road on is 600 vehicles/h) and the local temperature from weather
station data. Using this information the sensor module estimates the sound decay
and generates a new acoustic signature with decay. This new decayed acoustic sig-
nature will be the ground truth for this particular acoustic sensor. Now the sensor
module performs incident detection using the decayed amplitude versus time data.
If the crash is detected it will calculate the error rate in its estimation. Each acoustic
sensor module makes the necessary calculations of sound decay from the acoustic
signature of the crash and then uses a threshold detection value to sense the crash.
Each simulation starts first by creating traffic on a road network on a regular day
from 7:00 a.m. to 9:00 p.m., with a predefined location of sensors. The crash creation
module creates a random crash from the crash-weights given to each road segment
from past data. Once the crash generation module creates a crash and the necessary
calculation in the sensor modules is complete, the simulation collects the data from
nearby sensors and stops the simulation. The sensor data obtained from the nearby
sensors is analyzed using signal-processing techniques. The output from the signal
processing is the input for data fusion techniques. The output from the data fusion
creates an estimate of the crash location and other crash characteristics. To gather
statistically significant performance measures like percentage detection, error in es-
timation of the crash location and the quality of crash estimates, the simulation is
run multiple times generating random crashes in every simulation run and keeping
the position of the sensors unchanged. We note that the issue of false positives is
204 T. Geetla et al.

not specifically considered here, in that we assume that the necessary emergency
response units/resources are available to respond to any detected incident.

8.2.5.1 Simulation-Based Optimization

The next step in the solution methodology is to use a simulation-based optimization

procedure. Simulation-based optimization uses simulation as an evaluation tool and
integrates an optimization method into the simulation to reach better performing
solutions. The simulation evaluates the stochastic system and the optimization tool
uses a heuristic, approximation or response surface methodology. Further reading
material on simulation-based optimization see Law und Kelton (1999) and Dong
(2007).

8.3 Location of Acoustic Sensors to Detect and Characterize

a Crash

So far, we have discussed the use of sensors in incident management systems and
the sensor placement model for the stationary sensors. A hybrid solution method-
ology using explicit-implicit and a simulation-based optimization procedure is
proposed. In this section, we apply the proposed solution methodology only to the
omnidirectional sensors, for which acoustic sensors act as representative sensors.

8.3.1 Case Study for the Acoustic Sensors Placement Problem

We chose a case study to test the explicit-implicit-simulation-based optimization

methodology for the omnidirectional sensors. We still follow the methodology pro-
posed but use specific solution methodologies only applicable to acoustic sensors
and in general omnidirectional sensors.
We chose a road network that covers a 9-mile square area near the University of
Buffalo’s North Campus in Buffalo for the case study. This road network consists of
an urban-arterial road network with part of an interstate-highway running through
it. The area chosen has a university, shopping mall, three schools and other com-
mercial areas. This area represents a road network with moderate traffic flow with
peak traffic typically in the early mornings and evenings. The average annual traffic
data is a reference to recreate the traffic flow patterns in the simulation, see Greater
Buffalo-Niagara Regional Transportational Council (2010). The area chosen is with-
out a hilly terrain and large industries, which is ideal for understanding and testing
the placement of acoustic sensors (large industries and hilly terrain would create an
added level of complexity into the simulation model). We assume that road noise
8 Locating Intelligent Sensors on a Transportation Network . . . 205

from the traffic is the most influential background source in the signal processing
for the acoustic sensors, which act as the only incident detection sensor.
The simulation case study collects two important performance measures from
every simulation run. The first performance measure is, “percentage of crashes
detected” by the acoustic sensors with a given location strategy. The second perfor-
mance measure is “% crashes detected by two acoustic sensors.” This performance
metric measures the data fusion opportunity arising from having multiple sources of
crash sensor data.
There are many factors, which influence these performance measures. One such
factor is the placement of the sensors, which influences both the performance mea-
sures. So far, in this study the solution methodology produces three placements
of the acoustic sensors, the first from the explicit model, the second from the im-
plicit model and the third, from the solutions obtained through simulation-based
optimization.

8.3.1.1 Explicit Model Solution

The explicit model for the acoustic sensor placement problem is an approxi-
mate mathematical model. We utilize the mathematical formulation in 2.3 to suit
the acoustic sensor placement problem. The objective function remains the same,
maximizing the coverage achieved by a fixed number of sensors.
In the area chosen for the case study, there were 3600 reported crashes over a
5-year period. From observing the crash data, we identified 11 locations that have
20 or more crashes occurring in very close proximity. These 11 locations have 784
crashes in total and serve as nodes in the path and node model. The weight (demand)
of a node was set equal to the number of crashes that occurred at the node divided by
the total number of crashes. The remaining 2816 crashes occurred on the remaining
road network. The road network divides into road segments that serve as paths in
the explicit model. The weight (demand) of a path was set equal to the number of
crashes on the path divided by the total number of crashes. We chose to have 760
paths in the model and of these 760 paths, 208 had zero demand.
To construct candidate sensor locations, we first chose all the ends of a path as
potential sensor locations. This choice yielded 400 potential sensor locations. We
also generated 469 additional locations by first randomly selecting 469 paths and
then for each selected path we choose a random location in the interior of the path.
The total number of candidate sensor locations is therefore 869.
The GPAH algorithm presented in Sect. 8.2.3 can generate feasible solutions for
the acoustic sensor placement problem. In Fig. 8.2, we present the performance
of the GPAH algorithm with different number of sensors. The results presented in
Fig. 8.2 show that the percentage of covered crashes increases as we increase the
number of sensors. In Fig. 8.3, we show the sample solution of the explicit model
on the case study chosen near the University of Buffalo’s North Campus with an ‘n’
value of 55. The GPAH algorithm (developed in JAVA®) runs is under 15 min using
206 T. Geetla et al.

Fig. 8.2 Explicit model solution

Intel® CoreTM 2 Duo with clock speed of 3.17 GHz and 4 GB RAM for the case
study chosen.

8.3.1.2 Implicit Model Solution

The Implicit model proposes an alternate solution methodology to the sensor place-
ment problem. The implicit model is inherently a local search heuristic to improve
a solution using a different objective function from the explicit model.
Observing the objective functions in Fig. 8.4 gives an impression of how much
the implicit model improves over the explicit model. This behavior is expected as
the implicit solution uses the explicit solution as a starting solution and improves it
using the local search heuristic.
Figure 8.5 presents the implicit model solution on the same case study. Compar-
ing Figs. 8.5 and 8.3, observe that sensor placements in the implicit model solution
are clustered (i.e. several sensors located nearby) whereas the explicit model solu-
tion has scattered placement of sensors. This clustering of sensors in the implicit
model is the result of having a secondary coverage term in the objective function.
Such secondary coverage consideration is not part of the explicit model.
From a computational perspective, this implicit model (developed in JAVA®)
uses the explicit model solution as a starting solution and takes a computation time
of 3 h using Intel® CoreTM 2 Duo with clock speed of 3.17 Hz and 4 GB RAM,
compared to the computation time of less than 15 min for the explicit model. These
computation times reported here are for the case study presented in Sect. 8.2.5.
8 Locating Intelligent Sensors on a Transportation Network . . . 207

Fig. 8.3 Map of acoustic sensor location using explicit model strategy

Fig. 8.4 Comparing objective functions of explicit and implicit model

208 T. Geetla et al.

Fig. 8.5 Map of acoustic sensor location using implicit model strategy

8.3.1.3 Simulation-Based Optimization in Acoustic Sensors

As noted in the case study, there are 869 candidate sensor locations and we plan to
place 20–85 sensors in some of them. The number of possible ways to locate the
869
sensors is large, e.g., C20 solutions exist for 20 sensors, far too many to evaluate
using simulation. Note that each solution needs multiple simulation runs to reach
statistically significant results to make a comparison with other placement strate-
gies. Given these constraints, the goal of the simulation-based optimization is to use
neighborhood search to identify a few solutions with potentially better performance
than the initial solution (obtained by placing sensors from implicit model as shown
in Fig. 8.5 for the case study).
A neighborhood search procedure for the acoustic sensor location problem
should generate alternate placement strategies from the present solution and move
in the direction of improved performance measures. There is no guarantee that this
method will generate a better solution because the method is an exploratory search
method. To generate a feasible set of neighbor solutions we propose a cluster-based
exchange for the acoustic sensors.
A cluster of sensors refers to a geographically close set of sensors. In the cluster-
based exchange, we move a sensor from one cluster to an unoccupied potential
sensor location in another cluster as shown in Fig. 8.6.
8 Locating Intelligent Sensors on a Transportation Network . . . 209

Fig. 8.6 A demonstration of cluster-based exchange

To implement the cluster-based exchange we first divide the set of potential sen-
sor locations occupied by sensors into clusters. For example, in the case of 75
sensors we chose six clusters. The reason for choosing six clusters is that we ex-
pect roughly 10–15 sensors per cluster. We use the Lloyd’s algorithm to form the
six clusters, see Lloyd (1982). Figure 8.7 illustrates the six clusters obtained by ap-
plying Lloyd’s algorithm on the 75-sensor problem for our case study. We note that
the number of sensors in each cluster varies significantly. To proceed with gener-
ating a neighboring solution, we select a geographically distant sensor (from the

Fig. 8.7 Cluster-based exchange generating neighbor solutions

210 T. Geetla et al.

center of one cluster) and remove the sensor allocated to this location. A new po-
tential sensor location is chosen (preferably closer to the center of the cluster) from
another random cluster and a sensor is allocated to this new location.
There are other methodologies to create feasible neighbor solutions; however, the
cluster-based proposed here preserves the data fusion opportunities provided from
using implicit model.

8.3.1.4 Preliminaries Related to Computational Experiments using

Simulation

In order to compare statistically significant estimates for the performance measures

we use 20 experiments for each acoustic sensor placement considered. Each exper-
iment consists of 50 individual simulation runs. At the end of each experiment, the
performance measures are collected. For example, if the first experiment results in
the detection of 30 crashes by the acoustic sensors, the performance measure “% of
crashes detected” is 30/50 or 60 %. We repeat this experiment 20 times generating
19 more observations for the “% of crashes detected” performance measure. In to-
tal, we run 1000 simulation runs for an acoustic sensor placement strategy. Running
this Arena® simulation using Intel® Core2 Duo with clock speed of 3.17 Hz and 4
Gb RAM, takes on average 45 s to complete a simulation run and takes on average
12.5 h to complete evaluation of a single placement of sensors.
Three points of clarification are necessary regarding our experiments. First, since
the computation time for evaluation of a single placement strategy is 12.5 h only a
few more evaluations of additional sensor placement alternatives are possible during
the neighborhood search procedure. Second, each simulation run should always start
with a different random number seed in Arena; this ensures that the results collected
from the simulation runs are statistically independent. Third, the fact that we run 20
experiments with each run generating a statistically independent value for both the
performance measures studied, allows us to generate 95 % confidence intervals for
the performance measures.

8.3.1.5 Explicit Model Versus Implicit Model Using Simulation Evaluation

In this section, we evaluate the results of the explicit and the implicit model using
simulation. We use the simulation model as an evaluation tool for both the perfor-
mance measures in the explicit and implicit model. For comparison, we place 75
sensors using the explicit and the implicit model. Note that the solution method for
the implicit model starts by using the explicit model solution as the starting point.
For the “% of crashes detected” performance measure, explicit model placement
has a mean of 23.1 % and the implicit model has a mean of 43.5 %. From the 95 %
confidence intervals in Fig. 8.8, the explicit model has higher variance in “% of
crashes detected” performance measure. Similarly, for “% of crashes detected by
8 Locating Intelligent Sensors on a Transportation Network . . . 211

Fig. 8.8 Explicit versus implicit solution evaluation in simulation with 95 % confidence intervals

two sensors,” the explicit model placement has a mean of 0.3 % and implicit model
placement has a mean of 11.2 %.

8.3.1.6 Implicit Model Performance with Different Number of Sensors

The second factor that will affect the performance measures in the implicit model is
“number of sensors.” Ideally, the number of sensors available depends on budgetary
constraints. In this section, we present the results of changing the number of sensors
on both the performance measures.
In Fig. 8.9, we present the 95 % confidence intervals for the “% of crashes de-
tected” performance measure with different number of sensors (50, 55, 60, 65, 70
and 75). The mean of “% of crashes detected” performance measure is found to
increase with the number of sensors.

Fig. 8.9 Implicit solution with different number of sensors evaluated using simulation
212 T. Geetla et al.

Fig. 8.10 Neighbor solutions generated using two iterations of simulation-based optimization

Another observation from Fig. 8.9 is the mean of “% crashes detected” perfor-
mance measure is not diminishing as the number of sensors increase; in particular
going from 55 to 60 sensors yields less improvement than going from 60 to 65 sen-
sors. The solution the implicit model reaches is a function of the explicit solution;
we attribute this discrepancy to having a comparatively better starting solution go-
ing from 60 to 65 sensors, i.e. the explicit solution for 65 sensor is a better solution
than 60 sensors and the heuristic approach of the implicit model found better so-
lutions starting from 65 sensors, compared to starting at 60 sensors. We note that
other approaches like starting the implicit model from a random solution were not
considered.

8.3.1.7 Results of Simulation-Based Optimization Using a Neighborhood

Search Heuristic

Simulation-based optimization adds a local search method to the simulation to

search for better solutions in the neighborhood of a particular solution. For the
acoustic sensor placement problem, we proposed a cluster-based exchange to
provide neighbor solutions as described in Sect. 8.3.1.3.
We initiate the cluster-based exchange procedure with the implicit model solu-
tion and generate a set of five neighbor solutions that are different from each other.
Each generated solution has a different sensor placement strategy. In the next step,
simulation evaluates each placement strategy. In Fig. 8.10, the neighbor solutions
are around the implicit model solution.
Since the implicit model solution is still one of the efficient/non-dominated solu-
tions after the first step, we use this as the starting point to generate a new set of five
neighbor solutions for the second step. We always choose a non-dominated solution
to start the search, since the local search process plans to move only in the improved
direction.
8 Locating Intelligent Sensors on a Transportation Network . . . 213

Figure 8.10 shows two steps of neighborhood search. In every Step 5 new
solutions are generated and using simulation the two performance measures are
evaluated. In Fig. 8.10 the solutions from first step are marked using the red tri-
angles whereas the solutions from the second step are marked in blue. The goal of
any neighborhood search is to generate new solutions, which are better in both per-
formance measures. The newly generated neighbors are evaluated using simulation
to observe both the performance measures. From the results in Fig. 8.10, the implicit
model solution is still a non-dominated solution in both the performance measures
after two iterations. Other solutions from the neighborhood search are not efficient.
It proves our intuition that the implicit model is a good solution for the acoustic
sensor placement problem.

8.3.2 Exploration of Three Special Cases: Higher Detection

Radius, Solution Robustness and Single and Double
Coverage Weights

In this section, we explore three cases of special interest in the acoustic sensor place-
ment problem. These cases demonstrate the use of simulation as an experimental
tool to evaluate future scenarios or specific questions useful for a practitioner. In
Sect. 8.3.2.1, we present the change in performance measures if there is an increase
in the detection radius of the acoustic sensors. Section 8.3.2.2 presents a robustness
test to evaluate the implicit model solution. Section 8.3.2.3 presents the effects of
changing single and double coverage values chosen in the implicit model on the two
performance measures.

8.3.2.1 Evaluation of a Sensor with Higher Detection Radius

Detection radius is a key input for both the explicit and implicit model. A favorable
change for a sensor location problem is the improvement in the detection radius. An
improvement in detection radius for a sensor is an important factor as it usually has
a direct impact on system performance measures. Section 8.3.1 assumes a detection
radius of 50 m (which is current specification for commonly available acoustic sen-
sors). This low detection radius for commonly available acoustic sensor is attributed
to the decay of sound in atmosphere. In this section we study the impact of a higher
detection radius (60, 70, 80 and 90 m) on the “% of crashes detected” and “% of
crashes detected by two sensors” performance measures. A higher detection radius
is possible if the acoustic sensors start using noise isolation and noise filtering.
From Fig. 8.11, we notice the increase in “% of crashes detected” performance
measure, when the detection radius increases from 50 to 90 m is about 4 %. The
increase in “% crashes detected by two sensors” is 26 % when the detection radius
changes from 50 to 90 m. Using a higher detection radius, the sensors increase the
214 T. Geetla et al.

Fig. 8.11 Simulation results for increased detection radius of the two performance measures

overlap region but considering the 9-mile square area chosen, the sensors are unable
to cover a large part of the road segments previously uncovered by sensors with
50 m detection radius. This explains increased gains in the “% crashes detected by
two sensors” performance measure. This also raises the need to explore potential
savings (in terms of higher costs for sensors versus better detection of incidents or
false positives).

8.3.2.2 Solution Robustness

A key input for both the explicit and the implicit model is the past crash/demand
data. We use past crash data to construct demand for each road segment in the ex-
plicit and the implicit model. In this chapter we assume that the past crash data
serves as a good indicator for the future crash data that guides the placement of the
incident detecting sensors. There are equally good arguments made to support and
reject the above assumption. For example, accidents are more common on or near
entry and exit roads of a highway where large volumes of traffic merge. However,
not all entry and exit highways are equally likely to have an accident. The other
reasons like road construction and traffic flow rates play a major role in causes of
accident. A favorable argument to use the past accident data is statistically some
highway entry and exit paths have a higher probability of accidents. Law of large
numbers suggests that if there is a large volume of data, the mean number of ac-
cidents remains the same. Most accidents are attributed to driver distraction, this
argument assumes that driver distraction is the major cause and no amount of past
data can predict future crashes.
In order to alleviate some of the concerns from the assumption, we design a
robustness test, to evaluate the performance of the implicit model through an ex-
periment. In this robustness test, the implicit solution strategy is experimented via
changes in the crash generation module. In general, the crash generation module
8 Locating Intelligent Sensors on a Transportation Network . . . 215

Table 8.1 Shows the construction of new demand for nodes and paths for the robustness test
Path/node Probability of crash Robustness test (p = 0.9) Robustness test (p = 0.7)
Path a 0.2 0.2(0.9) 0.2(0.7)
Path b 0.3 0.3(0.9) 0.3(0.7)
Path c 0.1 0.1(0.9) 0.1(0.7)
Node d 0.4 0.4(0.9) 0.4(0.7)
Path e 0 (1 − 0.9)/2 (1 − 0.7)/2
Node f 0 (1 − 0.9)/2 (1 − 0.7)/2

uses past crash data to simulate crashes in the simulation. We modify the approach
that uses past crash data in the simulation and use the implicit model placement
strategy for the sensors.
In our study area, we found that close to 30 % of the road segments have no
crashes in the period 2004–2009, but may clearly have future crashes. We use a
parameter p to signify the likelihood that past crash data accurately predicts future
crash data, with p = 1 implying that perfect predictability. In this section we test
the performance of the placement suggested in Sect. 8.3.1 with different values of p
(p = 0.9 and p = 0.7). If the value of p is chosen as 0.9, in that experiment 90 % of the
crashes happen on road segments where the crashes have occurred previously and
10 % of the crashes occur on road segments that did not have an crash previously.
Table 8.1 shows the construction of the demand for the two cases of p = 0.9 and p =
0.7. We construct a new path and node demand values and evaluate the placement
suggested by the implicit model using 75 sensors.
From the results shown in Figs. 8.12 and 8.13, comparing the original (p = 1)
results with the robustness test (p = 0.9 and p = 0.7), both the performance measures
decrease. The percentage of crashes detected by a single sensor decreases by 10 % as
the value of p decreases by 0.1. However, double coverage decreases drastically and
as p-value approaches to 0.7 there are just 1 % of the crashes covered by two sensors.
These results also indicate to a user/practitioner that if 30 % of future crashes happen
on road segments, which previously have no crashes, the data fusion opportunities
shrink to zero.

8.3.2.3 Effects of Changing Single/Primary (α1 ) and Double/Secondary

Coverage (α2 ) on Performance Measures

The implicit model allows a practitioner/user to choose between primary/single cov-

erage and secondary/double coverage and assign values for (α1 ) and (α2 ) in Eq. 8.10.
This choice of single and double coverage weights affects sensor placement and af-
fects both performance measures. In the results shown in Sect. 8.3.1, we selected
α1 = 0.7 and α2 = 0.3. In this section, we experiment with various values of α1 and
evaluate the changes in both the performance measures.
216 T. Geetla et al.

Fig. 8.12 Results from the robustness test using simulation

Fig. 8.13 Robustness test results showing both the performance measures

Table 8.2 A table showing Single coverage α1 Double coverage α2

different weights chosen for
the primary and secondary Experiment 1 0.4 0.6
coverage split Experiment 2 0.5 0.5
Experiment 3 0.6 0.4
Experiment 4 (original) 0.7 0.3

In order to understand the effect of single and double coverage values on the
performance measures, we chose three experiments with α1 and α2 values as shown
in Tab. 8.2. For example, Experiment 1 assigns a weight of 0.4 to α1 and 0.6 to α2 .
Experiment 4 references the results in Sect. 8.3.1.
8 Locating Intelligent Sensors on a Transportation Network . . . 217

Fig. 8.14 Performance of implicit model with varying levels of primary and secondary coverage
evaluated using simulation

From Fig. 8.14 we observe a clear positive correlation (except for Experiment 3
and Experiment 4) between the α1 value and “% of crashes detected” performance
measure and similarly there is a positive correlation between “% of crashes de-
tected by two sensors” performance measure and α2 . Upon further inspection of data
from sensor placement strategy 3 and 4, there is very little variation in performance
measures and a large overlap in the 95 % confidence intervals of both performance
measures, attributing this exception to random nature of the simulation. Using this
knowledge of the relation between the weights for the implicit model and both the
performance measures should help the user/practitioner to select the appropriate
weights before using the implicit model.

8.3.3 The Impact of Mobile Sensors

In Sect. 8.3.1 and 8.3.2, we discussed the placement of acoustic sensors and the
evaluation of the placement in various scenarios. This section presents the uses of
stationary sensors working in tandem with mobile sensors to generate a situational
awareness of an incident. We use AACN sensors as a representative sensor for mo-
bile sensors. Both AACN and acoustic sensors use a signature crash characteristic
to detect a crash. Acoustic sensors use the unique crash sound signature to detect
a crash, whereas AACN sensors use accelerometers and contact sensors to detect a
crash. Both the sensors are capable of detecting the crash characteristics:
• Crash detection
• Crash location
• Crash time
• Rollover
• Number of impacts
218 T. Geetla et al.

For example, a motor vehicle involved in a crash very close to an acoustic sensor
and the automobile involved is equipped with an AACN sensor; there are two sets
of dissimilar crash data for the above crash characteristics. The acoustic sensor at-
tempts to extract a sound feature that consists of a high amplitude noise for a short
duration of time. If the acoustic signal processing succeeds in extracting the crash
sound feature, the signal processor accomplishes detection and estimates other crash
characteristics. Similarly, the AACN reports a crash when the accelerometers report
unusual acceleration/deceleration observed in motor vehicle crash. Since a single
acoustic sensor observes this crash, we assume that the acoustic sensor reports a
point estimate for the crash and sends its (own) GPS location as an estimate for
crash location. Every AACN includes a GPS sensor that estimates the position of the
crash. The acoustic sensor and the AACN both report different sets of information
to determine crash time, e.g., the observation times of a crash-like high-amplitude
sound and high accelerometer reading. An AACN sensor also may include a 3-d ac-
celeration sensor system that can record unusual readings to detect a rollover. Once
the two crash data sets are available, the next step in the data fusion schema is to
send this data from individual sensors to local data fusion motes (a mote is a data
processing unit that can receive data from sensors and use data fusion algorithms).
We assume that a backbone fusion network, capable of combining information
from both acoustic sensors and sensors in the AACN system exists. Suppose there
is a roadway incident involving two cars in a data fusion capable environment, pop-
ulated with acoustic sensors and cars enabled with AACN. The tree diagram shown
in Fig. 8.15 identifies the data fusion opportunities.
The tree diagram in Fig. 8.15 identifies the levels of data fusion needed when
acoustic sensors and AACN sensors are considered. Consider the case of a two-car
crash in which one car is equipped with AACN. There is just one acoustic sensor
near the crash, two independent crash data sets are available through the acoustic
sensor and AACN and there are two levels of data fusion possible, L0 and L2. L0
data fusion associates the data from acoustic sensor and AACN. After Level 0 data
fusion, the two data sets are combined using L2 level data fusion to generate crash
data useful for situational awareness.
Expanding upon Fig. 8.15, in Fig. 8.16 we present the data fusion opportuni-
ties with 75 acoustic sensors and assuming 10 % of the vehicles are equipped with
AACN. The probability values reported for various scenarios of detection are results
from simulation.
Table 8.3 summarizes the simulation results and shows that the combined sensor
system detects 53.5 % of the crashes. Seventy-five acoustic sensors (without the
aid of AACN sensors) could detect 43.5 % of the crashes in the same system. With
AACN sensors alone (in 10 % of vehicles), 16.9 % of crashes were detected.
By simulating the AACN sensors alone, the error in position in 169 detected
crashes was 4.23 m. Using acoustic sensors alone in a data fusion environment, the
error in position was 5.42 m in 435 crashes. Combining the AACN with the 75 acous-
tic sensors, the error in position in 535 detected crashes was 5.04 m. Finally, using
the combined AACN and the acoustic sensor system, the error in detecting multiple
impacts was 5 %.
8 Locating Intelligent Sensors on a Transportation Network . . . 219

Fig. 8.15 A tree diagram identifying the data fusion opportunities when acoustic sensors and
AACN act as incident detection sensors

Using both AACN and acoustic sensors together in an incident detection system
capable of data fusion reduces the error in position (relative to acoustic sensors
alone) to 5.04 m. Although this reduction may seem small, it can be significant
in areas containing multiple road segments over bridges and through underpasses.
Five-meter accuracy in the position of a crash is close to lane level accuracy and
very helpful for emergency response personnel. The simulation evaluation assumes
that in the case study, GPS sensors work with an error of 10 m (Hubrich and Cur-
ran 2009) and the road network in the system has no tunnels and other forms of
interruptions for GPS systems. Even in cases with GPS interruption, we believe
that a system with acoustic sensors and AACN is effective and will help detect and
characterize a crash.
Table 8.3 demonstrates the positives of using a data fusion capable environment
with different sensor types observing a crash. The simulation evaluation results point
towards the improvements in crash detection rate for the combined system. It also
shows a decrease in error for the combined system relative to the acoustic sensor
alone, in detecting both the crash position and number of impacts, which can be
critical in assessing the crash.
220 T. Geetla et al.

Fig. 8.16 Data fusion opportunity in an incident detection system with 75 acoustic sensors and
10 % vehicles equipped with AACN

Table 8.3 Summary of the advantages of using AACN and acoustic sensors in incident detection
Acoustic sensors AACN sensor Acoustic and AACN sensor
% of crashes detected 43.5 16.9 53.5
Average error in estimation of 5.42 4.23 5.04
accident location (m)

8.4 Concluding Remarks

Multi sensor data fusion is essential in future road-transportation systems. For ITS
systems used in incident-detection, data fusion is paramount and increases the us-
ability and reliability of data from individual sensors. If stationary sensors are a part
of the incident detection system, their detection capability and quality of data fusion
are dependent on the placement of the sensors. There is abundant literature on place-
ment of sensors for maximizing incident detection. However, little to no literature
exists on placement models that consider both quality of data fusion and detection
capability in placement of stationary sensors. This chapter describes a sensor place-
ment problem for multiple sensor classes, where both single coverage and double
coverage are important in placement of sensors. Single coverage is a surrogate for
indirectly measuring the detection capability and double coverage is a surrogate for
indirectly assessing data fusion opportunities.
This chapter proposes an explicit-implicit model, followed by simulation-based
optimization to incorporate single and double coverage in the sensor placement
problem. This is a unique approach developed for this chapter. The explicit-implicit
model is a mathematical model that approximates the real situation, and proposes
8 Locating Intelligent Sensors on a Transportation Network . . . 221

an objective that aims to cover crash demand from both road segments (paths) and
intersections (nodes) where crashes are more likely to happen. The explicit model is
a quadratic maximal coverage problem and uses a greedy paired adding heuristic to
generate a feasible solution. The implicit model is an alternate geometric approach
to solve the sensor placement model. The explicit model placement strategy acts
as the starting solution for the implicit model. The simulation-based optimization
generates additional feasible solutions using a local search and improves the best
sensor placement strategy. This combined approach has yielded good solutions to
the omnidirectional sensor placement problem.
The simulation model borrowed from Henchey et al. (2013) and developed in
this chapter imitates traffic movement in a real road network and the functioning
of omnidirectional (acoustic) sensors and mobile (AACN) sensors. The simulation
model has unique features that create a crash, measure the performance of the sen-
sors and has built-in weather-effects, congestion and road noise. The simulation
model developed here evaluates a placement strategy for the acoustic sensors. The
main performance metrics considered are “% of crashes detected” and “% of crashes
detected by two sensors”.
Stochastic systems use simulation as an evaluation tool built on abstract models
to understand and quantify the real-system. This evaluation process goes well with
an improvement procedure to an existing system by using a local search procedure to
evaluate local solutions and move in the direction of the improved performance. This
approach used here to evaluate the acoustic sensor placement strategy and provide
more local search results for a decision maker to optimize their criteria.
This chapter makes significant contribution in using simulation as a beneficial
and cost-effective evaluation tool to answer questions pertaining to sensor deploy-
ment in a road network environment. We demonstrate that by choosing the location
of just 75 acoustic sensors with a detection radius of 50 m, 43.5 % of the crashes are
observed by a single sensor and two sensors make independent detections in 11 %
of the crashes. The combined incident detection system of AACN and acoustic sen-
sors detects 53.5 % of the accidents and the average error in estimation of the crash
location is 5.04 m.
The chapter evaluates the data fusion capable system with both acoustic and
AACN sensors. Many papers discuss the capabilities of an advanced system involv-
ing multiple sensor types, but none has been able to evaluate the real world applica-
tions. This work demonstrates the effectiveness of such systems and quantifies the
improvements.
Several factors affect the overall performance of the omnidirectional sensor
placement. Three factors explored in this chapter are radius of detection, single
and double coverage weights for the implicit model and the change in crash de-
mand data. In Sect. 8.3.2, we evaluate all these factors individually to understand
the relation between the factors and performance measures. A future research direc-
tion would be to explore the combined effects of changing multiple factors on the
performance measures.
222 T. Geetla et al.

References

Administration FH (2009) Traffic detector handbook, 3rd edn., Vol I. US Department of Trans-
portation. http://www.fhwa.dot.gov/publications/research/operations/its/06108/. Accessed 4
Dec 2015
Berman O, Drezner Z, Krass D (2013) Continuous covering and cooperative covering problems
with a general decay function on networks. J Operat Res Soc 64:1644–1653
BMW USA (2008) BMW innovations, Real time traffic information. www.bmwusa.com/Standard/
Content/Innovations/Engineering/RTTI/Default.aspx. Accessed 4 Dec 2015
BMW (2010) BMW Insights safety. http://www.bmw.com/com/en/insights/technology/
connecteddrive/2010/safety/emergency_call/emergency_call_information.html. Accessed
4 Dec 2015
Cambridge Systematics & Texas Transportation Institute (2005) Traffic congestion and reliability:
trends and advanced strategies for congestion mitigation. U.S. Department of Transportation,
Federal Highway Administration
Church R, ReVelle C (1974) The maximal covering location problem. Pap Reg Sci 32:
101–118
Dhillon SS, Chakrabarty K (2003) Sensor placement for effective coverage and surveillance in
distributed sensor networks. Proceedings of IEEE Wireless Communications and Networking
Conference, pp 1609–1614
Dong G (2007) Simulation-based optimization. Ph. D. Dissertation, University of Wisconsin-
Madison
Erdemir ET, Batta R, Rogerson P, Speilman S, Blatt A, Flanigan M (2008a) Optimization of
aeromedical base locations in new mexico using a model that considers crash nodes and paths.
Accid Anal Prev 40:1105–1114
Erdemir ET, Batta R, Rogerson P, Speilman S, Blatt A, Flanigan M (2008b) Location coverage
models with demand originating from nodes and paths: application to cellular network design.
Eur J Oper Res 190(3):610–633
Gentili M, Mirchandani P (2005) Locating active sensors on traffic networks. Ann Oper Res
136:229–257
Greater Buffalo-Niagara Regional Transportational Council (2010) Geographic information sys-
tem. http://www.gbnrtc.org/index.php/planning/gis/. Accessed 4 Dec 2015
Hall D, Llinas J (1997) Introduction to multi-sensor data fusion. Proc IEEE 85:6–23
Henchey M, Batta R, Blatt A, Flanigan M, Majka K (2013) A simulation approach to studying
emergency response in an advanced transportation system. J Simula 8:115–128
Hubrich S, Curran K (2009) Optimizing mobile phone self-location estimates by introducing
beacon characteristics to the algorithm. J Loc Based Serv 3:55–73
Kelton D, Sadowski R, Sadowski D (1997) Simulation with arena. McGraw-Hill, New York
Law A, Kelton D (1999) Simulation modeling and analysis. McGraw-Hill, New York
Lloyd S (1982) Least square quantization in PCM. IEEE Trans Info Theory 28(2):129–137
Nagare A (2012) Brief introduction to intelligent transportation systems. www.freeway.gov.tw/
UserFiles/File/Traffic/A1%20Brief%20introduction%20to%20Intelligent%20Transportation
%20System,%20ITS.pdf. Accessed 4 Dec 2015
Nie X, Parab G, Batta R, Lin L (2012) Simulation-based selectee lane queuing design for passenger
checkpoint screening. Eur J Oper Res 219:146–155
Office of Highway Policy Information (2009) Highway statistics. (Office of Highway Policy In-
formation). http://www.fhwa.dot.gov/policyinformation/statistics/2009/tc202c.cfm. Accessed
4 Dec 2015
OnStar (2012) Automatic crash response. https://www.onstar.com/web/portal/emergencyexplore?
tab=1. Accessed 4 Dec 2015
Srinivasan K, Jovanis P (1997) Determination of number of probe vehicles required for reliable
travel time measurement in urban network. Transp Res Rec 1537:15–22
White J, Thompson C, Turner H, Dougherty B, Schmidt D (2011) WreckWatch: automatic traffic
accident detection and notification with smartphones. J Mob Netw Appl 16(3):285–303
Chapter 9
Location Models for Preventive Care

Vedat Verter and Yue Zhang

9.1 An Introduction to Preventive Care

Health conditions are much easier to prevent than to treat, and recovery is often more
likely if an illness is diagnosed at an early stage. The substantial savings in the costs
of diagnosis and therapy as well as the relatively lower capital investment associated
with preventive care programs have been recognized for a long time (Walker 1977).
Preventive care programs can save lives and contribute to a better quality of life
by reducing the needs for radical treatments. For example, there is evidence that
mammograms taken on a regular basis have the potential to reduce deaths from
breast cancer for women between the ages of 50 and 69 by up to 40 % (Health
Canada 2007).
According to World Health Organization (2002), although many diseases can be
prevented, the current healthcare systems do not make the best use of their avail-
able resources to support preventive programs. Most of these systems are based on
responding to acute problems, urgent needs of patients, and pressing concerns. Pre-
ventive care is inherently different from healthcare for acute problems, and in this
regard, current healthcare systems worldwide fall remarkably short. For instance,
only 5 % of the $ 1.4 trillion spent on direct health care in the United States goes to
preventive care measures and the promotion of general health (Falkenheimer 2004).
Preventive care programs can be categorized into three groups with regards to
their objective: (1) primary prevention aims at reducing the likelihood of diseases
in people with no symptoms, e.g., immunizations of healthy children; (2) secondary
prevention aims at identifying and treating people who have risk factors or are at

V. Verter ()
Desautels Faculty of Management, McGill University, 1001 Sherbrooke Street West, Montreal,
H3A 1G5, QC, Canada
e-mail: [email protected]
Y. Zhang
College of Business and Innovation, University of Toledo, 2108 West Bancroft Street, Toledo,
43615, OH, USA
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 223

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_9
224 V. Verter and Y. Zhang

very early stage of diseases, e.g., pap smears to detect early forms of cervical cancer;
(3) tertiary prevention aims at treating symptomatic patients in an effort to decrease
complications or severity of disease, e.g., sugar control in a diabetic in order to
mitigate vision and nerve problems. Flu shots, blood tests, mammograms, and anti-
smoking advice are among the most well-known preventive care services.
An effective way to improve the efficiency of a regional healthcare system with
limited recourses is to increase the number of people receiving preventive care ser-
vices, which has been an integral part of many healthcare reform programs within
the past two decades (Goldsmith 1989). However, the achievement of desired par-
ticipation level continues to be a challenge to many preventive care programs. For
example, every province in Canada had an organized program offering biennial
mammography screening by 2003. The average participation rate, however, reached
only 44 % in 2006, and none of the programs achieved the nationally established
target of 70 % participation (Public Health Agency of Canada 2006). Thus, maxi-
mizing the level of participation to preventive care programs has been a common
objective for policy makers.
In contrast to sick people who need urgent medical attention, people who seek
preventive services have more flexibility as to when and where to receive preventive
care services. The accessibility of the facilities is an important factor for the success
of a preventive care program. Zimmerman (1997) found through a survey that the
convenience of access to the facility was a very important factor in a client’s deci-
sion to have prostate cancer screening. Facione (1999) revealed that the perceptions
of lack of access to services were related to the decrease of mammography partici-
pation. Therefore, most studies in the literature focus on accessibility related to the
number, type, concentration, and location of preventive care facilities, as well as
transportation to services and availability of providers.
Preventive care has a number of differentiating features in comparison with other
types of care such as emergency, acute and long-term care. First, the number of peo-
ple who seek the services at the facility is not controlled by the policy maker, i.e.,
preventive care is a user-choice environment in terms of the allocation of clients
to facilities. Second, when people have to wait for a long time to receive the ser-
vices due to limited capacity, their willingness to participate in preventive programs
could decrease significantly. As a result, congestion at the facilities is a crucial fac-
tor considered in most studies in the literature. Third, it is essential to consider the
relationship between demand volume and quality of preventive care services. For
example, U.S. Food and Drug Administration (1999) requires a radiologist to inter-
pret at least 960 mammograms and a radiology technician to perform at least 200
mammograms in 24 months to retain their accreditations. As a result, in order to
ensure service quality, preventive care facilities cannot be operated unless the size
of their clientele fulfills a minimum workload requirement.
This chapter focuses on the problem of designing a network of preventive care
facilities so as to maximize accessibility and hence, the level of participation. We
provide an overview of the prevailing literature and report on a real-life application.
The remainder of the chapter is organized as follows. The next section presents
the common characteristics of the fundamental models. Sects. 9.3 and 9.4 discuss
9 Location Models for Preventive Care 225

two categories of models in detail, respectively. Sect. 9.5 presents an illustrative

example in order to provide the reader with a comparative understanding of the
alternative modelling frameworks. Sect. 9.6 discusses alternative solution methods.
Sect. 9.7 summarizes an application of location models for designing the network
of preventive breast cancer screening centres in Montreal. Several limitations of the
reviewed studies and extensions to the current literature are discussed in Sect. 9.8.
The final section concludes the chapter.

9.2 Overview of the Fundamental Models

As mentioned above, the objective is to maximize the level of participation by

determining the most appropriate configuration of the preventive healthcare facil-
ity network. This comprises determining the number of facilities to be established,
the location of each facility as well as its capacity. In order to ensure service quality,
preventive care facilities cannot be operated unless the size of their clientele fulfills
a minimum workload requirement. Therefore, the estimation of the user’s choices
as to which facilities they will patronize constitutes an essential component of the
model formulation.
In many cases, the problem involves the accreditation of a subset of the existing
facilities, rather than building a set of brand new facilities, so as to provide the
service. As a result, most of the existing studies do not incorporate a fixed setup
cost, and the minimum workload requirement actually works as the accreditation
criterion. In order to focus on the resource allocation strategy, the total number of
servers to be allocated to open facilities is often fixed. Consequently, the number
of facilities that could be opened is limited, mainly due to the minimum workload
requirement as well as the total number of available servers.
As mentioned earlier, the accessibility of the facilities is a crucial factor to influ-
ence people’s decision to participate. Several papers used travel distance or travel
time alone as a proxy for the accessibility of a facility, including Verter and Lapierre
(2002 ) and Zhang et al. (2012). In contrast, another stream of research also incor-
porates congestion into the model, such as Zhang et al. (2009 and 2010). They used
the total (travel, waiting, and service) time as a proxy for accessibility.
The user-choice concerning where to receive the service is also a critical part
of the model structure. There are typically two modeling alternatives in the litera-
ture to address the user choice: (1) the deterministic-choice model or optimal-choice
model; and (2) the probabilistic-choice model or stochastic-choice model. The for-
mer requires a fully-informed and rational set of clients, who patronize the facility
with the highest attractiveness. (In most location models, for example, it is common
to assume that each customer will seek services from the closest open facility.) The
latter is based on random utility theory and assumes that the user-choice depends
on a deterministic component comprising identified variables and a random compo-
nent that captures the impact of all unobserved factors. This amounts to assuming
that each client can visits any facility with a certain probability, which is increasing
226 V. Verter and Y. Zhang

with the identified attractiveness (utility) of the facility. Note that both modeling
alternatives could be realistic depending on the specific circumstances.

9.3 Deterministic-Choice Models

9.3.1 Distance-Based Models

To the best of our knowledge, Verter and Lapierre (2002) is the first paper on
location analysis in preventive care. With no capacity decisions considered, their
problem is therefore to locate an undetermined number of facilities to maximize
the level of participation subject to the minimum workload requirement. They used
travel distance or travel time alone as a proxy for the accessibility of a facility. With
the determination-choice modeling, this leads to the assumption that clients go to
the closest facility.
Let N (|N | = n) be a set of nodes, representing the neighborhoods of a city or
the population zones. There is a finite set of potential facility sites M(|M| = m).
Let S ⊂ M be a set of open facilities. The shortest travel time between nodes i and
j is denoted by tij . The fraction of clients residing at node i is denoted by hi . The
number of clients who require the service over the entire network denoted by λ, and
thus the number of clients from each node i is λhi . A facility cannot be established
at node j unless aj exceeds a minimum workload requirement denoted by Rmin .
Assume that the faction of clients at node i who would seek at site j (i.e., the
participation rate), denoted by aij , is a decreasing function of the travel time tij .
Further assume that this participation function is linear with an intercept Ai and a
slope γ, i.e.,
⎧
⎪ A
⎨ Ai − γ tij if tij ≤ i
aij = γ , i ∈ N, j ∈ M. (9.1)
⎪
⎩0 otherwise

Let Lij denote the set of alternative facility sites that are closer to population zone i
than site j, i.e., Lij = { : til < tij ,  ∈ M}.
Define two sets of binary decision variables:

1 if a facility is located at node j
yj =
0 otherwise;

1 if population zone i is served by a facility at site j
xij =
0 otherwise.

The problem is formulated as a mixed integer program (MIP), which is referred to

as Model 1:
9 Location Models for Preventive Care 227


n 
m
Max λ hi aij xij (9.2)
y,x
i=1 j =1


m
s.t. xij ≤ 1, i ∈ N (9.3)
j =1


n
λ hi aij xij ≥ Rmin yj , j ∈ M (9.4)
i=1

xij ≤ yj , i ∈ N, j ∈ M (9.5)
xij ≤ 1 − y ,  ∈ Lij , i ∈ N, j ∈ M (9.6)
yj ∈ {0,1}, xij ∈ {0,1}, i ∈ N, j ∈ M (9.7)

The objective (9.2) is to maximize the level of participation. Constraints (9.3) ensure
that at most one facility can serve a population zone. Constraints (9.4) impose the
minimum workload requirements on open facilities. Constraints (9.5) stipulate that
a population center can only be served by an open facility. Constraints (9.6) ensure
that each population center is assigned to the closest open facility. The integrality
requirements on the decision variables are expressed by constraints (9.7).
Zhang et al. (2012) presented a deterministic-choice model following Verter and
Lapierre (2002). They also assumed that clients would obtain the service from the
closest facility. In contrast, they extended the earlier model by considering: (1) con-
gestion at the facilities in a constraint of the model, i.e., a service level constraint;
and (2) capacity allocation among the open facilities subject to a fixed total number
of available servers.
Specifically, to incorporate congestion into the model, they assumed that the
number of clients who require the service over the entire network is Poisson dis-
tributed. They also assumed homogeneous servers and exponentially distributed
service times. Thus, each facility is modeled as an M/M/c queue, where c denotes
the number of servers. To guarantee a timely service, they assumed that the mean
waiting time at any facility cannot exceed a maximum acceptable level. Other types
of service level constraints can be applied in the model as well. Similar service level
constraints have been used in the location literature, such as Marianov and Serra
(2002), Wang et al. (2002), and Berman and Drezner (2006), as well as in practice.
Marianov et al. (2004) adopt similar constraints on the wait times in a model where
public service centres compete with private centres.
In summary, the problem is to maximize the total participation rate by finding
the optimal set of locations and the number of servers at each open facility, subject
to the service level constraint, the minimum workload requirement, and the given
number of total available servers. They formulated the problem as a mixed integer
programming problem, referred to as Model 2.
228 V. Verter and Y. Zhang

9.3.2 Time-Based Models

Zhang et al. (2009) extended Verter and Lapierre (2002) by incorporating conges-
tion at the facilities into the participation modeling. Specifically, using the expected
total (travel, waiting, and service) time as a proxy for accessibility of facilities, they
assumed that clients would patronize the facility with the minimum expected total
time. This is a major difference from Model 2 described above, where congestion
at the facilities is considered in a service level constraint. Moreover, they assumed
that clients at the same population zone would patronize the same facility. However,
this assumption may not be realistic in the context of preventive care, and it also
in fact prevents them from finding an equilibrium allocation of clients to facilities
in general. Zhang et al. (2010) further extended the previous model by (1) relaxing
the assumption that clients at the same population zone would patronize the same
facility, and (2) considering capacity allocation among the facilities given a fixed
total number of available servers.
In general, it is important to note the relationship between the arrival rate and
the congestion at a facility for this type of time-based models. As the congestion or
expected time at a facility is considered an attractiveness determinant, it influences
the participation rate, i.e., the arrival rate to a facility depends on its congestion. On
the other hand, the congestion, in turn, clearly depends on the arrival rate. Thus,
one must determine the the arrival rate and the congestion in the steady state. More-
over, since we consider a network of multiple facilities, it implies that one must
solve an equilibrium problem to determine the arrival rate and the congestion for
each facility. This is called a user-equilibrium problem in the operations research
literature.
Due to this nature, Zhang et al. (2010) formulated the problem as a bilevel model,
where the lower level model is to find a user-equilibrium satisfying that clients
choose the facility with the minimum expected total time, while the number, loca-
tions, and capacities of the facilities are determined at the upper level model. For
similarity, this section primarily introduces the lower level model, given the set of
open facilities S.
As each facility is modeled as an M/M/c queue, the general formula for the
mean waiting time is (Kleinrock 1975):
C(c, u) 1 1
W̄ (a, c) = + (9.8)
s μ(1 − ρ) μ
where a denotes the arrival rate and
c−1
ul
a a 1 − K(u) l=0
l!
u= ,ρ = , C(c, u) = , K(u) = . (9.9)
μ cμ 1 − ρK(u) c
ul
l!
l=0

Define two sets of decision variables for the lower level model:
9 Location Models for Preventive Care 229

sj = number of servers at facility j;

x  ij = fraction of clients from population node i who request service from
facility j.
Denote the arrival rate of clients at facility j by a j , i.e.,

n
a j = λ hi x ij , j ∈ S. (9.10)
i=1

Denote by T̄ij the average total time that clients from node i spend in order to receive
service at facility j. The average total time T̄ij comprises of two components: (1) the
travel time from node i to facility j denoted by tij ; and (2) the average time clients
spend at the facility possibly waiting and receiving service which we denote by
W̄ (a j , sj ), i.e.,

T̄ij = tij + W̄ (a j , sj ), i ∈ N, j ∈ S. (9.11)

Denote the total participation rate (fraction) at node i by pi , i.e.,


pi = x ij , i ∈ N . (9.12)
j ∈S

As clients choose the facility with the minimum expected total time, denote by Ti
this shortest time incurred by clients at node i. Assume that the total participation
rate pi at node i is a linear decreasing function of the shortest time Ti with an
intercept Ai and a slope γ  (the participation function), i.e.,

pi (Ti ) = Ai − γ  Ti , i ∈ N . (9.13)

Denote by Ti (pi ) the inverse participation function, i.e.,

Ai − pi
Ti (pi ) = , i ∈ N. (9.14)
γ
In fact, Ti (pi ) represents a threshold time, i.e., the total time clients at node iare
willing to incur for participating in the service, while the actual time incurred by
clients at node i to facility j is T̄ij .
As mentioned earlier, given the set of open facilities S and the number of servers
at each open facility sj , j ∈ S, the lower level model involves the clients’ facil-
ity choices so as to minimize their expected total time. This is a user-equilibrium
problem, and at equilibrium, no client wants to change her choice. This equilibrium
condition can be stated as: given S and sj , j ∈ S, for all pairs of (i, j ), i ∈ N, j ∈ S,

∗
= T i (pi∗ ) if xij ∗ > 0
T̄ij = tij + W̄ (aj , sj ) (9.15)
≥Ti (pi ∗ ) if xij ∗ = 0,

This condition (9.15) states that if there is a flow of clients from node i to facility
j, then the actual time incurred by clients at node i to facility j must be equal to the
230 V. Verter and Y. Zhang

threshold time (the longest time that clients would accept to go to the facility); and
if the actual time exceeds the threshold time, there is no flow.
To find xij ∗ in (9.15), we have to solve the following nonlinear complementarity
problem, where aj and pi are expressed by x ij ,
  Ai − x ij

n
j ∈S
tij + W̄ λ hi x ij , sj − ≥ 0, i ∈ N, j ∈ S
γ
i=1
⎡ ⎤
  A i − x ij
⎢ 
n
j ∈S ⎥
x ij ⎣tij + W̄ λ hi x ij , sj − ⎦ = 0, i ∈ N, j ∈ S
γ
i=1

x ij ≥ 0, i ∈ N, j ∈ S (9.16)

Alternatively, using vector-matrix notation, the complementarity problem (9.16) can

also be rewritten as a variational inequality (Zhang et al. 2010). The upper level
model with constraints (9.16) is called as a mathematical program with equilibrium
constraints (MPEC), which is referred to as Model 3. Please refer to Zhang et al.
(2010) for more details of the bilevel model.

9.4 Probabilistic-Choice Models

Probabilistic-choice models are represented by a variety of spatial interaction

models in the literature. The first spatial interaction model that represents the
probabilistic-choice behavior is by Huff (1962). Following Huff’s model in which
client utility is expressed by a simple gravity formula, a variety of studies in econo-
metrics and marketing focused on developing better representations of the utility
function, such as the multiplicative competitive interaction (MCI) model (Nakanishi
and Cooper 1974). Both Huff’s model and the MCI model were developed based
on aggregate flows between population zones and facilities. Another well-known
model, the multinomial logit (MNL) model (McFadden 1974), was originally pro-
posed at the disaggregate (individual) level for discrete choice modeling, but it has
been applied at the aggregate level as well (Gupta et al. 1996 ).
In general, the MNL model posits that the total utility of individual ichoosing
alternative j, j ∈ J, denoted by Uij , can be decomposed into two parts: (1) a deter-
ministic component denoted by Vij , which is usually explained by a linear function
of attributes or characteristics; and (2) a random component denoted by εij , which is
assumed to be independent, identically extreme value distributed,

Uij = Vij + εij . (9.17)

Denote the probability of individual i choosing alternative j by Pij . With utility-

maximizing, Pij is given by,
9 Location Models for Preventive Care 231

eVij
Pij = . (9.18)
eVip
p∈J

Note that the MNL model can only identify utility differences. Thus, one arbitrary
alternative has to be defined as the reference alternative, and the utility associated
with this alternative has to be set to an arbitrary value, usually normalized to 1.
In the context of preventive care, Gu et al. (2010) first applied a Huff-based loca-
tion model for allocating clients to the facilities. Zhang et al. (2012) presented a
probabilistic-choice model analogous to Model 2. They applied the MNL model for
the client allocation with the assumption that travel time (distance), tij , to be the
primary attractiveness determinant. In other words, Vij = −βtij , where β > 0 is
a parameter that denotes the sensitivity to travel time and can be estimated empiri-
cally. For simplicity, β is assumed to be identical for any population zone i or facility
j. Again, this section focuses only on the allocation part of the entire problem using
the MNL model.
To formulate the problem as a mathematical program, define two sets of decision
variables:

 1 if nodej has k or more servers
sj k =
0 otherwise,
qij = probability (or fraction) of clients at zone i who request the service from
facility j.

Note that qij is actually an auxiliary decision variable, and that s j  in fact denotes
the location decision at node j. According to the MNL model, define the choice of
not visiting any facility as the reference alternative, and normalize its utility to 1.
Now, qij can be expressed as:
e−βtij s j 
qij = , i ∈ N, j ∈ M. (9.19)
1+ e−βti s 1
∈M

This formulation ensures that clients must require service only from open facilities.
Expression (9.19) can be rewritten as:

qij + e−βti qij s 1 = e−βtij s j  , i ∈ N, j ∈ M. (9.20)
∈M

Since s l is a binary variable and qij is a continuous variable, expression (9.20) can
be linearized as follows by defining zij  as an artificial continuous variable,

qij + e−βti zij  = e−βtij s j  , i ∈ N, j ∈ M (9.21)
∈M
232 V. Verter and Y. Zhang

zij  ≤ qij , i ∈ N, j,  ∈ M (9.22)

zij  ≤ B1 s 1 , i ∈ N, j,  ∈ M (9.23)
zij  ≥ qij − B2 (1 − s 1 ), i ∈ N, j,  ∈ M (9.24)
zij  ≥ 0, i ∈ N, j,  ∈ M, (9.25)
where B1 and B2 denote two big numbers. For this problem, they can be set to 1, the
upper limit of qij .
The mathematical model with constraints (9.21) to (9.25) and others can then
be formulated as a mixed integer programming problem, which is referred to as
Model 4. Please refer to Zhang et al. (2012) for the detailed formulation.

9.5 An Illustrative Example

This section presents a small example to illustrate the differences between the above
models. Suppose there are 5 open facilities and 10 population zones. Table 9.1
gives the travel times tij in hours and the fractions of clients hi , and set λ = 10
clients/hour. The parameters used for Model 1 include Ai = 1.0 and γ = 1.4, those
for Model 3 include Ai = 1, γ  = 0.2, and μj = 2.5, and that for Model 4 is
β = 0.9. Note that we are only interested in the allocation results now, which are
independent of the other parameters. These parameter values are chosen so that the
total participation rates derived from these models are all approximately equal to
75 %.
Tables 9.2, 9.3, and 9.4 show the allocation results of Model 1, Model 3, and
Model 4, respectively. Although the total participation rates are close, the alloca-
tions of clients to the facilities are very different. In particular, the main difference
between Tables 9.2 and 9.3 is that clients at the same population zone, e.g., zones 4
and 9, may go to different facilities at equilibrium in Table 9.3. Note that although
they go to multiple facilities, the utilities visiting these facilities are identical at
equilibrium.
We also note that, in this example, the user-choice modeling does not have much
impact on the number of clients from each zone, whereas its impact on the number
of clients visiting each facility is significant. For instance, in Table 9.2, facilities 2
and 3 serve the smallest and largest number of clients, respectively; in contrast, the
two facilities interchange their roles in Table 9.4.

9.6 Solution Methodologies

As some of the models, e.g., Models 1, 2, and 4, are formulated as MIPs, they can
be solved directly using a professional solver, such as CPLEX. However, large-sized
problems may not be solved within a reasonable time. Therefore, the existing studies
9 Location Models for Preventive Care 233

Table 9.1 Travel times and fractions

Facility
Zone 1 2 3 4 5 Fraction
0.41 0.68 0.29 1.01 0.88 0.06
0.84 0.43 0.78 0.18 0.41 0.15
1.04 0.38 0.63 0.60 0.52 0.07
0.84 0.53 0.42 0.44 1.07 0.07
1.07 0.43 0.70 0.60 1.04 0.10
0.31 0.62 0.94 0.53 0.64 0.02
0.69 0.47 0.65 0.48 0.82 0.02
0.19 0.28 1.05 0.25 0.63 0.18
0.20 0.61 0.15 0.78 0.27 0.16
0.95 0.44 0.99 0.70 0.22 0.17

Table 9.2 Allocation result of model 1

Facility
Zone 1 2 3 4 5 Total
0 0 0.44 0 0 0.44
0 0 0 1.26 0 1.26
0 0.45 0 0 0 0.45
0 0 0.39 0 0 0.39
0 0.55 0 0 0 0.55
0.15 0 0 0 0 0.15
0 0.12 0 0 0 0.12
1.44 0 0 0 0 1.44
0 0 1.36 0 0 1.36
0 0 0 0 1.33 1.33
Total 1.59 1.12 2.19 1.26 1.33 7.48

often proposed improved exact solution methods or heuristic methods to reduce the
running time.
Verter and Lapierre (2002), for instance, presented two exact solution procedures
for Model 1. Both procedures are intended to take advantage of the problem struc-
ture during the solution process. Specifically, recall that constraints (9.6) impose
the closest facility allocations. However, majority of these constraints may not be
binding at the optimal solution, they therefore suggested to add only the necessary
constraints in each iteration of the procedure.
In contrast, other models with nonlinear formulations, e.g., Model 3, are dif-
ficult to be solved exactly. Hence, developing efficient heuristic methods are the
234 V. Verter and Y. Zhang

Table 9.3 Allocation result of model 3

Facility
Zone 1 2 3 4 5 Total
0 0 0.45 0 0 0.45
0 0 0 1.15 0 1.15
0 0.53 0 0 0 0.53
0 0.05 0.02 0.38 0 0.46
0 0.69 0 0 0 0.69
0.16 0 0 0 0 0.16
0 0.16 0 0 0 0.16
1.35 0 0 0 0 1.35
0 0 1.08 0 0.13 1.22
0 0 0 0 1.31 1.31
Total 1.51 1.43 1.56 1.53 1.43 7.48

Table 9.4 Allocation result of model 4

Facility
Zone 1 2 3 4 5 Total
0.11 0.08 0.12 0.06 0.07 0.45
0.17 0.25 0.18 0.31 0.25 1.15
0.07 0.13 0.10 0.11 0.11 0.52
0.08 0.11 0.12 0.12 0.06 0.49
0.10 0.18 0.14 0.16 0.10 0.68
0.04 0.03 0.02 0.03 0.03 0.16
0.03 0.04 0.03 0.04 0.03 0.17
0.34 0.32 0.15 0.32 0.23 1.36
0.29 0.20 0.31 0.17 0.27 1.25
0.19 0.30 0.18 0.23 0.36 1.25
Total 1.42 1.63 1.36 1.55 1.53 7.49

primary focus for these models. The heuristic methods are often based on the
location-allocation framework:
Given a set of facility locations and the associated capacities, identify client flows
from population zones to the facilities; determine the best set of locations and the
associated capacities.
In this framework, Alloc P serves as a sub-routine for Loc P. For any set of loca-
tions and the associated capacities, with Alloc P, the number of clients from each
population zone to each facility and the objective function value can be determined.
9 Location Models for Preventive Care 235

9.6.1 Allocation Algorithms

A variety of allocation algorithms have been proposed according to the nature of

the different problems. This section primarily introduces an allocation algorithm
developed by Zhang et al. (2010) for Model 3, which is to determine equilibrium
flows given a set of facilities S. For the ease of exposition, consider all the facilities
with one server, i.e., sj = 1, j ∈ S.
When sj = 1, each open facility becomes an M/M/1 queuing system. Thus, the
mean waiting time formula (9.8) reduces to
1
W̄ (a j ) = , j ∈ S. (9.26)
μ − a j
Zhang et al. (2009) initially studied this allocation problem and proposed an allo-
cation heuristic for Alloc P. Furthermore, Zhang et al. (2010) developed an exact
allocation method. Specifically, they showed that the nonlinear complementarity
problem (9.16) can be reformulated as the following optimization problem,
 
  n n  Ai

Min− ln μ − λ hi x ij + λ hi tij −  x ij
γ
j ∈S i=1 i=1 j ∈S
⎛ ⎞2
 n
1 ⎝  ⎠
+ x ij (9.27)
2γ 
i=1 j ∈S

s.t. x ij ≥ 0, i ∈ N, j ∈ S.
They proved that the above optimization problem is convex. Since this is almost an
unconstrained problem, it can be solved by a variety of optimization methods, such
as the gradient projection method (Kelley 1999). They further proved the existence
of the user-equilibrium, whereas the equilibrium in general may not be unique.
They showed that the above allocation method can directly be applied to find
exact equilibrium flows when sj ≤ 2, as the mean waiting time formula for an
M/M/2 queue has a similar structure to expression (9.26). However, the mean
waiting time formula becomes more complex when sj > 2. To make the alloca-
tion method applicable, they developed an approximation for the mean waiting time
formula when sj > 2, which also has a similar structure to expression (9.26). The
above allocation method can then be used to find approximate equilibrium flows.
Refer to Zhang et al. (2010) for details of the approximation.

9.6.2 Location Algorithms

Many heuristic algorithms have been developed for Loc P. Zhang et al. (2009)
described four location heuristics for location decisions only, i.e., no capacity deci-
sions considered, including a greedy-type heuristic and three meta-heuristics based
236 V. Verter and Y. Zhang

on a probabilistic adaptive search. Each of these heuristics applies three basic

neighborhood move procedures as follows.
Remove Procedure The procedure starts with all potential facilities open. In each
iteration, it removes one facility at a time until feasibility is obtained. The facility to
be removed is chosen so that the total participation served by the remaining facilities
is maximized.

Add Procedure The procedure is the opposite of Remove Procedure, starting with
a feasible facility set. In each iteration, it adds a new facility. The new facility is
chosen so that the total participation served by the located facilities is maximized,
while maintaining feasibility.

Exchange Procedure This procedure attempts to improve a given feasible solution

by swapping a facility in the current solution with a potential facility that is not
currently open and then executing Add Procedure.
The greedy-type heuristic applies Remove, Add, and Exchange procedures
in succession. Although it is fast to produce a feasible solution, it often results
in a local optimum. In an effort to overcome this disadvantage, they also intro-
duced three probabilistic adaptive search heuristics which are meta-heuristics. These
meta-heuristics run the Remove, Add, and Exchange procedures in succession
repeatedly. In each procedure, the heuristic is based on the randomized choice
of available alternatives (alternatives here can represent facilities to be removed,
added, or facility pairs to be swapped). The probability of selecting an alternative is
proportional to the change in total participation once this alternative is selected.
Zhang et al. (2012) implemented a genetic algorithm to determine the best
set of locations. They further compared the performances of one probabilistic
adaptive search heuristic above and the genetic algorithm. In their computational
experiments, they showed that the former exhibits slightly better accuracy but less
reliability than the genetic algorithm.
Zhang et al. (2010) developed a tabu search procedure to make both location
and capacity decisions for (Loc P), subject to a given number of available servers.
Assuming that a facility can be allocated to at most K servers, each potential facility
is divided into K pseudo-facilities. Similar to the genetic algorithm above, a solution
is then composed of a series of binary numbers that correspond to each pseudo-
facility. The tabu search heuristic starts from an initial feasible solution. Then, in
each iteration, it applies Remove and Add procedures in succession. The procedure
repeats until a stopping criterion is reached. When neighborhood moves are selected,
tabu restrictions are used to prevent moving back to previously investigated solu-
tions. Define a tabu list in which each value is associated with a pseudo-facility to
represent its tabu status. Once removed or added to the set of open pseudo-facilities,
a node is classified as tabu with a certain length, which represents the number of
iterations in which the node typically will not be selected for removing or adding.
However, even for a tabu node, it can still be selected for adding, if an aspiration
criterion is satisfied. A typical criterion states that if a move produces a feasible
9 Location Models for Preventive Care 237

solution which is better than the best known feasible solution, then the tabu status is
disregarded and the move is executed.

9.7 A Real-Life Application

Verter and Lapierre (2002) adopted their distance-based model (Model 1) in work-
ing with the Quebec Ministry of Health for determining the breast cancer screening
centers in Montreal that should be included in the Ministry’s program to subsi-
dize mammograms for women between the ages of 50 and 69. At the time, there
were 194,475 women in Montreal in this age group and 36 existing facilities with
mammographic equipment. The problem was to determine which facilities to be
accredited so as to maximize participation to the program. The Ministry made a pol-
icy decision to require a minimum of 4000 mammographies per year for facilities to
be accredited. The Ministry’s preference was to make their decisions based on the
assumption that women travel a maximum of 7 miles to take mammograms. There
were 497 population centers in representing spatial distribution of the potential
clients, and a maximum participation rate was assumed to be 0.95.
Verter and Lapierre (2002) formulated the problem using Model 1. They have
observed that the number of closest facility assignment constraints (9.6) proliferates
with the problem size. Since the majority of these constraints may not be binding
at the optimal solution, they experimented with alternative solution procedures that
add only the necessary constraints at each iteration. Indeed, further computational
experiments showed that the way the closest facility assignment constraints are for-
mulated has a significant impact on the computational effort required to solve large
scale problems. In tackling the Montreal application, they were able to reduce the
computational time significantly by replacing (9.6) by constraints of the form sug-
gested by Rojeski and ReVelle (1970). The basic premise of this formulation is as
follows: each population center should be assigned to its closest alternative site, if
there is a facility at that site. The assignment should be to the second closest alter-
native site, if the closest site is not open and there is a facility at the second closest
site. The argument can be extended to third, fourth etc. closest alternative sites.
The optimal solution of Verter and Lapierre (2002) suggested accreditation of 17
of the 36 existing mammography facilities. The maximum expected participation
would be around 50 %. They presented the solution to the Ministry, with the recom-
mendation that a follow up study is performed to better understand the behavioural
aspects of the clients’ participation to the preventive breast cancer screening pro-
grams. Under time pressure to implement a solution, however, the Ministry went
ahead and approached the proposed 17 mammography centres to recruit them to the
program. A few of the centres have refused to be part of the subsidy program, and
hence the Ministry accredited some additional centres, which were willing to work
with the subsidy system albeit they were not part of the optimal solution.
Unfortunately, significant wait times were observed at some of the facilities after
the system was running. One of the reasons for this is that Model 1, based on
238 V. Verter and Y. Zhang

distance only, does not take congestion into consideration. This motivated several
follow-up studies (Zhang et al. 2009 and 2010) concerning the incorporation of
congestion into the models. For example, Zhang et al. (2009) demonstrated that
ignoring congestion would lead to significant overestimation of the participation
rate, and that increasing the capacity of the accredited facilities would be a better
policy than accrediting a larger number of small facilities. They also suggested, to
some extent, to allocate resources on health promotion projects targeted to potential
clients rather than capacity expansion. These latter recommendations have not been
implemented due to the budgetary pressures in the health sector.
Turning to the big picture, note that breast cancer is the most common cancer
diagnosed in Canadian women and is second only to lung cancer as the most com-
mon cause of cancer deaths among women. Annually, around 25,000 women are
diagnosed with this condition and unfortunately around 5000 lose the battle to breast
cancer. Just over 50 % of all breast cancer cases occur in women between the ages of
50 and 69. Accordingly, screening recommendations in Canada include a biennial
mammogram of the breast and clinical breast examination in asymptomatic women
in this age group. It is important that the breast cancer screening programs offer
timely service and capture a significant proportion of the women in the target age
group. The national target in Canada is 70 %.

9.8 Extensions

From the modeling perspective, all of the studies introduced above considered
facility accessibility, such as distance and congestion, as the main determinant for
participation. None of these studies provides empirical support for their assump-
tions. However, past empirical studies have suggested that factors other than
accessibility may play a more important role in the decisions (Rimer et al. 1989; Kee
et al. 1992; Munn 1993). Moreover, the models taking congestion into considera-
tion represent walk-in facilities, where appointments are not needed. This enables
the effective use of queuing models. Many preventive care services, including flu
shots, anti-smoking advice, drug and alcohol abuse, nutrition advice, are offered in
walk-in facilities. Many other preventive care programs, such as preventive cancer
screening programs, however, are offered by the use of appointment systems. One
of the main challenge concerning appointment systems is the fact that wait times for
appointments are often not commensurate with the travel and service times at the
facilities.
Recently, aiming at optimizing the configuration of a preventive cancer screen-
ing facility network, Kucukyazici et al. (2014) proposed a novel methodology that
includes a variety of empirical and analytical methods. They first identified potential
facility attributes that affect the client choice of preventive cancer screening facilities
based on a focus group meeting and interviews. They then used stated preference
discrete choice modeling (SPDCM) to model clients’ preference over the identified
attributes. In other worlds, they designed a survey and investigated the client choice
9 Location Models for Preventive Care 239

from survey data with discrete choice modeling. Results from the SPDCM analy-
sis show that nursing staff’s manner and knowledge regarding the target disease,
waiting time for an appointment, and parking availability are the most influential
attributes, among other significant factors such as travel time, communication dur-
ing the screening process, and waiting time for result. Kucukyazici et al. (2014)
also conducted latent class analysis to explore the existence of client heterogene-
ity. Further, they made use of the results from the econometric analysis to develop
an empirically based simulation model, and incorporated meta-heuristic algorithms
into the simulation. This simulation optimization approach was used to optimize the
configuration of a facility network, by relocating a fixed number of servers within
the network. To the best of our knowledge, this is the first study that integrates
empirical analysis and simulation optimization to provide valuable insights into how
to improve preventive cancer screening system performance through network design
and configuration.
Aboolian et al. (2014) developed an ε-optimal algorithm for a variant of Model
3. In their model, the service rate at each facility is a decision variable rather than
the number of servers. They reformulated the nonlinear problem as an MIP. The
idea of their reformulation is to replace the decision variables, the service rates,
by the waiting times and then to approximate the waiting time functions using the
TLA method developed in Aboolian et al. (2007). Aboolian et al. (2014) present a
realistic example based on the hospital network of Toronto in order to highlight the
potential use of their model in gathering policy insights.
Although Model 4 is formulated as a mixed integer programming problem,
medium- or large-sized instances cannot be solved to optimality within a reasonable
time. Haase and Muller (2013) presented an alternative formulation and showed
that medium-sized instances can be easily solved to optimality, or at least close to
optimality, by a professional solver in a reasonable time. They further proposed an
approach to derive a tight lower bound to the problem.

9.9 Conclusion

This chapter reviews several studies that apply location analysis for designing pre-
ventive care programs. The generic problem is to design a preventive care facility
network so as to maximize the level of participation. The number of facilities to be
established, the location of each facility, and the capacity at each facility are the main
determinants of the configuration of a facility network. The existing models in the
literature typically address some of the differentiating features of preventive care,
such as elastic participation, congestion, and the user-choice environment. Acces-
sibility of facilities is usually considered to influence the participation rate. Some
papers used travel distance or travel time alone as a proxy for the accessibility, while
others also take congestion into consideration. Moreover, the deterministic-choice
and probabilistic-choice models are two alternatives in the literature to address the
client choice on where to receive the service. In terms of solution methodology, both
240 V. Verter and Y. Zhang

exact and heuristic solution methods have been proposed to solve the problems. Evi-
dently, this is a fledgling research stream and as discussed in the previous section
there are a number of ways that the state of the art can be improved.

References

Aboolian R, Berman O, Verter V (2014) Maximal accessibility network design in the public sector.
Forthcoming in Transportation Science
Berman O, Drezner Z (2006) Location of congested capacitated facilities with distance sensitive
demand. IIE Trans 38:213–231
Facione NC (1999) Breast cancer screening in relation to access to health services. Oncol Nurs
Forum 26:689–696
Falkenheimer SA (2004) The adequacy of preventive health care: does the health care provider mat-
ter? http://www.cbhd.org/content/adequacy-preventive-health-care-does-health-care-provider-
matter. Accessed 13 April 2015
Goldsmith J (1989) A radical prescription for hospitals. Harvard Bus Rev 67:104–111
Gu W, Wang X, McGregor SE (2010) Optimization of preventive health care facility locations. Int
J Health Geogr 9:17–32
Gupta S, Chintagunta PK, Kaul A, Wittink DR (1996) Do household scanner data provide
representative inferences from brand choices: a comparison with store data. J Mark Res
33:383–398
Haase K, Muller S (2013) Insights into clients’ choice in preventive health care facility location
planning. Working Paper. University of Hamburg
Health Canada (2007) Mammography. http://www.hc-sc.gc.ca/hl-vs/iyh-vsv/med/mammog-eng
.php. Accessed 13 April 2015
Huff DL (1962) Determination of intra-urban retail trade areas. Real Estate Research Program,
UCLA
Kee F, Telford AM, Donaghy P, O’doherty A (1992) Attitude or access: reasons for not attending
mammography in northern Ireland. Eur J Cancer Prev 1/4:311–316
Kelley CT (1999) Iterative methods for optimization. Frontiers in applied mathematics. SIAM,
Raleigh
Kleinrock L (1975) Queueing system I: theory. Wiley, New York
Kucukyazici B, Song L, Verter V, Zhang Y (2014) Incorporating client choice into facility network
design of preventive cancer screening programs. Working Paper. McGill University
Marianov V, Serra D (2002) Location-allocation of multiple-server service centers with constrained
queues or waiting times. Ann Oper Res 111:35–50
Marianov V, Rios M, Taborga P (2004) Finding locations for public service centres that compete
with private centres: effects of congestion. Pap Reg Sci 83:631–648
McFadden D (1974) Conditional logit analysis of quantitative choice behavior. In: Zarembkar P
(ed) Frontiers in economics. Academic Press, New York
Munn EM (1993) Nonparticipation in mammography screening: apathy, anxiety or cost? N Z Med
J 106/959:284–286
Nakanishi M, Cooper LG (1974) Parameter estimation for a multiplicative competitive interaction
model: least squares approach. J Market Res 11(3):303–311
Public Health Agency of Canada (2006) Organized breast cancer screening programs in
Canada—Report on program performance in 2005 and 2006. http://www.phac-aspc.gc.ca/
cd-mc/publications/cancer/obcsp-podcs05/index-eng.php Accessed 14 July 2015.
Rimer BK, Keintz MK, Kessler HB, Engstrom PF, Rosan JR (1989) Why women resist screening
mammography: patient-related barriers. Radiology 172/1:243–246
Rojeski P, ReVelle CS (1970) Central facilities location under an investment constraint. Geogr
Anal 2:343–360
9 Location Models for Preventive Care 241

Verter V, Lapierre SD (2002) Location of preventive health care facilities. Ann Oper Res 110:
123–132
Walker K (1977) Current issues in the provision of health care services. J Consum Aff 11:52–62
Wang Q, Batta R, Rump CM (2002) Algorithms for a facility location problem with stochastic
customer demand and immobile servers. Ann Oper Res 111:17–34
World Health Organisation (2002) Integrating prevention into health care. http://apps.who.int/
bookorders/anglais/detart1.jsp?codlan=1&codcol=15&codcch=739. Accessed 13 April 2015
Zhang Y, Berman O, Verter V (2009) Incorporating congestion in preventive healthcare facility
network design. Eur J Oper Res 198:922–935
Zhang Y, Berman O, Marcotte P, Verter V (2010) A bilevel model for preventive healthcare network
design with congestion. IIE Trans 42/12:865–880
Zhang Y, Berman O, Verter V (2012) The impact of client choice on preventive healthcare facility
network design. OR Spectr 34:349–370
Zimmerman S (1997) Factors influencing hispanic participation in prostate cancer screening. Oncol
Nurs Forum 24:499–504
Chapter 10
Modeling the Location of Retail Facilities: An
Application to the Postal Service

Anthony M. Yezer and James W. Gillula

10.1 Introduction

There are many well-developed theoretical models of facility location for various
types of retail activity. Given a spatial distribution of consumers with known demand
curves and a function that transforms distance into transportation cost, it is possible
to specify the spatial demand curve facing a firm located at any point in economic
space. A plant level cost function, usually a combination of fixed and variable cost,
must be added to the model along with assumptions regarding the nature of spatial
competition from producers of close substitutes. Once this is done, the model can
be solved, and the revenue and cost consequences of alternative facility location and
marketing strategies can be computed in a fairly straightforward fashion.
The model developed here is based on a well established theoretical literature
that began with Launhardt’s (1885) model of the spatial demand curve. Consumers
spread over space face two “prices” when they make a consumption decision. The
first is the price charged for the good or service at the facility where it is sold (this
is the free on board (FOB) or mill price). The second is transportation cost to and
from the facility. Therefore, the cost of acquisition rises with distance to the facility
and accordingly the quantity demanded falls. Subsequently, Reilly (1931) called this
the “law of retail gravitation” because demand declines with distance, although the
effect of distance on transportation cost is generally found to be either linear or con-
cave. Neidercorn and Becholod (1972) showed that the spatial demand curve could
be derived from consumers maximizing a utility function. These insights regarding
spatial demand formed the basis for modern location theory whose historical devel-
opment is reviewed nicely by Beckmann and Thisse (1986). The specific theoretical
issues involved in modeling facility location are considered in Ye and Yezer (1992).

A. M. Yezer ()
George Washington University, Washinton, D.C., USA
e-mail: [email protected]
J. W. Gillula
IHS Economics, Washinton, D.C., USA
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 243

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_10
244 A. M. Yezer and J. W. Gillula

For a recent statement of the theoretical problem of spatial competition see Braid
(2011, 2012).
Going from the available theory to an operational empirical model should not
be difficult given sufficient data, particularly micro data on individual household
demand and location. However, privacy concerns generally make public disclosure
of such data problematic. Accordingly, while there is little doubt that many fine
models of spatial competition for retail activities exist, they have been developed
for the private use of individual firms. Individual retailers can collect information
on the purchasing patterns and location of customers by providing “buyer loyalty”
discount cards. Public access to these models is limited. Furthermore the extreme
requirements for micro data needed to estimate and implement these models limit
the ability to use them as the basis for model calibration. Data relating consumption
behavior of households, including both what is purchased and precisely where it is
purchased, to their individual characteristics is seldom available to researchers or
practitioners.
Instead, researchers must generally be content with information on facility sales
and characteristics of the market area in which the facility is located as a basis for
estimating spatial demand functions. The study described in this Chapter uses this
readily available data to formulate a rigorous model of a spatial retail market that
is capable of serving as the basis for a business model of the costs and benefits of
alternative patterns of facility size, type, and location. The model can simulate a full
range of economic outcomes of alternative patterns of production that can be used
for both positive and normative economic analysis. The method proposed here ac-
commodates the possibility that demand is based on both the residential location and
the workplace of potential consumers. Perhaps the closest model, at least in terms
of spatial demand modeling, to the problem considered here is Davis’ (2006) effort
to characterize spatial competition among movie theaters. The specific application
here is to retail postal services, including facilities providing, stamps, package ser-
vices, post office boxes, and special services. The fact that the United States Postal
Service (USPS) is a spatial monopolist is an aid in constructing the model, but the
techniques developed here could be applied to other forms of spatial competition.
The model is complicated because the USPS has two different modes of delivering
retail postal services, retail postal stores (PSs) and limited-service contract postal
units (CPUs). The scale of the model is massive, applying to the entire contiguous
48 U.S. states.1

10.2 General Statement of the Postal Location Problem

The USPS retail facility location problem is a special version of standard problems
where facilities subject to increasing returns to scale serve spatially distributed con-
sumers. Based on the location of their residence or place of work as well as their

1 The retail postal store is commonly known as a main post office, branch, or station. It provides a

full range of postal services.

10 Modeling the Location of Retail Facilities 245

income and personal characteristics, individuals demand services provided by a fa-

cility. The facility has production costs that increase at a decreasing rate with output,
i.e., increasing returns to scale at the facility level. This suggests producing in a
small number of large facilities to minimize production cost. But consumers must
pay both the cost of production and the cost of accessing the facility, which suggests
producing in a large number of small facilities to minimize transportation cost. The
optimal facility location problem is to select the size and spacing of facilities which
maximizes the net willingness of consumers to pay for postal services, i.e., which
maximizes the net benefit produced by the location of retail facilities.
The general statement of the problem provided below could be applied to a vari-
ety of facility location problems. The price of services is fixed at all plants so that it
is possible to write total revenue per square mile, R, in the following form:
R = PQ = f (D, r), (10.1)
where P is the fixed price per unit service, Q is the quantity of services provided, D
is the density of postal customers (households and firms) per square mile, r is the
market radius of the store, and the function f (D, r) is an empirical relation to be
determined using USPS data on revenue and market characteristics. It is anticipated
that ∂f/∂D = fD > 0 because greater density of customers yields greater demand
density and that fr < 0 because a longer market radius makes access to the store
more expensive.
Retail postal services at a store are produced through “windows.” There is a fun-
damental production relation between the number of required windows and revenue
that is given by2
W = g(R). (10.2)
Note that the equations in this section measure relations per unit area so that R
is revenue per square mile and W is windows per square mile. Based on a priori
expectations, window requirements should increase at a decreasing rate in revenue.
Estimates of g(R) using USPS data on revenue and windows of facilities serving
different market areas confirmed the expectation, i.e., that the first derivative of g(R),
gR > 0 and its second derivative gRR < 0. There is a capacity constraint on the ability
to deliver retail services and generate revenue from a given number of windows.
However, there is an added complication because customers will not wait in long
lines for most postal services. This customer impatience is one reason that empirical
measurement of the relation between windows and revenue is necessary.
There is a relation between the number of windows in a postal store and the
required labor, so that labor per square mile of market area L is given by
L = h(W ). (10.3)

2 For facilities that are connected to the Point Of Sale (POS) system, a window is a register that
is generating significant amounts of revenue. The ability to observe revenue generation at the indi-
vidual register level is a major advantage in the calibration of cost functions. This will be discussed
in some detail below.
246 A. M. Yezer and J. W. Gillula

Windows generate staffing requirements and, based on USPS data, there is a clear
relation between the number of windows operating and the number of USPS person-
nel in the facility so that hW > 0. Once again the fact that the analysis is conducted
in terms of labor and windows per unit market area has no effect on the sign of the
relation between the two variables.3
Similarly, there is a relation between interior space per square mile of market
area S, at a store and the number of windows given by

S = i(W ), (10.4)

where i(W) must also be determined empirically using USPS data and design cri-
teria. But it is clear that iW > 0 and iWW < 0 because more space is required to
accommodate the additional windows.
Translating clerical labor and facility space into USPS costs is fairly straightfor-
ward. Labor cost per square mile, CL , is the product of the number of workers per
square mile and the earnings per worker, E, or CL = EL. The cost of space to serve
demand generated by each square mile of market area will depend on space per
square mile and on the size of the market area, i.e., on r, viz.,

CS = j (S, r) (10.5)

where: CS is the cost per square mile of market area, and j(S, r) is a facility rent
equation that gives the relation between the space per square mile of market and
the facility rent per square mile. Unlike labor cost, space rental costs vary with
location. Consistent with real estate literature, estimates from USPS data indicate
that jS > 0, jSS < 0, and jr < 0 because additional space costs more but larger
facilities that serve a larger market radius cost less per square foot.
One function of facility location models is normative, i.e., to solve for the optimal
spacing between facilities. In most cases, this maximization problem is straightfor-
ward once a criterion for judging optimal facility provision is selected.4 One stan-
dard choice is maximization of the surplus of revenue over cost. Note that this is not
a monopoly solution because price is fixed and the USPS is a non-profit producer. In-
stead this corresponds to a welfare criterion for consumers who ultimately must pay
a higher price for the cost of additional facilities.5 Of course, the model also func-
tions in a positive, i.e., non-normative, mode because it can simulate the implications
of changes in location and other operating decisions for revenue, cost, etc.
As in most facility location models, there is a revenue recognition problem. Be-
cause USPS is producing more than retail services all of the revenue should not be

3 Employees who staff windows have other duties. Some of these are directly related to revenue
generation, handling incoming mail. Others may involve sorting mail for delivery. These issues
will be discussed later in this chapter.
4 For an excellent discussion of the literature on location theory see Beckmann and Thisse (1986)

or Beckmann (1999).
5 For a specific discussion of the public facility location problem in continuous space see Ye and

Yezer (1992, 1993).

10 Modeling the Location of Retail Facilities 247

attributed to retail operations. Revenues collected at the retail level must also cover
costs of mail handling, processing, transportation, and delivery, hereafter referred to
as “mailing cost.” These operations impose costs on the USPS that are a substantial
fraction of total revenue. Mailing costs have been estimated as 75 % of USPS oper-
ating cost. It is not at all clear that this average cost figure should be deducted from
revenue for purposes of the analysis conducted here, particularly given that marginal
cost is likely well below average cost. At this point, a fraction, 0 < (1 − ) < 1 of
revenue is attributed to covering mailing cost, which leaves a fraction  of revenue
to be counted as net revenue associated with retail services.
Based on Eqs. 10.1–10.5 above, the welfare criterion may be stated as
!
ω = R − CL − CS = f (D, r) − Eh(W ) − j (i(w), r)
(10.6)
ω = f (D, r) − Eh(g(f (D, r))) − j (i(g(f (D, r))), r)

where ω is surplus per unit area and maximizing ω maximizes surplus for the entire
system. The important insight from Eq. (10.6) is that, given the density of cus-
tomers that determines demand per unit area, welfare maximization is based on the
selection of postal store spacing at intervals of 2r. This also determines facility size
because D and r determine postal store revenue from Eq. (10.1), R = f (D, r). Once
revenue is determined, all the other operating parameters of the model such as num-
ber of windows, facility size, and workers per facility follow recursively. Thus, the
facility location problem can be solved by choosing the appropriate plant spacing,
r, and tracing the implications of this for all other postal store characteristics using
the model.
For given demand density, surplus is maximized by setting the derivative of
(10.6) with respect to r equal to zero, i.e.,
dω
= ψfr − EhW gR fr − jS iW gR fr − jr (10.7)
dr
Equation (10.7) can be solved to determine the optimal market radius between stores
in areas where demand density is given. Recall from the above discussion that fr <
0, gR > 0, hW > 0, iW > 0, jr < 0, and jS > 0, which implies that the first term on
the right side of (10.7) is negative, while the other terms are positive. As expected,
increasing r tends to have a negative effect on surplus per unit area due to the fall
in revenue as average demand per square mile falls, but there is a compensating rise
in welfare as larger sales per postal store lowers space cost per square foot. It is
apparent that labor cost per square mile of market area falls as revenue falls. Clearly
the necessary condition to maximize welfare in Eq. (10.7) is a function of demand
density D. In these models, the effects of demand density are complex because a rise
in demand density tends to increase costs of additional r because the fall in revenue
with added market area is larger but the fall in cost is larger also.
248 A. M. Yezer and J. W. Gillula

10.3 Specific Statement of the Postal Location Problem

This section shows how the standard setup for a facility location problem can be
adapted to the specific retail location issues confronting the USPS. Much of this
adaptation is necessary because of the limited data availability, and other aspects
are necessitated because it is not appropriate to assume that current operations of
the USPS are technically efficient, i.e., that current facility location, size, and oper-
ation are designed to minimize operating cost. Put another way, current operating
costs of the USPS are not based on facility sizes and locations that minimize cost
or maximize net revenue. This means that the cost function for providing postal ser-
vices must be constructed and calibrated based on first principles rather than on the
actual costs observed at different facilities.
Going from the general statement of the postal store location problem in the pre-
vious section to a specific model that can be calibrated and applied requires that
the general functional forms be replaced by specific equations. There are two steps
in this process. First, functions must be replaced with equations that can be justi-
fied in terms of economic theory. Second, the parameters of the equations must be
estimated using USPS data on revenue and facility characteristics.
This section discusses the first of these tasks. As suggested in the general solu-
tion, there are several fundamental relations to be understood: the determinants of
revenue per square mile ( f (D, r)), and then the relations between revenue and win-
dows (g(R)), space and windows (i(W)), employees and windows (h(W)), and space
and rents ( j(S, r)). As noted above, each of these relations should be based on a fa-
cility operating at or near capacity. In this section, each of these issues is discussed
in turn.

10.3.1 Determinants of Revenue Per Square Mile of Market Area

Demand for retail postal facilities is based on local employment and residential pop-
ulation. Following standard practice in the literature discussed in Beckmann (1999),
assume that demand for retail postal services is given by

q = α − β(P + tx) (10.8)

where q is the quantity of services used by an individual household or firm, P is the

price of postal services, t is the cost of transportation to and from the facility, x is the
distance to the nearest facility and α > 0 and β < 0 are parameters. This is a simple
linear demand curve where the total cost of the services is the sum of the price and
the transportation cost, tx, to access the facility. The problem is simplified by the
fact that P is known to be constant across facilities.
Assuming that facilities serve a circular market area of radius r so that the market
area is πr2 = A, the distance between facilities is 2r, and that demand for postal
services comes from households or firms (the illustrative equation could apply to
10 Modeling the Location of Retail Facilities 249

either group although α, β, and t would be different for households and firms), total
demand at any facility is given by
" r
f (D, r)
QA = = 2π xD(α − β(P + tx))dx, (10.9)
P 0

where Q is revenue per square mile, QA is total sales per facility, and D equals the
density of households or firms measured as households or employees per square
mile. There is a direct connection between Eqs. (10.9) and (10.1) above, and
multiplying by P gives an expression for total revenue that is equivalent to (10.1):
" r
RA = PQA = Pf (D, r) = P 2π xD(α − β(P + tx))dx (10.10)
0

Evaluating (10.10) we find that total revenue per facility can be written as:
#    $
a bP brt
RA = Pf (D, r) = P 2π r 2 D − − (10.11)
2 2 3

Recognizing that P is a constant, this equation states that total revenue of a facility
is the product of the total number of customers in the market area (πr2 D) and a
constant, 2P[(a/2) − bP/2)], and a constant multiplied by the radius of the facility
service area, 2P(− bt/3)r. It is convenient to write Eq. (10.11) as
#  $
2bP t
RA = Pf (D, r) = (π r 2 D) P (a − bP ) − r (10.12)
3
or

RA = Pf (D, r) = (π r 2 D)[θ − τ r], (10.13)

where πr 2 D = N is the total number of households or employees (i.e. potential

customers) in the service area of the facility, θ is the constant (aP − bP2 ), and τ is an-
other constant, (2Pbt/3) which includes the effects of transportation cost on demand.
Note that the effect of increasing market radius on revenue depends on two effects.
First is the negative effect due to the τr term as higher transportation cost cuts de-
mand by effectively raising the cost of accessing retail services. As suggested above,
fr < 0 as average revenue per square mile falls when r rises. However, raising r in-
creases market area and hence raises the number of customers (dN/dr = 2π rD > 0).
The second effect on f (D,r)A dominates, and total revenue per facility rises when
r increases.6 In the next section, Eq. (10.13) will be shown to provide the basis for
the empirical estimation of the facility revenue equation that is vital to the exercise
conducted here.

6 There are technical reasons for this result that need not concern us here.
250 A. M. Yezer and J. W. Gillula

10.3.2 Relation Between Revenue and Retail Input Needs

Postal stores produce retail postal services principally using inputs of labor and
space.7 The challenge is to relate revenue generated, principally walk-in revenue,
at stores to required inputs of labor and space that can then be priced to form a
cost function. Fortunately, the USPS collects revenue data through the point of sale
(POS) system of wired registers in a substantial number of retail stores. It is possible
to observe the relation between the number of registers in a store and the amount of
revenue. There are obvious technical reasons to expect that a given quality of service
can only be maintained in a postal store experiencing additional walk-in demand by
increasing the number of operating registers. In the POS data some registers are
clearly dedicated to retail services and are termed “front” and others are “back,” al-
though they sometimes take in significant amounts of revenue. Some front registers
are located at formal front windows and others at counters. In the POS data it is pos-
sible to identify all registers that perform a significant revenue-generating function
and, as noted above, these are termed windows. If W denotes windows per square
mile, then total windows for a facility with market area A is WA. It is expected that
there should be a relation between total postal store revenue and required windows
that takes the following specific functional form

WA = κ(RA)λ , (10.14)

where κ is a constant approximately equal to the reciprocal of the annual revenue ex-
pected from operation of a single window and λ is the relation between additional
revenue and the need for added windows.8 It is expected that λ will be slightly
less than unity because adding more registers permits specialization in services and
should raise revenue per window slightly. Note that (10.14) expresses the relation
between annual revenue per store and number of windows per store assuming that
the windows are kept busy. Clearly, the causality runs from revenue to windows.
Doubling windows should have no effect on revenue unless the previous number of
windows was inadequate to deal with demand and significant lines and wait times
were experienced by patrons. Sometimes window services are used by customers
who are not generating revenue. For purposes of this model, it is assumed that these
activities are proportional to revenue generation and simply reduce the potential of
windows to generate revenue. It is possible in the USPS data to find cases in which
facilities with several windows generate very little revenue. It is important to eval-
uate (10.14) based on facilities where the registers are operating near capacity and,
given technical efficiency cannot be assumed, this creates an empirical challenge for
the model.

7 The emphasis here is on operating inputs and operating costs, and general and administrative
costs are ignored.
8 Equation (10.14) relates to the function in (10.2) as W = g(R) = κ(RA)λ /A
10 Modeling the Location of Retail Facilities 251

10.3.3 Relation Between Windows and Labor Requirements

Registers must be staffed. While employees have duties beyond direct revenue gen-
eration, the data reveal a reliable relation between windows and the number of
clerical employees operating the registers at each facility.9 The general form of this
relation was determined to be
χ
LA = ϕ(WA)χ = ϕ(κ(RA)λ ) (10.15)

where ϕ indicates the clerical employees required to staff a postal store with a sin-
gle window and 0< χ < 1 indicates the rate at which employment expands with
additional windows.10 Returns to scale in staffing arising from the ability to re-
lieve workers systematically suggests that doubling the number of windows does not
require doubling the number of employees, and the data confirm this expectation.
Translating labor inputs into labor cost requires an estimate of unit labor cost for
employees. Because workers can be reassigned across facility size categories in a
flexible manner, a constant labor cost was used for all units of clerical labor.

10.3.4 Relation Between Windows and Postal Store Space Needs

There are two ways to develop the relation between windows and required postal
store space needs. The first is to use the traditional economic approach of taking data
from actual USPS operations to estimate a production relation. Alternatively, design
criteria for facilities and operating standards developed for services can be used to
infer a relation between revenue and space needs. Current postal store designs relate
number of windows to overall area. These two approaches will be discussed in turn.
First, the traditional economic approach to relate inputs to outputs is through a
production function. In this case, the labor input is ignored, leaving only space input,
and the general form of the postal services production function may be written as
ν
SA = μ(WA)ν = μ(κ(RA)λ ) , (10.16)

where SA is the total size of the facility in square feet expressed as the product of size
per square mile, S and market area in square miles, A, W is windows per unit, and μ
and ν are parameters to be determined empirically.11 It is anticipated that μ reflects

9 The analysis relies on the assumption that the non-register activities of employees observed in
the data do not vary systematically with the number of windows. Estimation of equations designed
to explain the number of retail employees consistently revealed that the number of windows is the
dominant factor determining employment. To the extent that this is not true, some adjustment of
the relation between size and labor cost attributed to revenue generation is needed.
10 Equation (10.15) relates to the function in (10.3) as L = h(w) = ϕ(W A)χ /A
11 Equation (10.16) relates to the function in (10.4) as S = i(W ) = μ(W A)ν /A.
252 A. M. Yezer and J. W. Gillula

the space required for a facility with a single window and 0 < ν < 1 because space
needs increase at a decreasing rate with windows due to certain indivisible space
requirements for doorways, bathrooms, etc. The discussion here is in terms of total
space per facility because that is what is observed in the data and this section is
designed to set up the empirical analysis to follow.
The second method that could be used to relate postal store size and revenue
would be to use facility design criteria that have been established by the USPS along
with current criteria that relate window operation time to retail revenue. This might
be termed an engineering approach to the production function relation. Currently,
the USPS has postal store design criteria that relate the number of windows and
number of carriers assigned to a facility to the physical size of the facility. This is
done through a form of a table, in which different combinations of windows and
carriers are related to different designs.12
Both methods were employed and found to be generally consistent. However, the
final model calibration was based primarily on the second, engineering, approach
because it presumably reflects current USPS thinking about the relation between
space and windows. The economic approach gave larger space requirements, which
may reflect past policies of providing greater space per window.

10.3.5 Relation Between Postal Store Size and Cost

The USPS rents over 70 % of its facilities. The lease information allows estimation
of hedonic rent equations that give the expected rent per square foot in different
geographic locations as a function of facility size and other characteristics. Based
on this work, as well as the general literature on the relation between the size and
rent of office space, the costs of space should increase at a decreasing rate with
facility size. This leads to the following basic specification of a postal store rental
cost equation
ν φ
Cs A = ϒ(SA)φ = ϒ(μ(W A)ν )φ = ϒ(μ(κ(RA)λ ) ) (10.17)

where CS A is the total rental cost of the facility written as the product of rent per
square mile of market area (CS ) and market area (A), and ϒ and φ are parameters to
be determined. Because rent increases at a decreasing rate with size, it is expected
that 0< φ < 1. Rents vary by location and this effect is captured by ϒ, which varies
based on estimated rental hedonic equations for different locations.

12 At the same time, USPS has standards for labor productivity that relate window operating time
to revenue. Taken together, these standards would allow one to work backwards from retail revenue
expectations for each facility to the size of facility needed to meet those expectations.
10 Modeling the Location of Retail Facilities 253

10.3.6 Solving the Model

Recall from Eq. (10.6) above that surplus per unit area may be written as the differ-
ence of revenue net of mailing cost and the sum of labor and space cost per square
mile or that

ω = f (D, r) − Eh(g(f (D, r))) − j (i(g(f (D, r))), r) (10.6)

or

ω = R − CS − CL . (10.18)

The discussion presented above has demonstrated that

ν φ
RA =(π r 2 D)[θ − τ r], Cs A = ϒ(SA)φ = ϒ(μ(WA)ν )φ = ϒ(μ(κ(RA)λ ) )
λ ν φ
= ϒ(μ(κ[(π r2 D)(θ − τ r)] ) )

and that

CL A = LA = Eϕ(WA)χ = Eϕ(κ(RA)λ )χ = Eϕ(κ[(π r 2 D)(θ − τ r)]λ ))χ .

The market area of these circular markets is given by A = πr2 . Dividing the expres-
sions for RA, CS A, and CL A by A and collecting terms gives an expression for ω
that is not too messy:

ω =ψD(θ − τ r) − Eϕκλχ π λχ −1 r 2(λχ −1) (D(θ − τ r))λχ

(10.19)
− ϒ(μφ κλνφ π λνφ−1 r 2(λνφ−1) (D(θ − τ r))λνφ

The problem of choosing facility location and size to maximize surplus is thus re-
duced to choosing r to maximize ω. This is done by setting the derivative of (10.18)
equal to zero, i.e., dω/dr = 0. To get some insight into this result note that λ, ν, ϕ, χ
are all on the (0, 1) interval. When they are estimated empirically using USPS fa-
cility data for purposes of model calibration, they all lie on the [0.8, 0.95] interval.
This means that the exponents of r such as 2(λχ − 1), and 2(λνφ − 1), are both neg-
ative. Consider the three terms in (10.18). The derivative of the first with respect to
r is negative and the derivatives of the second two terms with respect to r are both
positive. The second derivative of the first term with respect to r is zero and the
second derivatives of the other terms with respect to r are negative. Taken together,
this means that, provided the first derivative is initially positive for lower values of
r, Eq. (10.18) produces an internal maximum of welfare for some positive level of
r. There is, however, no guarantee that surplus is maximized at a positive value of
254 A. M. Yezer and J. W. Gillula

ω. In these cases, given the business model underlying the postal store, there is no
positive r for which surplus is positive and the model minimizes the expected loss.13
In particular, when demand density is low, as it is in much of the U.S., there is
no value of r that produces a positive surplus. However, service must be universal
as an obligation of the USPS. In addition, there are significant transaction costs in
modifying the number and size of facilities in an area. Finally, there is a choice
between operation as a postal store or substituting some lower cost contract postal
units into the retail mix. For these reasons, the business model constructed is de-
signed to quickly simulate a variety of alternative configurations of PS and contract
postal units (CPUs) in a given market area so that the user can trace the pattern
of incremental change in surplus associated with changing the level and mix of re-
tail services. Given the nature of the model, the surface that maximizes surplus is
quickly traced and the mix associated with a maximum is readily apparent. It was
only after the model was constructed and operated that this feature of its perfor-
mance was apparent. Furthermore, operation revealed that there are often a number
of alternative configurations of retail facilities that generate similar levels of surplus.
In the presence of significant transaction costs, the configuration that involves the
least disruption to current service is likely preferred to another that promises a small
gain in surplus but causes significant disruption.

10.4 Calibration of the Location Model

The central challenge to constructing a business model of spatial location is calibra-

tion, particularly specification and estimation of the spatial demand relations. Based
on the above discussion, it is evident that estimates of a number of critical parame-
ters, specifically ψ, θ, τ, κ, λ, μ, ν ϒ, and φ will be needed to evaluate the model.
This estimation problem is solved by relying on the theory discussed in the previous
sections, and specifically on Eqs. (10.13), (10.14), (10.15), (10.16) and (10.17).
However, this discussion was based on fairly straightforward application of stan-
dard theory. In the case of postal services, as with many other retail and wholesale
activities, the actual market situation is not really accurately represented in the stan-
dard theory. This is particularly true of the demand for retail postal services, which
is discussed in some detail in the next section. Issues with the calibration of the sup-
ply side of the model are then considered, followed by a presentation of the results
of the entire calibrated model.

13 The possibility of a corner solution at r = 0 is not important here. As a practical matter, solutions
for sufficiently small radius are not feasible because they would imply operation with fractional
windows and employees. There is a minimum size of feasible facility due to the need to have at
least one window and retail worker.
10 Modeling the Location of Retail Facilities 255

10.4.1 Estimating and Calibrating a Model of Retail Demand

Consider the standard approach to spatial demand where consumers are located in
space and assumed to face a delivered price based on the FOB price plus transporta-
tion cost from the plant. In the case of many retail purchases, the origin of the trip
is not the residence, but it may be the place of employment instead. Furthermore,
shopping trips may have multiple purposes and the postal service may be relatively
unimportant compared to other reasons for travel. Therefore, the “origin” of a retail
postal service shopping trip may be a residence in the market area, an employment
location in the market area, or another location in the market area that motivated the
consumer to be in the vicinity of the postal store. However, as long as the origin of
the trip is a random location within the market area and density of trip origins is
constant across the area, the total demand for retail postal services can be expressed
in terms of Eq. (10.13) as

RA = Pf (D, r) = (π r 2 D)[θ − τ r], (10.20)

except that this equation can be applied to each individual type of customer. That
is, the expression for total revenue based on trips from home would involve DR
or residential density and could be written as the sum of two terms, θπr2 DR and
(−tr)θπ r2 DR . The first term is the total demand that would be experienced in the
absence of transportation cost, i.e., if all customers in the market were located ad-
jacent to the postal facility. The second term is the fall in demand that arises with
increasing market radius. Thus, demand is determined by the number of consumers
in the market and the radius over which they are spread. In addition to residential
demand, employment based demand will depend on the density of employment in
the market area, or πr2 DE (θ−tr). Further experimentation indicated, as might be
expected, that the income elasticity of demand for postal services is positive and
hence household demand is increasing in income of residential households in the
market area.
These special characteristics of retail demand, rather than serving as an imped-
iment to the estimation of a postal services demand relation, actually simplified
the estimation and calibration of the model. While it is very difficult to get indi-
vidual household, firm, or employee data on use of postal services, observing the
number of households and employees in the market area of postal stores is straight-
forward. Indeed, census data allow disaggregation of population by demographic
characteristics and income and employment by industrial sector by market area.
This allowed specification and estimation of a model of total revenue, and even
revenue by product type, using revenue observed by postal store.
One other consideration influencing total retail revenue at USPS facilities is the
presence of competition from other mail delivery services. Two different measures
of local competition are available. One is the total employment in private mail ser-
vices located in the ZIP codes served by the postal store, and the other is the number
of these facilities within the market area of the facility. Given that either measure
of competition could reduce effective demand, both are included in the estimated
256 A. M. Yezer and J. W. Gillula

revenue equation. Finally, transportation cost, t, differs geographically. The primary

determinant of variation in t should be density of population and employment, or
urban location. Accordingly, the estimation allows the effect of r on revenue, and
thus by implication the effect of t, to vary for facilities located in urban areas.14
Based on these considerations, the final form of the total revenue equation is
RA = α0 +α1 NE + α2 NE U + α4 NH + α5 NH U + α6 NH M + α7 NE IP + α8 NE IC
+ α9 NE r + α10 NE U r + α11 NH r + α12 NH rU + ε
(10.21)
where NE is total employment in the market area of the facility, NH is total house-
holds in the market area, U is a 0–1 dummy variable equal to unity in large urban
areas, M is median household income in the area, IP is an index of private mail
employment in the area, IC denotes an index of competing private mail establish-
ments in the market area, r is the market radius of the facility (half the distance to
nearby postal facilities),  is an iid random error term, and the α’s are parameters
to be estimated.15 There is particular interest in α9 , α10 , α11 , and α12 , because these
reflect the effect of increased radius on demand, that is, they reflect transportation
cost to the facility, represented by τ in Eq. (10.13). The terms multiplied by the ur-
ban dummy U, allow the effect of households and firms on demand to be different in
large cities and the effects of market radius to differ because of higher transportation
costs. The other estimated coefficients reflect the components of θ in Eq. (10.13).
A number of alternative statistical techniques were used to estimate Eq. (10.19)
using data for postal stores located in the 48 contiguous U.S. states. Alaska, Hawaii
and U.S. territories were excluded based on differences in topography and/or cli-
mate. Facilities reporting revenue under $ 100 and those with a market area less
than one-tenth of a square mile were eliminated. The preferred estimation was
accomplished using robust regression and is displayed below as Eq. (10.20):16
RA =13,244.7 + 971.9 NE + 13.6 NE U + 48.8 NH − 15.2 NH U
+ 0.0000344 NH M − 188.2 NE IP − 117.3 NE IC − 14.0 NE r (10.22)
− 39.9 NE rU − 4.7 NH r + 3.5 NH rU
The number of households and employees in the market area of the store has a major
effect on revenue. Additional employment is more important than additional house-
holds. This may reflect demand by workers based on their workplace or demand by

14 A variety of efforts to customize the estimation to differences in local characteristics that might
influence transportation costs have demonstrated some promise but are not discussed here in order
to focus on the main elements of the calibration effort.
15 The index I is 0 when there is 0 private mail employment in the service area of the facility, 1 if
P
employment is between > 0 and < 10, 2 if employment is at least 10 and < 20, 3 if employment is
at least 20 and < 30, etc. Similarly, the IC index is 0 if there are 0 competitors in the service area,
1 if there is 1 competitor, 2 if there are 2 competitors, 3 if there are 3–5 competitors, and 4 if there
are 5 or more competitors.
16 This equation was estimated using 21,898 observations with F(11, 21,886) = 38,000. The t-

ratios of the estimated coefficients were all larger than 4.0.

10 Modeling the Location of Retail Facilities 257

the firms themselves. Higher household income has a modest positive effect on de-
mand. As expected, the presence of competitors reduces demand, and the effect of a
larger market area also reduces demand. Location in a large urban area increases the
effect of market radius on demand for employment as expected, but it also reduces
the effect of households on demand. It is not surprising that, relative to household
demand, employment demand is more important in larger urban areas. In general,
the effect of households on demand for USPS services is lower in large urban areas.
All estimated coefficients are highly statistically significant. This will be used as the
basic demand equation in the simulation experiments performed in the next section.
The results in Eq. (10.20) provide some insight into the determinants of demand
at individual postal stores. It is encouraging that the signs of the estimated coef-
ficients and their general magnitudes agree with prior expectations. This effort to
understand demand is directed toward the purpose of a spatial model in which the
effect of proximity is of major importance. The results presented here should not be
confused with a more fundamental examination of the determinants of the demand
for retail services which should involve, at a minimum, greater disaggregation of
both household and employment by type as well as greater consideration of the
relation between distance and actual transportation cost.17
Note that demand is estimated at the facility level rather than aggregated over
market areas that involve several facilities. Actual operation of the model involves
prediction of outcomes for larger market areas, generally counties, because a change
in facility size and spacing generally involves transition of several facilities. How-
ever, aggregation must be done with some caution because comparisons will be
based on averages for a given area but averages are a linear aggregation of mag-
nitudes and spatial aggregation is often nonlinear. Indeed, the model itself is very
nonlinear.
Consider the follow illustration of a potential confusion that can arise in the spa-
tial aggregation process. Assume that there are two facilities and that each has a
market radius of 2 miles. The average radius will be 2 miles and the average market
area will be 3.14(22 ) = 12.56 square miles because each market area is identical.
Now assume that there are two facilities whose market radii are 1 and 3 miles. The
average market radius is still 2 but the market areas are 3.14 (for the 1 mile ra-
dius) and 3.14(32 ) = 28.26 (for the 3 mile radius). Thus the average market area is
(3.14 + 28.26) = 15.7. The average area in the “real world” where the radius varies,
will be larger than average area in the model solution where radius is constant. In
general, the greater the variance in the radius, the larger the excess market area
above that calculated based on the average radius.
When comparing results from the model simulations where radius is constant
with actual operating results, the difference between the market area of an average
radius when radius is uniform and the real world, where the variance in the radius
is significant, should be considered. The difficulty can usually be overcome, in this
case, by working in average area per postal store. Alternatively, analysis can only be

17 It would also be useful to consider the precise shape of the market area.
258 A. M. Yezer and J. W. Gillula

done for facilities which all have a similar market radius. In the calibrations reported
here, one or the other of these approaches is taken to avoid aggregation problems.

10.4.2 Estimating and Calibrating a Postal Services Cost

Function

Ordinarily a cost function relates physical output to costs of producing that output.
However, measures of the quantity of postal services produced at individual loca-
tions are not available. Therefore, costs must be expressed as a function of revenue.
The fact that prices are uniform nationally, means that there is potentially a linear
relation between prices but variation in output mix across facilities can be signifi-
cant and therefore there is a real possibility that revenue differences do not reflect
services.
The cost function calibration begins by estimating the relation between revenue
and number of windows in the postal store. Measures of the number of “front” reg-
isters and “back” registers along with their time of operation and revenue generation
are included in the available data along with the number of windows in the store.
The general form of Eq. (10.14) relating windows to revenue, WA = κ(RA)λ
appears straightforward. There are difficulties in estimating the values of the param-
eters κ and λ. First, the total number of windows, WA, is an integer, whose value at
RA = 0 should, of course, equal 0. The analysis should focus on values for WA = 1,
2, 3, or 4 as these constitute the relevant range for most retail stores. Second, this
is a frontier relation in that it reflects the revenue per window achievable under x-
efficiency, i.e., given the way postal stores operate, they are producing given services
at lowest cost.18 Observations from the many facilities that have more windows than
necessary, as evidenced by the fact that they are never, or hardly ever, shown to be
open should be ignored.
Because of the need to observe a frontier, where the relation between revenue and
windows is based on high rates of utilization, is so important, three approaches were
taken to eliminate observations from inside the frontier. First, the mean revenue per
window for different numbers of windows was computed using only observations in
which revenue per window was more than one standard deviation above the mean.
Second, only facilities in which all the windows were operating for most of the time
during selected high-demand weeks (the second week in December and April) were
used. Third, the demand equation in (10.21) was used to estimate the demand at
each facility, and those facilities, in which the demand per window was predicted
to be one standard deviation above the mean were used. In all three cases, a clear

18 The cost function used in the model is based on x-efficiency in that it takes the current rules

and procedures for operating postal stores as given. The term “technical efficiency” used here is
based on x-efficiency. It may be that different work rules or ways of deploying labor and capital,
perhaps on-line retail aids, could raise output per facility. These innovations are not considered
here, although they could be added to the model easily by recalibrating the cost functions.
10 Modeling the Location of Retail Facilities 259

pattern was observed in which, for facilities judged to be on the frontier, revenue per
window for facilities with one window was approximately $ 400,000 per year and
revenue per window increased slightly as the number of windows increased.19 As a
result of this, the value of κ was set at 0.000005 and the value of λ was set at 0.95.
Using these parameters, the relation between windows and revenue reflects a postal
store that is technically efficient.20
Next it is necessary to calibrate the relation between windows in operation and
labor requirements. The labor per window function, given in Eq. (10.15) above as
χ
LA = ϕ(WA)χ = ϕ(κ(RA)λ ) , is also a frontier concept. It reflects use of labor
in a technically efficient fashion, i.e., at facilities where revenue per employee and
window is high. There are some facilities for which the ratio of clerical workers
per window is quite high, i.e., greater than two. In such cases, it is likely that these
workers are performing a variety of duties unrelated to revenue generation. Fortu-
nately, the USPS point of sale data allow observation of the hours spent logged in
to a register by individual employee identification code by day for each facility. It
is possible to identify the number of different employees who spent a significant
amount of time at terminals during each of the 3 weeks that were observed (the
second weeks in December, April, and August). Employees who logged on for less
than 10 h/week were not counted as retail employees. Estimates of the parameters
in Eq. (10.15) were based on facilities where the same numbers of workers were
spending significant time logged on to a register during each of these 3 weeks.21
The estimates were all based on facilities, in which the retail operation appears to
be operating near capacity so that they were judged to be technically efficient. In
this case, facilities with revenue per window above $ 250,000/year were selected.
As anticipated, the number of employees per window falls with the number of
windows, and the fitted parameters of Eq. (10.15) were ϕ = 1.6 and χ = 0.9.
Thus, within the data on facilities with revenue per window greater than $ 250,000,
those with one window had an average of 1.6 employees per window and this ratio
fell slowly as the number of windows increased. Calibration of the parameters φ
and χ is accomplished based on a table of the mean values of retail workers for
facilities with different numbers of windows subject to the requirement that revenue

19 The pattern of revenue per window was also tested for the two high demand weeks, and the

small increase in revenue per window with number of windows was observed.
20 A further check on the calibration of revenue per window was done by using data from the USPS

Window Operations Survey (WOS) for FY 2008. The USPS has a standard for time per transaction
that determines the hours “earned” at a facility window. At facilities reporting actual hours equal to
earned hours, the average annual revenue per hour of actual window operation was computed and,
when this was multiplied by annual hours per window, the average annual revenue per full time
window in operation at these facilities was found to be $ 642,720, well above the $ 400,000 set
as efficient revenue per window in this model calibration. Evidently the standard for technically
efficient window operation adopted for calibrating the cost function is not particularly rigorous
compared to USPS standards for earned hours.
21 This does not mean that the employees in 1 month were the same as those in another month.

The employee counts were based on a single week in each month. It does mean that the number of
retail workers counted in each of the 3 months was constant.
260 A. M. Yezer and J. W. Gillula

per window exceeded $ 250,000. The value of φ is the mean for facilities with one
window, and χ followed based on the rate of decline in retail workers per window
as windows expanded beyond one.
The functional form selected for the relation between interior space and num-
ν
ber of windows was given in Eq. (10.16) by SA = μ(W A)ν = μ(κ(RA)λ ) . This
is also a frontier relation which should be measured for facilities operating under
x-efficiency. Empirical estimation of (10.16) produced estimates of space required
that were well in excess of the standards embodied in current USPS facility design
criteria for new facilities. Given the forward-looking nature of this analysis, an engi-
neering approach was taken to the evaluation of the parameters of (10.16) using the
standards for new retail facilities. These standards allow for P.O. Box sections and
carrier workroom space that is an increasing function of the number of windows.
While there are several alternative designs, it appears that 1800 square feet is an av-
erage standard for a postal store with a single window and that space increases at a
slightly decreasing rate with the number of counters. Accordingly, Eq. (10.16) was
parameterized by setting μ = 1800 and ν = 0.9. Thus, a facility with one window
occupies 1800 square feet of interior space, including the service lobby, P.O. Box
lobby, workroom for mail processing, bathrooms, etc. This increases to 3359 square
feet for two windows, 4839 square feet for three windows, etc.
Estimation of cost per square foot has been accomplished using the USPS Facility
Database (FDB) in support of P.O. Box fee-setting for over a decade. This provides
substantial evidence on the parameters of Eq. (10.17), viz.,
ν φ
Cs A = ϒ(SA)φ = ϒ(μ(W A)ν )φ = ϒ(μφ (κν (RA)λ ) ) . (10.23)
Specifically, it is well established that the cost per square foot falls with facility size
and that the elasticity of cost with respect to size is about 0.8, i.e. φ = 0.8.ϒ is
the cost per square foot appropriate for the real estate market under consideration
and ranges from under $ 5/square foot in rural areas to over $ 45/square foot in the
higher density portions of larger urban areas.

10.5 Demonstrating the Location Model

The calibrated model can be solved for operating revenue and cost as a function of
market radius and the levels of key variables such as population and employment
density, household income, competitors, urban character and rental price of space
in the area being served. Because of the recursive nature of the model, determi-
nation of the radius or facility spacing implies unique values for all the operating
characteristics of the postal stores serving the market area.
Investigation of the demand for USPS services indicated that the structure of de-
mand was different in large urban areas than in the rest of the U.S. This is evident in
the estimation results for Eq. (10.20) reported above where the effect of households
and employment on demand is different in large urban areas than in the rest of the
data. Compared to residential households, employment is a relatively more impor-
tant determinant of demand in large urban areas. Also, the effect of market radius
on demand for services is greater in large urban areas.
10 Modeling the Location of Retail Facilities 261

Table 10.1 Estimated revenue and cost at different radii for large urban areas
Radius (miles) Revenue ($) Windows Empl. Interior space Cost ($) NetRev
(sf) ($)/SqMi
0.5 146,186 0.4 0.7 795 60,711 15,773
0.75 302,700 0.8 1.3 1481 112,208 22,161
1 510,656 1.3 2.1 2316 174,552 25,725
1.5 1,055,126 2.6 3.8 4308 322,503 29,025
2 1,728,068 4.2 5.8 6568 489,733 29,801
2.5 2,477,950 5.9 7.9 8939 664,648 29,265
3 3,253,245 7.7 10.0 11,282 837,166 27,935
3.5 4,002,422 9.4 12.0 13,469 998,012 26,081
4 4,673,953 10.8 13.7 15,379 1,138,339 23,858
4.5 5,216,309 12.0 15.0 16,892 1,249,452 21,368
5 5,577,959 12.8 15.9 17,889 1,322,572 18,680
5.5 5,707,375 13.1 16.2 18,243 1,348,563 15,846
6 5,553,028 12.8 15.8 17,820 1,317,554 12,907
6.5 5,063,387 11.7 14.6 16,468 1,218,306 9900
7 4,186,925 9.8 12.4 13,998 1,036,897 6867
Employment density 355 workers/square mile, residential density 2259 households/square mile,
income $ 62,382

In addition to differences in the parameters of the demand equation, the costs of

space are higher in large cities. There are also differences in the density of house-
holds and employment and in the density of competition. These differences lead to
rather different implications for the allocatively efficient size and spacing of facili-
ties between large urban areas and the rest of the U.S., and the nature of postal store
operations is generally very different. All this is illustrated in Tables 10.1 and 10.2
and the accompanying figures below.
Table 10.1 and the associated Fig. 10.1 show the solution of the model for postal
stores located in large urban areas based on both the demand function in those areas
and the higher density of households, firms and high space cost. A total of 3423
postal stores fall into this category.22 The average household density in these areas
is 2259/square mile, employment density is 355/square mile, and median household
income is $ 62,382/year. The model shows that allocative efficiency is achieved
when net revenue per square mile is maximized at about $ 29,801 with a market
radius of 2 miles. The postal store associated with this maximum has approxi-
mately 4 windows, 6 retail workers, 6600 square feet of interior space and $ 1.73

22 The 3423 also reflects the number of facilities with sufficiently complete data to be used in

computation of characteristics of the actual condition of facilities in this group.

262 A. M. Yezer and J. W. Gillula

Table 10.2 Estimated revenue and cost at different radii for non-urban areas
Radius (miles) Revenue ($) Windows Empl. Interior space (sf) Cost ($) NetRev
($)/SqMi
0.5 32,799 0.1 0.2 222 16,505 − 134
1 89,932 0.3 0.5 525 38,814 1959
1.5 182,347 0.5 0.9 960 70,721 2895
2 307,748 0.8 1.3 1502 110,308 3469
2.5 463,839 1.2 1.9 2133 156,320 3852
3 648,324 1.7 2.5 2841 207,797 4118
3.5 858,907 2.2 3.2 3613 263,945 4303
4 1,093,291 2.7 3.9 4441 324,072 4430
4.5 1,349,181 3.3 4.7 5316 387,561 4514
5 1,624,280 4.0 5.5 6230 453,848 4564
5.5 1,916,292 4.6 6.4 7175 522,408 4587
6 2,222,922 5.4 7.2 8146 592,751 4589
6.5 2,541,872 6.1 8.1 9136 664,408 4572
7 2,870,847 6.8 9.0 10,138 736,934 4540
7.5 3,207,551 7.6 9.9 11,146 809,899 4495
8 3,549,687 8.3 10.8 12,155 882,889 4438
8.5 3,894,960 9.1 11.7 13,159 955,498 4373
9 4,241,073 9.9 12.6 14,153 1,027,335 4298
9.5 4,585,731 10.6 13.4 15,130 1,098,012 4216
10 4,926,636 11.4 14.3 16,087 1,167,151 4128
11 5,588,007 12.8 15.9 17,916 1,299,323 3934
12 6,206,816 14.2 17.4 19,600 1,420,905 3721
13 6,764,695 15.4 18.8 21,096 1,528,985 3493
14 7,243,276 16.4 19.9 22,366 1,620,654 3251
Employment density 26 workers/square mile, residential density 130 households/square mile,
income $ 45,269

million/year in revenue. The table also shows the performance characteristics of

choosing alternative patterns of postal store spacing.
The actual radius of all the postal stores in these urban areas was 2.1 miles and
revenue per facility was $ 1.25 million. The model estimate of an allocatively effi-
cient radius of 2.0 miles matches the average for this group of 3423 postal stores.
The fact that spacing of postal stores in this sample of urban areas is consistent
with the spacing suggested for allocative efficiency does not imply that the size or
spacing of these facilities is efficient for two reasons. First, the analysis is based on
“average” spacing and, in a sample that aggregates over all large urban areas, the
average may be appropriate but its detailed distribution on a city by city basis may
be flawed. However, the evidence suggests that, the spacing of these facilities is not
10 Modeling the Location of Retail Facilities 263

$35,000

$30,000
Net Revenue per sqm

$25,000

$20,000

$15,000

$10,000

$5,000

$0
0 1 2 3 4 5 6 7
Radius (Miles)
Employment/sqm = 355; HH/sqm = 2,259; Income/HH = $62,382

Fig. 10.1 Radius vs. net revenue per square mile: large urban areas

grossly inefficient. There may still be problems of lack of technical efficiency as the
analysis is disaggregated to the individual MSA level. Nevertheless, the fact that the
average size for all postal stores in the sample is about 7000 square feet, compared
to the optimal 6600 square feet associated with the radius of 2 miles indicates that
facility size is not, on average, inconsistent with allocative efficiency either.23
Second, the discussion thus far has not considered technical efficiency of the
actual postal stores found in these large urban areas. The results for allocative effi-
ciency above were based on the assumption that the actual size, in terms of interior
space, number of windows, and clerical staffing was technically efficient. Techni-
cal efficiency requires that the actual sizes be efficient given the actual demand for
services at each postal store. Measurement of technical efficiency for many postal
stores is limited because the number of windows in operation and the level of cler-
ical staffing are only observed for facilities with POS terminals. Nevertheless, a
rough estimate of technical efficiency can be obtained by using data from the POS
facilities to impute windows and retail workers for the non-POS postal stores. In
the case of the following analysis, this imputation was done by relating facility size
(interior space) to numbers of windows and workers for those facilities with POS
terminals and then using the resulting relation to estimate windows and workers for
non-POS postal stores.
The result of the imputation is that the average numbers of windows, retail work-
ers, and interior space can be substituted into the cost equations and the imputed
characteristics of the actual postal stores in large urban areas can be compared to
the technically efficient costs. In Table 10.1 below, a technically-efficient facility
serving the allocatively efficient average market radius of 2 miles has 4.2 windows,
5.8 retail workers, 6568 square feet of interior space. The total cost for the 12.56
square mile market area is $ 489,733 or $ 38,991/square mile. The imputed charac-
teristics of the current average postal store in large urban areas, where actual market

23 Specifically, average facility size for those facilities with < 20,000 square feet is 7000
square feet. Larger facilities likely include significant mail processing and/or vehicle storage and
maintenance activities.
264 A. M. Yezer and J. W. Gillula

radius is about 2 miles, are 2.95 windows, 4.63 workers, and 16,484 square feet
of interior space, yielding an imputed cost per square mile of $ 33,275. This is a
remarkable $ 5716 less than the technically efficient cost per square mile.
Compared to the technically efficient postal store, actual facilities have fewer
windows and retail workers but more interior space on average. Is it possible that the
actual operation of USPS retail postal stores in large urban areas has achieved some
kind of super efficiency? Are the standards for technical efficiency in the model too
low? Alternatively, it may be that the numbers of windows and employees provided
in these facilities in large urban areas are too small resulting in long lines and poor
service. Insight into these questions is gained from comparing the actual revenue
per facility with model estimates based on the demand equation. Actual revenue per
facility is $ 1.25 million compared to the demand model estimate of $ 1.7 million,
a difference of almost $ 40,000/mile of market area. The reason that the model re-
quires more windows and retail workers is that the model generates estimates of
significantly more revenue than the actual postal stores generate. Although these are
rough estimates based on averages over thousands of postal stores, the suggestion is
that, while the average spacing of postal stores (and hence the average number of fa-
cilities) is consistent with allocative efficiency, the number of windows and clerical
staffing per facility appears to be too low given the demand for postal services.
It may be that significant waiting time at postal stores in large urban areas deters
potential customers from using the USPS. A rough estimate of the net revenue loss
per square mile due to this under provision of window space would be $ 40,000 of
extra revenue per square mile less approximately $ 6000 in extra cost per square
mile or $ 34,000/square mile. This is slightly larger than the net benefit per square
mile of the average postal store in a large urban area under allocative efficiency, i.e.,
it is substantial. Some caution is warranted here because these estimates are based
on averages of very large and diverse facilities located across all large urban areas.
Still, the analysis of technical and allocative efficiency is consistent in the conclusion
that the number and spacing of postal stores in large urban areas is appropriate on
average and that the number of windows and workers or alternative mechanisms for
providing retail services should, if anything, be larger.
Table 10.2 and Fig. 10.2 illustrate the model solution in the average area out-
side large cities. This includes smaller cities, towns, and rural areas. The average
density of households is dramatically lower at 130/square mile as opposed to 2259
for large cities. Employment density is 26/square mile compared to 355 and there
are fewer competitors. The model is solved for the revenue, employment, and size
implications of alternative market radii separating postal stores under conditions of
technical efficiency. Allocative efficiency is achieved when net revenue per square
mile is maximized at $ 4589 when market radius is 6 miles. This is a substantial
contrast to the large city solution where net revenue per square mile is larger by
a factor of 6.5 and radius was one-third as great. The postal store that maximizes
net revenue per square mile has approximately 5 windows, 7 employees and 8000
square feet of interior space.
10 Modeling the Location of Retail Facilities 265

$5,000

$4,000
Net R evenue per s qm

$3,000

$2,000

$1,000

$0

-$1,000
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
R a dius (Mile s)
E mployment/s qm = 26; HH/s qm = 130; Inc ome/HH = $45,269

Fig. 10.2 Radius vs. net revenue per square mile: non-urban areas

Data on actual postal stores in this outside large city group indicates that the
average radius is 4.7 miles, revenue is only about $ 410,000/year, and the average
facility size is 3100 square feet.24 There are 22,811 postal stores in this group.
Comparing allocative efficiency with a radius of 6 miles with the actual radius
of 4.7 miles from Table 10.2 it appears that net revenue per square mile is $ 75
($ 4589 − $ 4514) per mile higher with a radius of 6 miles. This may seem like a
small margin but the 22,811 postal stores in question have an average market area of
106 square miles and this means that the total difference in annual net revenue per
square mile implies a total difference of $ 180 million/year. This is a rough estimate
of gains that could be achieved because it is an average over areas with very different
levels of service. In general, the greater the geographic disaggregation, the larger the
welfare gains from right-sizing and right-spacing the postal stores.
It is also possible to conduct an analysis of the technical efficiency with which
retail services are currently produced. As discussed above for the case of postal
stores in large urban areas, the challenge for measuring costs of current operations
is that numbers of windows and retail workers are only observed for POS postal
stores. Once again, the answer to the missing observations is to impute the numbers
of windows and retail workers. The imputation produced estimates, for the aver-
age facility, of 2.13 windows, 3.46 retail workers, and 4764 square feet of interior
space for the 22,811 postal stores in the non-urban sample. This produced a cost per
square mile of market area of $ 4486. These magnitudes are all substantially above
the technically efficient averages of 1.1 windows, 1.75 workers, 1964 square feet
of interior space, and $ 2264 cost per square mile for technically efficient facilities
located with a market radius averaging 4.5 miles. The difference in cost per mile, $
2222 = $ 4486 − $ 2264, when multiplied by the 22,811 postal stores and the 106
square miles of average market area produces an increase in total cost due to tech-
nical inefficiency of $ 5.4 billion/year. The standard cautions regarding estimates

24 Note that predicted revenue at a radius of 4.5 miles is much larger than actual revenue. This
illustrates the difficulty of applying the demand model to large diverse areas and also may point to
a further problem in revenue generation in low demand density area.
266 A. M. Yezer and J. W. Gillula

based on imputed values of windows and employees as well as averages computed

over all facilities located outside large urban areas are in order here. It may be that
the relation between the size and number of active retail windows in non-POS fa-
cilities is much lower than that for POS facilities. If true, the technical inefficiency
estimate would be substantially lower. Furthermore, the ratio of employees per win-
dow for non-POS facilities may be significantly lower than that for POS facilities.
This would also lower the $ 5.4 billion/year estimate of technical inefficiency. Even
if this cost figure were reduced by half due to systematic differences between POS
and non-POS facilities, the annual cost of technical inefficiency is impressive and
the question of differences between facility types should be investigated.
Clearly, location in large cities generates far larger net revenue per unit area than
in the rest of the country. Total surplus per postal store is the multiple of area and
surplus per unit area. Because market area increases with the square of the radius,
the total surplus of the large city postal store is $ 374,300 while the total surplus
of the optimal facility outside a large city is $ 518,740. Differences in surplus per
postal store are far smaller than differences in surplus per square mile, and total
surplus is actually larger for the facility outside large cities because the market area
is 113 square miles instead of 12.5 square miles.
Maximization of total surplus across the nation is achieved by maximizing net
revenue per square mile and then multiplying by the 4.1 million square miles of
U.S. territory. The dataset used in this study has 28,110 postal stores whose market
area totals 2,681,694 square miles. The mean overall market radius is 4.33 miles.
The mean market radius in large urban areas is 2.1 miles, virtually identical to the
optimal radius produced in the results in Table 10.1. The mean market radius in
the remainder of the sample is 4.7 miles, considerably smaller than the 6.0 miles
characterizing the surplus-maximizing solution in Table 10.2. These results suggest
that the actual market radius of postal stores, and by extension the number of postal
stores, in larger urban areas is close to the optimum and that there are too many
postal stores placed too close together in the rest of the country. Note that the im-
plication for the number of facilities of having a radius of 6 miles is dramatically
different than for 4.7 miles. The ratio of market areas is approximately 2 to 1 (ac-
tually 113/69 or 1.64 to 1), implying that the optimal number of facilities in the
portion of the U.S. outside the largest cities is 61 % of the current number.
The analysis of technical efficiency in large urban areas compared to the rest
of the U.S. (the “non-urban areas”), although requiring imputation of some very
important data, produces results that are quite complementary with those from the
analysis of allocative efficiency. For large urban areas, the number and spacing of
current USPS facilities appears consistent with allocative efficiency and the numbers
of windows and workers available appears to be slightly below that needed for tech-
nical efficiency. It may be that, in large urban areas, failure to provide adequate retail
inputs is costing the USPS significant amounts of revenue. Further study of waiting
times and revenue would be needed to make such a determination. For the rest of
the U.S., just as allocative efficiency requires fewer postal stores located farther
apart, technical efficiency implies a substantial reduction in the number of windows
and workers. The estimated gain from moving from the current level of operations in
10 Modeling the Location of Retail Facilities 267

these non-urban areas to a technically efficient condition is $ 5.4 billion/year, and the
further gain if these changes were associated with a move to an allocatively efficient
number and spacing of postal stores is $ 180 million/year. Although these are esti-
mates and better data on non-POS facilities might change them significantly, their
magnitudes suggest that using modern modeling of facility location could achieve
substantial gains for the USPS.

10.6 Modeling the Effects of Alternative Outlets

The USPS has recently been expanding the availability of retail postal services
through the use of Contract Postal Units (CPUs) as an alternative to postal stores
modeled above. The CPU provides a range of postal services including virtually
all mailing services but generally does not provide post office boxes.25 There are
approximately 3500 CPUs in operation, which means that there are many markets
without significant CPU penetration. Adding CPUs to the model provides an oppor-
tunity to demonstrate the manner in which the approach can be generally adapted to
a range of alternative retail environments.
The process of adding CPUs involves substantial negotiation and locations of
the facilities are generally predetermined based on other types of retail activity. In
some cases, agreements are signed with a single retailer for CPUs to be located
at a number of existing facilities. However, the intent of the USPS is to provide
additional retail postal service capacity without incurring the full cost of a postal
store. Accordingly, in contrast to postal stores where size and location are often the
result of a historical process that has little to do with current economic conditions,
CPU activity should be motivated by potential demand in the area. Furthermore,
given that compensation of CPU owners is based on a percentage of revenue, CPUs
are unlikely to remain in operation unless they generate significant revenue.26
Estimation of a revenue model such as that in Eq. (10.20) for CPUs, is therefore
problematic because the presence of a CPU in the area is likely to be a result as
well as a cause of the level of revenue generated in the market. CPUs are not likely
to open and continue in operation unless the market generates significant amounts
of postal revenue. In view of this, a modified instrumental variables approach was
used to estimate CPU revenue generation. The first stage involved estimating a mar-
ket revenue equation like Eq. (10.20) for all market areas that did not have a CPU.
The predicted value of market area revenue was then used in a second stage equa-
tion to predict the revenue generated by CPUs in markets that actually had one or

25 The CPU should not be confused with stamps sold on consignment by a range of ordinary

retailers who do not provide mailing services.

26 There is a substantial annual birth and death rate for CPUs both due to economic forces and

changes in the business model of the retailer hosting the operation.

268 A. M. Yezer and J. W. Gillula

more CPUs.27 Formally, the two stage estimation procedure involved the following
approach shown as the system (10.22) and (10.23):
RA = α+βE NE +βEU NE U+βH NH −βH U NH U+βH M NH M−βEP NE IP
− βEC NE IC − βEr NE r − βEU r NE U r − βH r NH r + βH U r NH U r + εRA
(10.24)

RCPU = ρCPU NCPU RA∗ + ρCPUr NCPU RA∗ r + εCPU (10.25)

Here the Eq. (10.22) is simply the facility revenue equation estimated for all markets
without a CPU, and the lower CPU revenue equation relates total CPU revenue in
the market area of the postal store to the product of the number of CPUs, NCPU , and
the estimated total revenue for the market from (10.22) noted RA∗ . The next term is
this same product multiplied by the radius of the market area, NCP U RA ∗ r, ρCP U ,
and ρCP U r , are parameters to be estimated and εCPU is an iid error term. Note that
(10.23) is specified without a constant term because logically RCP U = 0 when
RA* = 0.
Estimates of (10.23) reflecting the result of estimating the system of Eqs. (10.22)
and (10.23) are displayed in Table 10.3. In keeping with the general finding that
density of population has a significant effect, likely through transportation cost dif-
ferentials, on the spatial demand curve, different versions of the equation are shown
for areas with 4 different levels of population density, ranging from very low (less
than 50 households per square mile) to high (more than 500 households per square
mile). Estimates of the share of revenue going to a CPU, as indicated by estimated
values of the ρCPU , and ρCPUr parameters, tend to increase with density. As trans-
portation costs rise with density, CPUs become more attractive substitutes for postal
stores within a market area of given size.28
Of course, part of the revenue at CPUs is displaced from postal stores. Modeling
the effects of CPUs requires estimating both revenues at the CPU and declining
revenue at postal stores. For the reasons noted above, total market revenue, including
both postal stores and CPUs located in a given market area, is likely a cause of
CPU revenue. Therefore, the effect of adding CPUs on total market revenue was
estimated using the same type of two equation system described by Eqs. (10.22)
and (10.23) except that a total revenue equation was substituted for (10.23):

RT = ξS RA∗ + ξCP U r NCP U RA∗ r + εT (10.26)

27 One complication was the case in which a CPU appeared to fall within the geographic mar-
ket area of more than one postal store. In this case, the CPU was shared out to the market area
proportionally based on distance from each of the postal stores.
28 Of course market size falls significantly with density ranging from 5.2 miles at the lowest density

to 1.4 miles for the highest density areas. The share of revenue going to the CPU also increases
with market radius based on the second parameter in the CPU revenue equation.
10 Modeling the Location of Retail Facilities 269

Table 10.3 Estimates of equations (10.23) & (10.24): Determinants of CPU revenue and market
revenue with CPUs
Dependent ρCPU (ρCPUr ) in Eq. t-ratio Dependent ξS (ξCPUr ) in t-ratio
variable (10.23) variable Eq. (10.24)
Lowest density coefficients
NCPU RA∗ 0.122∗∗∗ 16.5 RA∗ 0.992∗∗∗ 246
NCPU RA∗ r 0.0019 1.13 NCPU RA∗ 0.027∗∗∗ 25
Modest density coefficients
NCPU RA∗ 0.115∗∗∗ 10.1 RA∗ 0.967∗∗∗ 129.7
NCPU RA∗ r 0.0083∗∗ 2.38 NCPU RA∗ r 0.045∗∗∗ 18
Medium density coefficients
NCPU RA∗ 0.093∗∗∗ 6.23 RA∗ 0.996∗∗∗ 88.6
NCPU RA∗ r 0.023∗∗∗ 3.81 NCPU RA∗ r 0.039∗∗∗ 7.9
High density coefficients
NCPU RA∗ 0.127∗∗∗ 25.6 RA∗ 1.00∗∗∗ 90.6
NCPU RA∗ r 0.0089∗∗∗ 5.04 NCPU RA∗ r 0.055∗∗∗ 6.7
All density coefficients
NCPU RA∗ 1.27∗∗∗ 25.7 RA∗ 1.064∗∗∗ 271
NCPU RA∗ r 0.0088∗∗∗ 5.04 NCPU RA∗ r 0.045∗∗∗ 27
∗ significant at the 10 % level;
∗∗ significant at the 5 % level;
∗∗∗ statistically significant at the 1 % level

where RT is total market area revenue, RA∗ is predicted revenue from areas with
no CPUs, NCPU is the number of CPUs in the market area, r, is the market radius
and t is an error term. The estimate of ξCP U r is the fractional change in total market
revenue of adding a single CPU in a market area whose radius is 1 mile.
Estimation results for Eq. (10.24) are shown in Table 10.3, so that they can be
compared to the results for Eq. (10.23). As was the case with the ρCPUr parameter
estimates in Eq. (10.23) the estimates of ξCP U r rise with density of the market area.
These results reflect the same increased substitution effect as transportation costs
rise. Note that the predicted revenue of CPUs expressed as a fraction of RA∗ based
on estimates of Eq. (10.23) is generally about twice as large as the expected effect
of adding a CPU on total area revenue using Eq. (10.24) estimates. This difference
reflects the revenue displacement effect, which appears to be close to 50 %. Put
another way, adding a CPU to an average sized market area for a given density
(where market area falls significantly with density) results in CPU revenue that is
about 12 % of postal store revenue, of which about 7 % is revenue that is displaced
from the postal store and 5 % is additional revenue. The ratio of the total revenue
effect to the displacement effect rises with the radius of the market area as expected.
These results are easily translated into a calibrated model of the effects of adding
or subtracting CPUs on net revenue because the cost function of CPUs is simply a
270 A. M. Yezer and J. W. Gillula

constant fraction of revenue. The model can be run consecutively with a changing
mix of postal stores and CPUs to determine the net revenue implications of changes.
The model does not solve for an “optimum” presence of CPUs because they do not
provide all postal services and the USPS has an obligation to provide a full range
of retail services to the public. Nevertheless, model simulations of alternative mixes
of stores and CPUs make apparent the tradeoff between net revenue and the mix
of postal stores and CPUs in a market area. Presumably, the additional revenue
generated by the CPUs reflects a gain in value of service to the public that should be
considered in evaluating the net position of the public when CPUs are substituted
for stores.

10.7 Conclusions for Modeling Retail Location

Applying standard location theory to retail activity is complicated because consumer

shopping trips do not necessarily originate from the primary residence. Unfortu-
nately, at most, information on consumers is available by place of residence. There
is information on places of employment and the location of other types of retail
activity that could influence shopping trips, but characteristics of these potential
customers or customers of rivals, is not available. This paper demonstrates that the
standard theory can be modified to provide implications for total facility revenue as
a function of the spatial distribution of residential population, employment, and ri-
val outlets in a market area. Empirical estimates of the revenue equation implied by
this theory produce results that agree well with prior expectations based on theory
and can be used to calibrate models that produce plausible results.
The model developed can be used to maximize net revenue per unit area, and
hence net revenue for any area being analyzed. It provides guidance regarding the
spacing, size, revenue, cost, and input requirements for a spatial retail activity. The
specific case of USPS activity refers to a case in which there are limits on entry be-
cause it is a public enterprise. The model could also be used to estimate the optimal
characteristics of a spatially competitive industry. In this case, free entry would gen-
erally be expected to produce a pattern of that departs from the optimum by having
too many facilities and smaller than optimal market areas. The model could be used
to measure the welfare effects of such departures from optimal size and spacing that
have been noted by Capozza and VanOrder (1980) in the theoretical literature on
spatial competition.
If the model is used to analyze a retail network that is already in place, it is helpful
to generate the type of net revenue surface, as a function of plant radius shown in
the figures. If, as is the case for the postal store network considered explicitly here,
the net revenue function is relatively flat, a range of plant size and spacing solutions
may produce roughly equivalent results. If there are significant transaction costs
associated with altering the network, a departure from net revenue maximization
may be warranted.
10 Modeling the Location of Retail Facilities 271

In the specific case of postal services, there are alternatives to postal stores
for serving consumers. The example of CPUs is considered here and the man-
ner in which this alternative technology for providing retail access modifies the
opportunities for net revenue is demonstrated.
In this case, all estimation and model calibration was accomplished using cross
section data from a single year. Clearly, panel data would have allowed adjustment
for unobserved heterogeneity that is always present in spatial models. In addition,
panel data would allow the research to identify either explicit or implicit quasi
experiments in which the pattern of retail service delivery was modified and the
consequences could be observed. These empirical techniques are well understood
in economics and their ability to inform the model calibration should be clear.
Nevertheless, the final results show that the model is capable of identifying cases
in which the spacing but not the size of facilities is optimal. It can distinguish areas
where both the size and spacing of facilities is suboptimal. Finally, it can allow users
to explore the consequences of implementing an alternative method of retail service
delivery and to relate these consequences for alternative market areas with different
demand and cost characteristics.

References

Beckmann MJ (1999) Lectures on location theory. Springer-Verlag, Berlin

Beckmann MJ, Thisse JF (1986) The location of production activities. In: Mills E, Nijkamp P (eds)
Handbook of regional and urban economics, vol 1. Elsevier Science, New York, pp 21–95
Braid R (2011) Bertrand-Nash mill pricing and the location of firms with different product
selections or product varieties. Pap Reg Sci 90(1):197–211
Braid R (2012) The location of firms on intersecting roadways. Ann Reg Sci 50:791–808
Capozza DR, VanOrder R (1980) Unique equilibria, pure profits, and efficiency in location models.
Amer Econ Rev 70(5):1046–1053
Davis P (2006) Spatial competition in retail markets: movie theaters. Rand J Econ 37(4):964–982
Launhardt W (1885) Mathematische Begründung der Volkwirtschaftslehre. Engelmann-Verlag,
Leipzig. (Translated by Schmidt H and edited and introduced by Creedy J as Launhardt’s
Mathematical Principles of Economics. Edward Elgar, Aldershot, 1993)
Neidercorn JH, Becholod VB (1972) An economic derivation of the law of retail gravitation: a
further reply and a reformulation. J Reg Sci 12(1):127–136
Reilly W (1931) The law of retail gravitation. Putnam, New York
Ye M-H, Yezer AM (1992) Location and spatial pricing for public facilities. J Reg Sci 32(2):
143–154
Ye M-H, Yezer AM (1993) Local government and supervised spatial multiplant monopoly.
Southern Econ J 59(4):733–748
Chapter 11
Rural School Location and Student Allocation

Ricardo Giesen, Paulo Rocha E Oliveira and Vladimir Marianov

11.1 Introduction

The School Location Problem (SLP) is a network design problem aiming at defin-
ing where to locate schools. Its solution requires also knowing how many and which
students should attend each one of the located schools in order to provide school ser-
vices to the population of a region, so it involves sizing of the schools, either as a
preset parameter or as part of the solution. The determination of what students will
attend which school can be seen as a districting problem, which adjusts the bound-
aries of the zone served by each school within a given school system or network
(Caro et al. 2004; Mandujano et al. 2012). In synthesis, the SLP involves location,
allocation, districting and sizing.
The SLP is rarely a straightforward p-median cost minimization problem as each
school system has a set of individual peculiarities and constraints that must be taken
into account, which vary significantly from case to case due to the different priorities
and policies of various school systems around the world. It is further complicated by
the fact that different students belong to different grades, which must be in general
treated separately, but in the same buildings.
Also, the school location problem in practice is rarely a “pure” cost minimiza-
tion location problem. Decision makers can belong to various government levels,
and the time horizon of the solution can vary significantly from project to project

R. Giesen ()
Department of Transport Engineering and Logistics,
Pontificia Universidad Católica de Chile, Santiago, Chile
e-mail: [email protected]
P. R. E Oliveira
IESE Business School, University of Navarra, Pamplona, Spain
e-mail: [email protected]
V. Marianov
Department of Electrical Engineering,
Pontificia Universidad Católica de Chile, Santiago, Chile
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 273

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_11
274 R. Giesen et al.

(from a horizon that ends with the next elections, to long-term horizons consider-
ing the benefit of the community). The objectives are also very different (racial or
social mix, cost minimization, more/less transportation involved, minimization of
the maximum distance to the nearest school, optimization of the balance between
school sizes, etc.). Therefore, there are many different formulations.
The problem of allocating children to schools first appeared as an optimiza-
tion problem triggered by the U.S. policy of desegregating schools applied in
1954 (Franklin and Koenigsberg 1973). Since then, many authors have studied
this problem considering different optimization techniques and constraints. Ini-
tial approaches used linear programming models (Franklin and Koenigsberg 1973;
Sutcliffe et al. 1984). Other authors addressed this problem using integer and mixed
integer programming models (Barcelos et al. 2003; Liggett 1973; Diamond and
Wright 1987; Church and Murray 1993; Pizzolato and Fraga da Silva 1997; Caro
et al. 2004; Pizzolato et al. 2004; Teixeira and Antunes 2008; Araya et al. 2012;
Mandujano et al. 2012) or heuristics (Pizzolato 1994). Researchers concerned with
the changes of students location and needs over time, addressed it using dynamic
optimization models (Antunes and Peeters 2000; Lemberg and Church 2000). Fi-
nally, there is literature that determines school location within a logit equilibrium
model, considering different types of providers, parental choices based on prices,
locations and other issues (Martínez et al. 2011)
Among the distinctive characteristics often found in the formulation of the school
location problem are the following:
1. Parental choice. In some school systems parents can choose in which school
they want to register their children. The rules governing how this choice takes
place vary significantly from one location to another and are often the subject
of heated political debates. For example, in New Orleans, Louisiana (U.S.), the
school system had to be completely planned from scratch after hurricane Kat-
rina, and one of the most heated political discussions in that context was whether
the city should have community schools or charter schools being families able
to choose sending their children to their preferred school.
2. School size constraints. School size constraints can enter the model for a num-
ber of reasons. Firstly, because of the lot sizes available for school construction
and expansion. Secondly, because of managerial and quality concerns: very
small schools are inefficient from a cost perspective, and very large schools can
be more efficient from an administrative point of view, but more impersonal.
Quality also may depend on the size of the school. Finally, because of peda-
gogical concerns, the number of students per classroom could be constrained.
There are conflicting theories as to whether it is preferable to have standard-
ized schools (all same size) or schools of different sizes to cater each specific
community’s needs. Depending on which side of the debate the decision mak-
ers at hand fall, the corresponding mathematical constraint can be strict or more
relaxed.
11 Rural School Location and Student Allocation 275

3. Number of schools. Restrictions as to the overall number of schools can be

a consequence of school size constraints but can also emerge for other rea-
sons. First, there can be a limit on the locations at which new schools can
be built, exogenously setting a maximum number of schools. But also, there
can be restrictions about closing schools, thus establishing a minimum num-
ber of schools. These decisions are often motivated by political (rather than
cost-related) reasons.
4. Total system capacity. School location studies are sometimes initiated due to
expected changes in demand (increase or decrease) due to demographic changes
or migrations. In occasions, school location studies are triggered by required
overhauls of scholar systems due to new policies regarding quality, the goal of
attracting more students from the private school system to the public schools,
and so on.
5. Shift allocation. Some school systems operate in single shifts, in that case there
is no need to allocate students to shifts, but in other cases school systems operate
in two or even three shifts. In the latter case, school location problems must
also decide which grade levels to operate in what shift and which students will
attend school in each shift. In some cases, a policy is established that assigns
older or younger students together to same shift, but in other cases this could
be a decision variable. In this case cost savings can be achieved by a better
utilization of the infrastructure and the transportation system.
6. Transportation. In many places it is common for local governments to be in
charge of meeting public school transportation needs. In these cases, the cost of
optimal location of schools must also take into account the transportation costs
of students to and from schools. This can bring significant complexity to the
problem, and often requires two separate problems being solved: a long-term
problem with approximate transportation costs, and a short-term problem that
defines the bus routes. Additional complexities that come up in these situations
include defining rules of access to transportation, which can be as simple as a
maximum distance students can walk; or as complex as specifying times of day
when students can walk, dangerous routes that cannot be covered by walking
(in our experience this can be due to a variety of reasons, from high traffic
to wild animals, including areas with high crime rates), and times of the year
when the route can be walked (typically due to weather conditions). Given these
constraints the set of bus stops and assignment of students to each one should
be defined.
7. Staffing levels. In most cases, it is beneficial to predefine student/teacher ratios,
administrative staff per school, and so on.
8. School types. Some systems have pre-defined types, e.g., Kindergarten, Basic,
Middle School, High School. Some countries have systems where the grade
levels of each school can differ within the same system.
9. School districting. Some school systems have very strict rules about school dis-
tricting. For instance, all students must attend the nearest school to their home.
Or the rules can include compactness and contiguity constraints, which are
desirable, in particular, when parents decide what schools are their children
276 R. Giesen et al.

to attend, based on distance to the school. Rules can also include maximum
distances from the residence of each student to the closest school.
10. Student profile balancing. The issue of racial or social balancing and mixed stu-
dent bodies was a particularly important issue in the United States after school
desegregation. In most of the world it is not relevant as such, but other types of
student profile balancing can be important when it comes to allocating resources
to students with special needs. These special needs can be learning disabilities
or language difficulties, which can be particularly important in school districts
with large immigrant or aborigine populations.
11. Rurality. In the case of the school location problem for rural areas, sparse
population further complicates the solutions, as it favors the location of many
schools, so avoiding long travel distances, but it also results in small size, less
efficient schools. Drop-out ratios also tend to be higher in rural schools, because
of quality and travel factors.
Five properties have been declared as desirable in the literature for school location:
• Students’ grades should be modeled separately. As Caro et al. (2004) pointed out,
ignoring students’ grades could only be reasonable if all schools have the same
grade structure and if the grade blend were homogeneous among bus stops. In
rural areas this is hardly the case.
• Schools should consider Continuous Serving Zones (CSZ). As pointed out by
Caro et al. (2004) and then restated by Teixeira and Antunes (2008) in a public fa-
cility planning context, users from neighboring centers should be prevented to be
assigned to different facilities, because it makes the solution harder to understand
and accept by users. This solution property was defined as CSZ.
• The model must allow schools’ closure, opening, and expansion. As the main ob-
jective is to propose more efficient schools’ networks, closing inefficient schools
and expanding others to gain efficiency has to be allowed. This assumes the
presence of economies of scale in schools costs.
• Schools’ instruction costs should be included explicitly. The reason for including
them explicitly is that administrative personnel and teachers’ wages (instruction
costs) are the major component of educational costs (White and Tweeten 1973).
• Multi-graded classrooms schools should be a possibility in rural areas. As pre-
viously pointed out, multi-graded classrooms schools are sometimes the most
efficient solution in rural areas.

Table 11.1 shows a summary of how previous researchers have addressed these five
desirable properties.
To the best of our knowledge, the only model that simultaneously considers
all of these five desirable characteristics is the one proposed by Mandujano et al.
(2012). In the next section, a case study is presented in which this model was used
on instances with data from Brazilian rural-municipalities.
In this chapter, we address two different experiences of school location problems.
The first case corresponds to Barao de Grajaú, in the northeast region of Brazil, and
includes a model that prescribes opening, closing and size of schools, allocates stu-
dents to them, and designs a transportation system for students, as in this case, the
authority in charge also deals with transportation. The second application, covering
11 Rural School Location and Student Allocation 277

Table 11.1 Desirable properties of the school location and sizing problem
Reference Grades Contiguity Closing and ex- Instruction Multi-grade
modeled constraint panding schools costs schools
Franklin and Koenigsberg (1973) Yes – – – –
Liggett (1973) – – – – –
Church and Murray (1993) – – – – –
Antunes and Peeters (2000) – – Yes Yes –
Lemberg and Church (2000) – – Yes – –
Pizzolato et al. (2004) – – Yes – –
Caro et al. (2004) Yes Yes – – –
Teixeira and Antunes (2008) Yes Yes – – –
Araya et al. (2012) Yes – Yes Yes Yes
Mandujano et al. (2012) Yes Yes Yes Yes Yes
(–) No

the whole rural school system in the country of Chile, uses a very similar model,
which minimizes total cost and students’ travel distance. In this case, the transporta-
tion is not included, and the model was intended first as a study not necessarily
to be implemented, but as an evaluation tool for the country’s rural school system.
However, after a very large earthquake in 2010, the model, with a few changes,
was applied to municipalities as a tool for finding the best solutions for school
reconstruction.

11.2 The Brazilian Application

In this section we present a case study using data from different rural municipalities
in Brazil, in which we compared the current situation versus an optimized proposal.
These proposals were obtained using the method developed by Mandujano et al.
(2012), which is presented in the next subsection. After explaining the optimization
model, we show some results of this methodology applied to some rural munici-
palities in Brazil. Then we discuss the main implementation difficulties, and some
thought on how to overcome these complications.

11.2.1 Optimization Approach for the School Location and Size

Problem

The main objectives of the School Location (and Sizing) Problem (SLSP) are
deciding which schools among the current ones should be closed, expanded or main-
tained, and which students based on the bus stops wherein they board should be
assigned to each school. The objective is to minimize the total cost of the school
278 R. Giesen et al.

Table 11.2 Set of parameters of the school location and sizing problem
Parameter Description
studb, g Number of students in educational grade g at bus stop b
distb Road distance between bus stop b and the possible school location 
mccape Multi-graded classroom capacity for educational level e
sccapg Single-graded classroom capacity for grade g
mnclasss Minimum number of classrooms for a school type s
mxclasss Maximum number of classrooms for a school type s
shiftss Number of operating shifts for school type s
class Current number of classrooms at location 

network, which includes: School Administrative Costs (SAC), Classrooms’ Operat-

ing Costs (COC) that correspond to teachers’ wages, and Classrooms’ Construction
Costs (CCC), which account for investments needed for expansions. In addition, a
Classroom Exceeded Capacity Cost (CEC) is considered in order to include flexibil-
ity in terms of classroom capacity. Using these CEC, opening a new class in cases
in which only a few students exceed the limit on the number of student in a class.
Furthermore, a Student Distance Cost (SDC) is considered as a proxy for the student
transportation cost.
The SLSP model decides for each current school location, among the previously
defined set of school types, if any of those types should be located there. If the
type of school to locate is larger than the current school, then new classrooms
will be needed. The definition of school type includes the school size, which is the
maximum number of classrooms, also whether multi-graded (MS) or single-graded
(SS) classrooms are operated, and the number of operating shifts. By modifying the
school assignment of the students in a particular bus stop, economies of scale can be
obtained in terms of school administrative and class operating costs, SAC and COC
respectively, since schools and classroom can be operated near to full capacity.
The mathematical programming model developed in Mandujano et al. (2012)
for solving the SLSP is presented below. Indices b and  represent the current lo-
cation of bus stops and schools respectively. The set BL represent all the possible
assignable pairs within bus stops and schools. Indices e and g represent school lev-
els and grades correspondingly. School levels are group of grades that can be taught
together in multi-graded classrooms, this relationship is represented by set GE[e].
Index s represents the school type. The decision variable Xs equals 1 if a school
type s is located at , and 0 otherwise. The decision variable Cs  is the number of
new classrooms to be built for school type s at location . The decision variables
ZMe and ZSg  are the number of multi-graded and single-graded classrooms used
daily for level e and grade g at location . The decision variables W Me and W Sg 
are the proportion of exceeded capacity on multi-graded and single-graded class-
rooms respectively. The decision variables Y Mb and Y Sb equals 1 if bus stop b
is allocated to a school located at , and 0 otherwise. Table 11.2 presents the set of
parameters used in the model.
11 Rural School Location and Student Allocation 279

The optimization problem can then be written as follows:


min (SACs Xs + CCC Cs ) +
∈L s∈S

+ (MCOCe ZMe + MCECe WMe ) +
e∈E ∈L

+ (SCOCg ZSg + SCECg WSg ) +
g∈G ∈L
 
+ [SDC studbg distb (YSb + YMb )] (11.1)
(b,)∈BL g∈G

s.t. Xs ≤ 1 ∀  ∈ L (11.2)
s∈S

(Y Mb + Y Sb ) = 1 ∀b ∈ B (11.3)
:(b,)∈BL

Y Mb ≤ Xs ∀ (b,) ∈ BL (11.4)
s∈MS

Y Sb ≤ Xs ∀ (b,) ∈ BL (11.5)
s∈SS
 
shiftss mnclasss Xs ≤ ZMe ∀  ∈ L (11.6)
s∈MS e∈E
 
shiftss mnclasss Xs ≤ ZSg ∀  ∈ L (11.7)
s∈SS g∈G
 
ZMe ≤ shiftss mxclasss Xs ∀  ∈ L (11.8)
e∈E s∈MS
 
ZSg ≤ shiftss mxclasss Xs ∀  ∈ L (11.9)
g∈G s∈SS
 
studbg YMb ≤ (ZMe + W Me )mccape ∀ E ∈ e,  ∈ L
b:(b,)∈BL g∈GE[e]
(11.10)

studbg YSb ≤ (ZS g + WS g )sccapg ∀ G ∈ g,  ∈ L (11.11)
b:(b,)∈BL

WM e ≤ ZM e e ∈ E,  ∈ L (11.12)
WS g ≤ ZS g ∀g ∈ G,  ∈ L (11.13)
Cs ≤ mxclasss Xs ∀ s ∈ S,  ∈ L (11.14)
280 R. Giesen et al.
 
ZMe ≤ shiftss (Cs + class Xs ) ∀  ∈ L (11.15)
e∈E s∈MS
 
ZSg ≤ shiftss (Cs + class Xs ) ∀  ∈ L (11.16)
g∈S s∈SS

Y Mb ≤ Y Mc ∀(b,) ∈ BL, c ∈ PBL[b, ] (11.17)

Y Sb ≤ Y Sc ∀(b,) ∈ BL, c ∈ PBL[b, ] (11.18)
Xs ≤ Y Mb ∀  ∈ L, s ∈ MS, b ∈ LZ[] (11.19)
Xs ≤ Y Sb ∀  ∈ L, s ∈ SS, b ∈ LZ[] (11.20)
Xs ∈ {0,1} ∀ s ∈ S,  ∈ L (11.21)

Cs ∈ Z+ ∪ {0} ∀ s ∈ S, l ∈ L (11.22)

Y Mb , Y Sb ∈ {0,1} ∀ (b,) ∈ BL (11.23)

ZMe ∈ Z+ ∪ {0} ∀ e ∈ E,  ∈ L (11.24)

ZSg ∈ Z+ ∪ {0} ∀ g ∈ G,  ∈ L (11.25)

W Me ∈ [0,1] ∀ e ∈ E,  ∈ L (11.26)
W Sg ∈ [0,1] ∀ g ∈ G,  ∈ L (11.27)
Constraints (2) to (5) are variations of the typical location constraints for the p-
median problem. Constraints (2) and (21) ensure that no more than one type of
school is located at each current location. For the allocation variables, constraints
(3) and (23) ensure that all bus stops are allocated to a school (demand covering).
As the allocation can be to multi-graded or single-graded schools, constraints (4)
and (5) ensure that the allocation (multi-graded or single-graded) corresponds to the
type of school located.
Constraints (6) to (16) define the number of operative classrooms (i.e., the num-
ber of teachers needed) at every located school. Because there is more than one shift
for school daily operation, operative classrooms differ from physical ones. Con-
straints (6) and (7) define the minimum number of operative classrooms depending
on the school type. If the school has multi-graded classrooms (MS) the comparison
has to be in terms of levels and if it has single-graded classrooms (SS) in terms of
grades. Constraints (8) and (9) define the maximum number of operative classrooms.
Constraints (10), (11), (12) and (13) balance the number of operative classrooms for
each level/grade with the allocated students at each level/grade. Constraints (14),
(15) and (16) determine the number of new classrooms to build if necessary.
Constraints (17) to (20) help the solution to be better understood by users and de-
cision makers. The set PBL[b, ] define all bus stops that may be in the path between
bus stop b and location . As was pointed out by Teixeira and Antunes (2008), if all
bus stops in the path between bus stop b and location  are also allocated to school
at , the solution makes more sense for users (students and parents) and decision
11 Rural School Location and Student Allocation 281

Table 11.3 Comparison of total costs of current versus proposed solutions

Municipality Year Current total costs USD Proposal total costs USD Variation
(%)
1 2012 3,521,350 2,828,450 − 19.7
2 2014 6,606,273 5,547,899 − 16.0
3 2014 28,854,974 34,012,037 17.9
4 2013 37,123,346 39,088,171 5.3
5 2014 17,682,912 19,391,104 9.7
6 2013 12,402,769 11,607,210 − 6.4
7 2013 8,600,210 8,467,699 − 1.5
8 2013 10,428,987 8,487,340 − 18.6
9 2013 3,411,440 2,996,249 − 12.2
10 2014 6,771,341 5,888,637 − 13.0

makers (educational board). The set LZ[] define a minimum serving zone for ev-
ery current location, and then all bus stops within this zone have to be allocated to
location  if the school at  is not closed. Finally, constraints (21) to (27) define the
nature of the decision variables involved.
This model for the SLSP had been applied, together with a heuristic to solve
the School Bus Routing and Shifts’ Programming Problem (SBRSPP) (Mandujano
et al. 2012), to instances from different municipalities in Brazil showing promising
results. In the next subsection the result of these case studies is discussed.

11.2.2 Analysis of Results

Table 11.3 summarizes results of proposals in ten municipalities in different re-

gions of Brazil, based on the results of the SLSP model presented in the previous
subsection combined with the SBRSPP model. As shown in most municipalities sig-
nificant gains are obtained. Gains reach up to nearly 20 %. However these savings
vary widely among municipalities with some of them without savings. Moreover, in
some instances the total costs of the proposal are higher than the current total costs.
This can be explained since it is common that municipalities operate under level of
service standards that are very different to those proposed by the model, e.g. more
students per classroom, bus routes exceeding its capacity, students traveling more
time than the maximum allowed. In those cases the process of collecting data to run
the model highlighted those inefficiencies. Thus the proposed solution might have
not reduced total costs, but for sure it improved the level of service.
Table 11.4 presents a more detailed comparison of instances 10 and 3 that per-
mits showing the trade off between the level of service and efficiency in design of
a school system in a municipality. For example in the instance 10, an increase in
282 R. Giesen et al.

Table 11.4 Comparison of current versus proposed solutions

Municipality 10 3
Number of schools 20 183
Number of students 3476 14,939
Current Proposed Current Proposed
Avg number of students per class 20.3 20.3 20.9 18.4
Avg distance to school [Km] 0.6 0.7 1.4 1.2
Avg distance traveled [Km] 5.7 5.5 10.9 9.7
Educational costs [USD] 6,303,632 5,218,096 25,609,616 27,424,236
Transportation costs [USD] 467,709 670,541 3,245,358 6,587,801
Total costs [USD] 6,771,341 5,888,637 28,854,974 34,012,037

transportation costs allows an important reduction in educational costs while keep-

ing the same number of students per class and slightly reducing the average distance
traveled by students. On the other hand, in the instance 3 all the performance indi-
cators associated with level of service to students increased, but transportation and
educational costs increased, since in many cases the maximum distances were not
respected, and in other cases the maximum number of students per class was not
enforced.
Even though from a technical standpoint the results are encouraging, there are
many implementation difficulties to translate a sound proposal into practice. These
difficulties and some ideas to overcome them are discussed in the next subsections.

11.2.3 Implementation Difficulties

Rarely is a school system planned from scratch. Therefore, the idea of planning an
optimal school network based on a mathematical model is a theoretical abstraction.
In typical applications, models are used to analyze the current system and propose
changes. In this sense, the optimal location problem can be used as benchmark to
identify opportunities for improving the status quo.
One example of the successful use of school planning model is the work done
in developing countries by UNESCO-IIPE (International Institute for Educational
Planning). They offer education on “school microplanning,” which includes school
mapping and tools for deciding school location. While it is not very sophisticated in
terms of optimization techniques, it is an example of a simple tool that is accessible
to decision-makers that can improve the status quo.
In our experience working with more developed school systems, where mapping
was not an issue, we have seen these models not being implemented for a number
of reasons. Indeed, the authors have never seen the implementation of the optimal
solution of an operational research method without modifications. Typical outcomes
include a meeting with decision makers in which the report is met with varying
11 Rural School Location and Student Allocation 283

levels of acceptance and one or two ideas are taken up for further discussion. The
study is typically requested by the department of education, and it implementation
often involves other departments (e.g., city planning, transportation, construction).
Furthermore, there are a number of additional stakeholders (politicians, teachers,
parents, etc.) each of which has their own agenda.
In mathematical terms, this means that there are too many variables to take into
account. So the way to proceed is to simplify the problem and make the required ad-
justments later. These adjustments, intended to overcome the barriers can sometimes
hinder implementation overall. It is essential for the operations research modelers
who want to see their solutions implemented to talk closely to the decision makers
to understand how these barriers may play out.

11.2.4 Overcoming Difficulties

In this subsection we discuss some key difficulties faced when implementing

optimization based proposals and some ideas to overcome them.
1. Transition plan. The outcome of the optimization model can be desirable, but
it is often not clear how to go from the current state of things to the optimized
proposal. Two aspects of a transition plan can help to make implementation more
likely. First, it is important to consider a financial model. Construction is often
required; these costs are typically not covered by the operational budget. There-
fore, it is fundamental to take into account how should municipalities pay for
these costs. Secondly, a logistics plan: if all the necessary construction cannot be
done during school holidays, which school will the students attend and how will
they get there during the transition phase? Decision-makers often do not know
how to answer these questions, and in absence of an answer implementation is
stopped.
2. Talking to communities. Implementation often involves closing schools. This can
require working with the communities, justifying the new schools. Politicians
may not want to close or open schools in certain locations for reasons that are
completely unrelated to cost optimization.
3. Dealing with unions. Teachers unions, when present, could object to anything
that changes the status quo in schools. In one situation we encountered, a very
strong union did not want teachers commuting long distances. As a consequence,
the students from the suburbs were being bused to schools in a central location
in order to keep the teachers’ commute within the acceptable level.
4. Political Cycles. Decision makers think in “election (e.g. 4-year) cycles.” De-
pending on the political structure, the decision is made by a major who does not
think beyond the next election. The benefits of the optimized solution are often
not seen in such a short term.
5. Educators in the way of the decision. Because of the biases in their training,
educators typically do not understand operations research models and methods
284 R. Giesen et al.

or their value. Rather than trying to engage with, and understand the models,
the reaction we most often encounter is a defense of the status quo based on
“pedagogical reasons.”
6. Renegotiations with suppliers. Changing schools to different locations can re-
quire renegotiating contracts with a number of external service providers, such
as caterers, security, and transportation. These negotiations can be politically
difficult, particularly when there are long-term contracts in place.
7. Too many decision makers. Modelers need to understand the various decision
makers involved and understand their role in the decision making process. School
location decisions often go beyond the education authorities. Thus, to guarantee
implementation, it is essential to understand the political environment.

11.3 Selected Experiences in Chile: Rural School Location

and Sizing

Answering to a requirement of the Ministry of Education, an analysis was done of

the effects of closing some rural schools, opening new schools and resizing existing
schools, over the total cost of the system, as well as over the distances traveled by
students to go to their closest school, in the whole country of Chile (15 regions).
Since most of rural schools are public, the project considered only this type of
schools. The results are reported in detail in Araya et al. (2012). Later, after the 8.8
magnitude earthquake that affected several regions of the country in 2010, the model
was partially used, together with a heuristic, to provide the Ministry of Education
with a tool for deciding what schools to reconstruct.
The Chilean school system is composed by public and private schools. The pri-
vate sector, in turn, includes schools that are financed strictly by student fees, and
subsidized schools. The last ones receive mostly middle and low-income students
and, in some cases, the subsidy is complemented by a small monthly fee paid by
students1 . Around a 9.5 % of school students in Chile live in rural areas. School
education in rural areas is provided by some 4000 schools, both public (managed
by municipalities) and subsidized, many of them small and run by one teacher who
teaches in a multigrade mode, i.e., to students belonging to several (up to four)
grades simultaneously. In multigrade schools, when the number of students is very
small, the quality of education can be good (as in home schooling). However, as
the student number increases, the quality deteriorates quickly, which is the case in
most of them in Chile. In regular schools, there are in all 13 grades: preschool, basic
primary school to the eighth grade, and four high school grades. In general, the stu-
dents must travel long distances to attend the school, which increases the drop-out
rate.

1 By the time this chapter is being written, the existence of these schools is being discussed.
11 Rural School Location and Student Allocation 285

Addressing these problems is complex and includes different actions, including

relocating and merging schools, as well as integrating transportation networks into
the analysis and the solution, as it was shown in the Brazilian example. Relocating
and merging schools was possible in Chile to an extent. However, unfortunately, as
opposed to the Brazilian case, the authority in charge of education in Chile does not
have the freedom to establish its own transportation system, or hiring third party
transportation, so the problem has to be addressed considering that students use
public transportation or their own resources, e.g., cars or hired minivans. This is the
main difference between the Brazilian and Chilean cases we discuss in this chapter.
The demand for schools comes from “rural entities”, which are small villages
or hamlets with populations that do not exceed 2000 people. There are around
26,000 of these entities. All the students from an entity must be assigned to the same
school. Entities close enough to towns with urban schools are not considered, as stu-
dents prefer attending urban schools when these are close enough. In our model, the
different types of schools include existing schools as they are, new schools with dif-
ferent compositions of grades (preschool, basic, high, preschool-basic; basic-high,
all-grade schools) and, existing schools that are to be closed as a consequence of the
optimization. In existing schools, the number of classrooms can change according
to the allocated students. Closures are limited in terms of percentage, as it is not
politically viable to have too many closures in a region. Budget is also limited. The
costs and parameters are very similar to the Brazilian case, as well as the model
(except for the transportation component).
Almost 300,000 students were considered, 26,000 rural entities, more than 4000
existing schools and each entity could be a candidate location for a new school.
With these numbers, the problem became almost intractable from the point of view
of obtaining optimal solutions. Furthermore, many rural entities were too close to
each other, which would result in many solutions with the same objective value
that unnecessarily delay obtaining an optimal solution and multiply the number of
alternate optima.
In order to reduce the number of potential locations of new schools, we first
solved a Location Set Covering Problem (LSCP, Toregas et al. 1971) formulated as
follows:

Min Cj
j
 
s.t. Ej + Cj ≥ 1 ∀ i,
j ∈Ni j ∈Ni

where Cj is 1 if j becomes a candidate location; Ej is 1 if there is an existing school

with at least 100 students at j, and Ni is the set of all sites within 10 km of location i.
Note that the solution to this problem minimizes the number of candidate locations,
so that all entities i have either a school or a candidate location within 10 km. The
number of candidates is reduced in almost 90 %. Furthermore, the administrative
division of the country enforces students from one region not attending schools
286 R. Giesen et al.

in a different region. Thus, the divisions between regions are a natural boundary
for dividing and conquering this very large problem. Still, the largest regions have
associated problems with millions of variables, out of which, hundreds of thousands
are binary. We divided these regions in two or three parts, using natural boundaries
such as rivers.
For the largest regions, we allowed closure of up to 40 schools. For the remaining
regions, closures are bounded above by 20 or a 20 % of the existing schools. We
also bounded the distance to be traveled by students to 30 km, except in the least
populated, extreme regions in the far south (fjords) and far north (desert) to 50 km.
A model was used that is very similar to that of the Brazilian case. Finding a
good solution required, in cases, a heuristic, as the optimal solution was not found
within 3 h, even using an integrality gap of 5 %. For more details, see Araya et al.
(2012).
A baseline solution was found, assuming all students assisting to their nearest
schools. The results of the new model, compared to the baseline are shown in
Table 11.5. We solve the problem in two versions: using the travel distance limit
of 50 km in extreme regions of the country, and relaxing this constraint for these
regions only, keeping, in both versions, the distance limit of 30 km in the remaining
regions. Note that even though the number of schools is reduced in the new situa-
tions, the average travel distance is significantly reduced, as well as the maximum
distance, when it is constrained in the extreme regions. Also, the cost decreases in
17 % when the total distance is constrained in the extreme regions, and in 18 % in
the unconstrained case.
It is interesting to remark what were the practical issues that had to be addressed
during the project. In the first place, the data required for carrying the project was
scattered among several different divisions at the Ministry of Education. There are
divisions in charge of financial aspects, quality aspects, political issues, and so on,
and at the time of the project there was no sharing of the data among them, so the
research group acted as a consolidation center for the information. Secondly, the ex-
istence of multi-grade schools had to be accepted, even though when such schools
increase in size, the quality of the educational results decreases significantly. On the
other hand, small multi-grade schools seem to provide a very good education, sim-
ilar to home schooling. Thirdly, the investment money and the operational money
come from different divisions, and each one seeks its own optimum. The opera-
tional expenses decrease importantly if the solutions are implemented; however,
there is investment to be done, which makes the corresponding divisions oppose to
any change.
Finally, closing a school is extremely unpopular, and a strong opposition appears
whenever a school is in danger of being closed. This makes the implementation of
the results of the model very difficult.
However, with a few changes, the model was used in practice after the Febru-
ary 27, 2010, 8.8 magnitude earthquake that affected a large area in Chile. After
this event, approximately 2600 urban and rural schools were lightly damaged; 2200
schools suffered moderate damage; 210 severe damage and 90 were destroyed.
11 Rural School Location and Student Allocation 287

Table 11.5 Comparison of the baseline and the proposed solutions

Baseline Solution (50 km) Solution (unconstrained
distance)
Number of schools Existing 4165 3851 3851
New 235 173
Closed 314 314
Reduced classrooms 471 522
Distance traveled by Average 21.9 8.5 13.7
student [km] Maximum 329.0 50.0 371.2
Minimum 0 0 0
Students per school Average 76 75 80
(Existing) Maximum 2059 1658 1658
Minimum 0 0 0
Average 64 78
Students per school (New)
Maximum 565 565
Minimum 1 1
Average unused capacity 28 % 30 % 15 %
Students per teacher 15.6 15.6 18.1
Total cost MMUS$ 415.8 352.6 348.8

This time, the model was applied to both urban and rural schools, not as a pre-
scriptive model, but as a helping tool for making decisions about which damaged
schools to reconstruct; what schools to move to a different place, or whether pairs
of schools were to be closed, and a new school opened in a new place instead. The
objective of the model (cost, average distance from school to students’ home) was
complemented with a quality objective. In this case, a proxy for quality was obtained
from standardized tests that are being applied to all schools in the country. The qual-
ity depends mainly on the personnel of each school (teachers, director). With these
changes, the model was used by the Ministry of Education to negotiate with the Mu-
nicipalities in charge of the schools the amount of investment to be assigned to each
damaged school (or a replacement school in a different place), depending on cost,
travel distance and quality. In some cases, the use of the model allowed the Ministry
of Education to convince both Municipalities and parents of replacing the damaged
schools by new schools at a different location.
The application of the model in this case, reduced the operational cost per student
by US $ 27, and the travel distances in 6 %, while increasing the quality, measured by
an average increase in 3.4 points in the standardized national tests, for the considered
schools. Additionally, there were savings of US $ 22 million in damage schools
repairs.
288 R. Giesen et al.

11.4 Conclusions

In this chapter, we presented a particular application of location models in public

policy, namely the problem of school sizing and location and student assignment
to schools in rural areas, discussing the main complexities and differences with
common location problems. We discuss in detail a methodology for optimizing the
number and location of public schools in rural areas which had been applied to dif-
ferent rural-cities in Brazil, and briefly describe a similar problem, but covering a
whole country, in Chile. This School Location and Sizing Problem (SLSP) includes
properties from school location models, school districting models and also peculiar-
ities of rural schools allowing multi-graded classrooms in schools. The results in
different regions in Brazil show gains in efficiency up to nearly 20 %, even though
the benefits had a lot of variability, which can be explained since the level of service
provided by current solutions vary widely among different cases and the level of ser-
vice of the proposed solution respected standards such as maximum travel time per
student. In the Chilean case, the average distance travelled by students is decreased
in a 37 % or a 61 %, depending on the distance constraints in extreme regions. The
total cost of the system is also significantly reduced.
In addition, a discussion on the main implementation difficulties and some ideas
on how to overcome them were included. Among the key recommendation for a
successful implementation of optimization-based proposal are: to provide a good
transition plan; to implement a communication plan to persuade different communi-
ties about the benefits; to establish an agreement with unions and educators so that
they become proactive and supporters of the plan; and understanding the window
of opportunity in the political cycle, so that the plan could be implemented during
the period of the champion of the project or with the support of different political
groups, so that it can be continued after new authorities take control.
In other cases, as after the Chilean earthquake, circumstances help the authorities
in the decision of using operations research tools.
We think that a critical challenge for operations research practitioners to have
an impact on school locations problems is to carefully understand the priorities of
all decision makers involved. In these politically motivated settings, the criteria for
making the decision often evades what operations research modelers consider to be
rational cost minimization. Yet, often these alternative criteria can be brought into
the model in one way or another.
Therefore, we think that there are essentially two ways for operations researchers
to have more impact in practice as far as the school location model goes. In the short
term, implementation is more likely to the extent that models are customized to each
decision setting. In the long term, it is in the interest of all involved parties that
operations research practitioners engage with professionals and politicians involved
in school planning in discussions about the relationship between school efficiency
and school quality, particularly in what has relation with size of the schools and
their cost. Thus, we see a big opportunity in exploring the relationship between the
operational efficiency of school systems and education quality, bringing efficient
11 Rural School Location and Student Allocation 289

operations to the political agenda, and finding effective ways to convey the important
of efficiency to the general public.

Acknowledgements Giesen and Oliveira gratefully acknowledge partial support by the Instituto
Alfa e Beto, Brazil. Marianov gratefully acknowledges partial support by the Institute Complex
Engineering Systems through Grants ICM MIDEPLAN P-05-004-F and CONICYT FBO16.

References

Antunes A, Peeters D (2000) A dynamic optimization model for school network planning. Socio-
Econ Plan Sci 34(2):101–120
Araya F, Dell R, Donoso P, Marianov V, Martínez F, Weintraub A (2012) Optimizing location and
size of rural schools in Chile. Int Transa Oper Res 19(5):695–710
Barcelos FB, Pizzolato ND, Lorena AN (2003) Localização de escolas do ensino fundamental com
modelos capacitado e não-capacitado: caso de Vitória/es. Pesqui Operacion 24(1):133–149
Caro F, Shirabe T, Guignard M, Weintraub A (2004) School redistricting: embedding GIS tools
with integer programming. J Oper Res Soc 55(8):836–849
Church RL, Murray AT (1993) Modeling school utilization and consolidation. J Urban Plann Dev
119:23–38
Diamond JT, Wright JR (1987) Multiobjective analysis of public school consolidation. J Urban
Plann Dev 113(1):1–18
Franklin AD, Koenigsberg E (1973) Computed school assignments in a large district. Oper Res
21(2):413–426
Lemberg DS, Church RL (2000) The school boundary stability problem over time. Socio-Econ
Plan Sci 34(3):159–176
Liggett RS (1973) The application of an implicit enumeration algorithm to the school desegregation
problem. Manage Sci 20(2):159–168
Mandujano P, Giesen R, Ferrer J-C (2012) An optimization model for location of schools and
student transportation in rural areas. Transp Res Rec 2283:74–80
Martínez F, Tamblay L, Weintraub A (2011) School locations and vacancies: a constrained logit
equilibrium model. Environ Plan A 43:1853–1874
Pizzolato ND (1994) A heuristic for large-size p-median location problems with application to
school location. Ann Oper Res 50(1):473–485
Pizzolato ND, Fraga da Silva HB (1997) The location of public schools: evaluation of practical
experiences. Int Trans Oper Res 4(1):13–22
Pizzolato ND, Barcelos FB, Lorena NAL (2004) School location methodology in urban areas of
developing countries. Int Trans Oper Res 11(6):667–681
Sutcliffe C, Boardman J, Cheshire P (1984) Goal programming and allocating children to
secondary schools in Reading. J Oper Res Soc 35(8):719–730
Teixeira JC, Antunes AP (2008) A hierarchical location model for public facility planning. Eur J
Oper Res 185(1):92–104
Toregas C, Swain R, ReVelle C, Bergman L (1971) The location of emergency service facilities.
Oper Res 19(6):1363–1373
White F, Tweeten L (1973) Optimal school district size emphasizing rural areas. Amer J Agr Econ
55(1):45–53
 Part III
Enforcement and First Responders
Chapter 12
Fire Station Siting

Alan T. Murray

12.1 Introduction

Location analysis and modeling has long been relied upon to support a wide array of
public facility siting efforts (ReVelle 1987; Marianov and Serra 2002; Church and
Murray 2009). As noted in Daskin and Murray (2012), public sector policy and plan-
ning often involves multiple constituents, various competing/conflicting objectives,
politically driven decision making and constraints on available resources, making
policy and planning problems “wicked” (see Liebman 1976). Addressing issues and
making good decisions therefore requires a delicate balance of skill and finesse that
is supported by data and quantitative methodologies (Brill 1979). In this chapter
the focus is on a particular aspect of public sector policy and planning, namely the
siting of facilities. Public facilities are necessary to provide services to people in
communities, towns and cities, and often this must be done with access, accessi-
bility, equity, fairness, fiscal prudence, etc., in mind. Examples of public facilities
along these lines include post offices, parks, schools, police stations, firehouses, etc.
Public financing and support of these services continues to be the norm because of
their widely held necessity and value.
The emphasis in many modeling and siting approaches is the design of a new
system of facilities. Often ignored or overlooked is the necessity of systematic
re-evaluation and assessment of existing facilities in relation to anticipated future
conditions. Public facilities in particular reflect evolving needs and incremental
changes, so an existing system of facilities is a byproduct of additions and clos-
ings over time. When new facilities are needed to complement existing facilities,
it is essential that systematic and comprehensive evaluation take place (Schilling
et al. 1980; Francis et al. 1992; Badri et al. 1998; Murray 2010, 2013). Most urban
areas have an existing system, referred to as a brownfield system (in contrast to the
Greenfield case where no existing facilities exist). This is noteworthy because an

A. T. Murray ()
Center for Spatial Analytics and Geocomputation, College of Computing and Informatics, School
of Public Health, Drexel University, Philadelphia, PA, 19104, USA
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 293

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_12
294 A. T. Murray

existing system may suggest a prevailing idea of expansion, yet broader integrated
planning may in fact be warranted that involves closing or relocating one or more
existing facilities. Doing so may well turn out to be more cost effective and efficient
over the long term, but cannot be attained if it is not sought or explored through
planning and analysis.
The purpose of this chapter is to address the need for fire and emergency response
in a region experiencing changing social, economic, residential, employment, etc.
conditions. In particular, the goal is to design a fire station response system in an
urban area. This need arises due to urban growth, development and/or decline, where
an area already has a number of existing fire stations and must plan for anticipated
future conditions. Thus, new facilities combined with a re-evaluation of existing
services not only makes sense, it is a prerequisite for sound public management of
scare resources targeted to provide essential services.
A number of things make the problem of locating fire stations interesting, spe-
cial and different from the broader context of facility location modeling. First, fire
stations are very much response oriented. It is extremely important that emergency
personnel arrive to a call for service within a stipulated response standard, typically
something in the range of 5–9 min. The reason for this is that quick response means
that lives and property can be saved (Murray and Tong 2009; Murray 2013). Second,
it is recognized that rapid response may not always be possible, but should occur
most of the time. Third, there are real fiscal implications for building, equipping
and manning a fire station. The costs per year to man a fire station are significant,
essentially that of the fixed costs of establishing a station (Plane and Hendrick 1977;
Badri et al. 1998; Murray et al. 2012). Care should therefore be taken to provide suf-
ficient services, now and over time, without bankrupting or financially crippling a
community. Finally, many important advances have been made in location model-
ing to accommodate the richness of spatial data, and application oriented studies
dealing with public facilities must recognize and exercise care associated with both
substantive context as well as technical considerations in carrying out analysis.

12.2 Review

As suggested throughout this book, there are a range of location models that have
been applied in both descriptive and prescriptive ways to address a variety of
public and private sector planning issues. There has, in fact, been an extensive
amount of research devoted to the use of location models in fire and emergency
response (ReVelle 1991; Marianov and ReVelle 1995; Badri et al. 1998; Sorensen
and Church 2010; Murray 2013).
Hogg (1968) was among the early efforts to advocate systems analysis in fire pro-
tection planning, suggesting the problem of minimizing loss through the selection of
a minimum number of stations (and where to locate them). This marked the begin-
ning of theoretical research on location models to support such a planning situation,
particularly oriented toward fire station siting. Toregas et al. (1971) is noteworthy
12 Fire Station Siting 295

for establishing a connection to the notion of coverage, where response from a fire
station is critical within a distance/time standard. They demonstrated that coverage
of a demand point could be assessed for a potential facility location given a response
standard (e.g., 6 min, 2 km, etc.). With this in mind, Toregas et al. (1971) detailed
the location set covering problem (LSCP) as an approach for identifying the min-
imum number facilities, like fire stations, and where they should be sited in order
to serve a region in a timely manner reflected in time/distance response standards.
Church and ReVelle (1974) recognized that it may not be economically viable to
expect all demand in a region to be served as imposed in the LSCP for emergency
response facilities such as fire stations. That is, while most of the time response
should be within the desired standards, on occasion slight delays in response may
be unavoidable and acceptable. To address this, Church and ReVelle (1974) pro-
posed the maximal coverage location problem (MCLP) seeking to maximize service
demand coverage provided through the location of a pre-specified number of facili-
ties. They demonstrated that considerable cost savings could be achieved compared
to the number of facilities identified using the LSCP, but a high percentage of
demand could still be ensured coverage within desired service standards.
A number of fire station siting studies have followed based on coverage mod-
els. Plane and Hendrick (1977) applied the LSCP to site fire stations in Denver,
Colorado. Schreuder (1981) also utilized the LSCP for fire station placement in
Rotterdam, Netherlands. Schilling et al. (1980) proposed an MCLP based approach
for fire response in Baltimore, Maryland. Reilly and Mirchandani (1985) developed
a combined median and coverage approach for fire service placement in Albany,
New York. Badri et al. (1998) structured a model to minimize the maximum dis-
tance demand was from a sited fire station, applying this in Dubai, United Arab
Emirates. They ensured that a maximum distance standard was not exceeded, which
is essentially what is imposed in the LSCP. In Singapore, separate studies by Liu
et al. (2006) and Huang and Fan (2011) relied on an MCLP based approach for fire
station siting. Coverage models were also used in Catay (2011) as well as Aktas
et al. (2013) for fire station planning in Istanbul, Turkey. Murray et al. (2012) and
Murray (2013) report the use of an MCLP oriented model in fire station system
design for the community of Elk Grove, California. Chevalier et al. (2012) applied
an extension of the MCLP to site fire brigades in Belgium.
What is evident then is that coverage models like the LSCP and MCLP are impor-
tant and widely applied in fire station siting because they explicitly account for
response time standards. Not only this, but such response standards are often artic-
ulated in community, state and/or federal guidelines regarding fire service provision
(Murray 2013). While clearly important and reflective of fire response needs and
goals, the LSCP and MCLP, as well as extensions and variants, have been shown
to be sensitive to scale and demand unit representation issues. Murray and O’Kelly
(2002) demonstrated that point based representations intended to reflect continuous
potential service demand across a region actually under-estimates spatial coverage
provided when the LSCP is relied upon. Alternatively, Murray et al. (2008) showed
that polygon based representations of a region will result in an excessive number
of facilities identified as needed using the LSCP. Similar findings are reported for
296 A. T. Murray

the MCLP in Tong and Murray (2009). The implication is that spatial represen-
tation using LSCP or MCLP based modeling approaches to support facility siting
could be a complicating factor capable of biasing analysis results in unintended
(and undesirable) ways. More importantly, this could lead to poor strategic planning
and decision making. Extensions to address bias/error issues in coverage models
have been developed in Murray (2005), Murray et al. (2010), Tong (2012), among
others, and coverage location modeling remains an active area of research.
Coverage models are important approaches that encapsulate a range of con-
cerns central to fire response, as well as other public facilities, and therefore are
an essential component of strategic planning involving the placement stations.
While accounting for existing facilities is recognized as important, few studies
have attempted to incorporate potential system changes in utilized approaches, yet
such considerations are fundamental for the long term fiscal management of a pub-
lic service system. Contemporary approaches that take advantage of geographic
information system (GIS) capabilities and spatial data richness have seen limited
use in fire station siting studies. The contributions of this chapter revolve around
addressing the above issues in support of planning an integrated system of fire
stations.

12.3 Location Planning Context

As reported in Murray et al. (2012), the author was commissioned by the Commu-
nity Services District (CSD) Fire Department in Elk Grove, California to provide
analysis supporting insight into how many future fire stations would be needed and
where they should be sited. When the study began, Elk Grove had a population of
approximately 144,000 people over 10,900 ha served by nine existing stations (three
stations were not yet operational). The city and fire stations are shown in Fig. 12.1.
Their established standard was response within 5 min of a call for service. Anal-
ysis of call history, performance and response conducted by the fire department
suggested that this length of time corresponded to an effective distance of 1.90 km
from a station. Overall, the Elk Grove CSD Fire Department strived to maintain a
5 min response standard for at least 80 % of service calls.1 This suggests that an
MCLP based location model to support analysis and planning is particularly appro-
priate as this is an approach that accounts for coverage based standards intent on
maximizing service to as much demand as possible.
The above conditions were to be assumed in the analysis. Further, the initial
study was only to focus on the portion of the city where growth and development

1 This standard is not unlike those found in many communities, though the time and performance

percentage may vary. The National Fire Protection Association (NFPA) (2010), as an example,
advocates that fire stations in urban areas be sited so that response times of nine minutes are
achieved at least 90 % of the time.
12 Fire Station Siting 297

Fig. 12.1 Elk Grove region with 9 existing fire stations

was anticipated. The population of Elk Grove is expected to more than double, with
most of this anticipated in the southern half of the region. Expansion of the fire
response system is seen as essential. New potential station sites were deemed possi-
ble anywhere in the area of expected growth. The fire department was not interested,
at the time, in altering existing stations. Based on stipulated conditions, the analysis
reported in Murray et al. (2012) indicated that 9 additional fire stations would be
necessary, bringing the total number of fire stations to 18.
An important question remains about whether a better long term plan is possi-
ble for the community. As noted previously, existing systems are a byproduct of
evolving needs over time. The resulting configuration may therefore become ineffi-
cient in a number of ways, potentially stressing or overburdening services in some
areas while leaving some stations underutilized. This of course creates inequities,
but also suggests system inefficiencies. It is precisely these issues that are important
to address, especially when there are long term financial implications for inefficient
system design. Most studies view (or are forced to view) existing services to be
fixed and given, with the analysis focused on where new stations should be sited.
This was essentially the case in Kanoun et al. (2010) and Murray et al. (2012), as
an example. Plane and Hendrick (1977), Badri et al. (1998), Schilling et al. (1980),
Catay (2011), Murray (2013) and Aktas et al. (2013), among others, discuss and/or
provide evidence that simply expanding an existing system is problematic in the
sense that system inefficiency may be introduced or perpetuated. Certainly there is a
fiduciary responsibility to ensure that expenditures on fire response services are kept
298 A. T. Murray

to a minimum, and doing so may mean that reconfiguration of the existing system
is paramount in order to realize increased efficiency.
As suggested above, some studies have attempted to address these issues within
the context of an existing system in various ways. Plane and Hendrick (1977)
accounted for a tradeoff in using a combination of new and existing facilities.
Badri et al. (1998) incorporated fixed and annual costs in their goal programming
approach. Schilling et al. (1980) detailed an approach to add a constraint restrict-
ing the number of existing facilities to utilize. This too was incorporated into the
approach of Murray (2013). A different sort of approach is to assume that no facili-
ties exist in the first place, then carry out the analysis to provide comparison of what
could be possible if a region were able to begin anew. This was done most recently
in the study reported in Aktas et al. (2013). Unfortunately, each of these approaches
is lacking in certain ways, detached from specifically addressing issues of long term
expenditures. Murray (2013), in fact, had to examine each of the various combi-
nations of keeping existing facilities, then calculate the financial implications for a
specific period of time. This was done following the analysis (and model applica-
tion), requiring many different scenarios to be considered. What really is necessary
is an explicit approach to address long term costs for a fire station response system,
combining existing and new station decision making.
The following notation will be utilized in the derivation and discussion of the
applied location model:
i index of demand areas to be served;
ai demand for service in area i;
Ni set of potential fire stations that serve any portion of area i within the
standard;
j index of potential future fire station locations;
cj annual staffing and maintenance costs for fire station j ;
fj fixed cost to build a fire station at site j ;
bij portion of area icovered by fire station j within stipulated service standard;
 set of sites where a new fire station could be built;
 acceptable level of regional demand suitably served;
λ temporal horizon for which cost considerations apply;

1 if fire station at site j is selected to be in system
Xj =
0 otherwise

Zi amount of area i demand served by any fire station

With the above notation, it is possible to formalize the measures and criteria relied
upon in the location model. To begin, a critical issue is system costs. In siting new
stations, the prevailing cost would be building the fire station and acquiring the
land. Given the set of potential locations for a new station, , the fixed cost for site
j would be fj . The total costs for new facilities would therefore be:

fj Xj (12.1)
j ∈
12 Fire Station Siting 299

where Xj denotes decisions about which facilities are selected. It is not uncommon
that such fixed costs would essentially be the same. Thus, in the LSCP the objective
is simply to minimize the total number of facilities, which is equivalent to minimiz-
ing total costs when the fixed cost at each potential fire station site is the same. Of
course, fixed costs are not the only concern. Plane and Hendrick (1977), Badri et al.
(1998) and Murray (2013) report that the annual costs of operating a fire station is
equivalent to the fixed cost. The total annualized operational cost of the fire response
system is the following:

cj Xj (12.2)
j

This is across all stations, existing or new. If even one less fire station is possible
through more efficient system design, taking into account existing and new stations,
there are tremendous savings possible over the long term. This is precisely why it is
important to explicitly consider existing and new stations when there are changing
demands for service.
The objective of the LSCP is minimizing total fixed costs while ensuring that each
demand is suitably covered. This is accomplished by imposing a service response
constraint for each demand i:

Xj ≥ 1 (12.3)
j ∈Ni

This guarantees that the minimum number of facilities found using the LSCP, objec-
tive (12.1), are configured so that at least one fire station is located within the
response standard for each demand. Noted above was that the Elk Grove response
standard is service within 5 min (1.90 km). In contrast to the LSCP, the MCLP rec-
ognizes that serving the entire region under strict response standards may not be
cost effective. The implication is that constraints (12.3) must effectively be relaxed
for select demand. To address this, the MCLP has as an objective to maximize
demand suitably served within the standard. Specifically, Elk Grove has stipulated
that response within 5 min should occur for at least 80 % of service calls. Based on
our notation here, the MCLP objective is:

Zi (12.4)
i

as Zi is the demand covered by the fire station siting configuration. The MCLP also
imposes a particular level of system investment:

Xj = p (12.5)
j

where p is the number of fire stations to site, specified a priori. Murray and Tong
(2009) highlight that analysis must therefore be conducted to identify the minimum
300 A. T. Murray

investment level, p, that provides the stipulated level of regional coverage, . Typ-
ically the tradeoff curve showing regional demand coverage for each value of p,
1 facility, 2 facilities, 3 facilities, etc., is consulted, as illustrated in Church and
ReVelle (1974). It is possible, however, to relate coverage in percentage terms as
follows:
 −1
 
100 ai Zi (12.6)
i i

The terms before the summation simply standardize Eq. (12.4) based on total
demand in the region, giving a percentage of demand covered by the fire station con-
figuration. One can then impose a bound on this percentage as a model constraint.
Murray and Tong (2009) demonstrate that it is possible to use this as a replacement
for objective (12.4) and constraints (12.5), representing an extension of the classic
MCLP that they refer to as the Threshold Coverage Model.
What remains to be discussed is tracking demand coverage using decision vari-
ables Zi . The LSCP and MCLP originally were conceived assuming that demand
was represented as a point. Coverage in this case is fairly straightforward. A point is
either covered or it is not. Some ambiguity arises, however, when demand is actually
a line, ployline, polygon or other geometric object. Assessment of coverage may not
be clear when the demand is only partially covered/served within the response stan-
dard. As noted previously, it is precisely this representational issue that research
has found to introduce bias, error and uncertainty in LSCP and MCLP application
(see Murray and O’Kelly 2002; Murray et al. 2008; Tong and Murray 2009). To
address this issue, Murray (2005) introduced the idea of complementary coverage
for the LSCP in the case where demand is a line, polyline, polygon, etc. The idea is
that there is an implicit expectation of demand being served based on spatial com-
plementarity, but this must be structured in the model. An alternative is to explicitly
account for portions of demand being covered by a facility siting configuration. This
was detailed for the MCLP in Tong and Murray (2009) and requires enumeration of
specific coverage configuration. A summary review of implicit and explicit coverage
approaches can be found in Murray et al. (2010). For the MCLP, Tong (2012) pro-
posed an implicit approach tracking the portion of demand area i served by selected
facilities:

bij Xj (12.7)
j ∈Ni

where Ni is the set of facilities within the coverage standard of demand i and bij
represents the amount of area i served by a facility sited at location j. While not
exact in terms of the actual portion of demand i served, empirical evidence suggests
that it can be used in an effective manner to track coverage. This basic approach was
relied upon in Murray (2013).
With the above details and discussion, it is possible to now formalize an MCLP
based model extension to support fire station siting, taking into account existing and
new services as well as spatial representation issues. The model is as follows:
12 Fire Station Siting 301

 
Min fj Xj + λ cj Xj (12.8)
j ∈ j
 −1
 
s.t. 100 ai Zi ≥  (12.9)
i i

bij Xj ≥ Zi ∀ i (12.10)
j ∈Ni

Zi ≤ ai ∀ i (12.11)
Xj = {0, 1} ∀ j (12.12)
Zi ≥ 0 ∀ i (12.13)

The objective, (12.8), minimizes total costs over time horizon λ, including fixed
and annual recurring expenses. Constraint (12.9) establishes the percentage of
regional demand that must be covered within the stipulated service standard. Con-
straints (12.10) implicitly track the portion of a demand served by sited stations.
Constraints (12.11) bound possible coverage of a demand area. Binary integer
and non-negativity requirements are imposed in constraints (12.12) and (12.13),
respectively.
This model is an extension of the MCLP, related to the Threshold Cover-
age Model detailed in Murray and Tong (2009). This model further extends the
so called Complementary Threshold Coverage Model detailed in Murray (2013),
enabling long term financial expenditures over time, λ, to be explicitly account for.
Unlike the MCLP, the number of facilities required is endogenously determined
based on achieving regional coverage requirements stipulated in constraints (12.9).
The model is also structured to accommodate alternative spatial representations of
demand, including points, lines, polygons, etc., addressing issues that could arise in
the application of the LSCP and MCLP for non-point demand.

12.4 System Insights

The analysis is carried out using a Windows-based computer (Intel Xeon processor
at 2.53 GHz) with 6 GB of RAM. The location model was structured using Python,
relying on GIS based libraries (Shapely), and solved using commercial optimization
software (Gurobi).
There are 638 neighborhoods representing potential demand for fire response
(all polygons in Fig. 12.1). As fire stations can technically be sited anywhere in
continuous space, processing to identify a finite dominating set of discrete potential
locations was carried out. Murray and Tong (2007) derived a finite dominating set
for coverage based problems like the LSCP and MCLP involving non-point demand
302 A. T. Murray

Fig. 12.2 Minimum long term cost siting configuration

objects. They proved that this set would contain an optimal solution for the con-
tinuous space context when facilities may be located anywhere in the region. The
resulting set in this case consisted of 1638 points. This set combined with the nine
existing station sites gives 1647 potential fire station locations for consideration. A
time horizon of 10 years was used (λ = 10). The assumed fixed cost for a new fire
station is $2 million (fj for j ∈ ), with the annual cost per station of $2 million
(cj ). The location model contained 2285 decision variables (1647 integer and 638
continuous) and 1277 constraints. The data processing time to structure the model
was 325.4 s and solution time was 3389 s.
As noted above, if the existing system is fixed and given with new stations to be
sited in order to cover currently unserved portions of region, then 18 fire stations (9
existing and 9 new) would be necessary to achieve at least 80 % system response
stipulated in constraint (12.9). That is, response within 5 min to at least 80 % of
calls for service would be ensured for a configuration of 18 stations. Considering a
time horizon of λ = 10, the total costs for Elk Grove to provide fire service over 10
years would be $378 million (18 stations costing $2 million a year per station over
10 years plus nine new stations at a cost of $2 million per station). The question of
interest is whether this total cost can be decreased.
Applying the location model in order to consider siting decisions involving
existing as well as new fire stations finds that 16 fire stations are sufficient. This
configuration of fire stations is shown in Fig. 12.2. Evident is that 2 existing stations
12 Fire Station Siting 303

Fig. 12.3 Minimum short term cost (1–4 year horizon) siting configuration

are maintained supplemented by 14 new stations. It is therefore possible to provide

suitable response to over 80 % of demand in Elk Grove using only 16 fire stations.
The cost to provide this service will be $348 million over 10 years. Needless to
say, this is a substantial reduction in costs over a rather arbitrary time horizon. Still,
$30 million over 10 years is a significant savings for any community. Of course, a
longer period of time would mean greater realized savings. Worth noting is that if
λ is varied and the model reapplied, alternative recommendations could be reached
based upon a shortened return on investment window. For example, 1–4 years would
suggest a 17 fire station system consisting of 8 existing stations and 9 new stations.
This configuration is shown in Fig. 12.3. While appealing because is makes use of
most of the existing system, after 5 years it becomes financially more prudent to
select the minimum number of stations (16) depicted in Fig. 12.2.

12.5 Future Steps

There are many issues that arise in any application oriented study. These issues
span articulation of the problem but also practical considerations regarding data,
projected demand for service, response, personnel availability, etc., not to mention
304 A. T. Murray

decision making context, political and social preferences. Bridging theoretical and
application concerns is no simple task, making it a challenging area to work in.
There is no doubt more to be done in this area given better supporting data and
greater insights. Highlighted below are some of the primary considerations driving
planned future research.
Somewhat unexpected was that the Elk Grove CSD Fire Department merged with
a neighboring community, the Galt Fire Protection District, shortly after the comple-
tion of the project. This combined with the economic downturn continues to thwart
plan implementation, with no changes to date to the fire response system. The future
remains uncertain at this point, yet the population has increased by over 10 %.
Irrespective of system implementation or change issues, the research did not
address timing considerations of bringing a new system online. Of course this is
not trivial as communities like Elk Grove have real budgetary limitations, but also
growth and development is a process that is controlled by economic and individual
decision making. A second issue involved expected demand for service. The cur-
rently served portions of Elk Grove are largely developed. However, the unserved
areas are mostly agricultural with residential development expected over the coming
years. Addressing expected demand required an assumption of uniformly distributed
demand for service. The CSD Fire Department made this assumption based on call
history, but there does remain a source of uncertainty regarding projected service
demand. This may or may not impact fire station location decisions. Extensions to
address projected growth over time combined with annual investment possibilities
over time are interesting areas for future research. Of course, this requires good
future demand predictions and cost estimates to better refine model parameters, like
ai , cj and fj , where temporal specificity is possible. If t represents an index of
time periods, it is conceivable that demand, annual cost and fixed cost could be
indicated temporally as ait , cjt , and fjt , respectively. No doubt there would need to
be subsequent modification of the model to account for this enhanced detail and
specificity.
Further location model attention would be spatial representation and detail. The
reviewed literature has discussed recognized issues in spatial representation known
to impact results identified using coverage models. While the complementary cover-
age approach reported here is certainly an improvement, it is not free of all potential
error or uncertainty (Murray et al. 2010; Tong 2012). This is precisely why research
continues on coverage location model approaches for reducing and/or eliminating
representation error and uncertainty, such as that by Murray and Wei (2013) and Wei
and Murray (2014, 2015). This too remains a promising area for future research.

Acknowledgements This work benefitted from assistance and comments by Dr. Ran Wei
(University of Utah).
12 Fire Station Siting 305

References

Aktas E, Ozaydın O, Bozkaya B, Ulengin F, Onsel S (2013) Optimizing fire station locations for
the Istanbul metropolitan municipality. Interfaces 43:240–255
Badri MA, Mortagy AK, Alsayed A (1998) A multiobjective model for locating fire stations. Eur
J Oper Res 110:243–260
Brill ED (1979) The use of optimization models in public-sector planning. Manage Sci 25(5):
413–422
Catay B (2011) Siting new fire stations in Istanbul: a risk based optimization approach. OR Insight
24:77–89
Chevalier P, Thomas I, Geraets D, Goetghebeur E, Janssens O, Peeters D, Plastria F (2012)
Locating fire-stations: an integrated approach for Belgium. Socio-Econ Plan Sci 46:173–182
Church RL, Murray AT (2009) Business site selection, location analysis and GIS. Wiley, New York
Church R, ReVelle C (1974) The maximal covering location problem. Pap Reg Sci Assoc 32:
101–118
Daskin MS, Murray AT (2012) Modeling public sector facility location problems. Socio-Econ Plan
Sci 46:111
Francis RL, McGinnis LF, White JA (1992) Facility layout and location: an analytical approach,
2nd edn. Prentice Hall, New Jersey
Hogg JM (1968) The siting of fire stations. Oper Res Q 19:275–287
Huang Y, Fan Y (2011) Modeling uncertainties in emergency service resource allocation. J
Infrastruct Syst 17:35–41
Kanoun I, Chabchoub H, Aouni B (2010) Goal programming model for fire and emergency service
facilities site selection. INFOR: Inf Syst Oper Res 48:143–153
Liebman JC (1976) Some simple-minded observations on the role of optimization in public
systems decision-making. Interfaces 6:102–108
Liu N, Huang B, Chandramouli M (2006) Optimal siting of fire stations using GIS and ant
algorithm. J Comput Civil Eng 20:361–369
Marianov V, ReVelle C (1995) Siting emergency services. In: Drezner Z (ed) Facility location: a
survey of applications and methods. Springer, New York, pp 199–223
Marianov V, Serra D (2002) Location problems in the public sector. In: Drezner Z, Hamacher HW
(eds) Facility location: applications and theory. Springer, Berlin, pp 119–150
Murray AT (2005) Geography in coverage modeling: exploiting spatial structure to address
complementary partial service of areas. Ann Assoc Am Geogr 95:761–772
Murray AT (2010) Advances in location modeling: GIS linkages and contributions. J Geograph
Syst 12:335–354
Murray AT (2013) Optimising the spatial location of urban fire stations. Fire Saf J 62:64–71
Murray AT, O’Kelly ME (2002) Assessing representation error in point-based coverage modeling.
J Geograph Syst 4:171–191
Murray AT, Tong D (2007) Coverage optimization in continuous space facility siting. Int J Geogr
Inf Sci 21:757–776
Murray AT, Tong D (2009) GIS and spatial analysis in the media. Appl Geogr 29:250–259
Murray AT, Wei R (2013) A computational approach for eliminating error in the solution of the
location set covering problem. Eur J Oper Res 224:52–64
Murray AT, O’Kelly ME, Church RL (2008) Regional service coverage modeling. Comput Oper
Res 35:339–355
Murray AT, Tong D, Kim K (2010) Enhancing classic coverage location models. Int Reg Sci Rev
33:115–133
Murray AT, Tong D, Grubesic TH (2012) Spatial optimization: expanding emergency services to
address regional growth and development. In: Stimson R, Haynes K (eds) Studies in applied
geography and spatial analysis. Edward Elgar, UK, pp 109–122
National Fire Protection Association (NFPA) (2010). NFPA 1710: Standard for the organization
and deployment of fire suppression operations, emergency medical operations, and special
operations to the public by career fire departments. http://www.nfpa.org/. Accessed 13 April
2015
306 A. T. Murray

Plane DR, Hendrick TE (1977) Mathematical programming and the location of fire companies for
the Denver Fire Department. Oper Res 25:563–578
Reilly JM, Mirchandani PB (1985) Development and application of a fire station placement model.
Fire Technol 21:181–198
ReVelle C (1987) Urban public facility location. In: Mills E (ed) Handbook of regional and urban
economics, volume II. Elsevier, New York, pp. 1053–1096
ReVelle C (1991) Siting ambulances and fire companies—new tools for planners. J Am Plan Assoc
57:471–485
Schilling DA, ReVelle C, Cohon J, Elzinga DJ (1980) Some models for fire protection locational
decisions. Eur J Oper Res 5:1–7
Schreuder JAM (1981) Application of a location model to fire stations in Rotterdam. Eur J Oper
Res 6:212–219
Sorensen P, Church R (2010) Integrating expected coverage and local reliability for emergency
medical services location problems. Socio-Econ Plan Sci 44:8–18
Tong D (2012) Regional coverage maximization: a new model to account implicitly for comple-
mentary coverage. Geogr Anal 44:1–14
Tong D, Murray A (2009) Maximizing coverage of spatial demand for service. Pap Reg Sci 88:
85–97
Toregas C, Swain R, ReVelle C, Bergman L (1971) The location of emergency service facilities.
Oper Res 19:1363–1373
Wei R, Murray AT (2014) Evaluating polygon overlay to support spatial optimization coverage
modeling. Geogr Anal 46(3):209–229
Wei R, Murray AT (2015) Continuous space maximal coverage: insights, advances and challenges.
Computers and Oper Res (66: 325–336)
Chapter 13
Locating Vehicle Identification Sensors for
Travel Time Information

Monica Gentili and Pitu B. Mirchandani

13.1 Introduction

Currently, transportation networks are significantly instrumented with various types

of sensors, ranging from simple inductive loop detectors embedded in the pavement
that produce a signal when a vehicle is over it, to image detectors which use cam-
eras machine vision to detect activities over the pavement, to mechanical gauges
to measure the strain on the pavement and associated infrastructure like bridges,
to GPS-equipment in the vehicles that is be able to track vehicles in the network
(commonly referred to as traffic probes when they are used to obtain a congestion
measure of traffic such as travel times). Machine vision and laser-based readers are
used to automatically collect tolls on a highway. Traffic sensors may also be used to
detect incidents on a road by observing a sudden change in traffic flow patterns. All
these uses are ultimately related to (a) managing traffic on a network by monitor-
ing traffic and making traffic control decisions to decrease congestion, energy use,
air pollution, etc. and (b) planning networks so that the demand using the road net-
work is appropriately served. Thus, locating sensors optimally improves our traffic
management and network planning decisions.
The problem of optimally locating sensors on a traffic network has been the
object of growing interest in the past few years. Sensor location decision models
differ from each other according to the type of sensors that are to be located and
the objective function that one would like to optimize. Many different models have
been proposed in the literature as well as corresponding solution approaches. The
proposed existing models could be classified according to two main criteria: (i) the

M. Gentili ()
Department of Mathematics, University of Salerno, Fisciano, Salerno, Italy
e-mail: [email protected]
P. B. Mirchandani
School of Computing, Informatics and Decision Systems Engineering,
Arizona State University, Tempe, AZ, USA
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 307

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_13
308 M. Gentili and P. B. Mirchandani

typology of sensors to be located on the network (e.g., counting sensors, image sen-
sors, Automatic Vehicle Identification (AVI) readers), and, (ii) the objective function
to be optimized, such as minimizing errors in estimating Origin- Destination (OD)
trips, route flows, link flows, or travel times. Among the first models, Lam and
Lo (1990) addressed the problems of locating sensors to estimate OD flows and
studied how different deployments of sensors on the network could affect OD flow
estimation. Since then, studies on locating sensors have grown rapidly due to the
development of new methods and technologies. Indeed, traditionally, counting sen-
sors have collected speed and flow volume data, but sensors based on technologies
such as vehicle recognition and GPS can give much more information about the
operation of the network. Hence new objective functions have been studied: route
flow estimation and observability, link flow estimation and observability, travel time
measurement and estimation, etc.
The authors have recently reviewed the literature related to sensor location for
flow observability and estimation (Gentili and Michandani 2011, 2012). The focus
of this chapter is on the optimal location of sensors to monitor travel time related
information.

13.2 Travel Time Measurement and Estimation

Real-time measurement and estimation of travel times is a useful tool for traf-
fic management and for providing information to travelers. For example, this
knowledge is used for dynamically updating travel time predictions. Tradition-
ally, collection of information for travel time estimation has been carried out by
transportation agencies through an extensive use of counting sensors (i.e., loop
inductance detectors) uniformly located on freeways and highways usually at dis-
tances of 0.5–1.0 miles. The origin of such a practice is based on past analysis
that has shown uniform spacing among detectors is useful for Automatic Incident
Detection algorithms (ADI) (the role of ADI algorithms based on loop detector data
is addressed in Chap. 7 of this volume). However, it should be noted that given
the massive use of GPS enabled smart phones, the data from the resulting probe
vehicles’ travel times are increasingly being used for incident detection as well
as travel time estimation. Nevertheless, for situations where this kind of informa-
tion is not readily available, either because of low market share of smart phones
or when data communication networks are not sufficiently advanced or when com-
munication of vehicle data is expensive, the location of fixed permanent sensors
is still an important and relevant approach to gather real time data for estimating
travel times. The effect of counting sensor deployment on the quality of the mea-
surements and estimation of travel times has been widely studied (see for example,
Ban et al. 2009; Bartin et al. 2007; Chan and Lam 2002; Chaudhuri et al. 2011;
Danczyk and Liu 2011; Edara et al. 2010; Fujito et al. 2007; Liu et al. 2006). Travel
time measurement can also be obtained using technology such as Automatic Vehi-
cle Identification (AVI) readers. Examples include sensors that use machine vision
13 Locating Vehicle Identification Sensors for Travel Time Information 309

Fig. 13.1 A simple network

example

to read license plates, sensors that read unique blue tooth signals from each vehicle
that passes in its vicinity, and sensors that read a radio electronic tag much like the
RFID tags used to track parts in manufacturing facilities. The remainder of the chap-
ter will review existing approaches to optimally locating AVI readers on networks to
estimate travel times and related measures.

13.3 Travel Time and Related Information on Networks

The use of AVI readers for estimating roadway travel times is a recent development.
Their primary application has been for electronic toll collection. AVI reader systems
rely on a combination of passive tags attached to vehicles and electronic readers
installed on the roads (above lanes, or on roadsides). A vehicle is detected and iden-
tified every time it passes a reader. Upon detection of a vehicle, the system transmits
the information to a data processing center where a measurement of the travel time
between two consecutive readers can be computed. For example consider the net-
work in Fig. 13.1. By locating two AVI readers, one on link (4, 7) and one on link (8,
9), all the vehicles that pass both locations are intercepted, and, by comparing the
interception times, information on the travel time on the portion of the intercepted
network can be determined. Obviously, one could be interested in determining travel
310 M. Gentili and P. B. Mirchandani

Table 13.1 OD pairs, OD trips and OD routes related to the example network in Fig. 13.1
OD pair Route ID Trips Links
1–3 1 100 (1, 2) (2, 3)
1–9 2 10 (1, 2) (2, 3) (3, 4) (4, 7) (7, 8) (8, 9)
7–11 3 10 (7, 8) (8, 9) (9, 10) (10, 11)
5–13 4 10 (5, 4) (4, 7) (7, 8) (8, 9) (9, 12) (12, 13)

times on links of the network, on particular routes in the network, or travel times
between OD pairs in the network. Optimal location decisions depend on what travel
times one wants to estimate.
Assume, for example, the network of Fig. 13.1 has the following OD pairs W =
{(1, 3), (1, 9), (7, 11), (5, 13)}, each of them connected, in this simple example,
by a single route (these are listed in Table 13.1). To get information about travel
times between OD pair w = (1, 3) one needs to locate AVI reader at the upstream
and downstream end-points of the route connecting w, that is one on link (1, 2) and
the other on link (2, 3). Hence, to have information about travel times among all the
OD pairs in the network, AVI readers should be located on links (1, 2), (2, 3), (7,
8), (8, 9), (10, 11), (5, 4) and (12, 13). This would provide knowledge of the travel
times on the entire routes connecting all the OD pairs in the network. However, in
case of multiple routes between OD pairs and also when there are budget limits to
acquire sensors, such detailed knowledge is unlikely to be collected. In this case, one
cannot get information about travel times on all entire routes but only on portions
of some of them. For example, by locating readers on links (4, 7) and (8, 9), travel
times on parts of routes two and four could be collected. Consider now the OD trips
listed in Table 13.1. If the purpose of determining travel times is to disseminate such
information among commuters, then the majority of users would not benefit from
the reader located on links (4, 7) and (8, 9). It would be preferable to locate one of
the readers on link (1, 2) and the other on (2, 3) to intercept as many commuters as
possible for as long as possible.
It is clear, then, that different factors should be considered when locating AVI
readers. Some of them are:
• Number of AVI readers: this is a measure of the installation cost of the system.
• Number of readings: a reading is obtained when the same vehicle is intercepted
at two different locations. This measure relates to the amount of travel time data
that is obtained from the AVI system.
• Length between readings: a reading obtained by two readers that are far apart has
more travel time information compared with readings obtained by two readers
that are close to each other.
• Number of OD pairs covered: For an O-D pair to be covered, at least one reading
should be obtained from one of the routes between the OD pair. This is useful
13 Locating Vehicle Identification Sensors for Travel Time Information 311

for estimating OD flows, and estimating OD travel times when the network is in
equilibrium (Sheffi 1984).
• Travel time reliability: If travel time along a link has little variance, then a real-
time travel time reading on that link has little value; if variance is large, then real-
time travel times are very useful. Hence, sensors can collect real-time information
to estimate travel time reliability.
The models reviewed in the next section take into account some of these factors in
developing the locational decision criteria and associated models. In particular, six
different mathematical formulations are presented to optimally locate readers on the
links of the network, which we refer to as AVI Location Problems or AVIL models.

13.4 Mathematical Models

We can represent a traffic network by a graph G = (V, A) where the set of nodes
V represents intersections in the network and the set of links A, joining node pairs,
represents roads. We denote by W the set of OD pairs of the network and by Rw the
set of routes connecting the OD pair%w ∈ W. The entire set of routes in the network
is denoted by R and we have R = Rw . We denote by hw the average number of
w∈W
trips connecting the OD pair w for the period of study, and by fr the average flow
volume on route r. Additional notation will be introduced as needed.
The first model (model AVIL1 ), presented by Sherali et al. (2006), works under
the assumption that there is a single route connecting each OD pair. This model
associates different weights (benefit) to OD pairs and seeks to maximize the total
benefit accrued by locating at most k readers, respecting a given budget constraint.
An OD pair is considered to be covered if the readers are located on the route con-
necting the OD pair at the upstream and the downstream end-points of the route
(particularly, at the beginning and ending links). In particular, let us consider the
OD pair w ∈ W and consider the shortest route rij connecting it, that starts with
link i and ends with link j. Due to the simplifying assumption of having one route
for each OD pair, we can use route rij to represent its corresponding OD pair. Route
rij is defined to be covered if readers are located both on link i and on link j. When
the route rij is covered, a benefit uij is obtained. Let cj be the cost of installing a
reader on link j. Define zj to be a binary variable associated with link j ∈ A that
assumes a value of 1 if a reader is located on the link and 0 otherwise. The AVIL1
model is the following:

AVIL1 : Max uij zi zj (13.1)
rij ∈R

s.t. zj ≤ k (13.2)
j ∈A

cj zj ≤ Cmax (13.3)
j ∈A
312 M. Gentili and P. B. Mirchandani

zj ∈ {0,1} ∀ j ∈ A (13.4)

The objective function (13.1) maximizes the total benefit accrued by locating the
readers. Constraint (13.2) limits the sensors to at most k readers. Constraint (13.3)
imposes the total budget to be at most Cmax . A possible definition of the benefit
factor is also proposed by Sherali et al. (2006). They defined the benefit factor uij to
describe the relevance of route rij (and hence, of its corresponding OD pair) in cap-
turing information about the variability of travel times in the network. Such factors
are computed according to the coefficient of variation associated with each link j,
defined as follows:
σj
COV j = , (13.5)
μj

where μj and σj are estimates of the mean and standard deviation of travel times
on link j, respectively. For each route rij , the corresponding benefit factor uij is
computed considering the coefficients of variation (13.5) associated with the links
that belong to the route. In particular, the suggested benefit uij is computed as:
&
'
' a∈rij σa2
uij = COV ij = ( . (13.6)
a∈rij μa
2

Sherali et al. (2006) linearized model AVIL1 and applied the Reformulation Lin-
earization Technique (Adams and Sherali 1986, 1990) to solve it. They applied the
model to data from Interstate-35 Freeway, San Antonio, Texas, USA.
The more realistic assumption of having multiple routes between each OD pair is
considered in the other models AVILi , i = 2, . . ., 6. These models assume an OD pair
to be covered if there are at least two readers located on each of the routes connecting
the OD pair, without any restriction on where on the routes they are located. Model
AVIL2 to follow is basically a set covering formulation aimed at minimizing the total
number of readers to be located on the network to ensure the coverage of all the OD
pairs. Let zj be the number of sensors located on link j. Model AVIL2 is as follows
(Chen et al. 2004).

AVIL2 : Min zj (13.7)
j ∈A

s.t. zj ≥ 2 ∀ r ∈ Rw ∀ w ∈ W (13.8)
j ∈r

zj ∈ {0,1, 2, . . .} ∀ j ∈ A (13.9)

The objective function (13.7) minimizes the total number of readers to be located
on the network. Constraints (13.8) ensure that each OD pair is covered. Since the
location variable zj associated with link j denotes the total number of readers that
can be installed on the link, the total number of readers located on a link can be
13 Locating Vehicle Identification Sensors for Travel Time Information 313

greater than 1. The maximum number of readers that can be installed on each link of
the network can be determined by the spacing requirement between pairs of readers.
Moreover, note the slightly different definition of coverage of an OD pair compared
to the one used in model AVIL1 : an OD pair w is considered to be covered when
there are at least two readers installed on each route r ∈ Rw connecting the OD
pair. Unlike model AVIL1 where for each route the only locations for readers that
are considered are on the first and last link of the route, readers for model AVIL2 can
be located anywhere on the route.
When the total number of readers is limited, a different objective function has
been proposed. The following model, AVIL3 (Chen 2004), is a variant of the well
known maximal covering location problem (Church and ReVelle 1974) where, for
a fixed number of available readers to be installed, the objective is to maximize the
total number of covered OD pairs.

AVIL3 : Max yw (13.10)
w∈W
⎧ ⎫
⎨  ⎬
s.t. yw ≤ max 0, zj − 1 ∀ r ∈ Rw ∀ w ∈ W (13.11)
⎩ ⎭
j ∈r

zj ≤ k (13.12)
j ∈A

zj ∈ {0,1, 2, . . .} ∀ j ∈ A (13.13)
yw ∈ {0,1} ∀ w ∈ W (13.14)

Variables yw , defined for each OD pair, assume value equal to 1 if the OD pair w
is covered and 0 otherwise; and, therefore, objective (13.10) maximizes the total
number of ODs covered. The coverage is ensured by the set of constraints (13.11)
which force yw to be equal to 0 unless there are at least two readers located on each
route connecting w. The total number of installed readers is limited to k through
constraint (13.12).
The main differences between model AVIL2 and AVIL3 with respect to the clas-
sical set covering problem and the maximal covering problem, respectively, are: (i)
multiple location of readers on the same link is allowed, and (ii) the definition of
OD pair coverage requires the installation of at least two devices on each covered
route connecting the OD pair.
The following model AVIL4 locates k readers on the links of a network to max-
imize the weighted sum of the total number of readings and of the total number of
covered OD pairs (Teodorovic 2002):
⎛ ⎞
h 1  ⎝ h2 
AVIL4 : Max fr z j − 1⎠ + yw (13.15)
H |W |
r∈R j ∈r w∈W
314 M. Gentili and P. B. Mirchandani
⎧ ⎫
⎨  ⎬
s.t. yw ≤ max 0, zj − 1 ∀ r ∈ Rw , ∀ w ∈ W (13.16)
⎩ ⎭
j ∈r

zj ≤ k (13.17)
j ∈A

zj ∈ {0,1, 2, . . . } ∀ j ∈ A (13.18)
yw ∈ {0,1} ∀ w ∈ W (13.19)

where (i) h1 , h2 , are the weights assigned to model the importance of the two com-
ponents in the objective and (ii) H is the total number of trips in the network (i.e.,
H = hw ). The criterion (13.15) maximizes the weighted sum of two compo-
w∈W
nents in the objective: the first component is the ratio between the total number of
readings obtained from the locations of the readers and the total trips in the net-
work; and the second component of the objective function is the ratio between the
total number of covered OD pairs and the total number of OD pairs in the network.
When a weight uw is assigned to each OD pair w as a measure of importance, then
the second component of the objective function could also be: h2u uw yw .
w
w∈W w∈W
The coverage of the OD pairs is considered by the set of constraints (13.16) which
force yw to 0 unless there are at least two readers located on each route connecting
w. Constraint (13.17) assumes that the total number of installed readers cannot be
greater than k.
Distances among two located readers (that is, the length of a reading) is not
considered in any of the above models. In AVIL5 this measure is defined as the total
vehicle-miles monitored. In AVIL6 it is a general benefit factor. These two models
are explained next.
Model AVIL5 solves the problem of optimally locating k readers on the links
of the network to maximize the number of vehicle-miles (or vehicle-travel times)
monitored. For each link j ∈ A and each route r ∈ R the parameter djr denotes
the distance from the origin of route r to a potential reader located on link j. The
following sets of variables are defined for AVIL5 : the binary variable zj equals to 1
if a reader is located on link j and 0 otherwise; the binary variable yrj equals to 1 if
a reader on link j is the most downstream reader located on route r and 0 otherwise;
the binary variable xrj equals to 1 if a reader on link j is the most upstream reader
located on route r and 0 otherwise. The mathematical formulation is the following
(Mirchandani et al. 2009):
) *
AVIL5 : Max yrj − xrj djr fr (13.20)
r∈R j ∈r

s.t. zj ≤ k (13.21)
j ∈A

yrj ≤ zj ∀ r ∈ R ∀ j ∈ A (13.22)
13 Locating Vehicle Identification Sensors for Travel Time Information 315

xrj ≤ zj ∀ r ∈ R ∀ j ∈ A (13.23)

yrj ≤ 1 ∀ r ∈ R (13.24)
j ∈r

xrj ≤ 1 ∀ r ∈ R (13.25)
j ∈r
 
yrj − xrj = 0 ∀ r ∈ R (13.26)
j ∈r j ∈r

zj ∈ {0,1} ∀ j ∈ A (13.27)
yrj ∈ {0,1} ∀ r ∈ R ∀ j ∈ A (13.28)
xrj ∈ {0,1} ∀ r ∈ R ∀ j ∈ A (13.29)

Objective function (13.20) maximizes the total vehicle-miles monitored. Note that
this function takes into account both the length of each reading and the total number
of readings. Constraint (13.21) requires locating no more than k readers on the net-
work. Constraints (13.22) and (13.23) are logical constraints linking the variables
and ensuring that there cannot be a most upstream (or a most downstream) reader
on a route if no reader is located on it. Constraints (13.24) and (13.25) ensure that
there is no more that one most upstream reader and no more than one most down-
stream reader on each route. Finally, constraint (13.26) states that if there is a most
upstream reader on route r, there must also be a most downstream reader on the
same route and vice versa. Similarly, if there is no most upstream reader on route r,
then there cannot be a most downstream reader on the same route and vice versa. It
is also true that if on a route there is only one reader installed, then it is both the most
upstream and downstream reader on that route, and the vehicle miles monitored on
that route is zero.
The objective function of AVIL5 can be modified so that the model tries to locate
AVI readers to obtain travel time statistics and improve predictability of estimated
travel times. This model, referred to as AVIL5 , is illustrated in Sect. 5 where its
application is discussed.
A generalized surveillance benefit (Li and Ouyang 2012) is maximized in model
AVIL6 . Consider a route r, and, without loss of generality, let us assume every link
of the route is a candidate location for a reader. If sr readers are located on the
route, then we can divide the route into sr − 1 sections where each section has a
reader on its upstream boundary link and its downstream boundary link. Without
loss of generality, we also assume there are two virtual locations b and t for each
route, located at the very beginning and at the very end of the route with two virtual
readers located at positions 0 and sr + 1, respectively. Denote the portion of the
route between link j and link h by rjh . We say that the portion rjh is monitored if (i) a
reader is located on link j and on link h, and (ii) there is no reader located in between
(that is, the two readers on j and h are consecutive). If rjh is a monitored portion,
then the two links j and h are said to be paired on route r. More specifically, given
316 M. Gentili and P. B. Mirchandani

Fig. 13.2 Monitored portions of a route and corresponding benefit factors

a route r and sr readers located on it, the link of the route where the i-th reader is
located is paired with the link where the (i + 1)-st reader is located, and with the link
where the (i–1)-st reader of the route is located. The virtual initial link b at position
0 is paired with link j1 where the first real reader is located; the virtual last link t
at position sr + 1 is paired with link jsr where the last real reader, with position sr ,
is located. Figure 13.2 depicts a route r with five links, that is j = 1, 2, 3, 4, 5. Two
virtual links b and t are added to the route, and two virtual readers are assumed to
be located on them. Suppose sr = 3 readers are located on the route, on links one,
three, and five. Then, the monitored portions of the route are: rb1 , r13 , r35 , and r5t .
The virtual link b is paired with link one which is also paired with link three, which
in turn is paired with link five, and so on.
Finally, if the i-th reader of the route is located on link j and the (i + 1)-st reader
on the route is located on link h, then we denote by εjr h = εjri ji+1 the benefit asso-
ciated with the monitored section rjh (see example in Fig. 13.2). The total benefit
accrued with the given locations of the readers is then:

sr
ε= εjri ji+1 (13.30)
r∈R i=0

To introduce model AVIL6 , which seeks to locate a limited number of readers to

maximize (13.30), some additional notation is needed. Given a route r and a link
j ∈ r, the set Fjr+ is the set of links of the route after link j where a reader can
be located, while the set Fjr− is the set of links of the route before link j where a
reader can be located. The set Fjrh is the set of links of route r between link j and
link h where a reader can be located. Hence, for example, if we consider the route
r = {2,3, 4,5, t}, the set F r = {b} and the set F r = {3,4}.
in Fig. 13.2, the set F1+ 1− 25
Finally, let us define the binary variable yjr h to be equal to 1 if the portion of route r
between two readers located on link j and link h is monitored and 0 otherwise. The
mathematical formulation for AVIL6 is as follows (Li and Ouyang 2012):
  
AVIL6 : Max yjr h εjr h (13.31)
r∈R j ∈Ft−
r h∈F r
j+

s.t. zj ≤ k (13.32)
j ∈A
13 Locating Vehicle Identification Sensors for Travel Time Information 317

yjr h = zj ∀ j = t ∈ r ∀r ∈ R (13.33)
h∈Fjr+

yjr h = zh ∀ h = b ∈ r ∀r ∈ R (13.34)
j ∈F r−
h

zb = zt = 1 (13.35)
zj ∈ {0,1} ∀ j ∈ A (13.36)
yjr h ∈ {0,1} ∀ h, j ∈ r ∀ r ∈ R (13.37)

Constraint (13.32) imposes that the total number of readers to be installed be less
than or equal to k. Constraints (13.33) pair link j of route r where a reader is located
(that is, when zj = 1), with only one link h in the route that follows j. Constraints
(13.34) pair link h of route r where a reader is located (that is, when zh = 1), with
only one link j in the route that precedes h. Finally, the objective function (13.31)
maximizes the total benefit accrued by the location as defined by (13.30). For pos-
sible examples of how benefit factors εjr h can be defined, the reader can refer to Li
and Ouyang (2012).
Models AVILi , i = 1, 2, . . ., 6 were introduced and developed in the last few
pages to indicate how these models have been evolving, introducing more and more
constraints and objectives. The first model (model AVIL1 ) assumes that there is a
single route connecting each OD pair, which happens only in special sparse net-
works, for example in rural areas. Model AVIL2 (which minimizes the total number
of readers to be located on the network to ensure that all OD routes are observed)
and model AVIL3 (which maximizes the total number of OD routes observed with
a limited numbers of available readers), should be applicable in the cases when the
observation effort does not increase as the volume of traffic on the network changes,
and hence applicability is limited. On the other hand, when the traffic manager or the
transportation planner wants to monitor all the traffic on the network or to maximize
the amount of traffic monitored, then model AVIL4 is more appropriate.
Distances among two located readers is not considered in any of the Models
AVILi , i = 1, . . ., 4. In AVIL5 this measure is defined as the total vehicle-miles moni-
tored. This is generalized in AVIL6 where a general benefit factor is considered for an
OD route monitored, which could be some function of travel distance, travel time, or
some other measure of benefit to the planner or manager. The application described
next focuses on the travel time reliability of the OD routes monitored, where here
“reliability” essentially means how well the travel time can be predicted. If there
is no variance in travel time, then it is highly predictable, whereas when the travel
time has large variance its predictability is low. Installing sensors in the latter case
will help the traffic manager predict travel times better as real-time measurements
are made.
318 M. Gentili and P. B. Mirchandani

Fig. 13.3 Study area of Harris County, TX

13.5 Applications

In this section we discuss two applications of model AVIL5 to a portion of the Harris
County (Texas) network (see Fig. 13.3). One of the authors was working with the
County on a traffic management research project, and as a by-product they were
interested in if they could use toll tag readers to monitor traffic in a certain area.
Consequently, we developed two different models for (i) the problem of optimally
locating the toll tag readers (these are effectively AVI readers) on the arcs of the net-
work to maximize the number of vehicle-miles (or vehicle-travel times) monitored
(model AVIL5 presented in the previous section) and (ii) the problem of optimally
locating AVI readers on the arcs of the network to maximize the ability to predict
travel times on the network. The motivation underlying the first problem is for traffic
managers to have an idea of the traffic volume in the region at any given time. The
motivation underlying the second problem is to monitor travel times and communi-
cate to commuters some predictions of the mean travel times of their routes. In this
case, the underlying model is assumed to be stochastic. That is, each arc (or route
segment) will have a mean travel time that changes slowly over time during the day,
while the actual travel time has a further fluctuation which is due to differences in
driving characteristics of the commuters; this latter fluctuation may be treated as
some sort of additive noise. In particular, let τj denote the travel time on link j. τj
13 Locating Vehicle Identification Sensors for Travel Time Information 319

is therefore a random variable with mean μj and variance σj2 . This variance σj2 is
assumed to be constant over time and known for all links in the network. It is also
assumed that for any time period, the a priori distribution of μj is known, which is
Normal with mean ηj and variance γj . Therefore, the variance of travel time on link
j has two components, that is, var(τj ) = σj2 + γj . For the short period of time when
travel time data is collected, the expected travel time on link j is ηj . Knowing the
values of ηj and var (τj ), a confidence interval for τj can be obtained. If n obser-
vations of, the average value of the observed travel times (denoted τ̄j ) are collected,
then according to the Central Limit Theorem τ̄j can be approximated by a normal
distribution with mean μj and variance n1 σj2 . With an observation of τ̄j , the a priori
distribution of μj can be updated by using Bayesian statistical theory. The updating
can be performed using the following formulas:

n ηj σj + τ̄j γj
1 2
η̂j = (13.38)
n σj + γ j
1 2

n γ j σj
1 2
γ̂j = (13.39)
n σj + γ j
1 2

Therefore, the expected value of τj is η̂j and var(τj ) = σj2 + γ̂j . Note that the
variance of μj is reduced after the updating:
γj2
γj = γj − γ̂j = (13.40)
1 2
n σj + γj
Equation (13.40) shows that the larger the sample size (n) of τj , the more the
variance is reduced and, hence, the more reliable the travel time estimate. The
above analysis can be extended to an entire route r (or a portion of it) under the
assumption of independence among the travel times on different links of the same
route. Consider the portion rjh of route r between link j and link h and assume
AVI readers are located both on link j and link h. The route travel time τrj h on
such a portion is then a random variable with mean μrj h = μi and variance
i∈rj h
var(τrj h ) = σi2 + γi . The updating on the distribution of μrj h would lead
i∈rj h i∈rj h
to the computation of the reduction of variance:
γrj h = γrj h − γ̂rj h (13.41)
where γrj h and γ̂rj h are the a priori and the a posteriori variances of the μrj h .
Model AVIL5 can be applied to solve the problem of locating a given number k
of AVI readers on the links of a network to maximize the reliability of travel time
estimates. The resulting model formulation, AVIL5 , is:

AVIL5 : Max 12 yri + zrj γrij (13.42)
r∈R i∈r j ∈r

s.t. (13.21) – (13.29).

320 M. Gentili and P. B. Mirchandani

Fig. 13.4 Existing AVI Readers in the study area of Harris County, TX

13.5.1 The Study Area

The study area is bounded by Barker Cypress Road, Interstate 10, US 290, and
Sam Houston Tollway (represented by the shaded area in Fig. 13.3) in Texas, USA,
with 470 total arcs and 230 total nodes. The total number of OD pairs is 930,
which amounts to a total of 1700 routes, 12,075 miles, 90,020 vehicles and 459,234
vehicle-miles. There are already a number of AVI readers installed on the traffic net-
work. Figure 13.4 shows the approximate location and direction of those devices. It
can be seen that most existing vehicle identification devices are located on freeways
and highways such as Interstate 10, US 290, and Sam Houston Tollway.

13.5.2 Results

Figures 13.6, 13.7, and 13.8 give the optimal locations for 5, 10 and 15 addi-
tional AVI readers, respectively, obtained when solving AVIL5 to maximize the total
13 Locating Vehicle Identification Sensors for Travel Time Information 321

Table 13.2 Performance of AVI reader location strategies in the study area of Harris County, when
maximizing total vehicle-miles
Additional Total vehicle-miles Total monitored flow Total route length
AVI readers
installed
Monitored % of total Monitored % of total Monitored % of total
0 181,776 39.6 50,296 55.9 2236 18.5
5 291,040 63.4 61,764 68.6 3040 25.2
10 348,973 76.0 64,940 72.1 4183 34.6
15 380,530 82.9 68,818 76.4 4986 41.3

Fig. 13.5 Total benefit of the new sensors installed, from 5 to 15 sensors

vehicle-miles monitored (that is, the objective function (13.20)) by using CPLEX
7.0. Performance results for installing different numbers of additional AVI read-
ers are given in Table 13.2 where: (a) the first column reports the total number of
additional AVI readers installed; (b) the second and third columns report the vehicle-
miles monitored and the percentage with respect to the total, respectively; (c) the
fourth and fifth columns report the total monitored flow and the percentage with
respect to the total, respectively; (d) and the two last columns show the route length
monitored and the percentage with respect to the total, respectively.
We can observe (refer to Table 13.2) that the performance of the existing readers
is such that only 39.6 % of the total vehicle mileage, 55.9 % of the total flow and
18.5 % of the total route length is monitored. By locating additional AVI readers,
these percentages, obviously, increase. However, we can notice that the marginal
increase in the total vehicle-miles monitored decreases with additional readers
installed. Indeed, the first five additional readers produce an increase of 23.8 %
points; by adding five additional readers the increase is 12.6 % points and, with
5 more additional readers, the increase is equal to 6.9 % points.
Figure 13.5 shows the total benefit of the new sensors installed, from 5 to 15
sensors. Note the decreasing slope of the %total vehicle mileage function.
322 M. Gentili and P. B. Mirchandani

Fig. 13.6 Optimal location of 5 additional AVI readers in the study area of Harris County, when
maximizing total vehicle-miles

Fig. 13.7 Optimal location of 10 additional AVI readers in the study area of Harris County, when
maximizing total vehicle-miles
13 Locating Vehicle Identification Sensors for Travel Time Information 323

Fig. 13.8 Optimal location of 15 additional AVI readers in the study area of Harris County, when
maximizing total vehicle-miles

Figures 13.9, 13.10, and 13.11 give the optimal locations for 5, 10 and 15 addi-
tional AVI readers, respectively, obtained when solving AVIL5 to maximize the
reliability of travel time estimates (that is, the objective function (13.42)), again
using CPLEX 7.0. We assumed that the travel time variance is proportional to travel
time on that link. That is, longer travel times have larger variances. The total vari-
ance of the mean route travel times in the study area is 12,579. The traffic monitoring
results of installing different numbers of AVI readers are summarized in Table 13.3
where the first column reports the total number of additional AVI readers installed,
the second column reports the total reduction in variance, and the last column reports
the percentage of the reduction with respect to the total variance. With the existing
AVI readers located, the total variance reduction is 3361, which is 26.72 % of the
total variance. By installing additional AVI readers, the reduction of the variance
increases to 7578.79 with 5 additional readers, to 8997.69 with ten additional read-
ers and to 9908.40 with 15 additional readers. Note again that the marginal decrease
in the total variance decreases with additional readers installed. Indeed, the first five
additional readers produce a decrement of 33.53 % points; by adding 5 additional
readers the decrement is 11.28 % points and, with 5 more additional readers, the
decrement is equal to 7.24 % points.
324 M. Gentili and P. B. Mirchandani

Fig. 13.9 Optimal location of 5 additional AVI readers in the study area of Harris County, when
maximizing travel time reliability

Fig. 13.10 Optimal location of 10 additional AVI readers in the study area of Harris County, when
maximizing travel time reliability
13 Locating Vehicle Identification Sensors for Travel Time Information 325

Fig. 13.11 Optimal location of 15 additional AVI readers in the study area of Harris County, when
maximizing travel time reliability

Table 13.3 Travel time reliability results of AVI reader location strategies in the study area of
Harris County
Additional AVI readers installed Variance reduction Percentage of total variance (%)
0 3361.08 26.72
5 7578.79 60.25
10 8997.69 71.53
15 9908.40 78.77

13.6 Conclusions

Although much has been written about sensor location problems, this chapter is the
first comprehensive summary of the use of vehicle identification sensors to mon-
itor OD route coverage, travel times and related information on traffic networks.
In particular, application of such models was demonstrated for (a) maximizing the
326 M. Gentili and P. B. Mirchandani

vehicles-miles monitored and (b) maximizing the predictability of travel time esti-
mates where the resulting variance reduction is especially useful for routes that are
well traveled and a priori travel time variances are high.
Further uses of such sensors and attendant models by planners and traffic man-
agers for the applications discussed, as well as future applications that one has yet
to discover, attest to the value of this research and these models.

Acknowledgements The authors thank Dr. Yang He, former Ph.D. student of the second author,
for the computational results. The authors acknowledge the partial support from TranStar, Hous-
ton, Texas and USDOT support to the ATLAS Center, University of Arizona. The authors also
acknowledge support from National Science Foundation Grant n.1234584. Any opinions, findings,
and conclusions or recommendations expressed in this material are those of the author(s) and do
not necessarily reflect the views of the above mentioned agencies. Finally, the authors thank the
co-editors of this book for inviting them to provide this contribution for their timely book.

References

Adams WP, Sherali HD (1986) A tight linearization and an algorithm for zero-one quadratic
programming problems. Manage Sci 32(10):1274–1290
Adams WP, Sherali HD (1990) A hierarchy of relaxations between the continuous and convex hull
representations for zero–one programming problems. SIAM J Discret Math 3(3):411–430
Ban X, Herring R, Margulici JD, Bayen AM (2009) Optimal sensor placement for freeway travel
time estimation. In: William H, Lam K, Wong SC, Lo HK (eds) Transportation and traffic
theory 2009: Golden Jubilee. Springer US, Golden, pp 697–721
Bartin B, Ozbay K, Iyigun C (2007) Clustering based methodology for determining optimal
roadway configuration of detectors for travel time estimation. Transp Res Rec 2000:98–105
Chan KS, Lam WHK (2002) Optimal speed detector density for the network with travel time
information. Transp Res Part A Policy Pract 36(3):203–223
Chaudhuri P, Martin PT, Stevanovic AZ, Zhu C (2011) The effects of detector spacing on travel
time prediction on freeways. Int J Eng Appl Sci 7(1):1–10
Chen A, Chootinan P, Pravinvongvuth S (2004) Multiobjective model for locating automatic
vehicle identification readers. Transp Res Rec 1886:49–58
Church R, ReVelle C (1974) The maximal covering location problem. Pap Reg Sci Assoc
32(1):101–118
Danczyk A, Liu HX (2011) A mixed-integer linear program for optimizing sensor locations along
freeway corridors. Transp Res Part B 45:208–217
Edara P, Smith B, Guo J, Babiceanu S, McGee C (2010) Optimizing freeway traffic sensor locations
by clustering global-positioning-system-derived speed patterns. J Transp Eng 137(3):155–173
Fujito I, Margiotta R, Huang W, Perez WA (2007) Effect of sensor spacing on performance measure
calculations. Transp Res Rec 1945:1–11
Gentili M, Mirchandani PB (2011) Survey of models to locate sensors to estimate traffic flows.
Transp Res Rec 2243:108–116
Gentili M, Mirchandani PB (2012) Locating sensors on traffic networks: models, challenges and
research opportunities. Transp Res C Emer Technol 24:227–255
Lam WHK, Lo HP (1990) Accuracy of o-d estimates from traffic counts. Traffic Eng Control
31:435–447
Li X, Ouyang Y (2012) Reliable traffic sensor deployment under probabilistic disruptions and
generalized surveillance effectiveness measures. Oper Res 60(5):1183–1198
13 Locating Vehicle Identification Sensors for Travel Time Information 327

Liu Y, Lai X, Chang G (2006) Detector placement strategies for freeway travel time estimation. In
Intelligent Transportation Systems Conference, ITSC ’06. IEEE, pp 499–504
Mirchandani P, Gentili M, He Y (2009) Location of vehicle identification sensors to monitor travel-
time performance. IET Intell Transp Syst 3(3):289–303
Sheffi Y (1984) Urban transportation networks: equilibrium analysis with mathematical program-
ming techniques. Prentice Hall, Englewood Cliffs
Sherali HD, Desai J, Rakha H (2006) A discrete optimization approach for locating automatic
vehicle identification readers for the provision of roadway travel times. Transp Res Part B
40:857–871
Teodorovic D, Van Aerde M, Zhu F, Dion F (2002) Genetic algorithms approach to the problem of
the automated vehicle identification equipment locations. J Adv Transp 36:1–21
Chapter 14
Shape and Balance in Police Districting

Victor Bucarey, Fernando Ordóñez and Enrique Bassaletti

14.1 Introduction

Districting is a classic design problem when attempting to provide an efficient ser-

vice to a geographically dispersed demand. There exists a natural trade-off between
aggregation, which allows the pooling of resources, and individualization, where
each individual demand has its own resources and response is as efficient as pos-
sible. Police and security providers are no strangers to this phenomenon. Having
a single set of resources to satisfy the demand over an entire service area avoids
the duplication of resources, coordination problems, and uneven workloads that can
occur in districting. However, when the service area is large, service times at cer-
tain locations can exceed acceptable levels, making it more attractive to service
this demand from distributed resources. Furthermore, districting allows for special-
ization of the resources to efficiently service a diverse demand in different areas.
Such specialization causes additional complexity if managed from a centralized
pool of resources. This creates the basic problem of separating a demand area into
subregions to organize the service process.
The problem of districting in general, and police districting in particular, consists
of dividing a geographical region in subregions (also referred to as districts or quad-
rants) in a way that improves some objective measure. Perhaps the most well known
version of this problem is the political districting problem, where due to population
changes political districts are frequently reshaped, sparking intense debate (Hess
et al. 1965; Fleischmann and Paraschis 1988; Hojati 1996). There are however a
number of other applications where districting is regularly used, including territory

V. Bucarey () · F. Ordóñez

Department of Industrial Engineering, Universidad de Chile, Av. República 701, Santiago, Chile
e-mail: [email protected]
F. Ordóñez
e-mail: [email protected]

E. Bassaletti
Department of Criminal Analysis, Carabineros de Chile, Av. Libertador Bernardo O’Higgins
1196, Santiago, Chile

© Springer International Publishing Switzerland 2015 329

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_14
330 V. Bucarey et al.

design for sales and services, health care districting, police and emergency service
districting, and districting in logistics operations (Kalcsics et al. 2005).
Important considerations when designing districts include balancing the demand
for resources and creating districts that have geographical contiguity and compact-
ness (Wright et al. 1983). Existing literature has addressed the need for demand
balance constraints in districting models. In particular, not considering demand bal-
ance in the police districting problem has been shown to create difficulties such as
lack of indicators to compare the performance of the different districts by police
management, staff imbalance, and morale problems, among others Kistler (2009).
Compactness of a district is important because response time, a key measure of
quality of service, is related to the distance that has to be traversed. This makes it
desirable that all points in a district be close together, as measured by travel time, to
be serviced within a small response time. Therefore, travel time can be used to mea-
sure the compactness of a given district. Finally, contiguity is desirable because in
addition to the need for fast response times simple districts are easier to administrate.
A natural mathematical model of the districting problem can be built on a graph
representation of the geographical area and, as we describe in the literature review
section, previous work has introduced such models before. However, although the
requirements of balanced, contiguous and compact districts have been dealt with
before, these are rarely combined in a single model. In this work we introduce such a
mathematical model for a specific police districting problem of the Chilean national
police force in urban areas, including the requirements of balanced, contiguous, and
compact districts.
To build this mathematical model we consider a graph where nodes represent a
city block (or groups of city blocks) and arcs connect adjacent nodes (blocks), or
nodes that can be accessed thought the road network, see for example Fig. 14.1. In
addition, this network should consider information regarding the demand of police
resources at every node and information regarding distances (or travel times) on
arcs between nodes. On this network structure, the districting problem becomes a
graph partitioning problem. The demand balance constraints can be represented as
capacity constraints on the total demand on each district, where the lower and upper
bounds are calculated as a percentage deviation respect to the average demand of
resources or workload. The information on distances between nodes can be used
to characterize compactness and contiguity. Hence, the proposed graph partitioning
model at the heart of a districting problem can take into account the features of
balance, contiguity and compactness of each district.
Regrettably, introducing the requirements of balance, contiguity and compact-
ness makes the districting problem more difficult to solve. For instance, consider
the example in Fig. 14.2. This example shows two districting solutions, one for a p-
median problem without balance, contiguity or compactness constraints and one for
a p-median problem with balance constraints. For this problem which defines 9 dis-
tricts out of 407 nodes, the solution to the p-median problem on the left is obtained
in a few minutes, while the problem with balance constraints can take hours to solve
exactly.
We note that the p-median districting problem results in districts that seem
reasonably compact and contiguous. The reason for this is because the p-median
problem forms clusters around centers so that the total distance of nodes to these
14 Shape and Balance in Police Districting 331

Fig. 14.1 Adjacency graph

Fig. 14.2 Comparison between the results of a p-median problem (left) and a p-median with
balance constraint of 1 % of deviation with respect to the average workload (right)

centers is minimized, see Daskin (2013).This total distance objective is minimized

when clusters are compact and contiguous around selected centers. If the p-median
problem has balance constraints, however, then it creates districts that are not con-
tiguous. For example districts 4, 8 and 9 include nodes that are in the interior of
district 3. It also creates districts with branches that protrude into neighboring dis-
tricts, see for instance the branches out of districts 5, 7 and 8. Districts that are not
contiguous or have long boundaries tend to be less compact according to existing
definitions (Taylor 1973; Horn 1995).
It therefore becomes important to explicitly include contiguity and compact-
ness constraints in optimization models where balance requirements are enforced,
to obtain acceptable districts. This is particularly relevant in the police districting
problem considered, because the current districts are built with a detailed manual
procedure that immediately discards solutions that significantly violate contiguity
and compactness requirements.
The main objective of this chapter is to present a mathematical model of a real-
istic police districting problem. We are inspired by the urban districting problem
faced by Chile’s national police force (Carabineros de Chile) in their Preventive
Security Quadrant Plan (PCSP for its acronym in Spanish). This districting prob-
lem has been gradually implemented in Chile since 2001. By the year 2013 it had
332 V. Bucarey et al.

been implemented in the 120 most populous municipalities. In this work we pro-
pose a mathematical optimization formulation of the districting problem that Chile’s
national police force solves as part of the PCSP. This optimization problem also
takes into consideration balance, contiguity and compactness constraints.
In the next section we present a brief literature review on previous related work
on districting problems. In Sect. 14.3 we describe the general characteristics of
Chile’s national police force PCSP program and in Sect. 14.4 we present a mathe-
matical formulation of this particular districting problem. Section 14.5 presents our
computational results evaluating the tradeoffs between the different model objec-
tives and shows how the districting solution of the proposed model can improve
current practice for an example corresponding to a real municipality in Santiago.

14.2 Literature Review

The police districting problem can be considered as a particular case of the terri-
tory design problem, which consists in grouping small geographic areas (such as
city blocks) into larger geographic clusters or districts (in our case called quadrants)
in such a way the clusters are acceptable according to relevant planning criteria
(Kalcsics et al. 2005). Some important applications of the territory design prob-
lem are political districting (Hess et al. 1965; Fleischmann and Paraschis 1988;
Hojati 1996), territory design for sales and services (Hess and Samuels 1971), and
the assignment of students to public schools (Ferland and Guénette 1990; Caro
et al. 2004), among others. This previous literature on territory design usually does
consider at least one of the following three concepts: balance of demand, geographi-
cal contiguity, and compactness. We separate this literature review in describing how
these three concepts have been addressed by previous work on territory design and
in previous work on police districting. We begin by pointing out that the concepts
of balance, contiguity, and compactness are present in political districting problems
and touched on by the practice of “gerrymandering”, which involves redesigning
districts to make electoral gains (Erikson 1972; Vickrey 1961).
The concept of balance in territory planning refers to building quadrants where
geographic measures or characteristics are similar across the quadrants formed.
These measures or characteristics could be population (important, for instance, in
political and school districting), workload or demand (important in services such
as police districting), number of buildings or linear kilometers of roads, to mention
some possibilities. Related to police districting, the work in Kistler (2009) consid-
ers a combined measure that takes into account the amount of historic calls, the
average response time, the linear kilometers of roads, the total district area, and the
population.
Geographical contiguity means that every quadrant is geographically connected.
Optimal solutions to standard p-median problems with a metric or distance func-
tion between locations have contiguous quadrants. Indeed, since the objective of
a p-median problem minimizes the sum of distances to each center, each node is
14 Shape and Balance in Police Districting 333

assigned to the center that it is closest to. Therefore, the triangular inequality implies
that the line between the point and the center it is assigned to must also belong to the
same quadrant, making it contiguous (Daskin 2013). This is not necessarily true for
every objective and in particular, optimal solutions to the p-center problem, where
the objective is to minimize the maximal distance to the corresponding center, need
not be contiguous. This because nodes whose distance to more than one center is
smaller than the maximal distance could be assigned to either without changing the
objective. Partitions of an area formed by having each node/block assigned to the
center it is closest to, for some distance metric, are known as Voronoi tessellations.
As discussed above, such partitions form naturally contiguous quadrants. However,
even without using a distance metric, the geographical contiguity can be enforced
by explicit linear constraints. These constraints ensure that for every node assigned
to a district, an entire path of nodes to that district’s center are assigned to the same
district. This is done for example in Williams (2002) for a land acquisition problem
and in Drexl and Haase (1999) for a sales force deployment.
There are a number of different notions of compactness that are used in terri-
tory planning, which do not refer to the standard topological definition. Some of
the compactness measures proposed include the ratio of the perimeter to the area,
comparisons to a related compact shape such as a rectangle or circle, or moment of
inertia. See Li et al. (2013) for a recent review of different compactness measures.
To illustrate the variety in measures of compactness we review a few definitions that
arise in some political districting examples.
For instance, Niemi et al. (1990) investigate how different measures of compact-
ness are related to racial and partisan political discrimination in a political districting
problem. These different compactness measures are treated as a multidimensional
property of a district capturing: the dispersion and the perimeter of the district, and
the geographic dispersion of the population. The dispersion of a district was quan-
tified by one of four measures: the value |Li − Wi | or Li /Wi , where Li and Wi are
the maximum length and width of district i, respectively, a comparison between the
area of the district and the area of a compact shape circumscribing the district (such
as square, circle or hexagon) and a measure of the distances within a district, given
by the sum of the distances of every block to the center of gravity of its district.
Two perimeter measures of a district are considered: the length of the boundary of
the district (Horn 1995), and the difference of the perimeter of the district with a
circumference with equal area, or conversely, the difference between the area of a
district with the area of a circumference of equal perimeter. Finally, compactness
measures that considered the population are: the moment of inertia of the popula-
tion (Papayanopoulos 1973), and the ratio between the population of the district and
the population existing in the maximum convex figure or circle circumscribed in the
district.
A different compactness measure of a district, proposed in Theobald (1970), is
the absolute deviation of the district area with respect to the average area. Young
(1988) proposes eight measures of compactness of legislative districts. Many of
these measures are similar to the ones listed above, except for two: a visual test, in
which a human evaluator gauges the compactness of a district, and the Taylor’s test
334 V. Bucarey et al.

(Taylor 1973). The Taylor’s test quantifies the indentations of the districts by the
difference between reflexive angles and non-reflexive angles on the border divided
by the total number of angles. For example a convex figure has ratio 1 (compact),
and a five pointed star has ratio 0 (non compact).
Summarizing, there are many definitions of compactness measures which can be
used, depending on the application considered. To the best of our knowledge, to date
there is no previous work that builds compactness measures for territory planning
from axioms that the measure should satisfy, even for a specific application. Iden-
tifying desirable first principles for compactness measures is an interesting area of
future research that could help simplify these definitions. Compactness measures in
use focus on different characteristics of districts that should be taken into account.
Which characteristic is most important depends on the application being considered.
For example, in political districting measures of how population is distributed in the
district are important, while in territory planning for emergency service applications
a compactness measure that bounds the longest distance, such as a deviations from
a circle of the same perimeter, can be a desirable objective. Another consideration
is the difficulty in computing the compactness measure from geographical informa-
tion. Recent work has shown that a moment of inertia measure of compactness can
be more efficiently computed than isoperimetric ratios (Li et al. 2013).
A different application where the shape and characteristic of an area is important
is the problem of designing habitat reserves, see Marianov et al. (2008) or Church
(2015) in this volume. Habitats that are suitable to hosting a given species of animals
must satisfy certain shape constraints that depend on the species considered. These
shape constraints can impose restrictions on compactness, size, and connectivity.
Similar issues arise in the problem of harvest scheduling for environmental reasons
in the forestry industry (Barahona et al. 1992).
Previous work on police districting has already considered the issues of demand
balance and districts shape constraints. For instance, in Mitchell (1972) an optimiza-
tion model is proposed that aims to minimize the moment of inertia of the expected
workload of each area subject to balance constraints. An applied police district-
ing model for the Buffalo Police Department is presented in Sarac et al. (1999).
The paper discusses a set partitioning problem formulation and a practical approach
based on census tacks. The optimization approach was abandoned because the mul-
tiple objectives (with regards to demand balance, contiguity and compactness of the
districts) made the problem computationally challenging for problems of real size.
This difficulty is surmounted by limiting the flexibility of the districts to existing
census tracks. In D’Amico et al. (2002) the problem of police districting is consid-
ered as a graph partitioning problem subject to compactness, contiguity, convexity,
and size constraints. The problem also considered a constraint which forced an
upper bound on the average response time in each quadrant. This consideration leads
to a non-linear approach, which is solved by local search techniques such as simu-
lated annealing. The Constraint-Based Polygonal Spatial Clustering (CPSC) method
is another heuristic solution method used in police districting (Joshi et al. 2009).
The CPSC consists of designing quadrants by adding blocks to a seed block until
an objective, such as a compactness score, is met. This districting method has been
14 Shape and Balance in Police Districting 335

used for the Charlottesville Police Department and evaluated with an agent-based
model (Zhang and Brown 2013).
Another optimization model is used in Curtin et al. (2010) to determine opti-
mal police patrol areas. This work considers a maximum covering model (Church
and ReVelle 1974) to formulate a first phase problem. Then, a second phase prob-
lem determines how to distribute resources to meet a set of options with backup
coverage, i.e. maximizing the blocks that are covered two or more times with-
out considering any geographic issue or balance. In Verma et al. (2010) Voronoi
Tessellations are used to design police districts testing the incremental changes
in the improvement of measures such as balance of workload, 911 response time,
geographical area fit, and citizen satisfactions among others.
The police districting problem considered in this chapter is represented as a mod-
ified p-median problem with multiple objectives and balance constraints. One of the
objectives is related to the police’s definition of preventive policing, which is non-
linear in the quadrants formed. The combination of balance constraints and this
non-linear objective cause the p-median problem to provide solutions that are not
contiguous or compact. We therefore include a compactness measure that aims to
minimize the quadrant boundaries (or nodes adjacent to a different quadrant) sim-
ilar to minimizing the length of the boundary in (Horn 1995). As noted in Sarac
et al. (1999), this multiple objective problem is challenging to solve for real size
instances, we therefore adapt a location-allocation heuristic (Teitz and Bart 1968) to
solve this problem.

14.3 The Chilean Police Quadrant Plan

The Chilean national police force, Carabineros de Chile, is the principal internal
police force in Chile. Its mission includes maintaining and re-establishing order and
security throughout the country and patrolling the borders. Starting in 2001 Cara-
bineros de Chile developed its own police districting methodology for urban areas,
known as the Preventive Security Quadrant Plan (PCSP). The objective of this plan
is to define smaller districts within police precincts in order to (1) quantify the need
for policing resources in each district, (2) establish a closer connection with the pop-
ulation by providing a point of contact and information for each district to residents,
(3) organize and facilitate deployment of police resources to the community.
The process of implementing the PCSP in a municipality begins with the division
of the precincts being analyzed into quadrants. This division is currently performed
by the expert judgment of the police officers conducting the analysis. For each of
the quadrants identified, different requirements for police resources are tallied and
converted to a uniform policing unit, referred to as an Equivalent Unit of Vigilance
(or UVE in Spanish). This exercise helps organizing the expected policing activities
that have to be conducted in each quadrant, and dimensioning the need for policing
resources in the area under consideration. The PCSP also considers a number of
novel policing practices aimed at establishing a closer connection between the police
336 V. Bucarey et al.

and area residents, reduce response time, as well as institutionalizing improvement

practices.
Our work seeks to automate the design process of the PCSP in a given area. In
particular we aim to develop mathematical models that will form optimized quad-
rants that can improve the utilization of the police resources. Having an aid in the
planning of PCSP can help streamline the territory design process allowing a more
frequent revision of the information in different municipalities where this districting
methodology is deployed. We begin by detailing different aspects that are consid-
ered by Carabineros de Chile in the PCSP methodology as described in the PCSP
manual Dirección General de Carabineros de Chile (2010). In particular, the process
of dividing a precinct into quadrants takes into account the following four criteria:
• Patrolling constraints: There should be enough quadrants so that each quad-
rant can be patrolled in one shift. The police has estimated that a police vehicle
patrolling a neighborhood can traverse 82 km during an 8 h shift. Therefore the
number of quadrants has to be at least the total linear kilometers in the region
divided by 82.
• Geographical considerations: Quadrants should not be divided by avenues,
roads, train lines, rivers, mountains or other elements that can make it difficult to
patrol.
• Activities related considerations: It is desirable that the design of quadrants
does not partition areas of activities, such as commercial or civic districts,
residential neighborhoods, etc.
• Concentric radial design: It is also desired that quadrants be formed following
a concentric distribution around a civic quadrant or central plaza.
Of these, the first two considerations are strictly enforced by experts during
the design of quadrants, while the last two are desirable features that may not
necessarily be implemented.
Over each of these quadrants the PCSP methodology determines the amount of
demand for police resources in a standardized unit of UVEs. Each UVEs corresponds
to the amount of patrolling that a police car with three police officers is able to
conduct in one 8 h shift. The demand for police resources in each quadrant is divided
in two components: a reaction component and a prevention component. The reaction
component measures the procedures and other operational functions that are fulfilled
by Carabineros on average during an 8 h shift on the quadrant. This quantity is made
up of four factors:
• Deployments: Quantifies an expected number of deployments to service events
such as theft, injury, damage to property, alcohol law, disorder on public streets,
drugs, traffic accidents.
• Court orders: The expected number of activities due to court orders, such as:
arrests, detentions, notices, citations, closures, protective measures, injunctions
and evictions.
• Monitoring Establishments: Includes all resources needed to monitor establish-
ments such as liquor stores, restaurants, banks, gas stations, nightclubs, among
others.
14 Shape and Balance in Police Districting 337

Table 14.1 Parameters for preventive patrolling in PCSP

Factor Constant Units
Km 82 km/UVE
RC 23,500 Yearly reported crime/UVE
P op 50,000 People/UVE

• Extraordinary services: This factor is caused by special events that generate a

non-repeated large request for police resources such as: sporting events, concerts,
protests, and visits of foreign dignitaries or other authorities.
The prevention component of the demand for police resources is defined as the max-
imum of three factors over the quadrant: a measure of the criminal activity (amount
of reported crime RC), the population (Pop), and the total kilometers of roads in
the quadrant (Km). Each of these factors influences the need for police resources
used for preventive patrols. For instance: reported crime can be used as proxy for
possible future criminal activity, the population can be proportional to the level of
criminal activity, and finally the total linear kilometers in a quadrant is related to the
number of resources needed to patrol this distance in a fixed period. The methodol-
ogy, described in the Dirección General de Carabineros de Chile (2010), considers
normalizing constants, given in Table 14.1, to translate these factors in terms of
UVEs.
Therefore, the demand for police resources for prevention according to the PCSP
is given by
+ ,
1 1 1
D = max Km, RC, P op . (14.1)
82 23500 50000
This demand is defined as the maximum because any one of these three factors can
be the cause for additional police resources for prevention. If a quadrant has a very
large population, or reported crime, or total kilometers of roads, it can require addi-
tional resources to conduct adequate preventive patrolling. For example, a large,
sparsely populated district with little reported crime can have a demand for preven-
tive police resources determined by the Km, while a small, densly populated district
with high reported crime will have the number of police resources for prevention
determined by either RC or Pop.
An automatic method of constructing quadrants would naturally assign the infor-
mation regarding demand for police resources to each block, then depending on
the resulting quadrants, compute the reactive and preventive components of demand
over them. We note that the reactive component of demand is defined as the sum
of the demand of each of the blocks. Therefore over the whole area the reactive
demand remains constant, independent of how the quadrants are constructed. The
same is not true for the preventive component of demand for police resources. Since
this quantity is defined by a maximization over three factors, the total number of
resources needed to cover a district depends on the shape of the quadrants the district
is divided on.
338 V. Bucarey et al.

(i,j,k) (Km,RC,Pop)

(1,2,3) (3,2,1) (1,2,3) (3,2,1)

(2,2,2) (3,2,1) (2,2,2) (3,2,1)
(2,2,2) (3,2,1) (2,2,2) (3,2,1)

DTOTAL =D1 + D2 =7+9 =16 DTOTAL =D1 + D2 =6+8 =14

Fig. 14.3 Two ways of splitting an area with 6 nodes showing differences in the prevention demand
component

To illustrate this consider Fig. 14.3. This example considers a region formed of
six building blocks that can be organized in two quadrants. Each cell represents a
building block, and each number represents the normalized amount of Kilometers,
Reported Crime and Population. Two ways of splitting the region are represented,
each district is identified by the different shades of gray and the demand for each dis-
trict is calculated as the maximum between the sums of each of the three components
of demand prevention. The split of the area on the left achieves a total prevention
demand component of 16, while the one on the right equals 14. The example shows
that the total number of resources required over both quadrants depends on how the
quadrants are formed.
The demand of the district p is
 !
  
Dp = max Km , RC , P op .
∈P ∈P ∈P

This example shows that, in addition to automating the police districting problem
for the PCSP methodology, it is possible to optimize the amount of resources used
to provide the preventive component of demand by adjusting the shape of the quad-
rants formed. This is a key component of the mathematical model of the PCSP
methodology presented in the next section and a unique feature of this districting
problem.

14.4 A Police Districting Model

The model is based on a p-median model with a balance constraint in the demand
of police resources. Given a set of blocks I and a set of candidates to be centers
of districts J ⊆ I . The p-median model locates p facilities and assigns blocks to
each facility, so as to minimize the sum of the distances of each block to its cen-
ter. We refer to this measure as the total distance to centers. The p-median model
with balance constraints considers the previous model with lower and upper bound
constraints on the demand allocated to each quadrant.
In order to avoid non contiguities and “bad shapes,” we consider the compactness
as a two dimensional factor which minimizes the total distance to centers and the
14 Shape and Balance in Police Districting 339

length of the boundary of each district. To minimize the boundary of each district
we define an adjacency graph in which each node represents a block, and an edge
(i, j) exists if the block i share a piece of boundary with the block j, in this case
we say that i ∈ N(j) and vice-versa. We call the set N(i) as the neighborhood of i.
Also, in this model, constraints could be easily added that avoid dividing conflictive
areas, e.g., hot spots, or quadrants being crossed by rivers or main roads. Finally we
include the preventive demand component that minimizes the sum of the maximum
between the reported crime level, the population, and the total kilometers of streets
in each quadrant.
We use the variables of the classical p-median problem: xj is a location binary
variable which takes the value 1 if the center j ∈ J is assigned and 0 otherwise, and
yij are variables which take the value 1 if the block i ∈ I belongs to the center is j
∈ J and 0 otherwise. Also we use variables zik which take value 1 if the block i ∈
I is allocated to a different center than k ∈ N(i). This N(i). This variable allows to
count the length of the boundary of each district. Finally we use a continous variable
Dj which takes the value of the prevention component of the demand. With these
definitions the districting model is as follows:

  |N
 (i)| 
Min θ1 ij yij + θ2 κm uim + θ3 Dj (14.2)
i∈I,j ∈J i∈I m=1 j ∈J

s.t. xj = p (14.3)
j ∈J

yij = 1, i ∈ I (14.4)
j ∈J

xj ≥ yij , i ∈ I, j ∈ J (14.5)

xj L ≤ yij demi ≤ xj U, j ∈ J (14.6)
i∈I

zik ≥ yij − ykj , i ∈ I, k ∈ N (i), j ∈ J (14.7)

zik ≥ −yij + ykj , i ∈ I, k ∈ N (i), j ∈ J (14.8)
uim ≤ 1, i ∈ I, m ∈ {1, · · · , |N(i)|} (14.9)
|N
 (i)| 
uim ≥ zik , i ∈ I (14.10)
m=1 k∈N(i)

Dj ≥ fpop yij popi , i ∈ J (14.11)
i∈I

Dj ≥ fkm yij kmi , j ∈ J (14.12)
i∈I
340 V. Bucarey et al.

Dj ≥ frc yij rci , j ∈ J (14.13)
i∈I

xj , yij ∈ {0,1}, i ∈ I, j ∈ J (14.14)

Dj , zik , uim ≥ 0, i ∈ I, j ∈ J, k ∈ N (i), m ∈ {1, . . . , |N(i)|} (14.15)

The objective (14.2) minimizes the multi-objective weighted function, where θ1 is

the weight of the total distance to centers (where ij represents the distance of block
i ∈ I to its allocated center j ∈ J), θ2 is the weight of the penalty function of size
of the boundary and θ3 weights the prevention demand component.
The Eq. (14.3) states the number of centers or quadrants. The set of constraints
(4) states that each block must be assigned to a quadrant. The set of inequalities
(5) establishes that a block could be assigned to a center only if the center is acti-
vated. The set of constraints (6) balances the demand between the quadrants. This
means that, denoting demi the demand from block i, the total demand assigned to
each center cannot be more than an upper bound U = (1 + α) i∈I demi /p and
cannot be less than the lower bound L = (1 − α) i∈I demi /p. Here, α is the max-
imum acceptable percentage of deviation from the average of the demand of the
reaction component (e.g., 5 %). Expressions (7) and (8) define zik as the absolute
value of the difference between yij and ykj for every pair of adjacent blocks. By
defining auxiliary variables uim , the sets of constraints (9) and (10) allow to express
 β
|N (i)|
i∈I k∈N(i) zik as a piecewise linear approximation i∈I m=1 κm uim ,
where κm = κm (β) depends of the convexity of the function. These variables basi-
cally count the number of adjacent blocks not assigned to the same median, so
disconnected blocks can be penalized. The set of constraints (11–13) defines Dj
as the maximum of three factors: the linear kilometers kmi , the amount of reported
crime for the block i, rci , and the population of the block, popi . All of them are
normalized by a factor fw . Finally, expressions (14) and (15) define the domain of
the variables.
Realistic instances of this optimization problem can be difficult to solve exactly.
If we consider a problem on n nodes, where each node has at most ρneighbors, then
this optimization problem would have O(n2 ) binary variables, O(nρ) non-negative
continuous variables, and O(n2 ρ) inequality constraints. The realistic case study pre-
sented in this chapter considers n = 1266 blocks and up to ρ = 7 neighbors, which
gives a problem with millions of constraints and binary variables. Even reducing
the problem size by aggregating variables or limiting the possible centers and pos-
sible block to center assignments we obtain problems whose size are a challenge
for existing exact solution methods. We therefore consider a heuristic solution algo-
rithm based on a Location-Allocation heuristic (Teitz and Bart 1968) to find good
solutions to this problem efficiently. This heuristic implemented is as follows:
14 Shape and Balance in Police Districting 341

1. Step 1: Generate an initial set of p centers. Any rule can be used to generate this
set, for instance random selection or solving the p-median problem.
2. Step 2: (Allocation phase): Solve the mixed integer optimization problem that
allocates every block to a given set of centers, so that the balance constraints
hold, and the objective is minimized.
3. Step 3: (Location phase): For every quadrant generated, find the best geometrical
center, and go back to the step 2 until there are no changes.
We repeat this heuristic with multiple random seeds or initial solutions to further
search the solution space. One of the best features of this heuristic is that in every
step a feasible solution is obtained. This gives several districting plans within a
reasonable computational time that could be evaluated by the decision maker with
other criteria, as desired. We note that this heuristic can be accelerated by solving
the linear relaxation in the allocation phase, assigning the fractional blocks using
a heuristic rule. This acceleration makes it more difficult to satisfy the balance
constraints.
In the next section we show how this model with the heuristic and the adequate
parameters of the objective function (θ, κ) gives compact and balanced districts for
a real instance.

14.5 A Case Study

In this section we show the most important results of the model applied to a real
instance in Ñuñoa, Santiago de Chile. The current districts have a configuration
as shown in Fig. 14.4. The data of each block was extracted from demographic
census data and historical data of Carabineros de Chile. This instance has a size of
1266 blocks. The current districting plan has a maximum deviation of the reactive
component of demand for police resources that deviates 77 % from the average value
in this region (Bustamante 2011).
This map, after a clustering of some blocks was represented in an adja-
cency graph as shown in Fig. 14.5. The clustering was performed with a grid of
200 × 200-m cells. Each node represents a cell of the grid that contains a block of
the original map. After this clustering the graph has a size of 407 nodes.
If a p-median is used to find quadrants, the results would be nice and compact
shapes. However, the maximum deviation from the workload balance could reach
large figures, similar to the 77 % shown by the current quadrants. On the other hand,
the p-median with balance constraints will find an optimal solution, but the shape of
the districts will be irregular forming non compact quadrants. Even non contiguities
could be observed in this graph.
We applied the model and the heuristic for different values of the parameters of
the objective function (θ = (θ1 , θ2 , θ3 )) and found that the quality of solutions and
the performance of the heuristic are very sensitive in this parameter. For instance,
Fig. 14.6 shows the output for different parameters θ.
342 V. Bucarey et al.

Fig. 14.4 Current quadrants

Figure 14.6a shows the result when the relative importance of the three factors
in the objective function is the same, i.e., all three are weighted uniformly. In this
case non-contiguities are observed and the shape of the districts is irregular. This
situation is improved when the total distance to centers is weighted with greater rel-
ative importance, which gives a solution closer to the p-median problem as in the
case of Fig. 14.6b. In that graph each district generated has contiguous blocks. The
case of Fig. 14.6c shows the result where the size of the boundary has greater rela-
tive importance in the objective function. Due to the symmetries in the optimization
model, using the proposed objective function, the allocation integer optimization
problem in the heuristic becomes quite difficult to solve. We therefore fixed a 300 s
limit for each iteration of the heuristic. Figure 14.6d shows the graph generated
by the heuristic where the most important factor is the prevention demand. As
expected, non contiguities and non compactness are present in the result. Finally
we weighted with greater relative importance both the total distance to centers and
size of the boundary. Figure 14.6e shows better shapes, while the balance between
the quadrants is maintained. In particular the improvement of this solution over the
one depicted in Fig. 14.6b is that because of the greater weight on the boundary,
nodes which have many neighbors tend not to be on the boundary.
14 Shape and Balance in Police Districting 343

Fig. 14.5 Adjacency graph that represents Ñuñoa

We analyze these solutions further with the results presented in Table 14.2. This
table shows the average solution time per iteration of the heuristic and the contribu-
tions to each part of the objective (distance to centers, boundary, prevention demand)
of the five solutions presented in Fig. 14.6. We note that the total distance to centers
component of the objective is not very sensitive to changes in the weights and varies
at most 6 % of the mean value, while the objective components of boundary and
prevention demand vary greatly (with a maximal variation of 60 % and 44 % with
respect to the mean value respectively). Another important aspect is the difficulty of
solving each iteration in the heuristic. When the objective is primarily driven by the
size of the boundary we observe that the heuristic reaches the run time limit. This is
due to the difficulty of solving the allocation step when the contribution of this non-
linear objective is significant. It is interesting to note that, for the best shapes found
(experiment e), the time required was just two seconds. This solution method also
provides many feasible solutions (one during each iteration) which can be evaluated
afterwards by decision makers.
Finally, in Fig. 14.7 we show a new districting plan which is the output of the
heuristic where the parameters of the objective function θ weight most importantly
the geometrical aspects as in the experiment (e). Notice that this districting solution
has a demand balance within 5 % of the mean demand value and forms contiguous
and compact districts.
344 V. Bucarey et al.

Fig. 14.6 Comparison for the output of the model for differents values of (θ1 , θ2 , θ3 )
14 Shape and Balance in Police Districting 345

Table 14.2 Average resolution time and contributions of each objective component for each
experiment
Experiment Parameters Time Distance Boundary Prevention
label [sec] to Centers Com- Demand
dij yij ponent γj
θ1 θ2 θ3 κm uim
a 1 1 1 5 229,528 4308 185,895
b 100 1 1 3 222,551 4173 288,158
c 1 100 1 300 235,451 2131 221,071
d 1 1 100 18 229,635 4186 185,895
e 100 100 1 2 222,639 3350 288,326

Fig. 14.7 Districting plan given for the heuristic using θ = (100,100, 1) and β = 3

14.6 Conclusions

Districting is a difficult problem arising in many contexts where resources must

be allocated and divided to provide an efficient service to a distributed demand.
In particular, police institutions are no stranger to this challenge, where the demand
for police resources must accomplish both preventive and reactive actions to provide
security to an area.
346 V. Bucarey et al.

In this work, done in collaboration with Chile’s national police force, we for-
mulated an optimization problem to represent Carabineros’ PCSP police districting
problem. This leads to a modified p-median model with a preventive demand non-
linear objective, balance constraints on the reactive demand component, and an
objective of minimizing the boundary. The balance requirements and non-linear
preventive demand objective not only make the districting problem much harder,
but give a problem that has optimal solutions that do not have contiguous or com-
pact quadrants. Adding the compactness measure that minimizes the boundary to
this multi-objective problem helps provide solutions that achieve efficient objectives
with acceptable quadrant shapes.
The proposed model and location-allocation heuristic provide several efficient
solutions that a planner can compare to trade-off additional less tangible criteria in
selecting the optimal districting plan. We applied the heuristic to a real case in San-
tiago, Chile, and obtain districts with a reactive demand imbalance of at most 5 %
from the average quadrant demand, as opposed to an imbalance of 77 % obtained
by the current manual solution. This benefit comes at the expense of districts which
have more complicated boundaries, which become operationally more challenging
to manage. Improvements of this model should include additional constraints on
the boundaries to achieve quadrant shapes that are simple to patrol. Ongoing work
includes evaluating the use of this methodology in other precincts of Santiago with
real police data and adding geographical and operational considerations of what
nodes can be used as boundaries of quadrants. Lieutenant Colonel Bassaletti, who
participated in this work, has stated that this model and solution method are part of
the techniques that the Chilean national police force is taking into account in their
ongoing revision of the PCSP districting plan.

Acknowledgments This work was partially supported by Conicyt through grant ACT87.

References

Barahona F, Weintraub A, Epstein R (1992) Habitat dispersion in forest planning and the stable set
problem. Operations Res 40(Suppl 1): S14–S21
Bustamante S (2011) Metodología para el rediseo de los cuadrantes utilizados por carabineros
de chile en el plan cuadrante de seguridad preventiva. Masters thesis. Universidad de Chile,
Facultad de Ciencias Físicas y Matemáticas, Santiago, Chile
Caro F, Shirabe T, Guignard M, Weintraub A (2004) School redistricting: embedding GIS tools
with integer programming. J Oper Res Soc 55(8):836–849
Church R, ReVelle C (1974) The maximal covering location problem. Pap Reg Sci 32(1):101–118
Curtin KM, Hayslett-McCall K, Qiu F (2010) Determining optimal police patrol areas with
maximal covering and backup covering location models. Netw Spat Econ 10(1): 125–145
Church RL, Niblett MR, Gerrard RA (2015) Modeling the Potential for Critical Habitat,
pp 155–172. Applications of Location Analysis, Springer NY
D’Amico SJ, Wang S-J, Batta R, Rump CM (2002) A simulated annealing approach to police
district design. Comput Oper Res 29(6):667–684
Daskin M (2013) Network and discrete location: models, algorithms, and applications, 2 edn.
Wiley, Hoboken
14 Shape and Balance in Police Districting 347

Dirección General de Carabineros de Chile (2010) Nuevo Manual Operativo del Plan Cuadrante
de Seguridad Preventiva. Santiago, Chile
Drexl A, Haase K (1999) Fast approximation methods for sales force deployment. Manage Sci
45(10):1307–1323
Erikson RS (1972) Malapportionment, gerrymandering, and party fortunes in congressional
elections. Am Polit Sci Rev 66(4):1234–1245
Ferland JA, Guénette G (1990) Decision support system for the school districting problem. Oper
Res 38(1):15–21
Fleischmann B, Paraschis JN (1988) Solving a large scale districting problem: a case report.
Comput Oper Res 15(6):521–533
Hess SW, Samuels SA (1971) Experiences with a sales districting model: criteria and implemen-
tation. Manag Sci 18(4):41–54
Hess SW, Weaver JB, Siegfeldt HJ, Whelan JN, Zitlau PA (1965) Nonpartisan political redistricting
by computer. Oper Res 13(6):998–1006
Hojati M (1996) Optimal political districting. Comput Oper Res 23(12):1147–1161
Horn MET (1995) Solution techniques for large regional partitioning problems. Geogr Anal
27(3):230–248
Joshi D, Soh L-K, Samal A (2009) Redistricting using heuristic-based polygonal clustering. In
Proceedings ICDM’09. Ninth IEEE International Conference on Data Mining, pp. 830–835,
IEEE
Kalcsics J, Nickel S, Schröder M (2005) Towards a unified territorial design approach: applications,
algorithms and GIS integration. Top 13(1):1–56
Kistler A (2009) Tucson police officers redraw division boundaries to balance their workload.
Geogr Public Saf 1(4):3–5
Li W, Goodchild MF, Church R (2013) An efficient measure of compactness for two-dimensional
shapes and its application in regionalization problems. Int J Geogr Inf Sci 27(6):1227–1250
Marianov V, ReVelle C, Snyder S (2008) Selecting compact habitat reserves for species with
differential habitat size needs. Comput Oper Res 35(2):475–487
Mitchell PS (1972) Optimal selection of police patrol beats. J Crim Law Crim Police Sci
63(4):577–584
Niemi RG, Grofman B, Carlucci C, Hofeller T (1990) Measuring compactness and the role of a
compactness standard in a test for partisan and racial gerrymandering. J Polit 52(4):1155–1181
Papayanopoulos L (1973) Quantitative principles underlying apportionment methods. Ann NY
Acad Sci 219(1):181–191
Sarac A, Batta R, Bhadbury J, Rump C (1999) Reconfiguring police reporting districts in the city
of buffalo. OR Insight 12:16–24
Taylor PJ (1973) A new shape measure for evaluating electoral district patterns. Am Polit Sci Rev
67(3):947–950
Teitz M, Bart P (1968) Heuristic methods for estimating the generalized vertex median of a
weighted graph. Oper Res 16:955–961
Theobald HR (1970) Equal representation: a study of legislative and congressional appointment in
Wisconsin, pp 71–260. The State of Wisconsin Blue Book, Wisconsin
Verma A, Ramyaa R, Marru S, Fan Y, Singh R (2010) Rationalizing police patrol beats using
Voronoi tessellations. In Proceedings of 2010 IEEE International Conference on Intelligence
and Security Informatics (ISI), pp 165–167
Vickrey W (1961) On the prevention of gerrymandering. Polit Sci Quart 76(1):105–110
Williams JC (2002) A zero-one programming model for contiguous land acquisition. Geogr Anal
34(4):330–349
Wright J, ReVelle C, Cohon J (1983) A multiobjective integer programming model for the land
acquisition problem. Reg Sci Urban Econ 13:31–53
Young HP (1988) Measuring the compactness of legislative districts. Legislative Stud Quarterly
13(1):105–115
Zhang Y, Brown DE (2013) Police patrol districting method and simulation evaluation using agent-
based model & GIS. Secur Inf 2:7
Chapter 15
Location and Sizing of Prisons and Inmate
Allocation

Vladimir Marianov

“People shouldn’t live better in jail than they do on the outside.

Here in my jails, they don’t.”
—Joe Arpaio, Sheriff, Maricopa County, AZ.

15.1 Introduction

According to the International Centre for Prison Studies (Walmsley 2013), 10.2
million people were held in jails and prisons worldwide in 2013. The total world
population at the beginning of 2013 was of 7.1 billion, which means that 144 people
out of every 100,000 were prison inmates. The US not only has the largest inmate
population in the world, according to those figures (2.24 million), but it also has the
highest prison population rate (716 people per 100,000 inhabitants.) The Russian
Federation has a healthy figure of 475 inmates per 100,000, and the lowest rates
(with the exception of San Marino, with 6,) are found in some European Northern
countries (58–88), Japan (51), and India (30). Some Western African countries have
remarkably low rates (28–75). Chile has a rate of 266 per 100,000.
As these figures show, unless incarceration ceases to exist—a not very likely
event—societies cannot live without penal facilities. Although their basic function
is debated (it is not clear whether it is to punish, keep inmates from committing
more crimes, maintaining them apart from society, reeducate them, or several of
the above), there is agreement in considering jails and prisons indispensable. Some
studies even conclude that they boost the economy of the areas in which they are
located (Cherry and Kunce 2001). Such positive externalities are disputed by other
authors, though (Hooks et al. 2010; King et al. 2003.)

V. Marianov ()
Department of Electrical Engineering, Pontificia Universidad Católica de Chile,
Santiago, Chile
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 349

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_15
350 V. Marianov

Some definitions are required at this stage. The term “jail” refers to places where
people are awaiting or under trial or serving sentences that last for less than 1 year.
A prison, or correctional, or penitentiary1 , is a place where convicted felons serve
their sentences. According to the National Institute of Corrections (Hall 2006), due
to their nature, a jail turns over its beds far more frequently than a prison: 36 times
a year against 0.75 times a year in average. Our application refers to penal facilities
that fulfill both requirements.
There are high, medium and minimum security closed prisons, from where in-
mates are not allowed to leave at any time. However, in some countries, industries
may contract inmates as employees. For example, the program Federal Prison In-
dustries (“Unicor”) provides prison labor in the US. (NBC 2012). In New Zealand,
there is a policy oriented to provide jobs to inmates with the involvement of the
private sector (New Zealand Department of Corrections 2001) Some of these pris-
ons could have forced labor, although this is a decreasing tendency2 . There are also
open prisons, allowing prisoners to be free to move with minimum supervision, and
even to leave. These include training, psychiatric and drug abuse treatment centers.
Finally, there are the halfway houses, intended for rehabilitation of inmates who
have recently been released from prison and are in the process of rehabilitation. In
this chapter we do not deal with open prisons or halfway houses, which, from the
point of view of location, have requirements that are different from those of prisons,
since they tend to require some contact by inmates with people in the surrounding
community. Interested readers are referred to Johnson (2006).
Penal facilities are generally considered undesirable. NIMBY (not in my back
yard), NIMTO (not in my term of office), LULU (locally unacceptable land use),
BANANA (build absolutely nothing anywhere near anything) are among the terms
that have been used to denote facilities that people do not want to be located in their
neighborhood. Landfills, jails and prisons, polluting industries and power plants are
examples of such facilities. Among locally undesirable facilities, penal institutions
have one very distinctive feature: they house people who do not like being there,
and most of the people in their surroundings do not like to have them around, either.
This feeling is not surprising: being locked with other inmates with not-so-clean
records and not being able to leave is not the best state of affairs for anybody. For
neighbors, living close to a place full of delinquents generates some uneasiness due
to the likelihood of escapes, decline in property values and the constant presence of
visitors who are relatives, friends and colleagues of those inside.
However, not everybody wants prisons and jails located far away: families of the
inmates, especially of those sentenced to long stays, naturally prefer closeness to
the place where their relatives are, so they can visit often. Facility employees and
service providing personnel want to be close to their workplace, which is relevant

1 Or papa’s house (Comfort 2006).

2 In reference with forced labor and other therapies, Joe Arpaio declares: “I want to make this
place so unpleasant that they won’t even think about doing something that could bring them back.”
Inmates in his prisons do work.
15 Location and Sizing of Prisons and Inmate Allocation 351

considering that, in 2005, there was one employee per each 3.2 inmates in US federal
correctional facilities (Stephan 2008).
The penal facility location problem could include different decision levels, which
are usually addressed separately. For a federal prison, a first level considers the
decision about the state, province and county in which the prison will be located
(Marianov and Fresard 2005; Hernández et al. 2012). At this level, the location
factors are related mainly to transportation costs and, in minor degrees, with public
opposition. After the first-level decision has been made, a second level concerns the
decision about the specific lot in which the prison will be built. At this level, the
main issues are availability and cost of appropriate and accessible land, as well as
existence and cost of utility services; public concerns and environmental impact. A
good description of the issues can be found in Ricci (2006), specific criteria have
been developed by a number of US counties, and are available on the web, and an
interesting example can be found in Chan (1996). A good literature review up to
that date, on impacts of a jail on a community, can be found in Fehr (1995).
A third level includes the architecture or layout design of the facility
(Chan 1996). When locating a federal prison, the three levels must be addressed,
but the problem of locating a county jail includes only the second and third levels.
An example of the site selection process for county jails can be found in San Ma-
teo County Sheriff’s Office (2009). A very good description of these levels and the
issues in prison location can be found in Engel (2007).
The academic literature on jail and prison location is very scarce, even though
mathematical methods have been used in crime and justice problems (Maltz 1996).
One of the few articles related to the subject is Korporaal et al. (2000), who propose
a model for assessing the needs for prison capacity in the Netherlands. They for-
mulate a queuing model of the prison population, which can be solved for different
scenarios of capacity expansion of the system, computing in each the blocking prob-
ability (probability of inmates sent home for lack of capacity). However, there is a
number of reports, manuals and leaflets developed by practitioners and users, about
criteria and considerations to be taken into account when locating a prison. Rather
than reviewing this literature, we describe the problem, and provide an example of
penal facility location problem, based on a real case in Chile in the early 2000s.
In our application we locate multiple facilities, and we include only the first deci-
sion level, i.e., the determination of what province or county should the facilities be
located in. Additionally, we determine the optimal expansion schedule for existing
facilities and allocate inmates to facilities. In the application, the facilities fulfill the
needs for jails and prisons.
352 V. Marianov

15.2 Issues in a Prison Location Problem

15.2.1 Location Criteria

As with other undesirable facilities, most of the public would want prisons and jails
as far as possible from their own locations. This criterion has indeed been used in the
past. Examples are Siberia for Russian inmates and Australia for English criminals.
This criterion is still valid for large prisons, up to a certain point, and in some coun-
tries (Moran et al. 2011; Lawrence and Travis 2004). However, as mentioned above,
prisons and jails need to be reachable by inmates’ families (contact with family is
one of the determinants of rehabilitation), and by people in charge of custody and
services (lawyers, drug rehabilitation personnel, prison personnel, catering, health
services, etc.) As jails house inmates under trial or expecting it, they need to be
not too far from courts, a requirement not necessary for prisons. In consequence,
there is a tradeoff between public concerns and operating (transportation) cost, and
the solution in each case is highly dependent on the context and budget availability.
For instance, in the case of a county jail, location in a rural area keeps the jail far
from the bulk of population, which is concentrated in urban areas, but also far from
families and courts and, in most cases, requires an extension of services to the loca-
tion. An urban location is closer to courts but requires more expensive land, poses
security risks, and faces more opposition.

15.2.2 Planning Horizon and Population Forecast

Locating a jail or prison is a strategic matter, i.e., the planning needs to consider a
long term horizon, since jails cannot be easily moved3 . Consequently, prison pop-
ulation must be forecast as accurately as possible, using the historical trends in
prison population, past and present crime trends and other influencing factors, as
well as all the uncertainties: future changes in incarceration policies, new crime
trends, changes in police efficiency and effectiveness, and so on. Modern methods
of population forecast are reviewed in Berk (2008), a recent method used in New
South Wales, Australia is described in Wan et al. (2013), while other examples of
forecast can be found in de Silva et al. (2006), Scalia (2004) and Ministry of Justice
UK (2013). Practical considerations in forecasting can be found in NAATAP (2014).

3 With the exception of the (in-) famous Tent City (for different points of view on

it, access http://www.mcso.org/MultiMedia/PressRelease/Tents%20Birthday.pdf and http://www.

motherjones.com/politics/2013/05/10-worst-prisons-america-joe-arpaio-tent-city), established by
the Maricopa County sheriff Joe Arpaio, whose quotes decorate this chapter. (http://thinkexist.
com/quotes/joe_arpaio/).
15 Location and Sizing of Prisons and Inmate Allocation 353

15.2.3 Overcrowding in Prisons

Many prisons in the world suffer from overcrowding or overpopulation. According

to the International Centre for Prison Studies4 , in 2013, in Haiti there were more
than four inmates locked in a space meant for one (an occupancy level of 416.3 %).
In Venezuela, almost three people were sharing the space built for one (270.1 %).
In Italy, the occupancy level was of 121.6 %. In Chile, 111 %. Although in the US
as a whole, the occupancy level in 2013 was officially 99 %, in California there was
an occupancy level of 146 %5 . Overcrowding is due to insufficient prison space and
it has several causes: fluctuating or unforeseen crime rates; changes in laws; and
increased public pressure on law enforcement bodies and courts, that makes them
act more harshly than expected by planners; new offenses and, last but not least, the
delay between the time at which prison population starts exceeding the capacity and
the time the decisions are made and actions required taken for the opening of new
prisons.
The effects of overcrowding are increasing inmate misconduct; insufficient ser-
vices such as education and drug treatment programs; and decreasing safety and
security of inmates and personnel (US Government Accountability Office 2012).
It is out of the scope of this chapter to analyze the causes and effects of overpop-
ulation and the required measures to decrease it. However, from the point of view
of prison planning, it is reasonable to expect some overpopulation because of the
uncertain nature of future population, and because, if large penal facilities are to be
built, it may not be reasonable to build 1 with a capacity of, say, a 1000 inmates, as
soon as the prison population exceeds the current capacity by 1 inmate.

15.2.4 Dealing with Public Opposition

Being undesirable facilities, prisons generate public opposition (Martin and My-
ers 2005.) As with other locally undesirable infrastructure, there are basically two
policies when it comes to dealing with public concerns. The first is the authoritarian
approach, consisting of maintaining secrecy about the location of the facility, and
once the location is decided, deal with the public by defending the location. This
approach has been termed DAD (Decide—Announce—Defend) by Kleindorfer and
Kunreuther (1994), see Eiselt and Marianov (2015). Naturally, elected officials do
not like this policy, as it may lead to losing a re-election.
A second policy considers including the public in the decision, and compensating
those most affected by the facility. A nice example of this policy can be found in
Armstrong (2012), and a discussion of the different policies in Farkas (1999).

4 See http://www.prisonstudies.org/highest-to-lowest/occupancy-level?field_region_taxonomy_
tid=All.
5 See http://www.economist.com/blogs/graphicdetail/2013/08/daily-chart.
354 V. Marianov

15.2.5 Location Models

Different levels of decision use different models. The location of a large federal
prison (which we termed a first level decision) could be decided by an optimization
model in which the cost is the main objective (land acquisition, building, opening,
operating, transportation, etc.), and will consider also other lower-level objectives,
e.g., minimization or penalization of overcrowding and penalization of travel dis-
tances covered by all the involved personnel and service providers, as well as by
visitors of the inmates. Note that this is a cost that is not incurred by the govern-
ment, but by providers, employees and relatives; a social cost. The constraints must
include satisfaction of the forecast demand, opening and closing schedule over the
horizon, number of facilities to be built, and others.
At the second level (determination of the specific community and the lot on
which the prison will bebuilt), the decision is usually made by comparing the
alternative sites one by one, using a multiobjective approach as the analytic hier-
archy process (Saaty 1980) or PROMETHEE (Brans and Vincke 1985). On this
level, costs are again important, but other objectives are also taken into account:
neighborhood socioeconomic characteristics as level of poverty, crime rate and
unemployment rate; site attributes including development cost, availability of ser-
vices and transportation; distance to the courts; environmental and public opposition
issues, etc.

15.2.6 Offender Flow and Inmate Segregation/Classification

Although the organization of the crime treatment system differs between countries,
the offender flow follows similar stages everywhere: once a crime is committed and
the offender is arrested, a short period of holding in a police premise is the usual
first stage. The second stage corresponds to an initial hearing, after which there
is either no case and the suspect is released, or there is a case, in which situation
the suspect is detained in jail or released on bail. A trial follows and, again, there
can be no case (the suspect is released), or the suspect may be either acquitted
and released or sentenced. The last stage corresponds to the classification of the
sentenced individual into one of several types of prisons or centers (prison, training
center, open prison, psychiatric center, drug-addiction treatment center, or similar
facilities). Furthermore, prisons can usually be of maximum, medium or minimum
security, and there could be segregation by other inmate features, including gender
and health status. Finally, defendants may go free or use a halfway house.
During the time they serve their sentences, prison inmates can go through several
different stages depending on the correctional system. In some cases, the whole time
includes induction, regular incarceration and pre-release stages, while in other cases
a sentenced individual can access some benefits that can include a supervised (or
unsupervised) time in open prisons or even an early release in the cases of a good
behavior.
15 Location and Sizing of Prisons and Inmate Allocation 355

15.3 An Application

15.3.1 The Setting

In 2000, Gendarmería de Chile, the institution in charge of the penitentiary system in

Chile, commissioned a study, whose main objective was to determine the capacities
required for housing an increasing number of inmates over a 20-year time horizon
in the whole country. The facilities in the country serve the needs of both jails and
prisons, and they also house short-term detainees. For this reason, from now on,
we indistinctly refer to the facilities as jails or prisons. These capacities could be
fulfilled by increasing the size of existing prisons and by building new, large-size
prisons. The locations (provinces) for these new prisons were to be determined, as
well as a general allocation of inmates to facilities. Furthermore, the schedule of the
new prison openings, the timing of each expansion of both new and old prisons, and
the final size of all prisons at the end of the time horizon (in terms of number of
inmates) was to be also determined. In regard with the locations, the main idea was
to solve only the first level of the problem, i.e., the provinces in which the facilities
should be located, as opposed to finding the exact lots for each facility or the archi-
tectural features of the establishments. The objective was to minimize the cost of
maintaining the penitentiary system, including opening and expansion costs, over-
crowding penalties, transportation costs and some social and environmental costs.
Finding the required capacities required a forecast of the demand over the whole
project horizon (20 years).
The prison population in Chile includes different types of inmates: individuals
under pre-trial detentions (staying in a facility for up to 5 days), defendants under
trial (ranging from a few weeks up to 5 years6 ) and sentenced offenders staying
in prison from a few months to life. In the year 2000, there were some 33,000 in-
mates, and the flow of detainees in the jails was of approximately 200,000 people.
Detainees and inmates under trial must be housed in facilities close to the respective
courts, because they have to frequently travel to and from court. Sentenced crimi-
nals do not need to be close to courts, but the rehabilitation rate increases if they stay
in contact with their families, so they must be kept not far from their relatives. All
types of inmates are kept in the same prisons, although there is internal segregation.
The system comprises small local jails as well as regional or multi-jurisdictional
large jails, each covering one of the jurisdictions of several courts. The capacities
range from a few tens to a few thousands of inmates. At the time of the study, the
country was divided in 13 regions and 51 provinces. Each province was composed
by the jurisdiction of one or several courts, depending on its population. The juris-
diction of one particular court falls always entirely within one province. Since large
prisons can cover more than one court jurisdiction and more than one province’s
demand, the problem is not separable by districts.

6 Since then, criminal justice went through a deep reform, and the trials are much shorter.
356 V. Marianov

No case Bail, release Parole or

diminution of
sentence

General Defendants IC(t) OC(t)

ID(t) IT(t)
Popu- Detainees under trial Convicted
lation SD(t) ST(t) felons
SC(t)

Fig. 15.1 Inmate flow

15.3.2 Demand Forecast

The demand for the 20-year horizon was estimated for each geographic unit
(province) by using a flow model in which each type of inmate (detainees, de-
fendants under trial, and convicted felons) is modeled as a process with an input
rate, a stock and an output rate, all changing in time according to demographic and
policies changes. Some authors use models in which recidivism is also considered
(Baker and Lattimore 1994); however, information on recidivism was not available
to us. Fig. 15.1 shows the diagram of the model we used.
Note that in Fig. 15.1, the number ID(t) of individuals entering the “Detainees”
category, is the number of people that are arrested, which is proportional to the pop-
ulation of the province. The proportionality factor depends on the police efficiency
and practices. Detainees stay as such for a maximum of 5 days, so in a year, the
output of the “Detainees” process IT(t) is basically equal to the input ID(t). How-
ever, there is a stock SD(t), which is computed as the total number of detainees in a
year (approximately 200,000) divided by 365 and multiplied by the average length
of stay of the detainees in the jail. A detained person can be released or become a
defendant under trial. As a result, the input to the process “Under Trial”, IT(t), is
proportional to ID(t). The proportion of defendants that are bailed is computed from
historical data. So are the proportions of defendants that stay for different periods
of time in the jail. Since at the time of the study, people could stay under trial up to
5 years, the stock of people under trial can be computed as
⎛ ⎞
4 5
⎝ bkt ⎠ λi+1
ST (t) = 1 I T (t − i),
i=0 k=i+1
15 Location and Sizing of Prisons and Inmate Allocation 357

where bkt are the proportions of defendants whose trials finish after k years. The
subscript t reflects the fact that these proportions vary in time. The parameter λ1 is
the probability of a person under trial not being bailed out. A similar reasoning is
used to compute the stock of convicted felons, except that IC(t) consists of people
who were convicted after being under trial for 1, 2, 3, 4 and 5 years, and the stock
of people comprises people who have been in prison for times that go from less than
a year up to life, although it is uncommon to find people that have been there for
more than 19 years. More specifically, we obtain

5
I C(t) = c bit I T (t − i)
i=0
⎛ ⎞

18 
19
SC(t) = ⎝ ckt ⎠ I C(t − i)
i=0 k=i+1

All of the coefficients and factors are computed by fitting the series to historical
data.
Once a forecast was obtained by fitting the series to the historical data, three dif-
ferent inmate population growth scenarios were obtained using different projected
values of the most significant parameters: police efficiency, duration of trials, sen-
tencing and custodial policies of courts, changes in crime rates due to new types
of crimes, and changes in the percentage of offenders sentenced to programs that
are alternatives to incarceration. Also, a fourth scenario was proposed by the de-
cision maker, which considered a strong decrease in the number of inmates. The
uncertainty and hence, the need for different scenarios, was mainly due to a penal
law amendment affecting sentencing policies and trial modalities, whose applica-
tion was starting at the time of the project. As a consequence, inmate population
could equally likely increase or decrease, because there are opposite effects: more
alternatives to incarceration and shorter trials would decrease the prison population,
while the public reaction to a more effective justice would increase the number of
trials and inmates.
Additional scenarios were required because of the fact that prisons were heavily
overpopulated at the time of the study. As a consequence, some of the inmates from
provinces with overpopulated jails were being sent to prisons in nearby provinces.
Thus, the base time series of some provinces did not necessarily reflect the real
number of inmates originating in that province. Rather, it reflected a saturated con-
dition of nearby provinces. In order to correct this distortion, the fact was used that,
because of administrative reasons, inmates were never moved out of their region of
origin (provinces are grouped in 13 regions), i.e., the inmates in a region belong
to that region only. Thus, aggregating all the inmates in a region and redistributing
them later among provinces in proportion to the provinces’ population, would tend
to correct this distortion, provided that the crime rate is the same across the region.
Then, four new scenarios were devised, equivalent to the basic ones, but in which
the inmate population of a whole region was first aggregated and then, distributed
among provinces, proportionally to their populations.
358 V. Marianov

Table 15.1 Eight population growth scenarios S1 to S8

Scenario description What is affected Effect
S1 Time series Total number of inmates –
S2 Time series, plus aggregation Total number of inmates Redistribution between
of inmates in a region and provinces
subsequent redistribution
among provinces,
proportionally to their
population
S3 Increase in police Number of detainees Increases in 15 %
effectiveness
Shorter trials Number of inmates under trial Decreases in 15 %
Less alternatives to Number of sentenced inmates Increases in 20 %
incarceration
S4 Same parameters as scenario See scenario 3 See scenarios 2 and 3
3, with province
redistribution as in scenario 2
S5 Increase in police Number of detainees Increases in 4 %
effectiveness
Shorter trials Number of inmates under trial Decreases in 4 %
Less alternatives to Number of sentenced inmates Increases in 6 %
incarceration
S6 Same parameters as scenario See scenario 5
5, with province
redistribution as in scenario 2
S7 More alternatives to Number of inmates under trial Decreases in 50 %
incarceration and sentenced
Increase in police Number of detainees Increases in 15 %
effectiveness
S8 Same as scenario 7, with See scenario 7 See preceding scenario
province redistribution as in and scenario 2
scenario 2

The model was solved for the eight demand scenarios, described in detail in
Table 15.1.

15.3.3 The Model (Marianov and Fresard 2005)

For modeling effects, demand was aggregated at the geographical centers of the
provinces. The same points are the locations of courts and existing facilities, and
candidates for locations of the new facilities. Each province has a known existing
capacity for inmates. Each existing facility has a known maximum growth capacity
(in number of inmates). New facilities have preset initial and maximum capacities.
15 Location and Sizing of Prisons and Inmate Allocation 359

Some of the provinces are candidates to location of new jails, and each candidate
province has a maximum number of new facilities. Construction activity (opening
of new facilities and expansions) is considered only in the 1st, 4th, 8th, 12th, 16th,
and 20th years, the “active” years.
The system was modeled as a network, in which the nodes are the geographical
centers of the provinces. The lengths of the arcs are the road distances between
province centers. Detainees or inmates under trial must remain in the same province
of their respective courts. However, if the capacity in their province of origin is
not sufficient, sentenced offenders can be sent to other provinces. In such a case,
inmates cannot be located farther than 150 or 400 km from their families, depending
on which part of the country they are being held. The model penalizes transportation
distances.
The variables of the model are the following:
Sentijt Number of sentenced inmates which are transported from courts in
province i, to facilities in province j, on year t. These variables are de-
fined only for provinces i and j closer to each other than the maximum
transportation distance (150 or 400 km).
Locjt Number of new jails that are opened (located) in province j in year t.
Openjt Number of new jails that are either opened or remain open in province j in
year t. This variable does not need to be declared integer, since the model
structure forces it to be integer.
Capjtt Capacity (number of inmates) of existing facilities in province j in year t.
CapNjt Capacity (number of inmates) of new facilities in province j in year t.
Expjt Expansions (number of inmates) of existing facilities in province j in
year t.
ExpNjt Expansions (number of inmates) of new facilities in province j in year t.
The initial capacity of the facility is included.
Overpjt Overpopulation in province j during year t.
The parameters and sets are the following:
FCjt Fixed opening cost of a new facility at province j in year t.
ECjt Expansion cost per inmate at province j in year t. This may be different
for new and existing jails.
TPij Transportation penalty per inmate sent from province i to province j.
OPj Overpopulation penalty per inmate in province j.
DemSentit Demand of space for sentenced inmates in province i in year t.
DemTrit Demand of space for detainees and defenders under trial in province i
in year t.
CurrCapj Current capacity in province j.
PotCapj Maximum potential capacity in existing facilities in province j.
IniCapj Initial capacity of new facilities in province j.
PotCapNj Maximum capacity of new facilities in province j.
SCij Social cost per inmate of visitors’ transportation, as a percentage of
the transportation penalty of sentenced inmates.
360 V. Marianov

pj Maximum number of new facilities in province j.

Ni Set of provinces j not farther to i than the allowed maximum distance
(150 or 400 km)
The model can then be written as follows:
 ) * 
min F Cjt Locjt + ECjt ExpNjt + ECjt Expjt +
j candidate t j t
  (15.1)
+ (1 + SCij )T Pij Sentijt + OPj Overpjt
i j t j t

s.t. Sentijt = DemSentit ∀i, t (15.2)
j ∈Ni

DemT rjt + Sentijt ≤ Capjt + CapNjt + Overpjt ∀j, t (15.3)
i

Capjt = Capj (t−1) + Expjt ∀j, t (15.4)

CapNjt = CapNj (t−1) + ExpNjt ∀j, t (15.5)
Capjt ≤ P otCapj ∀ j, t = last year (15.6)
CapNjt ≤ P otCapNj ∗ Openjt ∀ j, t (15.7)
CapNjt ≥ I niCapj ∗ Openjt ∀ j, t (15.8)
Openjt = Openj (t−1) + Locjt ∀ j, t (15.9)

Openjt ≤ pj ∀ j (15.10)
t

The first term in the objective (1) minimizes the fixed cost of opening new facili-
ties (cost of land, basic services; environmental and political costs, all depending
on the province, its characteristics and population) and the expansion costs of new
facilities, including common areas and services. The second term minimizes the ex-
pansion costs of existing facilities, which are lower than those of new facilities. The
third term penalizes inmate transportation costs and visitors’ distance costs. The last
term penalizes overpopulation. This cost, per inmate, is higher than the expansion
cost (otherwise, the model will always increase overpopulation, and there will be no
capacity expansions). It was set so that a new facility opens when the overpopula-
tion in its area has reached a level that makes its opening cost effective. In agreement
with the planners, we set this level to one third of the opening capacity. In this appli-
cation, the minimum capacities of new prisons were 1000 and 1650, depending on
the province. New prisons were opened when the overpopulation in their “service
areas” reached 333 and 550 inmates, respectively, meaning that the overpopulation
15 Location and Sizing of Prisons and Inmate Allocation 361

cost of 333 inmates is equal to the opening cost of a 1000 inmate prison, and the
overpopulation cost of 550 inmates is equal to the opening cost of a 1650 inmate
prison.
Constraint (2) requires that at any year t and province i, all the demand for space
must be satisfied. Constraint (3) states that at any year t and province j, the capac-
ity of existing and new facilities must be sufficient to satisfy the demand, and that
some overpopulation is allowed. Constraints (4) and (5) indicate for existing and
new facilities respectively, that for every year t, the capacity of in province j is their
capacity in the previous year, t–1, plus the expansions on year t. For the first period,
the previous capacity of existing facilities is CurrCapj and zero for new facilities.
Constraints (6), (7) and (8) set the maximum and minimum capacities for all facili-
ties. By virtue of constraints (9), in any candidate province, the new facilities remain
open once they are located (Openj0 = 0). Finally, constraint (10) sets the maximum
number pj of facilities that can be located at any candidate province.

15.3.4 Scenario Analysis

The model was solved for each scenario independently, obtaining for each sce-
nario the locations of the new facilities and the years in which they should be open;
the expansion schedule for all facilities; the inmate allocation to facilities, and the
overpopulation that could be expected in each facility for each active year.
In most scenarios, the existing jails are expanded to their maximal sizes during
the time horizon, because this is the least expensive way of increasing the capacity of
the system. However, as Table 15.2 (Marianov and Fresard 2005) shows, locations
and numbers of new facilities strongly depend on the scenario. In the Table, the
numbers correspond to the opening year (4, 8, 12, 16, 20.)
In different scenarios, different numbers of prisons are open, ranging from 3
to 12. As there is no information about what scenarios could be most likely, we
assumed equally likely scenarios and used a minimum regret analysis to recommend
the locations for the new facilities. For the subsequent runs, instead of solving the
problem with a free number of new facilities, we agreed with the decision maker that
it would be better to fix the number of new facilities at 10. This decision, although
political, is justified because, no matter what is the result of the exercise, it must be
repeated every few years to adapt the planning to the actual scenario as it develops
in time. If too few facilities are planned at the beginning of the horizon, some of the
locations could result highly suboptimal if a populated scenario actually happens.
Furthermore, in no scenario more than 10 new facilities are open in the first 10 years,
and there is always time to add new facilities. Because of political reasons, also, the
new facilities should be planned to be open in the eighth year.
362 V. Marianov

Table 15.2 Opening years of new jails, for all eight scenarios S1 to S8
S1 S2 S3 S4 S5 S6 S7 S8 Total
number
Province
24 12 8 20 12 4
27 4 8 8 3
29 20 1
33 16 12 2
34 4 4 4 4 4 4 4 4 8
35 20 20 2
37 20 8 8 16 16 5
38 12 16 20 3
39 4 4 4 4 4 4 4 4 8
40 12 1
43 20 16 2
44 12 16 2
45 4 4 4 4 4 4 4 4 8
46 16 16 2
47 16, 20 12 8, 12 12, 16 7
Total 5 6 12 12 9 8 3 3 58
number

15.3.5 Regret Analysis

At this stage, we required an approach to decide the location of the new prisons, that
would allow reducing as much as possible the decision error, no matter what sce-
nario happened in practice. This worst-case error minimization is achieved through
regret analysis.
The regret analysis is performed by solving an aggregated model that integrates
several scenarios simultaneously. In this model, all objectives, constraints and vari-
ables of the original model are repeated for each scenario, except for the location
variables corresponding to new facilities, which are left as common to all scenarios
and make the problem non-separable by scenarios. The objective to be minimized
is the highest cost among all scenarios.
Five scenario combinations (cases) were run separately:
• Aggregation of scenarios 1, 3, 5, 7, i.e., those that are generated using the current
prison population without redistribution among provinces.
• Aggregation of scenarios 2, 4, 6, 8, i.e., those in which the demand is redistributed
among provinces according to their population.
• Three combinations of all scenarios (1 to 8), with different penalties on overpop-
ulation: a “high overpopulation” case, with a low penalty on overpopulation, thus
15 Location and Sizing of Prisons and Inmate Allocation 363

Table 15.3 Locations that minimize the worst-case scenario

Province Scenarios 1, 3, 5, 7 Scenarios 2, 4, 6, 8 High over- Medium over- Low over-
population population population
24 * * *
27 *
29 * * *
34 * * * * *
36 * * * *
37 * * * * *
38 * * * * *
39 * * * * *
41 *
43 * *
44 *
45 * * * * *
46 * * *
47 − 1 * * * * *
47 − 2 * *
Total 10 10 10 10 10
The asterisks mark the provinces in which prisons are open for each scenario group.

allowing more overcrowding; a “low overpopulation” case, with high penalty on

overcrowding and resulting in low overpopulation; and an intermediate value on
overcrowding penalty (in this case, the value used for the penalty ensures opening
of a new jail when overpopulation reaches one third of its capacity).
The results of the analysis are shown in Table 15.3 (Marianov and Fresard 2005).
In Table 15.3, the first column shows the numbers of the provinces in which jails
are open in the different runs. The numbers 47 − 1 and 47 − 2 in this column repre-
sent a first and a second jail in province 47 (a very populated province). In the second
column, the asterisks mark the provinces in which prisons are open in year 8, when
regret analysis is applied to the aggregation of scenarios 1, 3, 5, 7 (the scenarios
designed using the current prison population, without correcting facility saturation).
The third column shows the provinces in which prisons are open when the regret
analysis is performed aggregating scenarios 2, 4, 6 and 8 (the scenarios with the
corrected population distribution). The remaining columns show the provinces with
prison openings, for regret analysis performed by aggregating all scenarios, with
different values of penalization on overpopulation.
In our final analysis, we maintain first those locations that seem to be robust,
i.e., locations that appear in every possible result in Table 15.3. These are the new
facilities in provinces 34, 37, 38, 39, 45 and 47. We also recommended a jail in
province 36, since it appears in all cases, except in the not very desirable case in
364 V. Marianov

which high overpopulation is allowed. For the three remaining prisons, we choose
those candidates that appear more frequently. Finally, the proposed set of locations,
minimizing the worst case cost, is in provinces 24 (or 29), 34, 36, 37, 38, 39, 43, 45,
47 and either 46 or a second jail in 47.
If these new prisons are open, the overpopulation is reasonable in the whole
country, except for some small provinces in the central zone, which could be an
artifact due to the current offenders-prison assignment. If the prison population in
these provinces shows to be large in practice, small, inexpensive jails (150 inmates)
could be easily open in the provinces 42, 19, 16, and 13.
Furthermore, the overpopulation is low in all cases until the middle of the time
horizon. Thus, there is time enough to correct the situation, if needed.

15.3.6 Final Outcome of the Project

Given the uncertainty in the future demand for capacity, it was strongly recom-
mended to update the study every active year or even more frequently. Furthermore,
once the project was completed and its results handled to the authorities, political
considerations not taken into account previously came into play, and the planning
was partly changed. Finally, the office in charge decided to plan for the building of
10 prisons in two stages. In the first stage, 5 prisons were to be built in 5 of the pro-
posed provinces: 34, 36, 39, 44, and 45. The construction for this stage started right
away. A second stage would include provinces 33, 38, 43, 46 and 47. This second
stage was not to be initialized until some years later, and only after checking the
new change rates in prison population, which development could show a new trend,
as a product of the recent law amendments.
A second recommendation addressed the availability of data. For this project,
data had to be collected from many different sources, and there were inconsistencies
that made the data analysis very complicated. A list of useful data to be recollected
in the future was suggested, which could be used in subsequent updates of the study.
Some years after the project was finished, a different approach to addressing un-
certainties in the problem was explored, in which a tree of scenarios was defined
(Hernández et al. 2012). In this tree, there is branching at each “active” period of
time. Branches correspond to different uncertain parameter realizations in that par-
ticular period. At each active period, decisions can be made regarding construction.
Thus, the number of possible scenarios to analyze grows significantly. Once the tree
is defined, a model is formulated which aggregates individual models for each sce-
nario (similar to the model used during the project,) and constraints are added to
ensure non anticipativity and consistency between scenarios; i.e., a given scenario
does not use information available only in a later scenario, and scenarios originating
in a common branch of the tree share those parameters and context that correspond
to the common branches. In order to solve the large model, a technique is used called
branch and fix coordination (Alonso-Ayuso et al. 2003).
This new approach requires far more computational power, but provides solutions
that reduce overpopulation and the expected cost (Hernández et al. 2012).
15 Location and Sizing of Prisons and Inmate Allocation 365

15.4 Conclusions

We here present an application aimed at designing the correctional system for a

whole country. As such, it considers the expansion of the current capacities, as well
as location and opening of several large prisons. Readers interested in location or
dimensioning of individual jails, can check some available reports, e.g., Ricci Green
Associates (2013).
A previous requirement was to forecast the prison population in the coming 20
years. In our application, the forecast was complicated by the high degree of un-
certainty introduced by a law amendment, affecting the behavior of police corps,
judges and population, passed shortly before the study was started. For the forecast
to be useful, a time series is not enough: it must be analyzed and changed according
to new trends affecting crime, e.g., the increase in drug-related crime in Chile at
the time of the study, as well as the demographic and sociological changes in the
country. In synthesis, all the factors that have an influence on the number of inmates
must be analyzed and embedded into the forecast. In our case, there was an addi-
tional difficulty: historical information about prison population was obtained from
the yearly figures of existing prisons. However, some of the facilities were saturated
and inmates in the province had to be sent to prisons in other provinces. This fact
introduces a distortion in the forecast that must be corrected by reassigning inmates
between provinces, as described here.
Another particular feature of this application is the fact that the prison population
is composed by three different types of inmates: detainees staying as inmates for
5 days or less; defendants under trial, in the system for up to a few years; and
sentenced offenders. Each type of inmate has requirements in terms of location and
closeness to courts or families. In other countries, these requirements are fulfilled
by opening at least two types of facilities: jails and prisons. Jails are close to court,
while prisons need to be reachable by visitors and employees and service providers.
Finally, the location model considered cost as an objective, plus a penaliza-
tion of long distance travelled by families and visitors of the sentenced inmates.
Additionally, instead of considering hard capacity constraints, we allowed some
overpopulation, also penalized in the objective. For the locations to be found, the
model also assigned inmates to existent and new facilities. It also provided, as part
of the solution, the scheduling of the opening of the new facilities and the expansion
works for existing facilities.
Together with the results of the study, some recommendations were offered to the
decision maker, regarding the size of the future facilities and the data requirements
for future studies. Regarding facility size, although large facilities can take advan-
tage of economies of scale in staffing, supplies and services, from the point of view
of overcrowding it seems better to open facilities with small opening sizes (150–400
inmates), since large facilities require waiting until overcrowding in existing pris-
ons justifies their opening; otherwise they would be operating with a few inmates
for some time, which is not an efficient use of resources. These sizes also make
rehabilitation more likely, as show some studies (Hall et al. 2013; Johnsen 2011).
366 V. Marianov

This is probably due to the fact that the likelihood of rehabilitation increases with
the quality of life in prison, which, in turn, tends to be better when the prison size
is smaller. In any case, these issues require much more discussion, as it seems that
there are no conclusive studies.

Acknowledgments Partial support by FONDECYT grant 130265, as well as by the Institute Com-
plex Engineering Systems through Grants ICM MIDEPLAN P-05-004-F and CONICYT FBO16
is gratefully acknowledged.

References

Alonso-Ayuso A, Escudero LF, Ortuo MT (2003) BFC, a branch-and-fix coordination algorithmic

framework for solving some types of stochastic pure and mixed 0–1 programs. Eur J Oper Res
151:503–519
Armstrong S (2012) Siting prisons, sighting communities: geographies of objection in a planning
process. Working Paper, Scottish Centre for crime and justice research, University of Glasgow.
http://ssrn.com/abstract=2117840. Accessed 4 Nov 2015
Baker J, Lattimore P (1994) Forecasting demand using survival modeling: an appli-
cation to US prisons. Australas J Inf Syst 2(1). http://journal.acs.org.au/index.php/
ajis/article/view/411/372. Accessed 4 Dec 2015
Berk R (2008) Forecasting methods in crime and justice. Annu Rev Law Soc Sci 4:219–238
Brans JP, Vincke PH (1985) A preference ranking organization method, the PROMETHEE method.
Manage Sci 31:647–656
Chan H (1996) Medium security prison. Master of Architecture Thesis, The Univer-
sity of Hong Kong (Pokfulam, Hong Kong). http://hub.hku.hk/bitstream/10722/26060/
1/FullText.pdf?accept=1. Accessed 4 Nov 2015
Cherry T, Kunce M (2001) Do policymakers locate prisons for economic development?
Grow Change 32(4):533–547. http://www.appstate.edu/ cherrytl/papers/prison_location.pdf.
Accessed 4 Nov 2015
Comfort ML (2002) ‘Papa’s House’: the prison as domestic and social satellite. Ethnography
3(4):467–499
de Silva N, Cowell P, Chow T, Worthington P (2006) Prison population projec-
tions 2006–2013, England and Wales. Report of the Great Britain home office
research development and statistics directorate. https://www.ncjrs.gov/App/abstractdb/
AbstractDBDetails.aspx?id=236652. Accessed 4 Nov 2015
Eiselt HA, Marianov V (2015) Location modeling for municipal solid waste facilities. Computers
and operations research 62:305–315
Engel MR (2007) When a prison comes to town: siting, location, and perceived impacts of correc-
tional facilities in the Midwest. PhD Thesis, The graduate college at the University of Nebraska.
Lincoln, Nebraska, August 2007
Farkas MA (1999) ‘Not in my backyard’: the issues and complexities surrounding prison siting.
Justice Prof 12:95–109
Fehr L (1995) Literature review of impacts to communities in siting correctional facilities. Re-
port, Washington council on crime and delinquency, Seattle, WA. http://www.tippecanoe.
in.gov/egov/docs/1203448928_100158.pdf. Accessed 4 Nov 2015
Hall D (2006) Jails vs. prisons. American correctional association website.
http://www.thefreelibrary.com/Jails+vs.+prisons.-a0184267124. Accessed 4 Dec 2015
Hall MJB, Liu WB, Simper R, Zhou Z (2013) The economic efficiency of rehabilitative
management in young offender institutions in England and Wales. Soc Econ Plan Sci 47:38–49
15 Location and Sizing of Prisons and Inmate Allocation 367

Hernández P, Alonso-Ayuso A, Bravo F, Escudero L, Guignard M, Marianov V, Weintraub A

(2012) A branch-and-cluster coordination scheme for selecting prison facility sites under
uncertainty. Comput Oper Res 39(9):2232–2241
Hooks G, Mosher C, Genter S, Rotolo T, Lobao L (2010) Revisiting the impact of prison building
on job growth: education, incarceration, and county-level employment, 1976–2004. Soc Sci
Quart 91(1):228–244
Johnsen B, Granheim PK, Helgesen J (2011) Exceptional prison conditions and the quality
of prison life: prison size and prison culture in Norwegian closed prisons. Eur J Criminol
8(6):515–529
Johnson MP (2006) Decision models for the location of community corrections centers. Environ
Plan B 33:393–412
King RS, Mauer M, Huling T (2003) Big prisons, small towns: prison economics
in rural America. The sentencing project, Washington, DC. http://prison.ppjr.org/
files/tracy%20huling%20prisons%20economy%20study.pdf. Accessed 4 Nov 2015
Kleindorfer PR, Kunreuther HC (1994) Siting of hazardous facilities. In: Pollock SM, Rothkopf
MH, Barnett A (eds) Handbooks in operations research and management science 6. Elsevier
North-Holland, Amsterdam, The Netherlands, pp 403–440
Korporaal R, Ridder A, Kloprogge P, Dekker R (2000) An analytic model for capacity planning of
prisons in the Netherlands. J Oper Res Soc 51:1228–1237
Lawrence S, Travis J (2004) The new landscape of imprisonment: mapping America’s
prison expansion. Research Report CPR04 0121, Urban Institute, April 2004, available at
http://www.urban.org/UploadedPDF/410994_mapping_prisons.pdf. Accessed 4 Nov 2015
Maltz MD (1996) From Poisson to the present: applying operations research to problems of crime
and justice. J Quant Criminol 12(1):3–61
Marianov V, Fresard F (2005) A procedure for the strategic planning of locations, capacities and
districting of jails: application to Chile. J Oper Res Soc 56:244–251
Martin R, Myers D (2005) Public response to prison siting: perceptions of impact on crime and
safety. Crim Justice Behav 32(2):143–171
Ministry of Justice UK (2013) Prison population projections 2013–2019 England and Wales. Min-
istry of justice statistics bulletin, 7th November 2013. https://www.gov.uk/government/uploads/
system/uploads/attachment_data/file/255574/prison-population-projections-2013-2019.pdf.
Accessed 4 Nov 2015
Moran D, Pallot J, Piacentini L (2011) The geography of crime and punishment in the Russian
Federation. Eurasian Geography Econ 52(1):79–104
NAATAP (2014) Population profiles, population projections and bed needs
projections. Native American and Alaskan technical assistance project.
https://www.bja.gov/Publications/PopProfiles.pdf. Accessed 4 Nov 2015
NBC (2012) Inside the secret industry of inmate-staffed call centers. http://usnews.
nbcnews.com/_news/2012/01/12/10140493-inside-the-secret-industry-of-inmate-staffed-call-
centers. Accessed 4 Nov 2015
New Zealand Department of Corrections (2001) Inmate employment policy. http://www.
corrections.govt.nz/__data/assets/pdf_file/0007/676087/inmateemployment.pdf. Accessed 4
Nov 2015
Ricci K (2006) Jail site evaluation and selection. Bulletin from the jails division
of the National institute of corrections, April 2006. https://s3.amazonaws.com/
static.nicic.gov/Library/021280.pdf. Accessed 4 Nov 2015
Ricci Green Associates (2013) Validation study of the dutchess county criminal jus-
tice system needs assessment May 28:2013. http://www.co.dutchess.ny.us/CountyGov/
Departments/CriminalJusticeCouncil/CJValidationStudyofNeedsAssessment.pdf. Accessed 4
Nov 2015
Saaty TL (1980) The analytic hierarchy process: planning, priority setting, resource allocation.
McGraw Hill, New York
San Mateo County Sheriff’s Office (2009) Site selection. http://www.smcsheriff.com/jail-
planning/site-selection. Accessed 4 Nov 2015
368 V. Marianov

Scalia J (2004) Federal prisoner detention: a methodology for projecting fed-

eral detention populations. Report of the office of the federal detention trustee.
http://www.justice.gov/archive/ofdt/rpd_methodology2004.pdf. Accessed 4 Nov 2015
Stephan J (2008) Census of state and federal correctional facilities 2005. bureau of jus-
tice statistics, U.S. department of justice, October 2008. http://www.bjs.gov/content/
pub/pdf/csfcf05.pdf. Accessed 4 Nov 2015
US Government Accountability Office (2012) Growing inmate crowding negatively sffects in-
mates, dtaff, and infrastructure. Bureau of prisons, Report GAO-12-743. Published September
12:2012. http://www.gao.gov/products/GAO-12-743. Accessed 4 Nov 2015
Wamsley R (2013) World prison population list (10th ed.) International centre for prison
studies. http://www.prisonstudies.org/sites/prisonstudies.org/files/resources/downloads/wppl_
10.pdf. Accessed 4 Dec 2015
Wan W-Y, Moffatt S, Xie Z, Corben S, Weatherburn D (2013) Forecasting prison populations
using sentencing and arrest data. Crime and justice bulletin, NSW bureau of crime statistics
and research, Nr. 174, October 2013
Chapter 16
Vessel Location Modeling for Maritime Search
and Rescue

Ronald Pelot, Amin Akbari and Li Li

16.1 Introduction

16.1.1 Locating Maritime Search and Rescue Vessels

The location-allocation problem serves to deploy assets effectively to respond to a

known or estimated geographically distributed demand for service, including pro-
viding emergency services. At first glance, the rescue vessel location problem in the
Search and Rescue (SAR) domain is similar to the emergency vehicle location prob-
lem such as for ambulance location. In both cases, mathematical location models
can be formulated to maximize the number of incidents that can be serviced by a
specified number of resources (vehicles) within a pre-specified amount of time, or,
alternatively, we can minimize the time it would take a vehicle to arrive at the scene
of the incident. However, several differences exist. First of all, in the case of emer-
gency vehicle location, all response units are generally assumed to have the same
capability and speed. Conversely, the Canadian Coast Guard (CCG) has many dif-
ferent SAR rescue vessel types that were designed or purchased with specific tasks
in mind, and not all are equally effective at handling different incident types. Also,
the ranges vary greatly among different types of rescue vessels, so rescue vessel
capabilities need to be considered in our study. Furthermore, the method of com-
puting distances to the incidents is different as rescue vessels are patrolling on the
sea, thus requiring a land-avoidance algorithm to calculate the travel distance rather
than Euclidean or Manhattan distance metrics. However, land-avoidance distance
is calculated before performing the optimization in this study, so this distinction is
moot in this instance.

R. Pelot () · A. Akbari · L. Li

Department of Industrial Engineering, Dalhousie University, Halifax, B2H 4R2, NS, Canada,
e-mail: [email protected]
A. Akbari
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 369

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_16
370 R. Pelot et al.

This research is based on earlier models in the literature, with some modifications
for our Maritime Search and Rescue situation. Planning the future location of rescue
vessels requires the following:
• Data on incident numbers for each grid cell, as well as the incident date and time
• Candidate sites for locating rescue vessels
• Stratification of incidents by severity and determination of associated response
time category
• Examination of differences between types of rescue vessels and categorization
of rescue vessels based on the capability to response to various incident classes
• Modification and application of Location-Allocation models for Maritime Search
and Rescue scenarios
The Canadian Coast Guard determines each incident’s categorization based on
severity and they also establish the response time requirements for different cate-
gories of severity. Categorization of rescue vessel types is based on their respective
capabilities (Cameron and Pelot 2005).

16.1.2 Research Scope

16.1.2.1 Study Area

The entire Atlantic Canada region serves as our research area, partitioned into 2263
grid cells (Fig. 16.1).

16.1.2.2 Study Period

The incident data from 2000 to 2003 were used for this study and checked and
cleaned for quality control. Two sizes of grid squares were generated in the GIS
software. The grid size cannot be too large or small. If it is too large, the calcu-
lated distances will not be very precise. On the other hand, if it is too small, the
distance calculation will be computationally expensive. However, greater resolution
is advantageous near the shore where the incident density is highest. Therefore, a
specified distance of 150 nautical miles (nmi) from shore was chosen to delineate
small and large grid square areas (see the dotted line in Fig. 16.1). For the area out-
side the line, a size of 1 × 1 degree square is used, while for the area nearer to shore
a smaller size of 0.25 × 0.25 degrees is used. The historical incidents within these
grids are used as reasonable representations of demand for SAR service.
16 Vessel Location Modeling for Maritime Search and Rescue 371

Fig. 16.1 Eastern Canada SAR incidents (2000–2003) by severity class: M1 (circle), M2 (triangle)
and M3 (square)

16.1.2.3 Choice of SISAR Incidents by Classification

All incidents are categorized into severity classes, of which the following three are
relevant for this study: Class M1 are distress incidents, class M2 are potential dis-
tress incidents, and class M3 are non-distress incidents. The study scope is shown
on the map in Fig. 16.1.

16.1.3 Objectives

This study examines the optimal locations of rescue vessels in Atlantic Canada. The
ultimate objective is to ensure the maximum likelihood of saving lives and secon-
darily of mitigating property loss using available resources. The access time is an
important performance criterion for emergency service systems and the likelihood
of saving lives and property is highly related to the access time. The access time is
defined in this study as the elapsed time from the departure of each rescue vessel to
the arrival at the scene of an incident. Since all of the vessels categorized as primary
type vessels (i.e. regular lifeboats) have similar speed, access time can be approx-
imately reflected by response distance. Moreover, primary rescue vessels have a
372 R. Pelot et al.

Partial Coverage Stochastic

MCLP

Capacity

GMCLP CMCLP MEXCLP

Secondary
Service
Consider Distance and No. of Vessels Consider Coverage and Distance

Coverage Model with Coverage Model with

Secondary Service Uncovered Demand
Service

Fig. 16.2 MCLP model and extensions

limited range, so the extent of coverage should also be a key concern in this system.
Therefore a surrogate objective is to maximize the coverage within a predetermined
access time limit or to minimize the average response distance by optimizing the
location of rescue vessels.

16.2 Literature Review

Classical mathematical programming approaches for discrete location problems fall

into three main categories: covering problems, median problems (minsum models)
and center problems (minmax models). The covering problem is concerned with
covering demands within a specified response time standard, while the median prob-
lem aims to minimize the system-wide average response time. In the center problem,
the objective is to minimize maximum distance from the facilities. Covering and
median problems are widely used in the area of emergency location analysis which
is the case in this study. The maximal covering location problem (MCLP) model
provided by (Church and ReVelle 1974) has proven to be one of the most useful
facility location models from both theoretical and practical points of view. It is a
tool for siting decisions where worst case performance is a primary concern. This
is often appropriate for emergency service systems such as police, fire, and ambu-
lance service. Berman (1994) describes the maximal covering location problem on
networks. Demands and potential facility sites are represented on nodes, so the prob-
lem space forms a network of nodes. In our case, potential rescue vessel stations are
limited and incidents are represented on gridded nodes which is a discrete problem,
not continuous.
The fundamental MCLP model and some of its possible extensions can be illus-
trated in Fig. 16.2 as follows. Based on the fundamental model, MCLP, three
extensions are produced by including considerations of partial coverage (GMCLP),
16 Vessel Location Modeling for Maritime Search and Rescue 373

capacity limits (CMCLP) and stochastic factors (MEXCLP) respectively. In order

to consider the service to uncovered demand, two coverage models with secondary
service are given. These two models will be modified to apply to our research, where
response units (rescue vessels) have different capabilities.
Berman and Krass (2002) presented a generalized maximal covering location
problem. They used the concept of partial coverage in their study and defined the
degree of coverage as a decreasing function of the distance to the closest facility.
They showed that this problem is equivalent to the general uncapacitated MCLP. It
is assumed that the demand is fully covered within the minimum critical distance S,
partially covered up to a maximum critical distance T, and not covered at all outside
of the maximum critical distance.
The location-allocation problem is an essential model for many applications,
each of which has distinguishing characteristics. The ambulance and fire station
location problems are “server-to-customer” type systems. Conversely, the distribu-
tion center and warehouse location problem is “customer-to-server” type system. In
the Maritime Search and Rescue situation, the rescue vessel location problem is a
“server-to-customer” emergency service system, which is similar to the emergency
vehicle location problem.
There is a very large number of studies in the literature dealing with analyzing
the location of emergency services facilities such as health centers, ambulances, fire
stations and Maritime Search and Rescue stations. Example applications are locat-
ing hospitals (Sinuany-Stern et al. 1995), emergency medical services (Pirkul and
Schilling 1988), blood banks (Jacobs et al. 1996), and ambulances (Ball and Lin
1993). Marianov and ReVelle (1996) discussed the problems and applications in sit-
ing emergency services. They proposed a model for a queueing maximal availability
location problem taking into account the randomness of servers’ availability. Hare-
wood (2002) formulated a bi-objective programming problem to locate ambulances
on the island of Barbados. Two objectives in the model are minimizing the cost
of serving customers and maximizing multiple coverage given a certain distance
standard.
A review of location models specifically applied to healthcare is found in (Daskin
and Dean 2004). Goldberg (2004) reviewed the literature of operations research
applications in emergency services vehicles. Another review study was performed
by (Brotcorne et al. 2003) on the evolution of models in the area of ambulance
location and relocation. They divided the studies into deterministic and probabilistic
models. Griffin et al. (2008) developed an optimization model to determine the best
location and number of new Community Health Centers in a geographical network
as well as what services each such center should offer at which capacity level. The
objective is to maximize the weighted demand coverage of the target population
subject to budget and capacity constraints.
Nguyen and Kevin (2000) incorporated maximal covering and p-median location
problems into a goal programming model to assess the level of service of the cur-
rent SAR system (in terms of location of SAR aircraft and helicopters) and compare
it to the optimal solution. They also used simulation and queueing theory to exam-
ine the performance of SAR aircraft in terms of average time that incidents spend in
374 R. Pelot et al.

Table 16.1 Incident class breakdown of SISAR data for 2000–2003

Incident class Description Incident count Percentage (%)
M1 Distress 658 4.83
M2 Potential distress 1094 8.02
M3 Non-distress 10,105 74.09
Subtotal 11,857
M4 False alarm/hoax 1733 12.71
M1P Serviced by others 48 0.35
Total 13,638

queue (i.e. waiting for a response unit) for both the current situation and their pro-
posed solution, with the latter showing a significant improvement over the current
situation. In a similar study for maritime SAR, Li (2006) employed both location
modelling and simulation to establish optimal vessel placement as well as produce
several system statistics related to utilization rates and workload balance. This work
forms the basis for some of the modelling in this paper.
Azofra et al. (2007) proposed a tool for assignment of sea rescue resources to
incidents. Their methodology, based on gravitational modeling, provides a coeffi-
cient for each possible assignment based on the appropriateness of the rescue vessel
to the incident. This study only evaluates different solutions but is not trying to
propose an optimal solution.
As shown in this section, many studies have been performed in the area of emer-
gency location analysis. However, there are still some gaps remaining, particularly
in the area of Maritime Search and Rescue location modelling.

16.3 Overview of Data

16.3.1 Incident Data

All maritime incidents in which Canadian Coast Guard Search and Rescue branch
is contacted are recorded in the SISAR (Search and Rescue Program Information
Management System) database.
There are 21 types of incidents, including: grounded, capsized, stranded, etc. The
incidents are also categorized into five classes based on severity (Table 16.1)
Different incident classes may have different service times. M1 is a distress inci-
dent, M2 is a potential distress incident, and M3 is a non-distress incident. Only
M1, M2, and M3 are included in the model objective as these classes of incidents
contribute to the busy time for response utilization computations. Conversely, M4
incidents correspond to false alarms for which no resource is dispatched or, if one is,
the rescue vessel is usually called back before arriving on scene. Thus, false alarms
16 Vessel Location Modeling for Maritime Search and Rescue 375

Table 16.2 Speed and range by rescue vessel type

Rescue vessel class Max speed Average speed Range
(knots) (km/h) (knots) (km/h) (nm) (km)
Lifeboat 25 46.3 22.5 41.7 100 185
Icebreaker 20 37.0 18.0 33.3 6518 12070
Offshore patrol vessel 16 29.6 15.5 28.7 6518 12070
Inshore rescue boat 45 83.3 45.0 83.3 4.35 8
(IRB)

are not included in the objective but sometimes contribute to the busy time of rescue
vessels. M1P comprises distress or potential distress incidents which are responded
to by some other non-Canadian Coast Guard (CCG) vessel.

16.3.2 Rescue Vessels’ Capabilities Study

The CCG has limited documentation available about their rescue vessels and their
rescue vessels’ capabilities. Cameron and Pelot (2005) conducted a preliminary
study on categorizing and analyzing the capabilities of CCG rescue vessels. The
results for four main types of rescue vessels are summarized in Table 16.2.
The icebreakers and offshore patrol vessels have relatively long range which
exceeds the maximum distance in our study area. Therefore, we can assume that
these two types of vessels have unlimited range for the purposes of this analysis.
For the MCLP model with consideration of incident severity presented in
Sect. 16.4.2, we consider two types of rescue vessels with limited ranges:
• Regular lifeboat with a range of 185 km,
• IRB with a range of 8 km.
For the two modified models in Sect. 16.4.5, we consider the primary rescue vessels
and special boats with unlimited range:
• Primary Rescue vessels: Regular lifeboats with a range of 185 km,
• Special boats: Icebreakers and Offshore Patrol vessels with an unlimited range.

16.4 Solving Location Models for Maritime SAR

In this section, we apply three models, notably MCLP (Church and ReVelle 1974),
CMCL (Pirkul and Schilling 1991), and MEXCLP (Daskin 1983) to our Maritime
Search and Rescue system. In order to consider different types of response vessels,
we also make some modifications to the coverage model by incorporating workload
376 R. Pelot et al.

capacities and backup service (Pirkul and Schilling 1988) and also extend the basic
coverage model with uncovered demand service (Pirkul and Schilling 1991).

16.4.1 Maximal Covering Location Problem Model

The maximal covering location problem (MCLP) model developed by (Church and
ReVelle 1974) has proven to be one of the most useful facility location models
from both theoretical and practical points of view. It is a tool for siting decisions
where worst case performance is a primary concern. This is often appropriate for
emergency service systems such as police, fire, and ambulance service.
The objective of the fundamental MCLP model is to maximize the total coverage
of demand (incidents) within a desired service distance S by locating a fixed number
of facilities (e.g., rescue vessels). The formulation is as follows:

Max z = ai yi (16.1)
i∈I

s.t. xj ≥ yi ∀ i ∈ I (16.2)
j ∈Ni

xj = p (16.3)
j ∈J

xj ∈ {0, 1} ∀ j ∈ J (16.4)
yi ∈ {0, 1} ∀ i ∈ I (16.5)

where
I: denotes the set of demand nodes (incidents),
J: denotes the set of candidate location sites for facilities as supplied by the CCG,
ai: demand in each demand zone (i.e., grid) i,
xi : 1 if a facility is allocated to site j; 0 otherwise,
yi: 1 if demand at node i is covered; 0 otherwise,
dij: distance between node i and facility at node j,
S: the specified maximum response distance,
Ni = {j ∈ J |dij ≤ S} is the cover set for each demand zone i within the desired
response distance S of a facility site j,
p: the number of facilities to be located.
Considering the incidents as demand, and regular lifeboats as the facilities to be
located, we apply the MCLP model in the Atlantic Canada area based on gridded
incident data from SISAR 2000 to 2003. We obtain the solution for coverage accord-
ing to the number of rescue vessels to be placed, as shown in Fig. 16.3. It is apparent
16 Vessel Location Modeling for Maritime Search and Rescue 377

Fig. 16.3 Coverage curve for maximum response distance = 185 km

that the total coverage increases as the number of rescue vessels used increases. In
addition, the figure reflects the fact that as additional rescue vessels are added to
the study area, the marginal rate of additional coverage declines. As seen in the
figure, all incidents can be covered by 57 rescue vessels with a range of 185 km,
which is the range of a regular lifeboat (the primary rescue vessel in the CCG). Note
however, that 95 % of the incidents could be covered by only 17 vessels optimally
placed. This model yields the best coverage given a specified number of vessels to
be located, but a limitation is that it does not consider the workload balance. Two
extensions are applied in the following sections which smooth the workload among
response units explicitly or implicitly.

16.4.2 MCLP Model with Weights on Incident Class

The above model weights each incident class equally and just considers the total
number of all types of incidents for each grid. However, in reality, incidents with
higher severity require more coverage and faster response time. Therefore, we want
to cover as many high-severity incidents as possible. One straightforward strategy
is to assign weights to different classes of incidents. Therefore the new objective is
given as follows:

Maxz = (w1 a1i + w2 a2i + w3 a3i )yi ,
i∈I

where w1 , w2 and w3 are weights assigned to incident class M1, M2 and M3

respectively, and a1i , a2i and a3i are the number of incidents of class M1, M2 and
M3 occurring in grid i.
The resulting solutions for locating 20 rescue vessels according to five sets of
weights are summarized in (Table 16.3)
378 R. Pelot et al.

Table 16.3 Solutions of MCLP with weights (locating 20 rescue vessels)

Weight Coverage
w1 w2 w3 M1 M2 M3 Total
1 1 1 591 992 9927 11510
6 3 1 607 995 9864 11466
20 3 1 607 995 9860 11462
50 3 1 608 986 9786 11380
1 0 0 608 955 9709 11272

Table 16.4 Solutions of MCLP with weights (locating 5 rescue vessels)

Weight Rescue vessels position Coverage
w1 w2 w3 1 2 3 4 5 M1 M2 M3 Total
1 1 1 747, 759, 797, 934, 1011 319 552 6968 7839
6 3 1 747, 759, 797, 934, 1078 327 566 6069 6962
20 3 1 747, 759, 797, 934, 1078 327 566 6069 6962
50 3 1 747, 762, 934, 1011, 1078 328 558 6673 7559
1 0 0 747, 762, 934, 1011, 1079 328 558 6681 7567
1 1 0 747, 759, 797, 934, 1078 327 566 6059 6952

As the relative weight on M1 increases, the number of M1 incidents which can

be covered is increasing slightly but the total coverage is reduced. In general, the
coverage of M1 incidents is not very sensitive to the weights in this example. For
comparison, the final solution with weight of (1, 0, 0) represents the maximum
possible number of M1 incidents covered for 20 vessels with the specified 185 km
range. The second-last solution maintains this optimal M1 coverage, but with better
coverage of M2 and M3 incidents.
This exercise is repeated using only 5 rescue vessels, with results given in
Table 4.4. Similarly, as more relative weight is placed on M1, the coverage of M1
incidents increases slightly but the total coverage reduced. The second-last solution
represents the maximum coverage of M1 possible. The final solution with weights
of (1, 1, 0) represents the maximum coverage of M1 and M2 combined.
The solutions for locating 5 rescue vessels are displayed on maps in Figs. 16.4
and 16.5. The circles represent M1, while the crosses denote both M2 and M3. In
Fig. 16.4, we put equal weights of (1, 1, 1) on M1, M2 and M3. The shaded area
represents the areas which can be covered. In this case, region I with many incidents
of class M1 but relatively few total incidents is not covered whereas region II with
more total incidents and proportionately fewer M1 is covered. In Fig. 16.5, we put all
the weight on M1 and this time region I is covered and region II is not. (Table 16.4)
From the above two Tables 16.3 and 16.4, we can see that the advantage of
putting weights on the incidents is that we can obtain comparatively higher coverage
of the incidents with higher severity. However, the drawbacks are also obvious. First
16 Vessel Location Modeling for Maritime Search and Rescue 379

Fig. 16.4 Coverage with equal weights on M1, M2 and M3

Fig. 16.5 Coverage with weight (1, 0, 0) on M1, M2 and M3

of all, we cannot quantitatively determine how much the coverage would improve by
increasing the relative weight without running simulations on all possible combina-
tions. Secondly, for a particular level of M1 and M2 coverage, we cannot guarantee
that the total incident coverage is optimized. For instance, according to the fourth
and fifth rows of Table 16.4, the total coverage given weights of (1, 0, 0) is higher
than when the weights are set to (50, 3, 1), even though the coverage of M1 and
M2 remains the same. The above drawbacks also cause the jump in coverage of M3
380 R. Pelot et al.

from the solution with weights of (20, 3, 1) to (50, 3, 1). Another possible reason
for this irregularity could be that there may exist some other optimal solution with
more M2 and fewer M3 which has the same total coverage.
A more general model given as follows can address the above drawbacks to some
extent. We split the total incidents in the objective function into three terms indicat-
ing the incident numbers for each class M1, M2 and M3. Constraints of type (16.7),
(16.8) and (16.9) allow us to specify the maximum response distance and create a
specific covering set for each incident class. Moreover, we can give specific cov-
erage requirements for M1, M2 and M3 as a percent of their respective totals in
constraint sets (16.11), (16.12) and (16.13) respectively.
  
Maxz = a1i y1i + a2i y2i + a3i y3i (16.6)
i∈I1 i∈I2 i∈I3

s.t. xj ≥ y1i ∀ i ∈ I1 (16.7)
j ∈N1i

xj ≥ y2i ∀ i ∈ I2 (16.8)
j ∈N2i

xj ≥ y3i ∀ i ∈ I3 (16.9)
j ∈N3i

xj = p (16.10)
j ∈J

a1i y1i ≥ r1 T1 (16.11)
i∈I1

a2i y2i ≥ r2 T2 (16.12)
i∈I2

a3i y3i ≥ r3 T3 (16.13)
i∈I3

xj ∈ {0, 1} ∀ j ∈ J (16.14)
y1i , y2i , y3i ∈ {0,1} ∀ i ∈ I (16.15)

where
I1: denotes the set of incident grids with incidents of class M1,
I2: denotes the set of incident grids with incidents of class M2,
I3: denotes the set of incident grids with incidents of class M3,
S1: maximum response distance for an M1 incident,
S2: maximum response distance for an M2 incident,
S3: maximum response distance for an M3 incident,
dij: the land-avoided distance from grid i to grid j,
16 Vessel Location Modeling for Maritime Search and Rescue 381

y1i: 1 if demand class 1 at node i is covered; 0 otherwise,

y2i: 1 if demand class 2 at node i is covered; 0 otherwise,
y3i: 1 if demand class 3 at node i is covered; 0 otherwise,
xi: 1 if a facility is allocated to site j; 0 otherwise,
a1i: number of incidents of class M1 in grid i,
a2i: number of incidents of class M2 in grid i,
a3i: number of incidents of class M3 in grid i,
p: the number of rescue vessels to be located,
T1: the total number of M1 incidents,
T2: the total number of M2 incidents,
T3: the total number of M3 incidents,
r1: the minimum percentage of M1 incidents to be covered,
r2: the minimum percentage of M2 incidents to be covered,
r3: the minimum percentage of M3 incidents to be covered.
N1i : {j ∈ J :dij ≤ S1 },
N2i : {j ∈ J :dij ≤ S2 },
N3i : {j ∈ J :dij ≤ S3 },
N1i is the set of rescue vessel location sites eligible to cover the M1 class incidents
in grid i, N2i is the set of rescue vessel location sites eligible to cover the M2 class
incidents in grid i, and N3i is the set of rescue vessel location sites eligible to cover
the M3 class incidents in grid i.
The objective is to maximize the total number of M1, M2 and M3 incidents which
can be covered within the respective maximum response distances S1 , S2 and S3 .
Constraints of type (16.7) allow y1i to equal 1 only when one or more rescue vessels
are located within S1 of incident grid i. Similarly for constraints of type (16.8) and
(16.9), y2i and y3i can equal to 1 only when one or more rescue vessels are located
within S2 and S3 respectively. Usually the more severe the incidents are, the shorter
is the required maximum response distance, so S1 ≤ S2 ≤ S3 . The number of res-
cue vessels to be located must equal p in constraint (16.10). Constraints of type
(16.11)–(16.13) specify the percentage of incidents M1, M2 and M3 which are
required to be covered respectively.
In addition, the model can also be extended to rescue vessels with different
ranges. As previously shown in Table 16.2, the lifeboats, special boats (Icebreaker
and Offshore Patrol Vessel) and Inshore Rescue Boat (IRB) have very distinct
ranges. However, since the model is based on the concept of coverage, the spe-
cial boats which can reach anywhere in the study area should not be considered,
otherwise all incidents could be covered by just one special boat. Therefore, only
lifeboats and inshore rescue boats are considered in this model. Then constraints of
type (16.7)–(16.9) in the previous model are modified as follows:

xj ≥ y1i ∀ i ∈ I1 (16.16)
j ∈N1i

zj ≥ y1i ∀ i ∈ I1 (16.17)
j ∈M1i
382 R. Pelot et al.

Table 16.5 Locating 5 rescue vessels (maximizing coverage of M1)

No. of r1 (%) Coverage of M1 Coverage of M2 Coverage of M3 Total coverage
lifeboats (%) (%) (%) (%)
5 0 48.48 50.46 68.96 66.11
5 48.5 48.63 50.18 66.55 64.05
5 49.0 49.70 50.64 66.38 64.00
5 49.8 49.85 51.01 66.12 63.82


xj ≥ y2i ∀i ∈ I2 (16.18)
j ∈N2i

zj ≥ y2i ∀ i ∈ I2 (16.19)
j ∈M2i

xj ≥ y3i ∀ i ∈ I3 (16.20)
j ∈N3i

zj ≥ y3i ∀ i ∈ I3 (16.21)
j ∈M3i

where xj denotes regular lifeboats: xj = 1 if a regular lifeboat is located in grid j;

xj = 0 otherwise. Similarly, zj denotes inshore rescue boats: zj = 1 if an inshore
rescue boat is located in grid j; zj = 0 otherwise.
N1i = {j ∈ J :dij ≤ min (R1 , S1 )} is the set of regular lifeboat location sites
eligible to cover the M1 incident class in grid i. R1 = 185 km is the range of a regular
lifeboat. S1 is the required maximum response distance of incident class M1.
M1i = {j ∈ J :dij ≤ min (R2 , S1 )} is the set of inshore rescue boat location sites
eligible to cover the M1 incident class in grid i, and R1 = 8 km is the range of an
inshore rescue boat. Constraints of type (16.16) ensure that for any grid where a
regular lifeboat is expected to respond to the M1 incident load, at least one lifeboat
unit is located in a proximate cell within a distance that is the lesser of that boat
type’s range of 185 km and the associated required maximum response distance.
Constraints of type (16.17) apply the same restriction for IRBs, and subsequent con-
straints of type (16.18)–(16.21) extend this logic to the other incident classes M2
and M3 respectively.
If we locate 5 lifeboats and want to cover as many incidents as possible of class
M1 which is the most severe incident type, then we can set r2 = r3 = 0 to obviate
constraints (16.12) and (16.13). Varying the value of r1 , the solution results are pre-
sented in Table 16.5. Initially letting r1 = 0 yields the solution in the first row. Then
increasing the M1 coverage a little bit at a time produces better and better coverage
of M1, and meanwhile the total incident coverage decreases. When r1 increases to
49.8 %, we obtain the best coverage of 328 M1 incidents for this example.
16 Vessel Location Modeling for Maritime Search and Rescue 383

Table 16.6 Locating 5 rescue vessels (maximizing coverage of M1 and M2 combined)

No. of rescue r (%) Coverage of M1 and Coverage of M3 (%) Total coverage (%)
vessels M2 (%)
5 0 49.71 68.96 66.11
5 49.8 49.89 68.20 65.50
5 49.9 50.29 66.38 64.00
5 50.3 50.57 66.12 63.82
5 50.6 50.74 65.36 63.20
5 50.8 50.97 60.04 58.70

Since the incidents of class M2 reflect potential distress, we may want to cover as
many as possible M1 and M2 incidents combined. Setting r3 = 0, the solutions are
summarized in Table 16.6. Initially letting r = 0, where r is the minimum coverage
required based on the percentage of M1 and M2 incidents combined, we get the
solution in the first row. Then increasing the coverage of M1 and M2 a little bit
each time, we can get better and better coverage of M1 and M2 combined, and
meanwhile the total coverage decreases. When r increases to 50.8 %, we obtain the
best coverage of 50.97 % for both M1 and M2 incidents combined. It is remarkable
that the slight increase in M1 + M2 coverage in the last row results in a dramatic
drop in the percentage of M3 incidents covered.

16.4.3 MCLP with Workload Capacities

The fundamental MCLP model does not consider balancing the workload among
facilities, therefore some response units may cover a disproportionate number of
the incidents while other units may not have much work to do. When designing a
model, capacity or workload limits can serve as a means of smoothing the workload
among the different response units. In order to impose the workload constraint on
each facility, a binary decision variable xij is introduced into the CMCLP model
objective function as follows (Pirkul and Schilling 1991):

Max Cij ai xij (16.22)
i∈I j ∈J

s.t. yj ≤ p (16.23)
j ∈J

xij = 1 ∀ i ∈ I (16.24)
j ∈J

xij ≤ yj ∀ i ∈ I, j ∈ J (16.25)
384 R. Pelot et al.

Cij ai xij ≤ Kj ∀ j ∈ J (16.26)
i∈I

yj ∈ {0, 1} ∀ j ∈ J (16.27)
xij ∈ {0, 1} ∀ i ∈ I, j ∈ J (16.28)
where
I: the index set of all demand points,
J: the index set of all potential facility sites,
ai: the demand at point i,
Kj: the workload capacity for a facility at site j,
p: the number of facilities to be sited,
S: the desired maximum response distance,
dij: the travel distance from facility j to demand i,
Cij: 1, if dij ≤ S; 0 otherwise,
xij : 1, if the demand at point i is assigned to a facility at j; 0, otherwise,
yj : 1 if a facility is located at j; 0 otherwise.
xij equals 1 if the demand at point i is served by a facility in grid j, otherwise
xij equals 0. Except for the capacity constraints, the CMCLP model is quite sim-
ilar to the original MCLP. However, the introduction of binary decision variable
xij makes the number of decision variables and constraints increase exponentially,
which makes the model substantially more difficult to solve.
Our problem has 2263 demand zones which means xij is of dimension
2263 × 2263. This is a large-scale problem which cannot be solved easily using
an exact approach. In actuality however, we do not need to consider all 2263 grids.
First of all, not all of the grids have incidents. Secondly, each incident can just be
covered by a limited number of grids given the lifeboat range of 185 km. Two meth-
ods to compress the data and reduce the problem size making it easier to solve are
given as follows.

16.4.3.1 Pick Certain Candidate Sites

Since our objective is to maximize the total number of incidents that can be covered,
we select 722 grids of the 2263 total in which one or more incidents occurred from
2000 to 2003 as the demand grid set. Usually rescue vessels would be better located
at a position which has more incidents. So we pick 183 grids which have 10 or more
incidents of the original 2263 grids as the potential location set. Then the problem
size becomes, 722 × 183 which is still large but can be solved by the CPLEX solver
(IBM ILOG CPLEX). This method is simple but not as precise because we assume
the rescue vessels are located in grids with 10 or more incidents. However select-
ing candidate sites according to some geographic or other environmental factors is
very practical in reality. Usually rescue vessels cannot be located just anywhere, and
most of time vessels are located near shore. So the candidate sites can be narrowed
down greatly.
16 Vessel Location Modeling for Maritime Search and Rescue 385

Table 16.7 Solution results of two data compression methods

No. of rescue Range Capacity No. of Coverage by Coverage by
vessels km constraints candidate sites CMCLP MCLP
20 185 ≤ 790 183 11152 94.05 % 11510 97.07 %
20 185 ≤ 790 2263 11454 96.60 %
50 185 ≤ 316 183 9235 77.89 % 11841 99.87 %
50 185 ≤ 316 2263 9505 80.16 %

16.4.3.2 Workload Capacity

We use the formula below to establish the maximum workload capacity:

Total incidents number
K = Average workload capacity = ×α
Number of vessels located
where α is the coefficient of capacity and α ≥ 1 as we assume that no rescue ves-
sel will be assigned a priori a lower than average workload. α = 1 means every
response unit is assigned a strictly equal workload. We set α = 4/3 in our problem.
Based on the data from 2000 to 2003 with a total number of 11,857 incidents, if
locating 20 rescue vessels:
11857 4
Workload Capacity = × = 790 incidents/vessel
20 3
If locating 50 rescue vessels:
11857 4
Workload Capacity = × = 316 incidents/vessel
50 3
Table 16.7 shows that the more candidate sites we choose, the better the solution
we obtain. However, selecting fewer candidate sites based on incident rate is fairly
efficient. Choosing grids with 10 or more incidents as candidate sites yields 183
options for placing the rescue vessels. However, the total coverage is only 2.55 %
less than choosing entire the 2263 grids as candidate sites for locating 20 rescue
vessels, and for locating 50 rescue vessels there is only 2.27 % difference in the
optimal CMCLP coverage.
In order to see how the total coverage is related to workload capacity balance
when varying the number of rescue vessels to be located, we solve the above model
varying the value of α from 1.1 to 3.5. The solutions are displayed in Fig. 16.6.
While all solutions converge to very high coverage when workload balance is
relaxed (except for the 10 vessel scenario), the more interesting result is when
greater parity is maintained, such as α = 4/3 as per the earlier assumption. In
this case, the effectiveness varies significantly, with 20 vessels providing the best
combination of coverage in this example.
386 R. Pelot et al.

Fig. 16.6 CMCLP solution for varying number of rescue vessels and α

16.4.4 MCLP with Stochastic Considerations

All of the above models assume that each rescue vessel has response capability to
handle any request for service at any time. However, this is not the case in reality
for Maritime Search and Rescue. The maximal expected covering location model
(MEXCLP) given by (Daskin 1983) extended the basic MCLP model by incorpo-
rating stochastic considerations of availability. The objective here is to maximize
expected coverage given that rescue vessels are busy (unavailable) with a probabil-
ity p. Given that a demand node can be covered by m facilities, then the probability
that this node is covered by at least one available response unit is 1 − pm .
There are three assumptions in this model:
• Response units operate independently
• Response units have the same busy probabilities
• Response units’ busy probabilities are invariant with respect to their locations
Under the above assumptions, the number of available facilities at any time follows
a binomial distribution:  
M
p(j facilities being available out of M located) = (1 − p)j pM−j ,
j
j = 0, 1, . . ., M.
The formulation of (Daskin 1983) is as follows:


N 
M
Max (1 − p)p j −1 ak yj k (16.29)
k=1 j =1
16 Vessel Location Modeling for Maritime Search and Rescue 387


N 
M
s.t. Cki xi ≥ yj k ∀ k (16.30)
i=1 j =1


N
xi ≤ M (16.31)
i=1

xi = 0, 1, 2, . . . , M ∀ i (16.32)
yj k ∈ {0, 1} ∀ j, k (16.33)

where
M: number of facilities to be located,
N: number of nodes in the entire study area,
S: the desired maximum response distance,
dki: the travel distance from facility i to demand k,
Cki: 1, if dkl ≤ S; 0, otherwise,
xi: the number of facilities located at node i,
yjk : 1 if node k is covered by at least j facilities; else 0, if less than j,
ak: the demand amount at node k.
The objective function (16.29) can be rewritten as:
 

M 
N
j −1
(1 − p)p ak yj k .
j =1 k=1

The solutions of locating 10, 20, 30, 40 and 50 rescue vessels respectively with
the busy (i.e., unavailable) probability ranging from 0.05 to 0.35 are graphically
compared in Fig. 16.7. As the number of rescue vessels located increases, the total
coverage is less sensitive to the probability of being busy.
Another useful tool for planning emergency response systems is the hyper-
cube queuing model, first proposed by Larson (1974). In this model, each server
in the system can be represented individually thus allowing for more complex
assignments, while accounting for temporal and geographic characteristics of the
area. An adjusted MEXCLP, model and a hypercube optimization procedure were
given by (Batta et al. 1989), which relaxed the three assumptions in the origi-
nal MEXCLP model. Based on their results, there is disparity among the original
MEXCLP, adjusted MEXCLP, and hypercube optimization procedure for large busy
probability p.
For our Maritime Search and Rescue system which does not experience heavy
utilization, the hypercube model is not necessary. For locating 20 vessels, the aver-
age probability of rescue vessel being busy is p ≈ 0.15; for locating 50 vessels,
p ≈ 0.05.
388 R. Pelot et al.

Fig. 16.7 Coverage percentage by varying the busy probability

16.4.5 Coverage Model with Response Units Having Unlimited

Range

The three models applied above are for considering regular lifeboats, the principal
type of rescue vessel. If we want to include the special boats with infinite range in
the model, we have to modify the formulations.

16.4.5.1 Coverage Model with Secondary Service

The coverage model with workload capacities and backup service given by (Pirkul
and Schilling 1988) ensures that each demand can be serviced by both a primary
and secondary response units using the following two constraints:

xij = 1 ∀ i ∈ I,
j ∈Ji

zij = 1 ∀ i ∈ I,
j ∈Ji

The first constraint requires that each demand point i can be serviced by a response
unit located at j within the maximum prescribed response distance. The second
constraint requires that each demand point can also be responded to by a backup
(secondary) response unit which can be beyond the maximum prescribed response
distance.
Combining the two concepts of secondary service and workload capacities as a
variation of the above model produces more realistic Maritime Search and Rescue
16 Vessel Location Modeling for Maritime Search and Rescue 389

scenarios. The regular lifeboat with a range of 185 km can be used to give primary
service, while special boats which can go to anywhere in our study area can provide
secondary service. If an incident can be responded to by both a regular lifeboat and
a special boat, the primary service is preferred. The above model requires both a
primary and a secondary service for each demand point which is suitable for heavily-
used systems. However, in our application, the average vessel utilization is less than
0.05 for locating 60 vessels, which is quite low. In addition, since the number of
rescue vessels is limited, a term is added to the objective function to represent the
respective costs of locating regular lifeboats and special boats. The modified model
can be expressed as follows:
⎛ ⎞
    
Min ⎝ cij ai xij + 
cik ai yik ⎠ + vj rj + wk sk (16.34)
i∈I j ∈J k∈K j ∈J k∈K

s.t. xij ≤ hij ∀ i ∈ I, j ∈ J (16.35)


xij ≤ Mrj ∀ j ∈ J (16.36)
i∈I

yik ≤ Msk ∀ k ∈ K (16.37)
i∈I

p1 − xij ≤ Mqi ∀ i ∈ I (16.38)
j ∈J

xij − p2 ≤ Mqi ∀ i ∈ I (16.39)
j ∈J

hij ≤ M(1 − qi ) ∀ i ∈ I (16.40)
j ∈J

1− yik ≤ Moi ∀ i ∈ I (16.41)
k∈K

1− hij ≤ M(1 − oi ) ∀ i ∈ I (16.42)
j ∈J
 
xij + yik = 1 ∀ i ∈ I (16.43)
j ∈J k∈K

xij ai ≤ KjR ∀ j ∈ J (16.44)
i∈I

yik ai ≤ KkS ∀ k ∈ ∀ K (16.45)
i∈I
xij ≥ 0 ∀ i ∈ I, j ∈ J ;yik ≥ 0 ∀ i ∈ I, k ∈ K; qi ∈ {0, 1} ∀ i ∈ I ; i ∈ I ;
oi ∈ {0,1} ∀ i ∈ I (16.46)
390 R. Pelot et al.

where
I: the index set of all demand nodes,
J: the index set of all potential positions for regular lifeboats,
K: the index set of all potential positions for special boats,
M: a large number,
S: desired maximum response distance (range of regular lifeboat),
p1 , p2: upper and lower bounds on the primary service percentage in a grid, which
can be covered by grids other than itself,
vj: the cost of locating a regular lifeboat in grid j,
wk: the cost of locating a special boat in grid k,
ai: the number of incidents in grid i,
KjR : the maximum demand that can be served by a regular lifeboat in grid j,
KkS : the maximum demand that can be served by a special boat in grid k,
xij: the fraction of incidents in grid i serviced by a regular lifeboat in grid j,
yik: the fraction of incidents in grid i serviced by a special boat in grid k,

1, if dij ≤ S
hij =
0 otherwise

1, if a regular lifeboat is located in grid j
rj =
0, otherwise

1, if a special boat is located in grid k
sk =
0, otherwise

The objective function (16.34) is to minimize both fixed and variable ‘costs’ such
that sufficient regular and special boats are located to provide each incident grid
with primary or secondary service.
The parameter cij is the cost of providing primary service by a regular lifeboat
 is the cost of providing secondary service by a special boat
from grid i to grid j. cik
from grid i to grid k. The cost can be interpreted in several ways. One simple and
practical way is to replace it by distance. To scale costs between 0 to 1, we define:

cij = dij / max dij


cik = dik / max dik

where
dij : land-avoided distance from incident grid i to grid j,
dik : land-avoided distance from incident grid i to grid k,
max dij : maximum distance from all incident grids to any potential position for
regular boat,
16 Vessel Location Modeling for Maritime Search and Rescue 391

max dik : maximum distance from all incident grids to any potential position for
special boat.
The constraints of type (16.35) state that incidents can only be serviced by a regular
lifeboat (primary service) which is within the specified range. K is the index set
of all potential rescue vessel positions for special boats. However, since the special
boat has unlimited range and can go anywhere in the study area, we do not have a
comparable set of constraints on yik , as any incident can be assigned to a special
boat.
Constraints of type (16.36) specify that a regular lifeboat must be placed at j if
it is to provide any primary service to incidents in grid i. Similarly, constraints of
type (16.37) specify that a special boat must be placed at k if it is to provide any
secondary service to incidents in grid i.
Constraints of type (16.38)–(16.42) specify the primary service percentage,
which can be logically represented as follows:
For all (i ∈ I )
{If

Cij > 0
j ∈J

then

p1 ≤ xij ≤ p2 ;
j ∈J

Else

yik = 1;
k∈K

End if;
};
For each incident grid, if there exists one or more candidate rescue vessel posi-
tions for a regular lifeboat, we require that the primary service percentage be
between p1 and p2 . Otherwise the incident will be fully serviced by special boats.
Constraints of type (16.43) state that the total service amount of each incident
grid by both regular lifeboat and special boat should be equal to 1, which means
every incident has to be assigned to exactly one response vessel.
Constraints of type (16.44)–(16.45) place bounds on rescue vessel workload
capacity. For a regular lifeboat we have a capacity of KjR and special boats have
a capacity of KkS .
Applying this model to the Canadian Search and Rescue Problem, we can iden-
tify the locations for rescue vessels such that all incidents around the Atlantic area
will receive either a primary or a secondary service or both. The demand network
consists of 2263 grids covering the whole Atlantic study area, 722 of which have
392 R. Pelot et al.

Table 16.8 Solutions for varying primary service percentage

Primary service No. of No. of Demand Avg. dist of Avg. dist of Total avg.
percentage (%) regular special serviced of primary km secondary dist. km
lifeboats boats primary (%) km
0–10 1 46 263 (2.2 ) 48.1 74.6 74.0
10 − 20 33 40 1983 (16.7 ) 55.0 71.7 68.9
20 − 30 34 35 3116 (26.3 ) 55.5 78.6 72.6
30 − 40 36 30 4414 (37.2 ) 51.5 83.1 71.3
40 − 50 38 26 5439 (45.9 ) 49.0 88.6 70.4
50 − 60 40 24 6385 (53.9 ) 50.1 95.1 70.8
60 − 70 42 18 7441 (62.8 ) 47.5 128.3 77.6
70 − 80 46 13 8690 (73.3 ) 46.9 157.4 76.4
80 − 90 50 10 9847 (83.1 ) 45.8 193.7 70.9
90 − 100 52 3 11264 48.7 407.1 66.6
(95.0 )

one incident or more. We select 183 of these nodes with more than 10 incidents as
potential candidate sites for regular lifeboats and special boats. Based on the 11,857
total number of incidents and assuming a fleet of about 60 rescue vessels, choos-
ing α = 43 to ensure equitable workload balance, the rescue vessel capacity can be
defined as:
Total incidents number 11857 4
Kj = ×α = × ≈ 263
Number of vessels located 60 3
Therefore, we set

KjR = KkS = 263 ∀ j ∈ J, k ∈ K.

We set fixed cost equals to 5 and vary the primary service percentage from [0, 0.1]
to [0.9, 1], yielding the location configurations shown in Table 16.8. For example,
when the primary service percentage is over 70 %, the total number of rescue vessels
is 59 of which the number of special boats equals 13.
To describe the solution more clearly, we display it graphically in Fig. 16.8. We
can see that as the required percentage of primary service increases, the number of
regular lifeboats is increasing while the number of special boats drops. The average
distance of service for regular lifeboats does not show much variation as their range
is restricted. However, the average distance of service for special boats increases
sharply as the service percentage increases beyond 60 %. The reason for this phe-
nomenon is that although the number of special boats and the fraction of secondary
service decrease as the primary service percentage increases, the secondary service
extent is not reduced as fast as the number of special boats needed. Therefore the
average response distance by secondary service increases dramatically, especially
when the primary service percentage is relatively high. (Table 16.8)
16 Vessel Location Modeling for Maritime Search and Rescue 393

Fig. 16.8 Number of regular lifeboats and special boats needed, and average distance of primary
and secondary service when varying the primary service percentage

The average distance of regular lifeboats (primary service) and special boats
(secondary service) can be calculated as follows:
 
Average distance per incident for regular lifeboats = dij ai xij / ai xij
i∈I j ∈J i∈I j ∈J

 
Average distance per incident for special boats = dik ai yik / ai yik
i∈I k∈K i∈I k∈K

Generally the more rescue vessels to be located, the lower the total average travel
distance obtained. When the primary service percentage is between 60 and 90 %,
according to the results, the total number of regular lifeboats and special boats are
close to the actual situation in Canadian Search and Rescue in that time period,
giving a fairly short average response distance of primary service, average response
distance of secondary service, and total average distance. Therefore, in the next
part, when we examine the tradeoff between the travel distance and the fixed costs
of locating rescue vessels, we will fix the primary service percentage to the range
70–80 %.
The objective function (16.34) in the preceding formulation serves to minimize
both the number of rescue vessels needed and the total response distance. In order
to trade off between these objectives, we include a weight W into the objective
function:
⎛ ⎞ ⎛ ⎞
     
Min ⎝ cij ai xij + cik ai yik ⎠ + W ⎝ vj rj + wk sk ⎠
i∈I j ∈J k∈K j ∈J k∈K
394 R. Pelot et al.

Fig. 16.9 Objective function tradeoffs between the number of rescue vessels located and the
average response distance

Table 16.9 Solutions for objective function tradeoffs

W No. of No. of Total no. Avg. dist of Avg. dist of Total avg.
regular special of vessels primary service km secondary service km dist km
vessels vessels
0.01 183 183 366 12.6 59.2 27.1
0.05 177 101 278 12.5 72.0 28.5
0.1 143 76 219 14.3 74.0 30.8
1 56 23 79 31.4 114.3 53.5
5 46 13 59 46.9 157.4 76.4
10 45 12 57 54.9 179.8 87.5
50 45 12 57 83.4 254.4 127.4

Requiring the primary service percentage to lie between 70 and 80 %, and set-
ting the fixed costs vj and wk to 1, seven location configurations are generated
by varying W (Table 16.9). The average travel distance is used here rather than
the total travel distance since the former is typically more meaningful when vary-
ing the number of rescue vessels located. As the weight on fixed cost W increases
from 0.01 to 50, the number of rescue vessels located decreases and the average
response distance increases. The first solution in Table 16.9, has the minimum
travel distance and assigns a vessel to all the potential positions as there is actually
little penalty for doing this (i.e. W = 0.01). The last solution represents the mini-
mum fixed cost investment, but at the expense of increased response distance. The
objective tradeoffs are graphically illustrated in Fig. 16.9.
16 Vessel Location Modeling for Maritime Search and Rescue 395

16.4.5.2 Coverage Model with Uncovered Demand Service

The model described in Sect. 16.4.5.1 considers the total travel distance in the objec-
tive. Sometimes, service within range or maximum response distance can be treated
having the same service level, which means if an incident can be responded to
by primary service, we would not care much about the exact distance. Therefore,
in the following model, we consider both the total coverage by primary service
and the total travel distance to “uncovered” incidents which cannot be covered by
regular lifeboats. For incidents inside the covering set which can be serviced by
both the regular lifeboats and special boats, a preference is given to using regu-
lar lifeboats for response, while incidents which cannot be covered by any regular
lifeboat must be responded to by special boats. We developed a modification to an
existing coverage model with capacity limits and uncovered demand service (Pirkul
and Schilling 1991) to make it suitable for Maritime Search and Rescue scenarios.
We have two objectives: maximize the coverage by primary service and minimize
the total distance to the incidents which cannot be covered by regular lifeboats. The
multiple objective functions would be written as follows:
 
Max Cij ai xij , Min ai d(i, k)yik
i∈I j ∈J i∈I k∈K

where:
ai: is the number of incidents in grid i,
xij: 1 if an incident in grid i is assigned to a regular lifeboat in grid j; and 0
otherwise,
yik : 1, if an incident in grid i is assigned to a special boat in grid k; and 0
otherwise,
Cij : 1 if grid i can be covered by a vessel in grid j; and 0 otherwise,
d(i,k): is the distance from incident grid i to potential special boat position k.
The second objective can be expressed by a more practicable definition as (Pirkul
and Schilling 1991):

Max vik ai yik ,
i∈I k∈K

where vik is the value of the service level for the incidents provided by secondary
service:
⎧
⎨ 1, if dik ≤ S
vik = - .
⎩α 1 − max{d dik −S
−S} , if dik > S
ik

S is the range of regular lifeboats and max dik is the maximum land-avoided distance
from an incident grid to a potential position for special boats in our study area. If the
396 R. Pelot et al.

vik

1 α =1

1
α=
2

α =0

d ik
S max dik

Fig. 16.10 Value of vik varying α from 0 to 1 (Pirkul and Schilling 1991)

distance is less than or equal to the range, we set vik = 1 representing full service.
For the situation where distance is greater than the range, we let
# $
dik − S
vik = α 1 − , vik ∈ [0, α), α ∈ [0,1].
max{dik − S}

The longer the distance is, the lower the service level vik will be. The constant
α acts as a relative weight tradeoff between the total incident number covered by
regular lifeboats and the service level of uncovered incidents which are serviced
by special boats. As α decreases from 1 to 0, less emphasis is placed on incidents
beyond range S, and therefore more relative weight is given to incidents which are
covered by regular lifeboats (Fig. 16.10).
To reflect that for an incident which can be served by both a regular lifeboat and
a special boat, preference is given to the regular one, a penalty term on special boats
is added to the objective function.
The model is as follows:
  
Max ai xij + vik ai yik − p sk (16.47)
i∈I j ∈J i∈I k∈K k∈K

s.t. xij + yik = 1 ∀ i ∈ I, j ∈ J, k ∈ K (16.48)

xij ≤ rj Cij ∀ i ∈ I, j ∈ J (16.49)
yik ≤ sk ∀ i ∈ I, k ∈ K (16.50)

ai xij ≤ KjR ∀ j ∈ J (16.51)
i∈I

ai yik ≤ KkS ∀ k ∈ K (16.52)
i∈I
16 Vessel Location Modeling for Maritime Search and Rescue 397
 
rj + sk = 60 (16.53)
j ∈J k∈K

xij ∈ {1,0} ∀ i ∈ I, j ∈ J ; yik ∈ {1,0} ∀ i ∈ I, k ∈ K;

rj ∈ {1, 0} ∀ j ∈ J ; sk ∈ {1, 0} ∀ k ∈ K (16.54)

where
I: the index set of all demand nodes,
J: the index set of all potential positions for regular lifeboats,
K: the index set of all potential positions for special boats,
ai: the number of incidents in grid i,
KjR : the maximum demand that can be served by a regular lifeboat in grid j,
KkS : the maximum demand that can be served by a special boat in grid k,
xij : 1, if an incident in grid i is serviced by a regular lifeboat in grid j; and 0,
otherwise,
yij : 1, if an incident in grid i is serviced by a special boat in grid k; 0, otherwise,

1, if dij ≤ S
Cij =
0, otherwise

1, if a regular lifeboat is located in grid j
rj =
0, otherwise

1, if a special boat is located in grid k
sk =
0, otherwise

p: Penalty for locating special boats.

The objective function (16.47) is to maximize the total service level including both
the incidents which can be covered by the regular lifeboats and the uncovered inci-
dents which are serviced by special boats. Constraints of type (16.48) state that
each incident grid must be responded to either by a regular lifeboat or by a special
boat. Constraints of type (16.49) guarantee that incident i can be assigned to regular
lifeboat j only if a regular lifeboat is located in grid j and the distance from i to j
is within the range. Similarly, constraints of type (16.50) state that incident i can
be serviced by a special boat at k only if there is a special boat located in grid k.
The requisite workload limits for all potential positions are imposed by constraint
sets (16.51) and (16.52). Finally, the total number of rescue vessels is restricted in
constraint set (16.53).
Computational experiments show that the results obtained are sensitive to the
value of p. Varying p from 0 to a very small value can make the number of special
boats decrease greatly (Fig. 16.11). For example, setting p equal to 0.001 and α = 4,
the number of special boats drops dramatically from 29 to 10 compared with setting
p equal to 0.
398 R. Pelot et al.

Fig. 16.11 Number of special boats by varying the value of p

Fig. 16.12 Regular lifeboats and special boats located as a function of α

The solution configurations obtained when varying the value of α from 0 to 1

are summarized in Table 16.10. As α increases, more emphasis is placed on inci-
dents beyond range S, hence less relative service value is given to incidents which
are responded to by primary service (regular lifeboats). Therefore, the number of
regular lifeboats located decreases and the number of special boats increases, while
the coverage by regular lifeboats gets worse and the average distance of special
boats improves. The solution results with p = 0.001 are graphically illustrated in
Fig. 16.12 and Fig. 16.13.
Based on the above results from the coverage model with uncovered demand
service, as the value of α increases from 0 to 1, the number of regular lifeboats
decreases from 58 to 47, while the number of special boats increases from 2 to 13.
The average distance of secondary service improves greatly without significantly
16 Vessel Location Modeling for Maritime Search and Rescue 399

Table 16.10 Solutions by coverage model with uncovered demand service (p = 0.001)
Alpha No. of regular No. of special Coverage by primary Avg. dist of
lifeboats boats service (%) special boats km
0 58 2 11467 (96.7 ) 1534.5
0.001 57 3 11449 (96.6 ) 528.7
0.01 52 8 11395 (96.1 ) 454.5
0.05 50 10 11366 (95.9 ) 437.5
0.1 49 11 11216 (94.6 ) 365.3
0.5 48 12 11035 (93.1 ) 304.4
1 47 13 11010 (92.9 ) 298.0

Fig. 16.13 Coverage by regular lifeboats and average distance of special boats

affecting the coverage of primary service. However, special boats are usually expen-
sive to purchase and use. Therefore, we can obtain the solution for a prescribed
number of special boats by adjusting the values of p and α. For example, if we have
12 special boats, we can set p = 0.001 and α ≤ 0.5.

16.5 Summary

Although the location–allocation problem has been thoroughly addressed by

researchers in many applications, there is scope for further developments relating to
resource allocation in the Maritime Search and Rescue field. Scenarios in Maritime
Search and Rescue are similar to the emergency vehicle location problem except
for the following two differences: a land-avoidance algorithm must be used to cal-
culate the travel distance rather than Euclidean or Manhattan distance metrics; and
response unit (vessel) capability needs to be considered. The review of previous
research provided a framework of applicable methods for the maritime resource
location problem.
400 R. Pelot et al.

The current study reviewed the methods and models in emergency vehicle
location problem and provided the following results:
• Summarizes previous research on emergency vehicle location problems which
can be useful for maritime resource location scenarios;
• Applies three optimization models of MCLP, MEXCLP and CMCLP to Maritime
Search and Rescue for locating lifeboats, the principal response vessels of the
Canadian Coast Guard;
• Modifies the coverage model with secondary service to consider special boats
with (essentially) infinite range, considering tradeoffs between the two objectives
of total response distance and the cost of locating and using regular lifeboats and
special boats;
• Modifies the coverage model with uncovered demand service to consider special
boats with unlimited range, tradeoffs between the two objectives of incident cov-
erage by regular lifeboats (primary service) and average response distance to the
“uncovered” demand (secondary service).
The models we applied and modified are either single objective or two objectives
combined by some weight or coefficient. In the first modified model with secondary
service, we put a weight W into the objective function to trade off between total
response distance and fixed cost of locating rescue vessels. For the second modified
model with consideration of uncovered incidents service, we use a coefficient α to
determine the service level by special boats for responding to uncovered incidents.
Also, in order to give the priority to primary service for the incidents which can be
responded to by both primary and secondary service, we add a penalty term into
the objective function associated with a penalty coefficient p. Although sensitivity
analysis was conducted in this study, a drawback is that it is hard to determine
reasonable values of weights or coefficients beforehand.
Typically, when some criteria get better, others will get worse. For instance, the
MEXCLP and CMCLP models can give better response time and workload balance
but worse total incident coverage than the MCLP model. Therefore, multi-criteria
approaches could be applied to consider several criteria such as coverage, response
time, workload balance, vessel utilization balance and budget simultaneously, with
weights provided by subject matter experts.
Generally, the approach used in this study was to develop different mathemat-
ical location models for the case of Maritime Search and Rescue. Starting with
simpler models that highlight the key elements of our SAR scenarios, the models
progressively become more realistic by adding features and additional constraints.
Of course adding constraints engenders a tradeoff of greater realism at the expense
of a poorer objective value. Also, all the models do not have exactly the same objec-
tive. Although maximizing coverage is the main objective in most of the models,
we have included other criteria such as cost and average distance from covered and
uncovered demands as well. Hence, it is not useful to directly compare the solutions
of different models to each other as they are not compatible. Overall, what we can
say about the models’ performance is that those incorporating multiple criteria and
considering additional constraints such as capacity and server availability work well
16 Vessel Location Modeling for Maritime Search and Rescue 401

in our case as they generate acceptable solutions with regards to the coverage com-
pared to the basic models like MCLP. Ultimately, the best model to use among these
alternatives could be selected based on the decision makers’ needs and preferences.
In this research, we assume that all the grid squares can be used to locate res-
cue vessels, which can yield optimal solutions in our study area. However, optimal
does not mean necessarily practical. In the actual system, rescue vessels are usu-
ally placed near the shore and around ports. The models in this research can be
modified to be more practical by adding vessel position constraints based on more
information about vessel location requirements from the Canadian Coast Guard.
While the model results described herein were not implemented directly by the
Canadian Coast Guard, they do form part of a suite of investigations performed
over many years by our MARIN (Maritime Activity & Risk Investigation Net-
work) research group to address diverse aspects of maritime accident prevention
and response. Our optimization models have been used to help guide decisions
on the location of new lifeboat stations across the country, planning for opening
of new waterways to remote areas, Search & Rescue Needs Analysis, and recom-
mended adjustments to coverage when CCG vessels are out of service for planned
or unplanned maintenance.

Acknowledgement The authors are grateful for the support provided by the Natural Sciences and
Engineering Research Council of Canada for this work.

References

Azofra M, Perez-Labajos C, Blanco B, Achutegui J (2007) Optimum placement of sea rescue

resources. Saf Sci 45:941–951
Batta R, Dolan J, Krishnamurthy N (1989) The maximal expected covering location problem
revisited. Transp Sci 23:277–287
Ball M, Lin F (1993) A reliability model applied to emergency service vehicle location. Oper Res
41(1):18–36
Berman O (1994) The p-maximal cover—-p-partial center problem on networks. Eur J Oper Res
72:432–434
Berman O, Krass D (2002) The generalized maximal covering location problem. Comp Oper Res
29:563–581
Brotcorne L, Laporte G, Semet F (2003) Ambulance location and relocation models. Eur J Oper
Res 147:451–463
Cameron A, Pelot R (2005) Search and rescue vessel capability study. MARIN group, Department
of Industrial Engineering, Dalhousie University, Halifax
Church R, ReVelle C (1974) The maximal covering location problem. Pap Reg Sci Assoc 32:101–
118
Daskin M (1983) A maximal expected covering location model: formulation, properties and
heuristic solution. Transp Sci 17:48–70
Daskin MS, Dean LK (2004) Location of health care facilities. In: Brandeau ML, Sainfort F, Pier-
skalla W (eds) Operations research and health care: a handbook of methods and applications,
Kluwer Academic Publishers, Drodrecht, pp 43–76
Goldberg JB (2004) Operations research models for the deployment of emergency services
vehicles. EMS Manage J 1(1):20–39
402 R. Pelot et al.

Griffin PM, Scherrer CR, Swann JL (2008) Optimization of community health center locations and
service offerings with statistical need estimation. IIE Trans 40(9):880–892
Harewood S (2002) Emergency ambulance deployment in Barbados: a multi-objective approach. J
Oper Res Soc 53:185–192
IBM ILOG CPLEX http://www01.ibm.com/software/commerce/optimization/cplex-optimizer/
Jacobs DA, Silan MN, Clemson BA (1996) An analysis of alternative service areas of american
facilities locations. Interfaces 26(3):40–50
Larson RC (1974) A hypercube queuing model for facility location and redistricting. Comput Oper
Res 1(1):67–95
Li L (2006) Rescue vessel location modelling. Nova Scotia. M.A. Sc. in Industrial Engineering,
Dalhousie University, Halifax
Marianov V, ReVelle C (1996) The queueing maximal availability location problem: a model for
the siting of emergency vehicles. Eur J Oper Res 93:110–120
Nguyen BU, Kevin YKN (2000) Modeling Canadian search and rescue operations. Faculty of
Administration, University of Ottawa, Ottawa
Pirkul H, Schilling D (1988) The siting of emergency service facilities with workload capacities
and backup service. Manage Sci 37(7):896–908
Pirkul H, Schilling D (1991) The maximal covering location problem with capacities on total
service. Manage Sci 37:233–248
Sinuany-Stern Z, Mehrez A, Tal A-G, Shmuel B (1995) The location of a hospital in a rural region:
the case of the Negev. Locat Sci 3(4):255–266
Chapter 17
Military Applications of Location Analysis

John E. Bell and Stanley E. Griffis

17.1 Introduction

Over time, the decision-making needs of military commanders have had a strong
influence upon the development of the field of operations research and analytic
problem solving. The challenge of correctly positioning military units and resources
within a geographical setting has vexed commanders and their staffs for thousands
of years. However, it is only in the last 70 years that optimization methods have
developed to the point where analysts can apply them to accomplish such goals.
Along the way toward solving these narrowly defined military-focused problems,
the advancement of the field has benefitted as generalizable techniques are extended
beyond their origins to countless non-military applications. For example, military
analysts first solved complex problems regarding routes for convoys of ships, code
breaking, and materiel allocation mathematically during World War II ultimately
leading to the development of linear and mathematical programming techniques by
wartime scientists. Location analysis knowledge was similarly benefitted by the war,
as commanders required the ability to spatially position munitions to destroy a tar-
get and covering a search area to find the enemy. The benefit is reciprocal however,
as military strategy and planning since World War II have literally been redefined by
the operations research field’s ability to solve larger and more complex problems.
The dawn of the computer age and the explosion of the operations research field
are understandably tied. Since the 1950s there has been a boom in problem solving
ability, and a wide range of applications of operations research and location analysis
methods have benefitted from the computing power, and increased analytic tools

J. E. Bell ()
Department of Marketing & Supply Chain Management, University of Tennessee,
Knoxville, TN 37996, USA
e-mail: [email protected]
S. E. Griffis
Department of Supply Chain Management, Michigan State University,
East Lansing, MI 48824, USA
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 403

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2_17
404 J. E. Bell and S. E. Griffis

to take advantage of this power. While military problems germinated many of the
original techniques, over time we have seen these military problems fade into the
background as industrial and urban planning priorities have fueled the research and
applications of location modeling. Despite this general trend, military applications
of location analysis remain an important and often unique contextual area for the
application of solution methods. The military application area has seen a renewal in
interest with increased number of publications over the past 15 years. In large part
this is the result of geopolitical shifts brought on by the end of the cold war, and
the post 9/11 realities for militaries. In particular, the current set of vexing problems
possesses challenging decision variables, parameters and constraints unique to the
current realities of military location modeling needs.
In this chapter, we will review the key military applications of location analy-
sis published over the last several decades in the academic literature. In doing so,
we will first outline the brief history and classic academic publications in the area.
Then more recent military applications will be reviewed in order to highlight newer
constraints and problems faced by 21st century military commanders. This review
will lead to the development and presentation of a typology of common features
that characterize military applications in location analysis. We will conclude with
recommendations for further research and opportunities for location analysts.

17.2 A Brief History of Military Application of OR Techniques

The location of combat resources and the effect of geography on those military
positions have played a critical part in military strategy and operations throughout
history. Military theorists such as Sun Tzu, Clausewitz and Eccles have studied and
emphasized the strategic importance of location decisions in the operations, tactics,
and logistics of warfare (Griffith 1963; Howard and Paret 1989; Eccles 1959). In one
famous historical example of the role of location in military strategy comes from the
war between Rome and Carthage. The Carthaginians successfully maneuvered their
forces across the desolate North African deserts to secure locations where vital food
and water resources existed to sustain their armies (Roth 1998, p. 308). This was
not chance on the part of the Carthaginian General Hannibal, he actively sought to
induce the Romans into battle on a location of his choosing. The end result a rout
of the Romans at Cannae (B.C. 216) in one of the most famous victories in mili-
tary history. Location science and geography continued to create challenges for the
Romans as the decline of the empire was brought on, in part, by the inability of the
military to cover and secure its vast geographical areas. As a result, Rome failed to
protect its holdings and hard-earned territories were abandoned (e.g., Britain). The
basic covering problems that plagued Emperor Constantine and his military leaders
have been revisited by modern location scientists including Stewart (1999), ReVelle
and Rosing (2000), Henning (2003), and Cockayne et al. (2004), and have shown
that solutions to the Roman maximal covering problem could have been achieved.
Against this backdrop, we consider, “how have analytical models been used in recent
history to make military location decisions?” Answering this question necessitates a
17 Military Applications of Location Analysis 405

review of some of the prominent military applications of location analysis over the
last several decades. This review reveals unique features of military application of
location analysis, and highlights opportunities for continued research in this area.
Prior to the 20th century, strategic decision-making (including location analysis)
was conducted primarily by the most senior military commanders. (Dantzig 1963,
p. 12). In the 1830s, military science was advanced greatly by the Prussian theorist,
Carl von Clausewitz. Despite being one of the most acclaimed military theorists in
history, in his classic work, On War, Clausewitz specifically denied the usefulness
of “geometric analysis” for the positioning of bases of operation and the locations
needed to provide supplies to military forces, and indicated that such analysis could
lead to “wrong tendencies”(Howard and Paret 1989, Ch. XV). However, by the time
of World War I, planning for modern armies broadened to include a General Staff
rather than a single leader to conduct planning processes and to help commanders
make complex strategic decisions in shorter periods of time. This planning pro-
cess, often called “programming”, allowed broader views on planning issues than
Clausewitz held, and grew significantly in the US military during World War II.
These scientists and military analysts engaged in wartime programming that laid
the groundwork for the development of the academic field of operations research.
During World War II, many of the operational problems studied contained locational
aspects including decisions for combat unit deployments to various geographic the-
aters around the globe (Dantzig 1963, p. 14). In addition, wartime analysts tackled
problems to include developing methods to decide where to drop depth charges
(location, depth, and pattern) to defeat enemy submarines, and the best methods
to convoy merchant vessels across the Atlantic Ocean to avoid attacks (Morse and
Kimball 1951). In addition, wartime scientists applied mathematical techniques and
search theory for the “coverage “of a continuous space in order to find enemy tar-
gets within distinct geographic area (Morse and Kimball 1951, p. 86). However, this
work was only a prelude to the advancement of operations research methods and the
military applications in location analysis that were to come.
Following WWII, the further development of linear programming by the US
Air Force, and the creation of the simplex method for generating solutions for
linear programming by Dantzig were major steps forward for military problem
solving. However, most of the early military applications of linear programming
were primarily for dynamic program planning (Wood and Geisler 1951), routing of
aircraft and ships, labor and equipment scheduling, and contract bidding (Dantzig
and Fulkerson 1954; Jacobs 1955; Natrella 1955). To understand this apparent lack
of attention on location analysis, an understanding of the focus of the Cold War
military from the 1950s until the late 1980s is needed. As described by Bonder
(2002), the focus of US Army operations analysis during this time was on tactical
military operations and weapons system analysis. The locational aspects of future
conflicts were considered relatively well known and stable for the US military and
its scientists, the battles of the Cold War were expected in central Europe. In the
US Navy, important work did continue through the 1960 and 1970s on problems
such as screening for submarines and the proper escort positioning for ship convoys
(Hughes 2002). But even then, the military’s mission focused on countering the
Soviet Union’s larger forces with superior tactics and systems; while the location
406 J. E. Bell and S. E. Griffis

of expected conflicts and needed facilities could be predicted with relative certainty
(Bonder 2002). Given all the assumptions about where the fight would occur were
‘resolved’ advancement of location science within the US military stagnated as
other, more pressing issues, were addressed. Reflecting this stagnation, relatively
few military applications of location analysis appear in the literature in this time
period.
Interestingly, while the military operations research scientists became less con-
cerned with the location problem, academic location analysis picked up where they
had left off, and the area developed rapidly. Access to the first computers in the early
1960s helped to spawn an explosion in location analysis techniques and many of the
problem formulations, and algorithms were developed, and seminal articles pub-
lished. Building on the economic works of von Thunen, Weber and Hoover and
inspired by the seminal works of Hakimi (1964), Cooper (1963), Toregas et al.
(1971) and others, location analysis methods exploded from the1960s through the
1980s. Through the end of the cold war period a wide range of location modeling
applications in industrial, geographical science, urban planning, and communica-
tions emerge. As we will see in Sect. 17.3, commercial applications drive the
location analysis agenda in the latter stages of the Cold War era, however, this would
change with the close of the Cold War era.
Within a relatively short period of time between 1989 and 1992 the military focus
of the world shifted (Bonder 2002). After the Cold War, geopolitical shifts resulted
in renewed interest within the military to location analysis. Primary among these
changes was uncertainty regarding the location of future conflicts. Throughout the
Cold War, the location of the next potential conflict was fairly well assured, the
evaporation of the USSR resulted in a dispersion of military might, and a resulting
need to reconsider where countervailing forces needed to be positioned. Dozen of
potential scenarios for future conflicts as well as the need to simultaneously support
multiple conflicts (itself a significant shift in planning complexity) became priories
for military planners (Bonder 2002; Bell 2003; Amouzegar et al. 2004a). The sec-
ond major shift centered on the need to pre-position stores of military equipment
and munitions as a way of scaling back the investments associated with standing,
ready forces in foreign installations around the globe. This change in philosophy
was most clearly seen in the European theater where dozens of facilities were per-
manently closed during the 1990s. At the same time, not only did the drawdown in
forces affect the location problems confronting analysts, the objectives changed as
well. In response to the “peace dividend”, analysts were required to find more fis-
cally motivated solutions, unlike the “win at all costs” perspective of the Cold War.
These factors combined to create the last shift, toward time-based contingencies as
a priority for military planners. Now that fewer permanent locations were available
around the globe, and the location of the next threat was unknown, and pools of sup-
ply were pre-positioned to address anticipated needs, analysts were confronted with
developing models to facilitate the generation of rapid response forces to address
short-notice ‘needs’ anywhere on the globe. As an example, the 1990s conflicts in
Kosovo required US forces deploy to a new combat location, with relatively no
notice, no logistical planning in place, and extremely short time windows. Further
17 Military Applications of Location Analysis 407

adding to the complexity faced by analysts, the new realities of location analysis
required greater consideration of the redeployment phase, that time period following
the military engagement, when all the manpower and materiel needs to be returned
and reset to their original locations (Bonder 2002). Taken together, and made even
more acute after the terrorist attack of 9/11, all of these factors combined to create a
need for new and different location analysis techniques, thereby reinvigorating the
military’s interest in the location sciences.
As we will show in Sect. 17.3, these new paradigms have resulted in a significant
increase in the number of military location analysis applications in the literature
since the late 1990s. As we review the primary military applications during the last
several decades, this chapter will identify and describe those special attributes of
military applications and the ever-changing conditions that create them.

17.3 Review of Military Applications

The common features and attributes of military operations research applications

published in the recent literature appear in Table 17.1. Although, not all articles
possess each attribute, trends in the research that distinguish military location appli-
cations from other areas become clear. For example, the multi-objective nature of
military applications must consider not only the cost minimization and equitable
service objectives seen in other location research (ie., the location of police and
fire stations), but also the ability to maintain a secure and resilient operating loca-
tion. This is because military locations themselves may also become the targets of
opposing forces, and the capability and survival of the network also depends on the
defense of locations within the network. Military models often treat factors with
different severity than commercial models, stock-out costs being an example. While
business models care about inability to cover or serve an area from the perspec-
tive of terms of lost sales or profit, for a military application an inability to serve a
“demand” might translate into death for fielded troops or failed national objectives.
Accordingly, a high premium is often put on achieving near perfect service levels
in military applications, even at the expense of increased financial costs. Table 17.1
lists several other unique aspects of military applications, and classifies the articles
reviewed in this section, based on the appearance of these aspects in the article.

17.3.1 The Cold War Years (1960 to 1990)

Military applications of location problems during the Cold War were fairly limited,
compared to the preceding decades. This was due in part to military planners rela-
tive certainty as to the location of expected future conflicts with communist nations.
Given the “base in place” finish to World War II, the location of forces in West-
ern Europe and the Korean Peninsula were well known. Therefore, from the early
Table 17.1 Classification of military applications literature
408

Articles Deterministic Stochastic Single Multiple Multiple layers Median Coverage Dispersion Security
Period Period (Hierarchical) (Min (Max (Avoidance) (Protection)
(Static) (Dynamic) cost) service)
Levin and Friedman (1982) X X X X
Yamani (1990) X X X
Loerch et al. (1996) X X X X X
Murty and Djang (1999) X X X X X
Nguyen and Ng (2000) X X X X
Segall (2000) X X X X X
Johnstone (2002,2004) X X X X
Leinart et al. (2002) X X X X X
Bell (2003) X X X X
Gue (2003) X X X
Flint et al. (2003) X X X
Kierstead and DelBalzo X X X
(2003)
Amouzegar et al. X X X X X X
(2004b, 2006)
Brown et al. (2005) X X X X
Farahani and Asgari (2007) X X X X
Dawson et al. (2007) X X X X
Ghanmi and Snow (2008) X X X X X
Overholts et al. (2009) X X X X
Dell (2008) X X X X
Killingsworth (2008) X X X
J. E. Bell and S. E. Griffis
 17

Table 17.1 (continued)

Articles Deterministic Stochastic Single Multiple Multiple layers Median Coverage Dispersion Security
Period Period (Hierarchical) (Min (Max (Avoidance) (Protection)
(Static) (Dynamic) cost) service)
Chan et al. (2008) X X X
Kress and Royset (2007) X X X X
Rappold and Van Roo X X X X
(2009)
Salmeron et al. (2009) X X
Ferrer (2010) X X X X
Burks et al. (2010) X X X X
Murphy et al. (2010) X X X X X
Military Applications of Location Analysis

Dong et al. (2010) X X

Ghanmi (2011) X X X X
Toyoglu et al. (2011) X X X X
Bell et al. (2011) X X X X X
Brown et al. (2011) X X X X
Plastria (2012) X X X X X
Tanerguclu et al. (2012) X X
Qiu and Sharkey (2013) X X X X
Dillenburger et al. (2013) X X X X
409
410 J. E. Bell and S. E. Griffis

1960s until 1990, the focus of much of the operational research modeling in the mil-
itary was on other topics. In an overview of 40 years of military modeling, Hughes
(1984) outlines the major application areas of military modeling during the Cold
War period. Hughes outlines seven primary military application areas during this
time period including (1) battle planning (2) wartime operations (3) weapon pro-
curement (4) force sizing (5) human resource planning (6) logistics planning and (7)
National policy analysis (Hughes 1984, p. 23). It is interesting to note that although
location modeling and spatial decision-making are not considered major applica-
tion areas within Hughes’ framework, he does mention the strategic importance of
facility location and layout decisions throughout this text. For instance, within the
area of logistics planning, Hughes notes that the location and capacity of military
maintenance and repair facilities is an important part of logistics modeling applica-
tions. In addition, he goes on to point out that much of the effort in facility location
planning in these early years was geared towards the objective of peacetime effi-
ciency (cost minimization) of military facilities and infrastructure, and that future
military modeling would need to consider analytical models that focus on wartime
realities including the achievement of the right level of “decentralization” of forces
being threatened by an enemy, and the need for building-in excess capacity to facil-
ity networks in order to prepare for the uncertainties created by war (Hughes 1984,
p. 32).
Given the reduced focus upon location analysis, resources were spent in other
areas in this period. Inventory management research dug into issues related to
improving repairable inventory decisions. Work by Zangwill (1966), Sherbrooke
(1966), Muckstadt (1973), Hillestad and Carrillo (1980) and others focused on
making the better decisions for inventory purchases to minimize the expected back-
orders in an inventory system, and to get the highest return for the inventory
system given limited resources. Although inventory research was originally tacti-
cal in nature, over time, work by Muckstadt (1980) and Hillestad (1982) broadened
to consider multiple levels or hierarchies of locations in the network, and then fur-
ther extended to include fixed costs of facilities and variable capacities at different
locations, resulting in more strategic models with integrated inventory and location
decision-making (Drezner and Hillestad 1984, p. 211).
Transportation problems also moved to the foreground during this era. Follow-
ing the work of Dantzig (1963) and Ford and Fulkerson (1954), military modeling
projects addressed needs for algorithms to handle location-specific bottlenecks in
transportation networks and to improve the network of locations (and routes) needed
to transport military material and personnel (Garfinkel and Rao 1971; Berman
et al. 1981). Such modeling highlighted the expense of building such transportation
systems with the determination of the needed capacity to reach distant locations
(Drezner and Hillestad 1984, p. 212). Foretelling the near-future, Drezner and
Hillestad (1984, p. 215) observe that military planning in the cold war period “de-
emphasized tactical conventional capabilities because of the nuclear umbrella. . .”
and that the objectives, constraints and structure of future military models would
have to take into account a much higher degree of uncertainty and that wars might
17 Military Applications of Location Analysis 411

occur at any location on the globe. Additionally, the increased integration of vari-
ous military models would be needed in the future, in order to understand important
tradeoffs in military decisions, such as those that simultaneously account for loca-
tion, inventory and transportation decisions by military decision makers (Drezner
and Hilliard 1984, p. 218.)
In addition to direct military applications, several theoretical pieces were pub-
lished on location modeling during this time period that were motivated by or
directly related to military problems. First, it is interesting to note that in their review
of private and public location models, ReVelle et al. (1970) stop short of classifying
military applications in the public sector, along with civil defense, police stations
and other emergency services. However, many of the aspects of the public loca-
tion models discussed in their article do apply to military applications. For example,
similar to other public location models the “the goals, objectives, and constraints” of
military applications may not be “easily quantifiable . . . or easily defined.” (ReVelle
et al. 1970, p. 694). Additionally, policy, laws, political constraints and government
budget restrictions often place unclear constraints on the feasible region of military
problems. Such challenges cause military applications to often be multi-objective in
nature, where according to ReVelle et al. (1970, p. 694), the hope is not “to be able
to define solutions, but to at the least gain more information about the system under
analysis.” Therefore, understanding the tradeoffs in objectives and identifying the
range of feasible solutions available to military commanders and political leaders is
often an aspect of military location applications.
Another research area inspired somewhat by the military resource allocation
challenges is the work of Kaplan (1974) to limit use scarce resources such as
fuel, water, and ammunition to meet anticipated demands. Similarly, Litwhiler and
Aly’s (1979) theoretical work on the spherical Weber problem that consider non-
Euclidean distances appears motivated by the need for improved mathematical
methods for naval deployment and communications over large geographical areas,
faced by the US military. Additionally, Woolsey’s (1986) editorial contained men-
tion of facility location modeling, and the Pentagon and the US military specifically.
Although none of these articles utilize military-sourced data, they are motivated
by military operations and continue to illustrate the need for improved location
modeling techniques to improve real world military decision-making.
Levin and Friedman (1982) specifically apply dynamic warehouse location meth-
ods to deploy military units to the most effective locations and minimal cost
locations. Their approach steps beyond theoretical mathematics to address a specific
military problem and develop solutions for this problem. In doing so they specifi-
cally addressed how to determine the maximum number of support units needed to
supply military forces, and where those units would need to be located. Interestingly,
this application also considered military unit effectiveness decreases due to “expo-
sure” to various hazards or threats when moved to a particular location. This aspect
of the model also makes dispersion and survival of units a part of the objective.
Towards the end of this period, Yamani et al. (1990) extended work on the spher-
ical Weber problem conducted by Drezner and Wesolowsky (1978), Litwhiler and
Aly (1979), using US Air Force aircraft refueling data. This paper built a model of
412 J. E. Bell and S. E. Griffis

single aircraft mid-air refueling with the objective of selecting refueling locations
in order to minimize aircraft fuel consumption. Yamani et al. (1990) show that the
problem faced by the military was actually a nonlinear convex location problem on
a sphere, and presented an optimal procedure for generating solutions to the prob-
lem. This paper was a prelude to a large number of future military applications that
would more deeply examine military oriented problems and apply actual military
data in the process of generating solutions to those problems.

17.3.2 Transition in Focus: 1990–2001

The period from the beginning of Operations Desert Shield/Storm in 1990 until the
terrorist strikes of 9–11 in 2001, represent a transition in military strategic planning
that ushered in a renewed need for improved military location analysis. The sheer
size of the forces and amount of equipment deployed to the Middle East during
Desert Storm and the follow-on no-fly zones in Iraq during this period led to a re-
examination of global military asset positioning. Given positioning of US military
forces in Europe and Asia was done in anticipation of conflicts against anticipated
Cold War opponents, most of the forces and equipment needed in the 1990s were
poorly positioned to respond to needs in the Middle East. In addition, increasing
pressure to decrease military budgets incentivized the nation and its military to focus
more on smaller, more mobile forces, needed to respond quickly to a wide range
of conflicts. During the 1990s, military researchers, including those at the RAND
Corporation, focused on creating flexible and responsive approaches to location
selection for the military. Global positioning and global reach became impor-
tant aspects of military strategy, especially given constraints associated reducing
the mobility “footprint” of forces to be moved during a future conflict. Accord-
ingly, military location modeling applications during this period helped reflected an
increased interest in location modeling in the US military.
One area the military began to reassess in the post Cold War environment was
better ways to pre-position and transport critical wartime material. Loerch et al.
(1996) studied the location strategy and positioning of US Army’s facilities in
Europe, making recommendations as to which bases to deactivate and close in
the European theater of operations in the 1990s. Their integer programming model
sought to minimize costs of the network subject to constraints related to minimum
support and proximity (coverage). Their results provided insight into the cost trade-
offs of opening and closing facilities in an pre-existing defense network, as well as
resource utilization insights. Loerch et al. (1996) analysis was complicated by the
fact that military units under consideration were required to located within a certain
distance of their superior units, which are also being located in the model. Because
of this requirement, the problem is more difficult to solve due to the multi-level hier-
archy of locations at play simultaneously. Further complicating the analysis, support
units in the model have different coverage requirements compared to combat units,
17 Military Applications of Location Analysis 413

and capacities exist for both shared resources (equipment) and funding at the loca-
tions. Using multiple scenarios and data from the US Army the research was able
to identify four alternate solutions to reach $ 49–58 million in savings annually in
comparison to a manually generated solution created by Army staff experts.
Similar to efforts by the US Army, modeling efforts at the US Naval Postgraduate
School by Anderson (1998) and Sentlinger (2000) were aimed at improving the
locations for the storage of munitions by the US Navy. Anderson (1998) looked to
improve weapons positioning across the Pacific region for the US Navy. Sentlinger
(2000) considered heterogeneous weapons mixes with a mixed integer program to
determine the best mix of weapons to preclude shortfalls in a number of potential
combat scenarios. Both of these studies are limited-distribution to those working
within US military circles.
Also during this time period, Murty and Djang (1999) developed a multiple layer
strategy for the US’s National Guard to locate mobile training units. In the first stage
of their model, a p-median approach is used to select the best set of home bases for
the mobile trainers. Then in the second stage, a set covering approach is used to
select secondary training locations within a maximum distance of units needing
training. Finally Murty and Djang (1999) determine routing to the secondary train-
ing locations. This study, the results of which were implemented by the US military,
achieved a 70 % reduction in the expected annual mileages driven by the trainers
and a significant cost savings of over $ 8 million yearly. The work stands as a clas-
sic example of the benefits achieved from applying location-modeling techniques to
operational military problems.
Segall (2000) responds to a more mobile military and considers hospital location
modeling, specifically the locations and capacity of a network of military hospitals
needed to respond to chemical agent attacks. Segall’s (2000) work describes a set
of optimization models for determining the location of facilities when faced by six
different cases of how medical facilities will be distributed over a planar region.
Patient flows to military emergency facilities and the vulnerability of the hospi-
tals and their staffs to the actual chemical weapons attacks that create the patients
are also considered, as well as access for patients to care and cost minimization of
facility damage; all while taking into account the spatial differences in the demand
created by military combat scenarios.
Nguyen and Ng (2000) apply p-median and coverage techniques within a goal
programming model to analyze search and rescue operations for the Canadian mil-
itary. This approach was used to determine the service capability of the current
military system and to make suggestions to the Canadian military for improving the
non-optimal locations then being used by search and rescue aircraft, while priori-
tizing guaranteed service to as many distress incidents as possible within a critical
coverage distance (similar to solving the maximal set covering problem). Addition-
ally they sought to minimize the average distance to incidents that occur outside of
the coverage areas of the selected bases, similar to a p-median approach. Using a
simulated annealing metaheuristic Nguyen and Ng (2000) generate solutions to the
problem and analyze the tradeoffs between maximizing coverage and minimizing
414 J. E. Bell and S. E. Griffis

distances and costs. Then Nguyen and Ng (2000) use simulation and queuing the-
ory to analyze the dynamic aspects of search and rescue operations by the Canadian
military over large spatial areas. Military planners in Canada used the study’s results
to propose changes to their existing network and also applied findings from the
follow-on simulation study to realize significant improvements in waiting time for
rescues.
By the end of this era, work by Tripp et al. (1999) and others at the RAND
Corporation began to consider issues related to the need for more agile combat
logistics support framework for the US military. This research foreshadowed a con-
tinued shift in the US military away from permanent overseas facilities and toward
a network of much smaller forward support locations and forward operating loca-
tions, positioned where potential conflicts might occur. This research specifically
addressed the uncertainty of future conflict locations and the need for a flexible set
of locations and operating capacities to handle multiple military combat scenarios
at the same time. This type of flexible and agile thinking, and the changed-state of
conflicts after 9–11 would help to trigger a significant increase in the number of
military applications of location modeling.

17.3.3 Recent Military Applications of Location Modeling

(Post 9–11 era)

The last few years have seen a veritable explosion of military application arti-
cles in the literature, as compared to the period prior to the new millennium. The
relative uncertainty and instability of military operations in the 1990s renewed dor-
mant interest in location analysis in contrast to the Cold War, when locations were
considered more fixed (Bonder 2002). However, the last decade has seen the devel-
opment of new strategic approaches such as “Flex-Basing” and “Forward Support
Locations” proposed to the US military by researchers at the RAND corporation
(Tripp et al. 1999; Killingsworth 2000; Geller et al. 2004). In addition, the US
government’s Base Realignment and Closure Commission (BRAC) sponsored and
motivated location research to analyze which US military bases should remain open,
and which should be closed. Changes to military basing strategies and the efforts to
improve the global network of military locations led to an increase in military appli-
cations of location analysis by researchers. The remainder of this section has been
broken into several sub-categories that detail the studies of this time period based
on their scope and objectives. Similar to the previous sections, the common features
and attributes of specific military applications are listed in Table 17.1.
17 Military Applications of Location Analysis 415

17.3.3.1 Strategic Global Networks

Building a global network of locations to respond to uncertain conflicts became a

priority for the US military at the beginning of the millennia. Conceptual pieces
began to call for an improvement in the selection of forward operating locations
around the globe for the military and for an improvement in the positioning of
difficult to move military equipment (Tripp et al. 2000; Underwood and Bell 1999).
In order to consider building such an improved global network, modeling
research by Johnstone (2002) (and Johnstone et al. 2004) examined the location
and deployment decisions for Air Force munitions inventories used to respond to
crises around the world. The objective of this model was primarily to minimize
the time to move munitions, but it also considered and selected the best supply
points for the pre-positioning of assets needed in potential conflicts. In addition, the
work by Johnstone (2002) emphasized the continued need for afloat positioning of
wartime material on ocean going vessels. In essence, afloat prepositioning allows for
mobile supply points that can be moved in international waters towards areas where
demands may occur. This model examined the selection of ship locations and ports
for off-loading material at the regional level, and also looked at a lower hierarchy
for determining the subsequent inland movement of the material of selected ports.
Like many effectiveness based military models, the model aimed to maximize the
primary objective, service (minimize delivery times), and ignored cost, the stochas-
tic nature of demands, and the various dynamic time-phased issues with moving
munitions over long periods of time.
Bell (2003) built upon Johnstone’s work, and used facility location modeling
techniques to develop a strategy for constructing a global network of munitions
locations for the US Air Force. Bell (2003) developed a combined minimal cost
and maximum coverage solution to address the multi-item facility location prob-
lem. Bell accomplished this by selecting warehouse locations from a discrete set of
potential locations, while simultaneously determining multi-item inventory levels at
the selected locations. Due to the size of the problem, simulated annealing was used
to find near optimal solutions to the problem. Using this approach, it became evident
that both cost and coverage objectives could be improved compared to the initial
solution provided by Air Force and efficient frontiers of solutions were mapped in
order to more fully understand the tradeoffs in cost and coverage.
RAND Corporation researchers have also been involved in the reconsideration of
military basing decisions. Amouzegar et al. (2004a) conducted a study of locations
for future forward support bases that rely upon a wide range of war reserve materials
stored around the globe. The study considered both land and sea-based locations
around the globe and also considered co-locating U.S. Army and U.S. Air Force
support facilities. The developed multi-item optimization model considered various
capacities at different locations and aims to minimize global system costs while
meeting dynamic time-based service requirements for varying conflict scenarios.
The model includes actual transportation constraints and distances and considers
the allocation of demand points to selected locations (location-allocation).
416 J. E. Bell and S. E. Griffis

Additionally, this military application of location modeling was one of the first to
include the vulnerability and defense of the basing options in the model. Aspects of
this stream of research and the resulting resource allocation modeling were also
detailed in Amouzegar et al. (2004b). Additionally, the same location modeling
approach was later reapplied to an expanded set of scenarios and military planning
considerations (Amouzegar et al. 2006).
A separate study by Killingsworth et al. (2008) considered the selection and
costs of modifying an existing supply chain network for a military helicopter sys-
tem in use by the U.S. Army. This analysis considered the repair capabilities of
the current system across the globe and considered multiple scenarios for altering
the network and selecting a new repair center location in the network. The problem
was formulated as a discrete minimal cost network flow problem and the cost sav-
ings and payback period, for investment in a new repair facility, were considered.
This effort identified significant annual cost savings ($ 9–18 million), with payback
periods ranging from 2 to 6 years.
The Canadian military also faces the challenge of rapidly deploying military
capabilities to virtually any location around the globe with minimal time and costs.
Ghanmi and Shaw (2008) developed a multi-objective model minimizing cost and
response time for deploying Canadian forces around the globe. Like their American
colleagues, the Canadians also considered the pre-positioning of military assets at
forward locations in advance of demands, while also trying to reduce problems with
military airlift. Ghanmi (2011), further developed a discrete facility location model
to select “support hub” locations with the multiple objectives of cost and capabil-
ity. Formulated as a mixed integer non-linear program, the work of Ghanmi (2011)
is based in the facility location problem literature; however it uses a unique objec-
tive function to achieve the maximum relative cost avoidance when determining the
optimal number and location of hubs in the final solution.

17.3.3.2 Regional Facility Networks

In addition to assessing global networks in military applications of location mod-

eling, a number of regionally focused models have appeared in the literature in the
last decade as well. Primarily, these applications have addressed the need to build
more-detailed repair and distribution networks to be controlled by a single mili-
tary commander. An interesting feature of many of the regionally focused models is
that they simultaneously solve routing or inventory problems as part of the facility
location problem.
Gue (2003) addresses building a network of discrete supply locations to support
the small and highly mobile forces of the U.S. Marine Corps. Gue (2003) addresses
multiple time periods, and seeks to minimize the actual quantity of inventory that is
positioned on seaborne supply points that are mobile and are consistent with their
seaborne logistics support. The problem is a dynamic mixed period facility location
and multiple commodity flow problem, and it is formulated as a mixed integer pro-
gram. Decision variables for this model include the candidate list of locations in a
17 Military Applications of Location Analysis 417

region, the inventories to position at land and sea locations, and the inventory flow
of items to combat locations. The model is tested against demands created from
actual combat scenarios expected by the U.S. Marine Corps. In addition, the anal-
ysis allows the military users to consider tradeoffs between the use of land and sea
based supply locations, and the impact of the distance between the seaborne sup-
ply points and the shoreline. As a military application, Gue (2003) represents an
early dynamic location model, with the analysis of the tradeoffs between mobile
(sea-based) and fixed (land-based) supply locations in the same regional network.
Farahani and Asgari (2007) use a multi-objective set covering approach to deter-
mine the best location for military distribution centers from a discrete set of current
warehouse locations while minimizing cost and improving the utility (quality) of
selected locations. The utility objective is unique in that it has twenty-four sepa-
rate attributes. To address this Farahani and Asgari (2007) adapt multiple attribute
decision-making (MADM) techniques to assess this objective within the model.
Additionally, the model is multi-echelon in nature considering the location of sup-
ply points (factories), warehouses, and supported locations. In doing so, the model
aims to create sub-groups of supported locations that can only be supplied by one
warehouse; while still allowing direct shipments from the factory that by-pass the
warehouse (Farahani and Asgari 2007, p. 1841). Results are presented using an effi-
cient frontier of the two objectives as minimal cost and maximum utility objectives
are in conflict, and tradeoffs must be considered. The results of the research show
that the military was able to reduce their critical coverage distance in comparison to
existing military practice, and were able to improve the balance of workload among
locations.
More recently, Rappold and Van Roo (2009) build a non-linear integer model for
a multi-echelon engine repair network for the US Air Force. They combine facil-
ity location and inventory allocation problems and attempt to minimize the cost
of the repair system and simultaneously determine the correct level and allocation
of inventory across locations for the entire system. Using a two-step optimization
approach, this study initially selects the minimal cost locations for the repair net-
work, and then determines the finite repair capacities to set for each location in
order to balance supply and demand in the system. Although the study is formu-
lated for a small region, the work could be extended to global networks, within
general complexity limits. Rappold and Van Roo’s (2009) work is an extension of
multi-echelon inventory problems in the military first studied by Sherbrooke (1966)
and Muckstadt (1973), and it makes unique contributions in that the demand for the
system is stochastic and finite capacities are put on the total inventory in the system.
This application provides a tool for military decision makers to improve acquisi-
tion decisions for purchasing optimal numbers of expensive inventory items (such
as aircraft engines) to stock at a network of repair locations.
Dell et al. (2008) consider manpower issues related to military units and person-
nel to be assigned to hundreds of U.S. Army locations around the United States.
The Base Realignment and Closure (BRAC) commission in 2005 sponsored this
research, and it resulted in the closure of 13 active duty Army installations and liter-
ally hundreds of Army Reserve facilities around the nation. Dell et al. (2008) model
418 J. E. Bell and S. E. Griffis

aims to minimize costs for the US Army while also considering the military value
(service) achieved by the optimal solution, similar to previous work by Ewing et al.
(2006). The model is able to open, close or modify the capacity of current facilities,
while simultaneously determining how to allocate Army units across the selected
locations, and uncovered potential savings to the military of $ 7.6 billion over two
decades.
Similar savings was sought in a study by Ferrer (2010), who considered improv-
ing the location of spare aircraft engines for the US Air Force’s F-16 fighter aircraft
within the continental United States. The application of the p-median approach by
Ferrer (2010) included a separate analysis for the two different engine types used
on F-16 aircraft (Pratt & Whitney and G&E) and identified the best set of locations
for storing each type of engine that met the distance constraint of the model. An
open architecture approach was considered that combined the two types of engines
into one inventory pool, assuming interchangeability of the engines. Ferrer’s (2010)
analysis identified potential direct savings by the US military of $ 58 million and a
48 % reduction in safety stock.
Burks et al. (2010) studied the Location Pickup and Delivery Problem with Time
Windows (LPDPTW) that simultaneously considered the location of depots for a
number of vehicles that would provide time definite delivery and pickup of mate-
rials throughout a geographic region. Simultaneously, vehicle assignments to depot
locations and allocation of customers to those depots was determined. The study was
motivated by the US Army’s wartime experiences in the Persian Gulf region and the
need to minimize its logistics costs in future operations while maintaining wartime
capability. Additionally, the model minimized delivery “lateness”, and system-wide
costs. The problem is also organized based on a set of tree layers (acyclic directed
graphs) and therefore analyzes and captures the hierarchical nature of the Army’s
distribution network (Division, Brigade, and Battalion support levels). Making it
even more challenging, Burks et al. (2010) considered multi-commodity, temporal,
and facility capacity issues in addition to time factors. Solutions for this complex
hybrid problem were generated using an adaptive Tabu search methodology, based
upon Glover and Laguna (1997), which provides dynamic updates within the search
heuristic to ensure both intensification and diversification during the search of the
feasible region. The approach was demonstrated on six different scenarios to provide
a realistic range of demands that might occur in actual military operations.
Toyoglu et al. (2011) also considered a dynamic environment, addressing a hier-
archical location routing problem based on the daily distribution of ammunition to
combat military units in the field of battle. This detailed problem is formulated as a
multi-layer, dynamic location routing problem. It considered a heterogeneous fleet
of vehicles with capacitated as well as fixed and mobile supply locations, Within
their model, Toyoglu et al. (2011) further consider exposure of delivery vehicles to
enemy fire and modify driving distances to take this into account. Additionally, sev-
eral unique aspects of this paper make it stand out. First, the model not only includes
three echelons of facility locations, but it also simultaneously considers location
decisions at two of these echelons; a feature rarely seen in the literature. In addition,
the model allows for multiple sourcing where more than one military supply point
17 Military Applications of Location Analysis 419

or transfer point can fill the demand at a particular combat location. Lastly, hard
time windows are imposed, making it even more realistic for military combat oper-
ations where success and life and death can be determined by the on-time delivery
of ammunition to front line combat units.
Overall, as a group, the military applications of military modeling in recent years
(including these regional applications) have begun to answer the call for studies of
Hughes (1984) to be more focused on wartime rather than peacetime objectives.
Recent military location studies are more focused on wartime combat objectives
and are also more integrated with the related routing and inventory decisions being
made by military commanders. Additionally, dynamic and hierarchical problems
and multi-sourced problems are adding to both the complexity and realism of the
military applications of location analysis.

17.3.3.3 Coverage Models in the Military

The primary objective of most studies in Sect. 17.3 so far has been minimizing
costs or distances (facility location problem or p-median type formulations) in the
model. While coverage or improved service has appeared as a secondary objec-
tive or constraint in those models they have generally not been the focus. The next
section considers those military applications that have focused primarily on the cov-
erage objective and have used models from the classic location literature such as
the maximal location covering problem, set location covering problem or p-center
problem.
Chan et al. (2008) determine the optimal location for signal receiving stations
used for military search and rescue operations. In addition to selecting locations,
the problem also allocates receiving equipment to each location and determines the
listening frequencies to use. Because coverage areas during search and rescue need
to be able to triangulate the position of a distress signal, simple single coverage is
insufficient, and all areas require triple coverage. The study uses a multi-objective
linear integer program (MOLIP) to approximate the original non-linear formulation
of the problem, and a network flow approach is used to generate solutions for two
cases studies provided by the US Department of Defense. Analysis of solutions to
this problem finds significant improvement (four standard deviations) in comparison
to randomly generated baseline solutions.
Bell et al. (2011) studied basing locations for military fighter aircraft used for
homeland defense. This research was motivated to examine which military airfields
the U.S. should consider critical for homeland defense against airborne threats, like
those that struck on 11 Sept 2001. This study was intended to help the U.S. Depart-
ment of Defense and Department of Homeland Security make positioning decisions
and to ensure that key locations were not closed by BRAC. The first stage of the
modeling approach was based upon on the set covering problem, where aircraft
launch times and speeds where also used to determine critical coverage distances
in the model. This stage identified the minimum number and location of sites for
positioning military response aircraft to cover a set of sixty-six critical locations.
420 J. E. Bell and S. E. Griffis

Sensitivity analysis during this stage made it evident that alternate optimal solutions
existed to the covering problem, and therefore the second stage of the model used
a p-median approach to find the minimum aggregate distance solution from among
those alternate optimal solutions to the location set covering problem. Results of the
study were briefed to decision makers across the US Air Force; however, the names
of locations and impact on actual decision making cannot be reported due to the
sensitive (classified) nature of the problem.
Following this study, Plastria (2012) offered a third stage to Bell et al. (2011)
work, based on a p-center-sum approach (Hansen et al. 1994). Plastria (2012) points
out that original formulation allows the critical distances to differ based on different
aircraft speeds and levels of protection. Plastria (2012) also creates a new critical
“protection” measure substitutable for the critical distance measure to account for
these differences in the model.
In another interesting military application, Tanerguclu et al. (2012) used location
modeling techniques and GIS data to improve the position of air defense weapons
and radar systems. This problem maximizes the expected coverage of potential air
attack routes (demands) by air defense systems while considering the probability of
a defense system successfully shooting down an airborne threat. A spatial decision
support system was developed to help quickly make decisions about weapon loca-
tions. The problem is complicated because vertical geography can obscure a defense
systems view of the sky it is designed to protect. This is why the link to GIS data and
their topographical mapping capabilities are important and unique to this problem.
Tanerguclu et al. (2012) base their maximal expected coverage model upon Daskin
(1983), however it differs in two important ways. First there are two different types
of facilities being located (weapons and radars) and a demand location is only con-
sidered covered if it is visible and within the range of both type of facilities. Second,
the model is hierarchical in nature and multiple weapon locations may be assigned
to more limited set of radar locations. Tanerguclu et al. (2012) uses LINGO opti-
mization software to generate solutions and dramatically improved upon the paper
maps solutions previously used by military commanders.
Finally, in one of the most recent military applications of coverage models is
by Dillenburger et al. (2013), who consider dropping equipment for military and
humanitarian aid missions by air. They seek to find the optimal location of drop-
ping equipment based on minimizing collateral damage and maximizing recovery
of equipment. Collateral damage is a concern for airdrop operations; especially in
humanitarian aid operations where dropping large amounts of supplies by air into
densely populated areas can create higher levels of damage on an already belea-
guered population. Therefore, the model in this research assigns weights to areas
that should be avoided. A number of search algorithms for generating solutions to
the problem are tested by Dillenburger et al. (2013), and these algorithms are vari-
ations of either a differential evolution or response surface algorithms. In order to
aid military decision-making, the paper presents the results of 14 different poten-
tial military scenarios, which lead to a number of suggested guidelines for military
planners conducting airdrop operations.
17 Military Applications of Location Analysis 421

17.3.3.4 Mobile or Dynamic Supply Locations

In addition to the more traditional location models, several military applications

have considered mobile demand or supply locations in their model in order to more
accurately replicate military operations. The facility location research of Dawson
et al. (2007) considers military security teams positioning in order to protect peri-
odic repair operations on the military’s nuclear intercontinental ballistic missile
(ICBM) fleet. The security (supply) locations that provide protective coverage for a
repair (demand) location in the problem are mobile and are repositioned as needed.
In the problem the ability of the military to repair an ICBM at a demand location
is dependent on the current location of security teams in the missile field. Dawson
et al. (2007) use a combined p-center and p-median approach for the problem and
employ a two-stage heuristic method to generate solutions. Their first stage finds
the p-center solution that minimizes the maximum distance to any demand location,
and the second stage attempts to find the minimum aggregate distance solution that
maintains the maximum distance found in the first stage p-center solutions.
Extending the work of Dawson et al. (2007), Overholts et al. (2009) studies
the ICBM maintenance problem and presents a different modeling approach that
changes the position of security coverage each day to allow for scheduling of repair
of nuclear missiles over multiple time periods. One of the authors of this chapter
took part in this study, and therefore we are able to provide a more detailed look at
this particular work.
The work of Overholts et al. (2009) is particularly interesting as it considers the
dramatic changes in military operations that have occurred in the last 15 years and
shows how these changed constraints may lead to different analytical approaches to
problems. For example, before 2003 maintenance schedules for ICBMs located in
the vast missile fields of Wyoming, Montana and the Dakotas were developed by
hand. Though not guaranteed to be overly efficient, these schedules were feasible
due to plentiful security personnel to cover operations. However by 2004, secu-
rity requirements for nuclear facilities increased, simultaneously with decreasing
security personnel availability due to budgets cuts and deployments to the Middle
East. This resulted in challenges for maintenance schedulers who needed to build a
schedule that could support increased maintenance requirements despite fewer secu-
rity personnel required to escort the maintenance crews. Further, complicating this
problem is the fact that not all missile maintenance actions are equivalent, there-
fore the research team had to work with the Air Force technicians in Wyoming to
develop a new weighting scheme with 18 separate categories to give priority in the
model to the most important repair work. Though constrained by a need to develop
a model useable by non-academics (necessitating an EXCEL model) Overholts et al.
(2009) developed a two-stage model to first select security alert locations, and then
build a daily maintenance schedule. The resulting first stage of the model is based
on the maximal covering problem formulation of Church and ReVelle (1974) and
aims to cover the maximum amount of weighted demand by locating security teams
at different discrete locations throughout the missile field. The second stage of the
422 J. E. Bell and S. E. Griffis

model, also based upon Church and ReVelle (1974), attempts to assign or sched-
ule the highest amount of weighted maintenance actions during the day at different
missile locations. In order to test its ability, the new approach is compared to the
actual maintenance schedules that were utilized by the US Air Force in Wyoming
during a 26-day period in 2005, and significant improvements in maintenance effec-
tiveness are observed with this modeling approach. Following the completion of the
analytical study, the results of the study were presented to Air Force leaders, and
transfer of the modeling capabilities to Air Force personnel was accomplished. As
an applications effort, this study highlights the need for close communication and
involvement of practitioners in the actual location modeling work, and the eventual
transfer of control of the newly developed analytical method back to the decision
makers who will use it.
A recent study by Qiu and Sharkey (2013) addressed an integrated location and
inventory planning problem motivated by mobile military sea-based supply ships.
The objectives of this problem include trying to minimize costs subject to minimum
service levels, measured by centrality of the mobile supply locations (ships) among
customers in different time periods. Complicating the problem, the demand points
(ships) are mobile across a time horizon, with variable inventory demand in each
discrete time period. Qiu and Sharkey (2013) employ dynamic programming and
tested on a number of problems based on actual support of humanitarian crises in
Haiti in 2010. Overall, Qiu & Sharkey provide one of the newest military applica-
tions that is grounded heavily in classic literature related to dynamic facility location
problems, and that highlights the mobile nature of supply locations in many military
problems.

17.3.3.5 Target Detection in a Spatial Area

Another area of military operations that has benefitted from operations research
involves spatial and location decision- making, includes finding and detecting mil-
itary targets over a large geographic area. The roots of this military application go
back to World War II, when mathematical models were first used to aid in the
detection of aircraft, submarines and other enemy threats in a combat zone. It is
important to note that this line of research continues today, and given the global-
ization of threats, is as important as ever. Flint et al. (2003) conduct a study on
how to coordinate the search for enemy targets with multiple unmanned aerial vehi-
cles (UAVs). The stochastic nature of target detection in a bounded spatial area is
addressed in Flint et al. (2003) using dynamic programming methods to formulate
solutions to this problem. Kierstead and DelBalzo (2003) also consider finding mov-
ing targets while searching for enemy submarines. Although general techniques for
the stationary detection problem have been around since the 1940s, Kierstead and
DelBalzo (2003) consider dynamic targets in an environment where space and time
are continuous. Applying a genetic algorithm Kierstead and DelBalzo (2003) find
solutions that provide improvement of up to 46 % over standard search techniques.
17 Military Applications of Location Analysis 423

More recently, Kress and Royset (2007) provided a location oriented search prob-
lem, to help special operations teams find the maximal number of targets they can
serve using UAV’s to help them conduct the local search. The problem is based on
the location-routing problem where the team locations are determined prior to the
search routing of the UAVs. However, this model includes additional constraints
such as the topography of the region, communications problems between vehicles,
and the location of potential threats in the region that are not included in most
LRP formulations. The results of the model were tested in actual US Army field
operations in 2006, and resulting in up to a 50 % improvement in target detection
compared to manual solutions generated by experienced military field commanders.
There are a growing number of studies related to searching for targets using spa-
tial modeling techniques, and these studies are closely related to facility location
modeling. This is one area where location modelers, especially those interested in
dynamic and stochastic techniques, might find interesting problems to study that
have direct military applications.

17.3.4 Military Location Modeling with Dispersion and Security

Military applications of location modeling sometimes also need to consider enemy

threats to resources and personnel. Although some of the models reviewed up to this
point included these constraints in their models, work has been done that considers
dispersion and security the primary focus of the research. While security and pro-
tection are always a focus of military commanders, this objective has received more
attention since 9–11. McGarvey and Cavilier (2003) study facility placement under
constraints that include “forbidden” regions. This problem was motivated in part by
the need of military commanders to avoid opposing military forces and to keep haz-
ardous materials out of civilian areas. Additionally, the work of Church and Scaparra
(2007 extends the r-interdiction median problem to include fortification of the most
important locations in a network that can potentially be disrupted or attacked by out-
side forces. Similarly, the work of Scaparra and Church (2008) uses a bi-level mixed
integer program for critical infrastructure protection planning, while Liberatore et al.
(2011) presents a model to optimally allocate defensive resources among existing
facilities in order to minimize the impact of terrorist attacks. This stream of model-
ing research, which includes facility network interdiction and protection of locations
has widespread application. Dong et al. (2010) considers multiple coverage lay-
ers of demand locations faced with limited protection resources in a network. The
model seeks to maximize the maximum coverage that protected facilities provide to
demand locations, when the unprotected locations in the network are destroyed.
Salmeron et al. (2009) considers the military planning of distribution operations
when certain locations in the transportation network are threatened. Using a stochas-
tic mixed integer model, Salmeron et al. (2009) specifically looks at ocean ship
scheduling and routing when some of the seaports in the network are vulnerable to
attack by the enemy, such as the deployment of material during Operation Desert
424 J. E. Bell and S. E. Griffis

Storm in 1990–1991. The model is primarily a dynamic routing problem that con-
siders both multiple ship and inventory types that must flow from supply to demand
locations in the network. However, attacks or disruptions may occur stochastically
at a location in the network, and the objective function of the model aims to min-
imize the resulting total disruption in the network. Although, levels of protection
are not yet considered in this problem, it may form the basis for continued work on
hybrid location routing problems that consider enemy attacks and needed protection
at critical nodes in the network.

17.3.5 Competitive Location Modeling: Attackers and Defenders

Military commanders not only consider how to defend their own networks, but also
consider how to compromise a competitor’s network. For example, Leinart et al.
(2002) developed a vertex cut-set algorithm on the transformation of a graph rep-
resenting a network to attack and intentionally disrupt the command, control and
communications network of an opposing enemy force. In particular, they considered
the best way to disrupt the facilities, radio towers, equipment, satellites, that support
the flow of information needed by the opposing forces commanders. The objective
of the model is to disrupt a set of target locations that possesses the maximum value
(cost vs. benefit) for the military operation.
Other military applications take into account dynamic actions by both attackers
and defenders to compete against each other in a geographical (spatial) area. For
example, research by Brown (2005) creates a decision tool for the prepositioning of
missile defense assets by the US military and its allies in Japan (defenders), taking
into account the potential attack of locations in Japan by ballistic missiles in North
Korea (attackers). This problem is an instance of a Stackelberg game, which the
authors represent as a bi-level integer linear program (e.g. Moore and Bard 1990).
The optimization model consists of an inner objective where the attackers attempt to
maximize the damage at enemy locations and the outer objective where the defend-
ers attempt to minimize the maximum amount of damage the attacker can achieve.
This bi-level program is converted to a mixed integer linear program for solution
purposes, and solved using CPLEX 9.0. The location aspects of this problem are
interesting, as the model optimizes the positioning of a limited number of defender
weapons at a candidate set of locations in Japan, based on the positioning decisions
of the attacker who locates his weapons in North Korea. The model is further com-
plicated by the inclusion of multiple weapon types with different coverage ranges,
as well as considering the various probabilities that a defender’s missiles actually
successfully intercept an inbound attacker missile. Additionally, the study analyzes
the value of secrecy in six different scenarios where the attacker and defender have
more or less information about the location of the other side’s weapons prior to
the actual attack. This military application was successfully presented to the US
Navy and US Strategic Command and was integrated into their actual planning and
operations for ballistic missile defense.
17 Military Applications of Location Analysis 425

Murphy et al. (2010) also consider the dynamic missile defense topic, but for
U.S. homeland defense. This model considers the simultaneous moves of both
attackers and defenders and also considers the positioning of a limited set of
resources to protect a larger number of critical locations dispersed across a large
geographic area. The formulation of the model follows von Neumann and Mor-
genstern (2004), where the attacker attempts to maximize expected gains and the
defender attempts to minimize expected losses. Murphy et al. (2010) present two
computational methods for reducing the size of the problem to make it more man-
ageable, and they include in their analysis the ability to reduce costs in the system
through improved strategic solutions. This military application highlights the dif-
ference between Cold-War (large detectable bombers) and post-Cold War (small
stealthy cruise missiles) threats. The paper analyzes the effectiveness and cost of
different military defense strategies for defense scenarios ranging from 5 to 40
US cities. One of the main contributions of this work is to point out the difficulty
involved with assigning “value” to different military targets. It also highlights that
the solutions selected are sensitive to different risk tolerances of the decision maker.
For instance, using a risk-tolerant Nash approach, the effectiveness of the models
solutions may be improved in comparison to more static and risk-averse approaches
that only consider the most critical locations.
Finally, in another military application that included competitive actions, Brown
et al. (2011) use an attacker-defender model to study a problem faced by the US
Navy, Coast Guard and Department of Homeland Defense as they attempt to defend
seaports from terrorist vessel attacks. This problem analyzes both the path that ves-
sels take as they travel the ocean in an attempt to attack vital seaport operations,
and the defenders capability to detect those attacking vessels through the optimal
positioning of its own defending vessels and on-shore radar sites. The problem is
stochastic and the objective is to minimize the probability that one or more attack
vessels will evade detection by the defender. An interesting aspect of this problem is
the geographic structure of individual ports is included with restrictions on port nav-
igation and observation based on the physical layout of the port area. Solutions to
this model show that despite the attacker having complete knowledge of the defend-
ers vessels and facilities, the defender can still detect nearly 100 % of the attacking
vessels in scenarios representing defense of the ports in Hong Kong and Bahrain.
The military applications in this section show that facility and resource location
decisions are imbedded in the attacker defender models being used by militaries to
defend against terrorist attacks and potential missile attacks in the modern post-cold
war error. These models contain aspects of coverage and protection of locations and
often contain stochastic aspects including the probability of detecting an adversary
or the probability of a location surviving an enemy attack. Overall, these models
represent some of the most complex and advanced wartime application of location
and spatial modeling.
426 J. E. Bell and S. E. Griffis

17.4 Military Applications Features

The articles that appear in Table 17.1 reveal several trends in military location anal-
ysis. To better understand these applications, a typology of features of military
applications is presented, including a list of some dominant trends and attributes
seen in location models in the literature to date.

17.4.1 Military Applications Typology

1. Multiple Objective. The majority (twenty five out of thirty six) of the loca-
tion models reviewed pursued more than one objective. Different combinations
of cost, coverage, dispersion, and security objectives were commonly present.
While many classic problems address a single objective it is evident that modern
military location problems are more complex, and often require a multi-objective
approach.
2. Global Attributes. Many of the applications reviewed in the literature have con-
straints or attributes related to the global nature of military problems. Both land
and sea-based facility locations were common in the applications. Addition-
ally, the complexity of facility location problems for the military is higher than
ever, considering cross border transportation bottlenecks, multiple transporta-
tion modes, short customer response requirements vs. long distances, unknown
construction costs in foreign countries, and potential cross-loading of materials
being moved between facilities in location routing applications. The global envi-
ronment present-day militaries operate within are far more complex that of the
middle 20th century when an assumption that an occupied area would be fully
controlled by the occupying military. The modern military locations look more
like islands within a given geography, rather than points within a secure border.
3. High Stock-Out Costs. In the military applications reviewed, it was clear that
meeting demand from a supply facility was critically important, and doing so
was assured through the outright prohibition of stock-outs, or the establishment
of artificially high stock-out cost when not filling demand. This makes sense for
combat support models, given stock-outs can result in failed missions, lost lives
and undermining political objectives of the nation. Additionally, the location of
facilities was also impacted by time sensitive nature of demand requirements
from military forces in the field.
4. High Demand Variability. Military operations are often characterized by high
demand variability. Sometimes there are long periods of zero demand, followed
by spikes in demand caused by combat operations or readiness training. Such
demand is atypical in commercial operations, and therefore military models need
to consider forecasts of demand at locations where previous demands have never
occurred. Such demand has been called “lumpy” by supply chain academics and
can exhibit high variability in both spatial and temporal dimensions. As such,
17 Military Applications of Location Analysis 427

thirteen out of the thirty-six articles reviewed employed stochastic models to

consider these highly variable demand patterns.
5. Minimize Logistics Footprint. A common objective in modern military operations
is to limit the amount of equipment and logistics support moved to a location
to support a military conflict. The weight and scope of such logistics support
is often called the “footprint” and a common objective is to minimize this. As
military applications seek to select facilities for peacetime operations, they must
also consider the footprint of logistics support that would have to be deployed
abroad to other locations in the case of a new conflict. This is especially important
as military cargo is often bulky, hazardous, and restricted by political factors from
being transported through certain locations when en route to a conflict (e.g. the
French prohibiting over flight to NATO aircraft during the bombing of Libya in
the 1980s).
6. Competition. A key feature of many military applications is that enemy forces
often compete against and attempt to destroy or disrupt the location network
created by military decision makers. Therefore, military decision makers must
be concerned with not only where the facilities are best located for operations,
but also how they will be protected and secured. This may involve considering
the location of enemy forces. Surprisingly, only a limited number of the studies
described in this chapter (eight out of thirty-six) included security or protection
features as part of their location model.
7. Dynamic. Military applications must constantly deal with changing demand
and supply characteristics within their models. As political interests and threats
change in the global environment, the dynamics of how to locate, adjust, or
close facilities across multiple time periods becomes of interest. Similarly, in a
more tactical combat environment, military applications that look at the dynamic
movements of friendly and enemy forces through a geographic area impacts the
decisions related to the placement of facilities, equipment and other resources.
Twelve out of the thirty six papers reviewed in this chapter had dynamic aspects
included in their location model.
8. Hybrid and Hierarchical Models. As the ability to solve more complex models
has increased, added realism has been added to military applications, includ-
ing combining location decisions with the closely related transportation routing
and inventory management decisions common to a supply chain network. Our
research sees a heavy emphasis on hybrid models that include routing and
inventory optimization being accomplished simultaneously with military loca-
tion decisions. Often these models consider location decisions at multiple layers
in a hierarchical network where units and commanders have a span of control in
their own operations that is tied to higher-level decisions made at the regional or
even global network level. Eleven of the thirty-six military applications reviewed
here contained some sort of hierarchical relationship in their location model.
These eight areas are representative of the attributes of military applications in loca-
tion modeling and provide a baseline for those initiating any new military location
428 J. E. Bell and S. E. Griffis

analysis effort. Together they represent a unique typology that separates this applica-
tion area from other areas, and highlights the complexity and challenges of solving
location problems in a military context.

17.5 Conclusion and Future Opportunities

It is believed that the dynamic and changing nature of the global environment and
military operations will continue to make this application area an important con-
text for location analysts in the coming years. Although, the typology presented
in Sect. 17.4 has listed several important features and trends in military location
modeling, the research reviewed in this chapter also points to several areas that are
underdeveloped and that can be expanded upon. In addition, we believe that the
expanding nature of military missions in a changing world provides several emerg-
ing opportunities not yet included in published military applications. Below we
describe several of the areas that offer opportunities for future research on military
applications of location analysis.
• Increasing Multiple Objectives. Although many of the military applications
reviewed in this chapter take a multiple objective approach, most look at only
a combination of two objectives, and usually this includes cost and coverage.
Opportunities still exist to include more objectives of protection and dispersion
in military location applications. In addition, work that includes and balances
three or more of these objectives simultaneously are much more realistic and
complex problems and are currently missing from the literature.
• Humanitarian Missions. The past decade has seen an increased number of mil-
itary operations related to providing humanitarian aid in the event of natural
disasters around the globe. This instance brings unique challenges for location
modelers who must rapidly design response supply chain networks when some
or the entire existing facility infrastructure and transportation network may have
been damaged by the disaster. Additionally, this means that the locations that are
selected and the resources deployed have to maintain a self-sustaining capabil-
ity since the underlying supply infrastructures in the disaster area may also be
destroyed or degraded. This topic area offers unique opportunities for military
and non-profit aid organizations to work together to make location decisions in
the face of a crisis.
• Homeland Defense. Following the attacks on September 11th, 2001, the US mil-
itary went through a major paradigm shift as it refocused its efforts on homeland
defense. In the Cold-War homeland defense often included external detection
capabilities in and around the coastal areas of the United States. Now, loca-
tion studies for homeland defense must consider border protection, but must also
consider the positioning of defense assets within the nation to protect important
national resources. It is believed that this topic will grow in importance.
• Urban Planning and Environmental Impacts. Military location analysis and
problem solving may often interact or even disrupt location modeling efforts
17 Military Applications of Location Analysis 429

by urban planners. For example, Westervelt and White (2009) and Johnson et al.
(2011), have shown that hazardous military operations can overlap with urban
development areas and actually disrupt those efforts and the underlying envi-
ronment. Therefore, there is an expanded need to look at joint planning between
military and urban planners to understand how urban growth and the need for new
public facilities such as hospitals, police stations, parks and housing areas may
overlap into military locations where hazardous military materials and training
operations already exist.
• Reverse Logistics Network Decisions. The military has a long history of design-
ing multi-echelon repair networks for heavy equipment items such as aircraft,
tanks and ships. However, as the need for environmental protection grows,
the location of facilities that creates a more sustainable network is growing in
importance. Military location modeling efforts that attempt to minimize fuel con-
sumption levels and carbon emission levels for operations in the network will
grow in importance as fuel prices and scarcity continue to grow in the future.
• Resource Security. As the global population and economies continue to grow
over the coming decades, military location planners must consider the poten-
tial competition for limited natural resources needed by the nation. Although
competition with the enemy has been considered at a tactical level in a num-
ber of military applications, future military applications should also consider the
competition between opposing military supply chains seeking to secure valuable
resources desired by others. It is believed that this resource security objective
will continue to grow in importance in military applications as decision mak-
ers attempt to secure access to ever more precious resources such as minerals,
fresh water, fuel, and strategic metals that are dispersed in a non-uniform patterns
across the world.
In conclusion, we believe that many of the classic military problems still persist,
though with added complexity factors previously unseen. Additionally, we believe
new classes of problems, with objectives unlike the classic military problems, are
emerging, and merit consideration by researchers interested in this area.

References

Amouzegar MA, Tripp RS, McGarvey RG, Chan EW, Roll, Jr RC (2004a) Supporting air and
space expeditionary forces: analysis of combat support basing options. RAND Corporation,
Santa Monica CA
Amouzegar MA, Tripp RS, Galway LA (2004b) Integrated logistics planning for the air and space
expeditionary force. J Oper Res Soc 55(4):422–430
Amouzegar MA, McGarvey RG, Tripp RS, Luangkesorn L, Lang T, Roll RC Jr (2006) Evaluation
of options for overseas combat support basing. MG-421. RAND Corporation, Santa Monica CA
Anderson EB (1998) Optimizing ammunition movement in support of US Pacific fleet’s position-
ing plan. Master’s Thesis, Naval Postgraduate School
Bell JE (2003) A simulated annealing approach for the composite facility location and resource
allocation problem: a study of strategic position of US Air Force munitions. Doctoral
Dissertation, Auburn Univ, AL
430 J. E. Bell and S. E. Griffis

Bell JE, Griffis SE, Cunningham WA III, Eberlan JA (2011) Location optimization of strategic
alert sites for homeland defense. Omega 39:151–158
Berman MB, Halliday JM, Carrillo MJ, Moore NY (1981) Combat benefits of a responsive logis-
tics transportation system for the European theater, R-2860-AF. RAND Corporation, Santa
Monica CA
Bonder S (2002) Army operations research-historical perspectives and lessons learned. Oper Res
50(1):25–34
Brown G, Carlyle M, Diehl D, Kline J, Wood K (2005) A two-sided optimization for theater
ballistic missile defense. Oper Res 53(5):745–763
Brown G, Carlyle M, Abdul-Ghaffar A, Kline J (2011) A defender-attacker optimization of port
radar surveillance. Nav Res Logist 58:223–235
Burks RE, Moore JT, Barnes JW, Bell JE (2010) Solving the theater distribution problem with tabu
search. Mil Oper Res 15(4):5–26
Cooper L (1963) Location-Allocation Problems. Oper Res 11:331–343
Chan Y, Mahan JM, Chrissis JW, Drake DA, Wang D (2008) Hierarchical maximal-coverage
location-allocation: case of generalized search-and-rescue. Comput Oper Res 35:1886–1904
Church RL, ReVelle CS (1974) The maximal covering location problem. Pap Reg Sci Assoc
32:101–118
Church RL, Scaparra MP (2007) Protecting critical assets: the r-interdiction median problem with
fortification. Geogr Anal 39(2):129–146
Cockayne EJ, Dreyer PA Jr, Hedetniemi SA, Hedetniemi ST (2004) Roman domination in graphs.
Discret Math 278:11–22
Dantzig GB (1963) Linear programming and extensions. Princeton University Press, Princeton
New Jersey
Dantzig GB, Fulkerson DR (1954). Minimizing the number of tankers to meet a fixed schedule.
Nav Res Logist Q 1(3):217–222
Daskin MS (1983) A maximum expected covering location model: formulation, properties, and
heuristic solution. Transp Sci 17(1):48–69
Daskin MS (2008) What you should know about location modeling. Nav Res Logist 55:283–294
Dawson MC, Bell JE, Weir JD (2007) A hybrid location method for missile security team
positioning. J Bus Manag 13(1):5–19
Dell RF, Ewing PL, Tarantino WJ (2008) Optimally stationing army forces. Interfaces 38(6):
421–435
Dillenburger SP, Cochran JK, Cammarano VR (2013) Minimizing supply airdrop collateral
damage risk. Socio-Econ Plan Sci 47:9–19
Dong L, Xu-Chen L, Xiang-Tao Y, Fei W (2010) A model for allocating protection resources in
military logistics distribution system based on maximal covering problem. 2010 International
Conference on Logistics Systems and Intelligent Management, Harbin China, 98–101
Drezner SM, Hillestad RJ (1984) Logistics models: evolution and future trends. In: Hughes WP Jr
(ed) Military modeling. The Military Operations Research Society, Alexandria
Drezner Z, Wesolowsky GO (1978) Facility location on a sphere. J Oper Res Soc 29:997–1004
Eccles HE (1959) Logistics in the national defense. The Stackpole Co, Harrisburg
Ewing PL Jr, Tarantino W, Parnell GS (2006) Use of decision analysis in the army base realignment
and closure (BRAC) 2005 military value analysis. Decis Anal 3(1):33–49
Farahani RZ, Asgari N (2007) Combination of MCDM and covering techniques in a hierarchical
model for facility location: a case study. Eur J Oper Res 176:1839–1858
Ferrer G (2010) Open architecture, inventory pooling and maintenance modules. Int J Prod Econ
128:393–403
Flint M, Fernandez E, Polycarpou M (2003) Stochastic models of a cooperative autonomous UAV
search problem. Mil Oper Res 8(4):13–32
Ford LR Jr, Fulkerson DR (1954) Maximal flow through a network, RM-1400-PR. RAND
corporation, Santa Monica
Garfinkel RS, Rao MR (1971) The bottleneck transportation problem. Nav Res Logist Q 1(3):
465–472
17 Military Applications of Location Analysis 431

Geisler MA, Wood MK (1951) Development of dynamic models for program planning. In:
Koopmans TC, Cowles Commission Monographs (eds) Activity analysis of production and
allocation, vol 13. Wiley, New York, pp 189–215
Geller A, George D, Tripp RS, Amouzegar MA, Roll RC Jr (2004) Supporting air and space expe-
ditionary forces: analysis of maintenance forward support location operations. MG-151. RAND
corp, Santa Monica CA
Ghanmi A (2011) Canadian forces global reach support hubs: facility location and aircraft routing
models. J Oper Res Soc 62:638–650
Ghanmi A, Shaw RHAD (2008) Modelling and analysis of Canadian forces strategic lift and pre-
positioning options. J Oper Res Soc 59:1591–1602
Glover F, Laguna M (1997) Tabu search. Kluwer Academic Publishers, Norwell
Griffith SB (1963) Sun Tzu The art of war. Oxford University Press, Oxford
Gue KR (2003) A dynamic distribution model for combat logistics. Comput Oper Res 30:367–381
Hakimi SL (1964) Optimum locations of switching centers and the absolute centers and medians
of a graph. Oper Res 12(3):450–459
Hansen P, Labbe M, Minoux M (1994) The p-center sum location problem. Cah du C.E.R.O.
36:203–220
Henning MA (2003) Defending the roman empire from multiple attacks. Discret Math 271:
101–115
Hillestad RJ (1982) Dyna-METRIC: dynamic multi-echelon technique for recoverable item
control, R-2785-AF. RAND corporation, Santa Monica CA
Hillestad RJ, Carrillo MJ (1980) Models and techniques for recoverable item stockage when
demand and the repair process are non-stationary, part I: performance measurement, N-1482-
AF. RAND corporation, Santa Monica CA
Howard ME, Paret P (1989) Translated version of Clausewitz C (1832) On War. Princeton
University Press, Princeton
Hughes WP Jr (1984) Military modeling. The Military Operations Research Society, Alexandria
Hughes WP Jr (2002) Navy operations research. Oper Res 50(1):103–111
Jacobs WW (1955) Military applications of linear programming. In: Antosiewicz HA (ed) Pro-
ceedings of the second symposium in linear programming, and Directorate of Management
Analysis 2. National Bureau of Standards, Washington DC, pp 1–27
Johnson S, Wang G, Howard H, Anderson AB (2011) Identification of superfluous roads in terms
of sustainable military land carrying capacity and environment. J Terramech 48(2):97–104
Johnstone DP (2002) Modeling the pre-positioning of air force precision guided munitions.
Master’s Thesis, Wright-Patterson AFB OH: Air Force Institute of Technology
Johnstone DP, Hill RR, Moore JT (2004) Mathematically modeling munitions prepositioning and
movement. Math Comput Model 39(6):759–772
Kaplan S (1974) Application of programs with maximin objective functions to problems of optimal
resource allocation. Oper Res 22:802–807
Kierstead DP, DelBalzo DR (2003) A genetic algorithm applied to planning search paths in
complicated environments. Mil Oper Res 8(2):45–59
Killingsworth PS, Galway L, Kamiya E, Nichiporuk B, Ramey TL, Tripp RS, Wendt JC (2000)
Flexbasing: achieving global presence for expeditionary aerospace forces. MR-1113-AF.
RAND corporation, Santa Monica
Killingsworth WR, Berkowitz D, Burnett JE, Simpson JT (2008) The application of supply network
optimization and location analysis to a DOD repair supply chain. Def Acquis Rev J 15(3):
277–291
Kress M, Royset JO (2007) Aerial search optimization model (ASOM) for UAVs in special
operations. White paper report. Naval Postgraduate School, Monterey CA
Levin KD, Friedman Y (1982) Optimal deployment of logistic units in dynamic combat conditions.
Eur J Oper Res 9:41–46
Liberatore F, Scaparra MP, Daskin MS (2011) Analysis of facility protection strategies against an
uncertain number of attacks: the stochastic r-interdiction median problem with fortification.
Comput Oper Res 38:357–366
432 J. E. Bell and S. E. Griffis

Leinart JA, Deckro RF, Kloeber JM Jr., Jackson JA (2002) A network disruption modeling tool.
Mil Oper Res 7(1):69–77
Litwhiler DW Jr, Aly AA (1979) Large region location problems. Comput Oper Res 6:1–12
Loerch A, Boland N, Johnson E, Nemhauser G (1996) Finding an optimal stationing policy for the
US army in europe drawdown. Mil Oper Res 2(4):39–51
McGarvey RG, Cavalier TM (2003). A global optimal approach to facility location in the presence
of forbidden regions. Comput Ind Eng 45(1):1–15
Melo MT, Nickel S, Saldanha-da-Gama F (2009) Facility location and supply chain management—
a review. Eur J Oper Res 196:401–412
Moore JT, Bard JF (1990) The mixed integer linear bilevel programming problem. Oper Res
38:911–921
Morse PM, Kimball GE (1951) Methods of Operations Research. Peninsula Publishing, Los Altos
California
Muckstadt JA (1973) A model for a multi-item, multi-echelon inventory system. Manag Sci
20(4):472–481
Muckstadt JA (1980) Comparative adequacy of steady state versus dynamic models for calculating
stockage requirements, R-2636-AF. RAND corporation, Santa Monica
Murphy EM, Payne MD, van der Woude GW (2010) Strategy alternatives for homeland air and
cruise missile defense. Risk Anal 30(10):1507–1519
Murty KG, Djang PA (1999) The US army national guard’s mobile training simulators location
and routing problem. Oper Res 47(2):175–182
Natrella JV (1955) New applications in linear programming. Comput Autom 4(11):22
Nguyen BU, Ng KYK (2000) Modeling Canadian search and rescue operations. Mil Oper Res
5(1):5–16
Overholts DL II, Bell JE, Arostegui MA (2009) A location analysis approach for military
maintenance scheduling with geographically dispersed service areas. Omega 37:838–852
Plastria F (2012) A note towards improved homeland defense. Omega 40:244–248
Qiu J, Sharkey TC (2013) Integrated dynamic single-facility location and inventory planning
problems. IIE Trans 45:883–895
Rappold JA, Van Roo BD (2009) Designing multi-echelon service parts networks with finite repair
capacity. Eur J Oper Res 199:781–792
ReVelle CS, Rosing KE (2000) Defendens imperium romanum: a classical problem in military
strategy. Am Math Mon 107(7):585–594
ReVelle CS, Marks D, Liebman JC (1970) An analysis of private and public sector location models.
Manag Sci 16(11):692–707
Roth JP (1998) The logistics of the Roman Army at war (264 B.C.—A.D. 235). Brill, Leiden
Salmeron J, Wood K, Morton DP (2009) A stochastic program for optimizing military sealift
subject to attack. Mil Oper Res 14(2):19–39
Scaparra MP, Church RL (2008) A bilevel mixed-integer program for critical infrastructure
protection planning. Comput Oper Res 35:1905–1923
Segall RS (2000) Some quantitative methods for determining capacities and locations of military
emergency medical facilities. Appl Math Model 24:365–389
Sentlinger BK (2000) Optimal positioning of naval precision guided munitions. Master’s Thesis,
Naval Postgraduate School
Sherbrooke CC (1966) METRIC: a multi-echelon technique for recoverable item control. Oper
Res 16(1):122–141
Stewart I (1999) Defend the Roman Empire! Scientific American 281:136–138
Tanerguclu T, Maras H, Gencer C, Aygunes H (2012) A decision support system for locating
weapon and radar positions in stationary point air defence. Inf Syst Front 14:423–444
Toregas C, Swain R, ReVelle C, Bergman L (1971) The location of emergency service facilities.
Oper Res 19(6):1363–1373
Toyoglu H, Karasan OE, Kara BY (2011) Distribution network design on the battlefield. Nav Res
Logist 58:188–209
17 Military Applications of Location Analysis 433

Tripp RS, Galway L, Killingsworth PS, Peltz E, Ramey TL, Drew JG (1999) Supporting expe-
ditionary aerospace forces: integrated strategic agile combat support planning framework.
MR-1056-AF. RAND corporation, Santa Monica CA
Tripp RS, Galway L, Ramey TL, Amouzegar M, Peltz E (2000) Supporting expeditionary
aerospace forces: a concept for evolving the agile combat support/mobility system of the future.
MR-1179-AF. RAND corporation, Santa Monica CA
Underwood DK, Bell JE (1999) AEF munitions availability. Air Force J Logist 23(4):12–17
Westervelt JD, White MJ (2009) Identifying future locations for noise-generating military training
opportunities within urbanizing landscapes. Mil Oper Res 14(1):53–60
Wood MK, Geisler MA (1951) Development of dynamic models for program planning. In: Koop-
mans TC (ed) Activity analysis of production and allocation. John Wiley & Sons, New York,
pp 189–192
Woolsey G (1986) The fifth column: on the minimization of need for new facilities, or space wars,
lack of presence, and Delphi. Interfaces 16:53–55
von Neumann J, Morgenstern O (2004) Theory of games and economic behavior. 60th anniversary
edition. Princeton University Press, Princeton
Yamani A, Hodgson TJ, Martin-Vega LA (1990) Single aircraft mid-air refueling using spherical
distances. Oper Res 38(5):792–800
Zangwill WI (1966) A deterministic multi-product, multi-facility production and inventory model.
Oper Res 14(3):486–507
Index

1-median, 66, 81 definition of, 199, 313

of demand, 376
A Covering, 1, 10, 280
Access, 13, 28, 72, 98, 293 Critical habitat, 161, 175, 177
Accessibility, 4, 13, 224, 226, 238, 293 and WUI, 178
Adjacency, 89, 99, 127, 184, 341 Cutting unit, 90
definition of, 339
Analytic hierarchy process (AHP), 11, 29, 35 D
Anti-covering, 156 Data fusion, 192, 193, 220
Automatic vehicle identification (AVI), 308, benefits of, 194
310, 318 Decision-making, 12, 29, 81, 403, 410, 411,
number of, 310, 323 420
Demand, 30, 64, 65, 70, 93, 205, 275, 356,
B 376, 407, 420
Balance, 293, 330, 332, 346, 400 agglomeration, 62
Bank branch, 25, 26, 28, 41 point, 27, 28, 62, 64, 415, 422
location of, 29, 30, 32, 33, 51 District, 2, 35, 174, 181, 330, 331, 342
Banking, 31 Districting, 273, 275, 329
telephone-internet, 25
types of, 33 E
Breast cancer screening centers, 237 Ecodistrict, 86, 87, 98
Emergency, 6, 43, 224
C Emergency vehicles, 13, 373
Canadian coast guard (CCG), 369, 370, 375, Equity, 8, 12, 13, 293
400, 401 Explicit model, 197, 198, 205, 206, 221
Center, 13, 57, 61, 64, 74, 140, 156, 165, 169, solution of, 206, 210
359 Explicit-implicit model, 220
Choice, 64, 96, 191, 226
of SISAR, 371 F
Compactness, 275, 330, 332, 334 Fire, 88, 104, 173, 175, 376, 418
definition of, 334 Fire station, 2, 13, 20, 26, 155, 294–296, 373,
Congestion, 81, 191, 192, 224, 228, 238, 307 407
Contiguity, 330, 332, 334 First responders, 20
Core habitat, 156–158 Forecast, 196, 352
definition of, 157, 163 of demand, 356, 357, 426
Coverage, 8, 13, 27, 175, 196, 200, 412, 426 Forest, 85, 89, 101
application of, 420 fire, 104

© Springer International Publishing Switzerland 2015 435

H. A. Eiselt, V. Marianov (eds.), Applications of Location Analysis, International Series in
Operations Research & Management Science 232, DOI 10.1007/978-3-319-20282-2
436 Index

in US, 162, 174, 182 Mid integer programming, 144, 227, 239
Forestry, 87, 89, 102 Military, 58, 66, 403
challenges in, 103 application of, 404, 407, 411, 414, 420, 429
coverage models in, 419, 420
G history of, 404
Global positioning system (GPS), 192, 219, in US, 415, 418
308 Military bases, 66
Graph-theoretical heuristic, 125, 142, 144 in US, 414
application of, 133 Mixed integer programming, 102, 274
Grid, 168, 195, 377, 384, 391 Multi-graded, 276, 278, 280, 284, 288
Multicriteria decision making, 4, 9, 10, 12
H Multiple criteria, 26, 29, 400
Harvest, 85, 86, 88, 89, 97, 98, 100, 102, 334 Multiple objectives, 80, 334, 416, 428
Health care, 148
in US, 223 N
layout planning problems in, 109, 110 Natural Reserve, 156
Heuristics, 13, 28, 99 NP-Complete, 28
development of, 195 NP-Hard, 13, 26, 28, 30, 31, 41, 42, 46, 51,
Highways, 13, 57, 66, 69, 196 129, 195
Hospitals, 6, 12, 35, 87, 125, 127, 133, 413,
429 O
management of, 133, 135, 140 Optimality gap, 100
Outlet, 12
I Overpopulation, 353, 360–363, 365
Implicit model, 197, 200, 205, 206, 210, 211,
213 P
Industrial forestry, 85, 334 p-median, 6, 12, 28, 330, 338, 341, 413, 421
Infrastructure, 57, 61, 67, 70, 76, 194, 353, 428 Penalty, 8, 26, 40, 400
Inmate, 349, 350, 354, 364 cost of, 40
Intelligent sensors, 191 Planning horizon, 90, 110, 174, 178, 181, 184,
Intelligent transportation systems (ITS), 191 186, 352
definition of, 191 Police, 6, 27, 329, 331, 335, 356, 372, 376
Post offices, 26, 27, 293
J Preventive care, 223, 224, 226, 231, 239
Jails, 13, 349, 350, 356, 365 Prison, 349, 350, 352, 354, 362, 365
Private sector, 284, 294, 350
L Profit, 2, 4, 33, 407
Law enforcement, 353 PROMETHEE, 10, 11, 354
Layout, 109, 127, 140, 142 Public sector, 293
development of, 144 Public services, 20
single-floor, 143 Pulpmill, 87, 88, 94, 96, 105
Location analysis, 10, 11, 13, 20, 85, 102, 239,
293, 374 Q
application of, 2, 405, 414, 419, 428 QAP See Quadratic assignment problems, 109
Location-allocation heuristic, 335, 340, 346 Quadratic assignment problems, 111, 145
Logistics park, 56, 62, 81
definition of, 56 R
location of, 67 Response standard, 294–296, 299, 300
Retail, 3
M Revenue, 90, 92, 96
Maritime search and rescue, 370, 373–375,
386, 387, 399, 400 S
Maximum capture problem, 4, 27 Sawmill, 86, 94
Median, 2, 4, 61, 340 Scenarios, 3, 186, 361, 399
Index 437

School location, 275, 284 incidents, 191, 197

Schools, 276 Training sites, 69
Search and rescue (SAR), 369, 370, 373, 375, Transportation network, 35, 40, 191, 307, 410,
400 428
Sensors, 192, 204 Travel distance, 80, 126, 146, 225, 226, 276,
impact of mobile, 217 277, 287, 354, 399
number of, 211 Travel time, 308, 311
Shape, 301, 329
Simulation, 144, 202, 210 U
Single objective, 7, 12, 196 Uncertainty, 103, 104
Single-graded, 278 US forest service (USFS), 174, 175, 183, 189
Spotted owls, 156
Stands, 86, 90, 100
V
location of, 98, 99
Vehicle identification, 320, 325
Strategic level, 88
Vessel, 375
Student allocation, 280, 285
Supply chain, 56, 76, 86, 88, 89, 91, 96, 100, Voronoi, 333, 335
427
Supply location, 88, 89, 96 W
Sustainable forest management, 87, 90, 93 Weapons positioning, 413
Weiszfeld, Endre, 2, 63
T Wildfire, 173, 189
Tabu search, 41, 42, 418 Wildlife, 156
Taxonomy, 2, 111 Workload
TOPSIS, 9, 29 with
Traffic, 197, 307 MCLP, 383