You are on page 1of 19

European Journal of Operational Research 165 (2005) 1–19

www.elsevier.com/locate/dsw

Invited Review

Location analysis: A synthesis and survey


a,* b
C.S. ReVelle , H.A. Eiselt
a
Department of Geography and Environmental Engineering, The Johns Hopkins University, 313 Ames Hall, 3400 N. Charles Street,
Baltimore, MD 21218-2681, USA
b
University of New Brunswick, Fredericton, NB, Canada
Received 15 January 2003; accepted 17 November 2003
Available online 4 June 2004

Abstract

The mathematical science of facility siting has attracted much research in discrete and continuous optimization over
nearly four decades. Investigators have focused on both algorithms and formulations in diverse settings in the private
sector (e.g., industrial plants, banks, retail facilities, etc.), and the public sector (e.g., ambulances, clinics, etc.). Each
formulation has differences and similarities relative to the others, but the peculiarities of each problem provide the fuel
for the hundreds of investigations. The presence of non-linearities and requirements for zero–one variables head the list
of challenges that have occupied researchers in this active and expanding field. We review here the many facets of this
exciting and centrally placed field through reference to both seminal works and current reviews.
 2004 Elsevier B.V. All rights reserved.

Keywords: Facility location; Siting models; Survey; Network location; Continuous location

1. Introduction layout problems, in contrast, the facilities to be


located are fairly large relative to the space in
The term Location Analysis refers to the which they are positioned, and interaction between
modeling, formulation, and solution of a class of facilities is the norm rather than an exception. The
problems that can best be described as siting bibliography by Domschke and Drexl (1985) lists
facilities in some given space. The expressions more than 1500 references dealing with location
deployment, positioning, and siting are frequently and layout problems, and many more contribu-
used as synonyms. We distinguish between loca- tions have appeared since then. There are four
tion and layout problems by noting that the components that characterize location problems;
facilities in location analysis are small relative to these are (1) customers, who are presumed to be
the space in which they are sited and interaction already located at points or on routes, (2) facilities
between the facilities may or may not occur. In that will be located, (3) a space in which customers
and facilities are located, and (4) a metric that
indicates distances or times between customers
*
Corresponding author. Tel.: +1-410-516-7095; fax: +1-410-
and facilities. Applications of location problems
516-8996. abound: they range from gas stations and fast food
E-mail address: revelle@jhu.edu (C.S. ReVelle). outlets to landfills and power plants.

0377-2217/$ - see front matter  2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2003.11.032
2 C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19

In some non-traditional location problems such be further subdivided into continuous or discrete
as product positioning, each facility represents a location problems. Distances in jd are most often
product, which is mapped into a feature space, each derived from Minkowski distances, which are de-
of whose dimensions represents a relevant feature fined as a family of distances with a single
of the product. Similarly, each customer (or, typi- parameter p. In particular, the ‘p distance between
cally, group of customers) is mapped into the fea- a point (ai , bi ) and a point (aj , bj ) with i 6¼ j is
ture space, and the distance between customers and defined as
products will give an indication which products the
dijp ¼ ½j ai  aj jp þ j bi  bj jp 1=p :
customers are going the purchase, thus allowing
estimates concerning the sales of the products. A A large part of the literature focuses on three
similar example from political sciences deals with special cases: for p ¼ 1, we obtain the rectilinear
the optimal positioning of candidates for office; for (or rectangular or Manhattan or ‘1 ) distance
details see Niemi and Weisberg (1976). A survey of dij ¼j ai  aj j þ j bi  bj j, the Euclidean (or
many distinct applications of location models is straight line or ‘2 ) metric with
provided by Eiselt (1992), ranging from traditional qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
applications involving newspaper transfer points dij2 ¼ ðai  aj Þ2 þ ðbi  bj Þ2 ;
(Jacobsen and Madsen, 1980), solid waste transfer
points (Marks and Liebman, 1971; Wirasinghe and and the Chebyshev (or ‘‘max’’, or ‘1 ) metric with
Waters, 1983), bank branches (Hopmans, 1986), dij1 ¼ maxfj ai  aj j; j bi  bj jg:
and motels (Kimes and Fitzsimmons, 1990) to the
more unusual location problems such as the loca- Other functions such as gauges have been sug-
tion of a church camp (Huxley, 1982), the deter- gested by various researchers; see, e.g., Plastria
mination of apparel sizes (Tryfos, 1986), ingot sizes (1995).
(Vasko et al., 1987), and the location of rain gauges For the estimation of road distances between
(Hogan, 1990). For a list of new location appli- points based purely on the coordinates of the
cations, readers are referred to Current et al. endpoints, Love and his co-researchers have
(2002). worked with a number of additional metrics, see,
Whereas routing problems are typically found for instance, the text by Love et al. (1988) or
on the lower end of the strategic–tactical–opera- Brimberg and Love (1995). The main reason for
tional continuum (meaning that they are typically calculating distances between points whenever
reasonably well defined and measurable), location needed rather than storing them is the sharply
problems are likely to have multiple objectives; decreased need for storage space: the coordinates
fuzzy and ill-posed, making it much harder for the of n given points require OðnÞ storage space, while
analyst to model them. One approach that appears all point-to-point distances require Oðn2 Þ memory.
to be promising is the use of tools from decision One of their main results is the recommendation
analysis; see, e.g., Rey et al. (1995), who use the of a weighted ‘p distance function for practical
ELECTRE outranking method to arrive at a problems.
solution that reconciles the many different criteria In contrast, distances in network location
included in the problem. problems are measured on the network itself,
typically as the shortest route on the network of
arcs connecting the two points. Using the shortest
2. The space of location decisions path usually requires a preprocessing phase to find
all shortest paths between all relevant pairs of
Location scientists often use the space in which points, typically a Dijkstra (1959) based O(n3 )
facilities are located to distinguish between classes algorithm. Both planar problems and network
of location problems. We distinguish between problems can be further subdivided into continu-
location problems in d-dimensional real space jd ous and discrete location problems. In continuous
and network location problems, each of which can problems, the points to be sited can generally be
C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19 3

placed anywhere on the plane or on the network. n new facilities as a fixed cost of cf each. At the ith
An example of a continuous location problem in facility, they will charge a price pi , at which the
the plane might be the placement of a helicopter demand is di ðÞ, where  denotes a variety of fac-
for trauma pickup, while the location of an tors that determine the price. Let tij denote the unit
ambulance or a tow truck along a stretch of transportation cost, and let cv denote the variable
highway is a typical application of a continuous production cost (which, for simplicity, is assumed
problem on a network. In discrete problems, in to be equal at all sites). Defining variables xij that
addition to the points to be positioned, the facili- denote the proportion of the demand at demand
ties can conceptually be placed only at a limited point i that is satisfied at the facility at point j, we
number of eligible points on the plane or network. can write the firm’s profit as
Thus, discrete location models have gone through X X
an additional preprocessing phase that has pre- P¼ pi di ðÞ  cf n  cv di ðÞ
i i
selected candidate sites at which the facilities may X X
be sited. An example of a discrete location prob-  di ðÞ tij xij :
i j
lem in the plane is the positioning of transmitter
stations that are to be placed at some permissible The first term is the revenue, the second describes
points within a region, such as mountaintops, the fixed location costs, the third denote the vari-
while an application of discrete network location ables production costs, and the last term expresses
problems is found in the location of retail facilities the transportation costs. Consider now two pricing
that can only be sited on lots that are zoned for policies. Suppose that the firm uses mill pricing.
them. This means that the transportation costs are borne
A result of the above discussion is that, typi- by the customers, so that the last term in the above
cally, continuous location problems, which are for profit function vanishes. For any given number of
the most part planar problems, tend to be non- facilities n and given set of prices pi , the first two
linear optimization problems, while discrete loca- terms are a constant, so that the maximization
P
tion problems, which most often are network P is equivalent with minimizing cv i di ðÞ ¼
of P
problems, involve zero–one variables that result in cv i di ðpi þ minj ftij gÞ, assuming that the cus-
integer programming/combinatorial optimization tomers purchase from the facility that quotes the
problems. Clearly, there are countless hybrid lowest full price, defined as the purchase price plus
models, many of which do not fall into these neat the transportation costs. It becomes apparent that
categories. a firm that applies mill pricing will attempt to
maximize the demand that it satisfies. This type of
problem was first mentioned by ReVelle (1986),
3. Classes of location objectives who referred to it as capture problem. Consider
now the case, in which the firm decides to pursue a
A variety of factors other than customers, policy of delivered pricing, i.e., it will pay for the
facilities, space, and distance functions play transportation costs which it includes in the prices.
important roles in location modeling. One such As a result, customer demand is now di ðpi Þ, and for
factor is the objective that is employed by the fixed n and pi , the profit function reduces P to the
decision maker. Traditionally, the facilities that minimization of the transportation costs j wi tij xij
were to be located were assumed to be desirable in with weights wi ¼ di ðpi Þ. A point that minimizes
the sense that the closer they are to the customers, transportation costs is typically referred to as
the better the value of the objective function. As median in a network setting (or sometimes as
Eiselt and Laporte (1995) pointed out, these are points of production or of distribution), while such
fall into the category of ‘‘pull’’ objectives. Gab- a point is generally called a Weber point in planar-
zewicz and Thisse (1989) have derived two of the continuous models. The latter name is in honor of
three main ‘‘pull’’ objectives from a general profit the German economist Alfred Weber (1909) who
function. Suppose that a firm has decided to locate worked on these problems at the beginning of the
4 C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19

20th century and summarized his findings in his Other features of interest to the modeler include
classical treatise; for a summary see Wesolowsky the number of facilities that are to be located. The
(1993). Another important pull objective, one that simplest of the models attempt to locate a single
cannot be derived from the above general profit facility, whereas more complex models might
maximization function, are so-called center prob- locate p facilities, where the parameter p may
lems, whose objective is to minimize the largest be fixed by the decision maker or imposed by a
customer-facility distance. budget limitation. Alternatively, the number of
Since the late 1970s, researchers have also facilities may be unknown at the outset and
considered the location of noxious or obnoxious determined endogenously through the elements of
(or simply undesirable) facilities, see, for instance, the model’s objective. The best-known in this
Erkut and Newman (1989). In contrast to facilities category of models with ‘‘free entry’’ is the plant
where nearness is desirable, customers wish, as one location problem that minimizes the sum of plant
of their objectives, to ‘‘push’’ undesirable facilities opening costs and distribution costs, and allows
as far from them as possible. By themselves and the number of facilities to be an output or conse-
without any further restrictions, ‘‘push’’ objectives quence of the minimum cost solution. Another
will attempt to locate towards infinity. This is why type of model in this category is the location set
they are often coupled with other, predominantly covering model first introduced by Toregas et al.
‘‘pull’’, objectives. An example of a problem with (1971). The idea is to find the least number of
these two conflicting objectives is the location of a facilities so that all demand points are covered
sanitary landfill, a facility most people would within a prespecified distance standard.
consider undesirable. The customers of the landfill Other enhancements of location models include
and the general population would prefer to push uncertainty or risk with respect to the ability to
the landfill as much away from them as possible, service demands within desirable distances.
while, in an attempt to minimize costs, the com- Sometimes, barriers may exist that cannot be
pany in charge of the garbage collection will at- crossed in continuous location models. As well
tempt to locate as close to those who generate the there is a category of facilities referred to as
garbage as possible. (A closer examination will, of ‘‘extensive facilities’’ that cannot be modeled as
course, reveal that, since the public is ultimately points but have length as well as position charac-
paying for the service, one the public’s objectives is teristics, thus providing a link between location
also of the ‘‘pull type’’.) Any solution to the models, layout models, and network models.
problem will ultimately require explicit or implicit An early characterization of model forms and
information about the tradeoffs of the objectives objectives that is still in use was provided by
involved. ReVelle et al. (1970) in which location problems
Finally, a third class of objectives is the were divided into private sector problems and
achievement of equity. Models with such objec- public sector problems. The private sector prob-
tives attempt to locate the facilities, so that the lems seek the sites for plants and or warehouses;
customer-to-facility distances may be as similar to these sites are those that optimize some function of
each other as possible. This equalization gives rise the monetary value associated with the location. In
to the term ‘‘balancing objectives’’. Alternatively, contrast, public sector problems seek facility sites
the distances from clients to the nearest facility that optimize the population’s access (measured in
may be bounded by some generally recognized the various ways discussed above) to those facili-
distance standard. Where the standard is an upper ties. Clearly, there are many shades of gray be-
bound, this is done so as to provide relatively tween the extremes of ‘‘private’’ and ‘‘public’’.
equal access to a desirable facility, or, equivalently Consider, for instance, a decision maker in the
in the case of undesirable facilities, where the public sector who accepts the measure of the
standard is a lower bound on the separation dis- average customer-facility distance as a proxy for
tance, the standard is imposed to distribute the accessibility. The resulting model minimizes the
unpleasantness as equally as possible. sum of distances, i.e., a minisum objective that is
C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19 5

most often associated with cost-minimization in earliest true location algorithm was offered by the
the private sector. Typically, defining objectives in mathematician Weiszfeld (1937). He solved a
public sector models is much more complicated minisum planar problem with one facility and
than in the private sector. Most often, the private Euclidean distances. His algorithm was written in
sector maximizes profits or minimizes costs, French and published in a Japanese journal of
whereas the public sector objectives are less tan- mathematics, thus eluding detection by other
gible. For instance, whether a public decision authors, resulting in the rediscovery of the algo-
maker who attempts to maximize accessibility of rithm by Miehle (1958), Kuhn and Kuenne (1962),
the facility in question will choose to minimize the and Cooper. His main ideas are as follows.
average distance or the longest distance between The objective function of the two-variable
the facility and any of its customers is a matter of model (the only variables are the two coordinates
opinion. While Rawls (1971) made a case for of the new facility) is
minimax objectives, Harsanyi (1975) demonstrated X X qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
the desirability of minisum objectives. Here, the Min z ¼ di wi ¼ wi ðai  xÞ2 þ ðbi  yÞ2 ;
concept of multiple objectives may well apply, and i2I i2I

the tradeoff between average and maximum dis- where customers are located at ðai , bi Þ, the facility
tance developed and displayed. Whichever objec- is to be sited at ðx; yÞ, wi is the known demand (or
tive may be chosen, the difficulty of measurement, weight) of customer i; and I is the set of customers.
particularly in public decision making, remains. This non-linear function can be differentiated,
resulting in two non-linear equations which are set
equal to zero, but for which no closed form solu-
4. Location problems in the plane tion exists. The first-order conditions are first
written in the form
Minisum planar location problems dominated P wi ai
the location discussion until the mid-1960s. Al- i2I qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
though those with a bent for history will argue ðai  xÞw2i þ ðbi  yÞ
x¼P wi ;
over firsts (see, e.g., Drezner et al., 2002), appar-
i2I qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ently the first location problem in the literature is
ðai  xÞw2i þ ðbi  yÞ2
due to the mathematician Torricelli (1608–1647).
His problem involves three customers and one with a similar equation for y. A clever iterative
facility, with the task being the minimization of the technique then uses a present guess of ðx; yÞ for x
sum of Euclidean distances between the customers and y on the right-hand side of the equation, and
and the facility. Torricelli determined that the calculates new values ðx; yÞ for the two variables
facility would be located at a point, such that on the left-hand side, which are then used again as
the angles formed by the lines that connected the input in the next iteration. It must be observed,
facility and any two of its customers are 120. This though, that the objective function is not differ-
result applies as long as the angle formed at each entiable everywhere, as the derivative does not
customer location by connecting it to the other exist at the demand points. Eyster et al. (1973)
two customer points, is no larger than 120. If any have proposed a remedy called hyperbolic
one of these angles exceeds 120, the optimal approximation, a technique based on perturba-
location coincides with that point. From the tions. With or without approximation, the com-
beginning of the 20th century until the 1960s, putational scheme, which often begins with the
economists, geographers and regional scientists centroid as a starting point, tends to converge
dominated the field. Their models tended to be quickly. However, Drezner (1992) reports very
either geometric or exhaustive computational slow convergence for some problems; in fact, a 4-
methods to locate facilities in the 2-dimensional node problem failed to converge in 100,000 itera-
plane. The earliest and, arguably, most prominent tions. Kuhn (1973) was the first to point out that
example is Weber’s (1909) seminal work. The there may be local optima at demand points.
6 C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19

Drezner and Simchi-Levi (1992) showed that for those facilities. Here, we typically distinguish be-
single-facility problems with equal weights and tween customer choice (or user-attracting) models,
demand points being randomly distributed on a and allocation models (or delivery systems). In
disc, the likelihood of a demand point being customer choice, the customers decide which
optimal is 1=n, with n being the number of demand facility to patronize; a system typically appropriate
points. Furthermore, they demonstrated that the in the retail sector. On the other hand, allocation
expected deviation of the objective value at the models are appropriate when the facility planner is
least-cost demand point from the true optimum is also in charge of deciding which facility supplies a
3=n. given customer. It is quite natural in allocation
In contrast, when the objective of the minimum models that customers are allocated to their closest
weighted squared distances is differentiated with facility (except in capacitated models). However,
respect to the x and y coordinates of the facility, many customer choice models also allocate cus-
the resulting equations are separable and a simple tomers to their closest facility. Such assumption
closed-form solution exists, in which makes sense only if (1) customers engage in sepa-
P P rate trips to the facility, (2) the facility character-
wi ai w i bi
x ¼ Pi and y ¼ Pi ; istics (such as waiting times) are all the same, and
i wi i wi
(3) customers are rational (i.e., utility maximizing)
providing the well-known center-of-gravity (also planners with utility functions that include only
known as the centroid) of the given demand transportation costs, which are assumed to be
points. For an excellent exposition of the history proportional to customer-facility distances. Pre-
of continuous problems on the plane, readers are suming these assumptions are satisfied, we can
referred to Wesolowsky (1993). Another recent define a set J of facility locations, where the point
reference on planar location is due to Drezner ðxj ; yj Þ is the location of the jth facility for all j 2 J .
et al. (2002). Similarly, the single-facility location Defining uij as the proportion of customer i’s de-
problem in the plane using the sum of rectilinear mand that is served by the jth facility, the multi-
distances also has a separable objective function facility location problem can then be written as
and can be solved easily by a simple search as XX
pointed out by Vergin and Rogers (1967). Min z ¼ wi dij uij ;
i2I j2J
As far as center problems (or minimax distance X
problems) are concerned, the model with Eucli- s:t: uij ¼ 1 8i 2 I;
dean distances is typically solved by either the j2J

approach of Elzinga and Hearn (1972), or, more uij 2 j8i 2 I; j 2 J ;


efficiently, by methods from computational xi ; yj 2 j8j 2 J ;
geometry, involving farthest-point Voronoi dia-
grams. For an early, but excellent reference, see where the variables xj and yj , j 2 J appear only
Shamos and Hoey (1975). Elzinga and Hearn have implicitly in the formulation in the guise of the
also shown that the 1-center problem in the plane distances dij .
with rectilinear distances has a remarkable closed- Except for the case of rectilinear distances,
form solution that results from the reduction of a where the formulation can be solved by a linear
large linear programming formulation. The prob- programming problem, the problem is NP-hard
lem begins with a number of constraints equal to for most distance functions. Hence, most exact
four times the number of demand points and re- algorithms can only solve problems of moderate
duces by dominance to just four constraints that size. One interesting approach that uses set parti-
can be solved graphically. tioning and a form of column generation is due to
As is the case in all multi-facility location Rosing (1992). His method first enumerates all
problems, the solution of planar location problems feasible partitions of the customers, solving within
with multiple facilities requires not only the siting each of these subsets of customers a single-facility
of facilities but also the allocation of customers to problem. The outputs of these subproblems supply
C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19 7

the parameters of the larger problem which is remarkable result from his study was that when-
solved as a set partitioning problem. As the num- ever one of the demand points had more than half
ber of partitions is exponential in the number of the total system weight, the optimal solution
customers, the approach is probably best used as a placed the facility at that point. This result is
heuristic by simply considering only a subset of remarkably similar to the optimality conditions for
potentially promising or spatially reasonable par- medians in 2-dimensional space with ‘1 distances.
titions of customers. Such spatially reasonable However, no such results are available for the
partitions are easy ‘‘to see’’, because clusters that general p-median problem. In fact, Kariv and
include spatially disjoint sets of demand areas are Hakimi (1979b) demonstrated that the general p-
unlikely to be in the same partition. A well-known median problem is NP-hard. Nonetheless, invok-
heuristic is Cooper’s (1963) ‘‘alternate’’ method for ing Hakimi’s theorem, we know that it is sufficient
location–allocation. The algorithm switches back to search for the optimal solution on the set of
and forth between a siting phase and a re-alloca- nodes. Using this result, ReVelle and Swain (1970)
tion phase. Many other heuristics exist, including were the first to formulate the p-median as a zero-
greedy methods and genetic algorithms. Multi- one programming problem. A variant of their
facility minimax problems frequently use exten- original formulation is
sions of the Elzinga–Hearn procedure mentioned XX
above. Min z ¼ wi dij xij ; ð1Þ
i2I j2J

s:t:
X
5. Location problems on networks xij ¼ 1 8i 2 I; ð2Þ
j2J
Although a few papers on network-based plant xij 6 yj 8i 2 I; j 2 J ; ð3Þ
and warehouse location appeared in the opera- X
tions research literature in the 1950s, serious yj ¼ p; ð4Þ
j2J
attention to network-based problems waited until
the middle 1960s. Hakimi (1964, 1965) in a pair of xij ¼ 0 _ 1 8i 2 I; j 2 J ; ð5Þ
seminal papers investigated the minimum weighted yi ¼ 0 _ 1 8j 2 J ; ð6Þ
distance location of p facilities on a network of n
demand nodes, a problem class he termed the p- where the locational variables yj are one, if a
median problem. Though he did not provide a facility is located at node j and zero otherwise. The
solution method for the p-median problem, he allocation variables xij denote the proportion of
proved the existence of at least one optimal solu- the demand of the customer at node i that is as-
tion which has all p facilities located solely at the signed to a facility at node j. Finally, p denotes the
nodes of the network. His contribution was similar number of facilities that are to be located; the
to that of Dantzig for linear programming, in that remaining parameters are as defined above. To be
it reduced the set of optimal solutions from a sure that the optimal solution is truly found re-
potentially infinite set to a finite (albeit astro- quires that the set of facility nodes include not only
nomically finite) set. This result is commonly re- demand nodes but also ‘‘empty nodes’’, i.e., nodes
ferred to as the ‘‘Hakimi theorem’’ or ‘‘the node without demand that are simply network inter-
property’’. section points. These network intersection points
Goldman (1971) showed that 1-median prob- would not, however, be included in the set of de-
lems on trees have very simple properties some- mand nodes I as they do not require assignment.
what reminiscent of single-Weber problems with ‘1 It should be noted that it is sufficient to require
distances (a result that was independently derived that xij P 0 8i 2 I, j 2 J rather the zero–one con-
by Hua Lo-Keng et al. (1962)). He also showed straints (5), as there is at least one solution in
that 1-median problems on general graphs can be which the allocation variables are zero–one. It is
solved by a brute-force search in O(n3 ) time. One also worth mentioning that the jIjjJ j constraints
8 C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19

(3) can be aggregated to jJ j constraints of the type


P tions and the true optimal increased as well with
i2I xij 6 Myj 8j 2 J , where M is a number no the frequency of these violations. In addition, the
smaller than n  p þ 1. However, as typical in linear program achieved an all zero-one solution
integer programming, such aggregated formula- less frequently and the amount of branching and
tions would be expected to increase the solution bounding required was more extensive. Genetic
time dramatically because of the fractions intro- algorithms have also assumed their place in the
duced to the LP optimum and the subsequent need context of solving p-median problems. A promis-
for branch and bound. ing approach was reported by Bozkaya et al.
Among the exact algorithms is a procedure due (2002).
to Galv~ ao (1993) that employs Lagrangean relax- The simple plant location problem (SPLP),
ation and the linear programming relaxation of sometimes also referred to as the uncapacitated
ReVelle and Swain (1970). It is also possible to facility location problem (UFL), is similar to the p-
apply iteratively the dual descent techniques median problem, except that the number of facil-
developed for the plant location problem by Bilde ities to be located is endogenous to the problem.
and Krarup (1977) as well as Erlenkotter’s (1978) This is achieved by deleting the constraint (4) in
‘‘dualoc’’ procedure. For an insightful description, the
P above formulation, and adding the term
readers are referred to Daskin (1995). Among j2J fj yj to the objective function (1), where fj
traditional heuristic methods applied to the p- denotes the fixed costs that are incurred if a facility
median problem is the location–allocation heuristic is established at the jth node. Although Balinski
of Maranzana (1964), which is actually a co-tem- (1965) was the first to formulate the problem, he
poraneous nodal version of Cooper’s ‘‘alternate’’ abandoned the formulation because his colleague
method mentioned above. Another of the older Gomory demonstrated a small (2 · 3) counter-
but widely known heuristics is the more powerful example which terminated with fractional vari-
vertex substitution technique of Teitz and Bart ables. A good survey is provided by Cornuejols
(1968), one of the earliest of the heuristics known et al. (1990).
as ‘‘1-opt’’ for the number of items being simul- Morris (1978) and ReVelle and Swain (1970)
taneously exchanged. The simplex procedure is a have demonstrated through experiments that p-
1-opt and leads to an optimal solution always, median problems and the simple plant location
because the solution space is convex. However, the problems are ‘‘integer-friendly’’ (ReVelle, 1993);
Teitz and Bart procedure does not always con- that is, their linear programming relaxations tend
verge to the global optimum because the solution to have integer or near-integer solutions that re-
space of the p-median is not necessarily convex. A quire but a few branchings in a subsequent branch-
speedier recent adaptation of the exchange heur- and-bound procedure. Although many have
istic is due to Densham and Rushton (1992). Ro- worked on plant location solution methods; see,
sing and ReVelle (1997) incorporated the exchange e.g., Galv~ao (1989, 1993), the methodology that
heuristic in a metaheuristic, known as heuristic currently reports best results to the plant location
concentration, which, to date, has performed well, problem is that of K€ orkel (1989). The related
even against tabu search (Rosing et al., 1998, capacitated plant location problem is identical to
1999). Schilling et al. (2000) recently showed that the simple plant location problem with the addi-
the performance of both exact and heuristic algo- tion of capacity constraints on the facilities. Work
rithms on p-median problems depends on the ex- in this area seems to offer ambiguous results. These
tent to which the triangle inequality is satisfied in additional capacity constraints destroy the prop-
the network underlying those problems. The au- erty that all demand of a customer is satisfied from
thors showed that the frequency of the Teitz and a single facility. To see this, consider a simple
Bart algorithm finding the optimal solution de- scenario in which, as usual, the closest facility
creased as the number of violations of the triangle satisfies the demand of a single customer, until the
inequality increased, and the extent of the differ- supply runs out, and the remaining demand must
ence (the error) between the Teitz and Bart solu- be satisfied from a more distant facility. The loss of
C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19 9

the ‘‘single-facility property’’ makes the capacit- longest distance to any of the given nodes requires
ated plant location problem much more difficult taking the maximum over all such piecewise linear
than the corresponding simple plant location concave functions. Finally, the best solution is
problem. Experiments appear to suggest that the one that chooses the edge that minimizes the
frequency of integer solutions and the perfor- maximum. In other words, the 1-absolute center
mance of algorithms for capacitated problems are is really a min-max–max-min problem. It is no
related to the closeness of the available system surprise that the general case of p-centers is con-
capacity to the demand being distributed; see siderably more difficult. Kariv and Hakimi (1979a)
Heller (1985). have formally demonstrated that p-center prob-
ReVelle and Laporte (1996) provide a review lems on networks are NP-hard. These results hold
that exposes new and challenging problems in for both vertex centers and absolute centers.
plant location. Brimberg and ReVelle (2000) offer In some instances, particularly when emergency
a solution to one of the many problems posed facilities are to be located, neither the concept of p-
there, the maximum return on investment prob- median nor that of a p-center may be satisfactory.
lem. However, many practical plant location Instead, decision makers may wish to ‘‘cover’’
problems remain unsolved, or, indeed, unap- customers. A customer or demand node is said to
proached. An example is a problem in which be covered by a facility, if the distance or time
plants ship goods to warehouses in an out-and- between a client and its closest facility is no greater
back movement and warehouses supply demand than a prespecified value of S, the distance or time
nodes, e.g., supermarkets, with a delivery tour. standard. As an example, it may be required that
In contrast to the p-median problem and the no tract of houses in a municipality is farther than
plant location problem, both of which focus on 112 miles (or 5 minute) from the nearest fire station
aggregate access or the aggregate cost of delivery, or ambulance.
the p-center problem on the network seeks the set To illustrate the models, we again define zero–
of facilities that minimizes the maximum distance one variables xj that equal one if a facility is lo-
that separates any demand node from its nearest cated at node j and zero otherwise. We denote dji
facility. It is easy to demonstrate that solutions to as the distance between the jth facility site and the
the p-center problem typically do not have at least ith customer. Then, we can define coverage sets
one solution in which the facilities are all at nodes Ni ¼ fj : dji 6 Sg. The set Ni includes all potential
of the network. That is, centers are usually located facility locations which can dispatch to and reach
somewhere along an edge of the network. As a customer i within the standard. The location set
result, location scientists distinguish between node covering problem, or LSCP (see Toregas et al.,
centers (in which the eligible facility locations may 1971; ReVelle et al., 1976), can then be formulated
be restricted to vertices of the network) and as
absolute centers, in which the facility location X
could be anywhere on the network. It is apparent Min z ¼ xj
j2J
that a 1-node center can be found by examining all X
nodes as potential candidates. Determining a sin- s:t: xj P 1 8i 2 I;
gle absolute center is more involved. First, each j2Ni

edge (ni ; nj ) of the graph is examined, and the xj ¼ 0 _ 1 8j 2 J :


distance from a node nk to all points on the edge
are determined. Each point on (ni ; nj ) can be While the problem is formally NP-hard, large
reached from node nk either via ni of nj , resulting instances of network-based location set covering
in linear functions on the edge (ni ; nj ) with slope +1 problems have been solved easily by linear pro-
and )1, respectively. The minimum, i.e., the lower gramming and the occasional addition of solution-
envelope indicates the closest distance from a point derived cutting plane constraints to achieve all 0–1
on the edge to nk . The point on this edge that is solutions. In brief, the often successful cut that is
most distant is the maximum of this function. The added simply rounds the objective value from an
10 C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19

integer plus a fraction to the next highest integer. by Church and ReVelle (1974). Denoting weights
Toregas and ReVelle (1973) describe a set of wi as the number of customers located at node i,
reduction rules that use the principles of essential the maximum covering location problem seeks to
columns plus notions of row and column domi- locate an economically feasible number of facili-
nance. The repeated application of the reduction ties, say p, in such a way that the number of cus-
rules is often capable of reducing the size of loca- tomers covered within the standard by the
tion set covering problems dramatically and even facilities, is maximum. The reduction rules for the
solving problems ab initio. When the sequence of location set covering problem no longer apply
rules fails to reduce the problem to an answer, the easily, making the MCLP potentially more chal-
remaining matrix is termed ‘‘cyclic’’, describing the lenging to solve. Nonetheless, linear programming
relationship between variables and constraints. supplemented by branch and bound as well as
Experience with location set covering problems greedy and exchange heuristics have all been ap-
solved by both linear programming and by plied with success. A thorough review of the
reductions/essential columns suggests that those application of covering problems in location set-
problems that reduce to cyclic matrices will also tings as well as to many other situations is offered
fail to be solved without a cutting plane being by Schilling et al. (1993).
added when a linear programming solver is ap- Another possibility to get around the complete
plied. coverage requirement is to maximize the expected
A consequence of the ability of the location coverage of demand nodes, a problem that Daskin
scientist to solve the LSCP easily is that the p- (1983) formulated and solved, although this for-
center or minimax distance problem on a network mulation requires knowledge of the probability of
with a prespecified set of eligible facility sites and a a system-wide level of servers being busy. Still a
given number of facilities can also be solved effi- third possibility is to solve the p-center problem
ciently. In particular, Minieka (1970) suggested a given a number of facilities that reflects a realistic
solution of the minimax distance problem for p budget limitation. That is, in lieu of minimizing the
given facilities by solving a series of location set number of facilities needed to cover all demand
covering problems with successively decreasing nodes within the preset standard S, a realistic
distance or time standards S. As the distance number of facilities is examined to determine the
standards are tightened (i.e., decreased), the value of the smallest distance standard can be
number of required facilities will remain the same achieved for all demands with p facilities. The
for a number of values of S and then suddenly problem can also be extended into the probabilistic
increase. At the smallest distance value before the realm with chance constraints on the reliability of
number of facilities increases from p to p þ 1 (or coverage being available within a distance or time
higher), the maximum distance from any demand standard. Further, multiple types of interrelated
node to its nearest facility is, by definition, a facilities may be modeled. A recent review of
minimum––that is, the p facilities and their posi- location covering problems that apply principally
tions at this distance value minimize the maximum to emergency services is provided by Marianov
distance. and ReVelle (1995).
The location set covering problem requires that The siting of facilities through time to serve
each customer must be covered, a requirement that evolving demands (as opposed to cyclic or periodic
is very restrictive. In a problem with many spa- demands through a day or year) has been a topic
tially dispersed population nodes, the requirement of recurring interest beginning in the 1960s. Pio-
of complete coverage of all demands may produce neering work in this field was done by Ballou
solutions with a number of facilities that is unre- (1968), whose contribution focused on the private
alistic from a budgetary point of view. Several sector problem of warehouse location. By the
possibilities for problem restatement emerge 1970s, applications in discrete as well as continu-
immediately. The first is known as the maximum ous location in both plant and warehouse location
covering location problem, or MCLP, introduced as well as public facilities location were being
C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19 11

offered. Roodman and Schwarz (1975), and mands that cycle through time. The work by Re-
Wesolowsky and Truscott (1975) considered the pede and Bernardo (1994) on cyclic demands for
possibility of plants phasing in and phasing out of emergency services is one of the very few refer-
the solution set as demands evolved through time. ences on the subject. However, informal discus-
Their ideas were later picked up in public sector sions indicate that many researchers have thought
location (in particular, in set covering models) by about these problems. It is well known in emer-
Gunawardane (1982), and Chrissis et al. (1982). A gency services that demand changes in a relatively
survey of the history of contributions in this field predictable pattern through the day week, and
in the private and public sector is found in the season. In particular, calls for ambulance service
review by Owen and Daskin (1998). peak during rush traffic hours and on Friday and
The field of dynamic location to serve evolving Saturday nights, when accidents and violence in-
demands slowed considerably by the late 1980s, crease. Although it may be possible to retire some
perhaps because the field had been mined of ideas, ambulance crews during slack hours, while
or possibly because the difficulty of solving prob- engaging additional crews during peak times, it
lems with an inordinate number of variables im- seems less likely that crews will move to new
peded progress. The increase in the number of locations for the slack times. This requires that
variables occurs because both, facility variables locations of crews during rush hours must be ro-
and assignment variables, require the time sub- bust, so that their locations during non-peak hours
script. Another potential reason for the slowdown are still reasonably efficient. Another setting in
is that researchers may have realized that the which cyclic demands may be encountered is in
forecasting of demand through time is a very dif- plant or warehouse location, where demands
ficult task. The last of these reasons is suggested by change by season for goods such as air condi-
the present research focus in siting facilities to tioners, beer, and clothing. Pertinent problems are
serve demands that evolve through time. This challenging in the context of siting, operation,
current approach uses the idea of scenarios and sizing, and storage, making them highly complex.
eliminates the need for precise demand prediction, A few researchers have begun to focus on these
thus obviating, at least to some extent, the need for problems; see, e.g., Ratick et al. (1987) and Osleeb
an excessive number of variables. In particular, the and Ratick (1990).
notion of generating a number of scenarios or Closely related to dynamic location models are
snapshots of what future demand might look like problems that involve uncertainty––since many
seems to have replaced the idea of responding to aspects of uncertainty deal with the details of fu-
predicted demand through time. These scenarios ture conditions. These details become certain as
are reminiscent of the ‘‘states of nature’’ used in time passes into the present and past. The most
decision analysis, see, e.g., Eiselt and Sandblom obvious uncertainty in location models is probably
(2004). The earliest location work to take the the uncertainty of demand. The first to deal with
scenario approach to siting facilities appears to be demand uncertainty was Frank (1966). Owen
Schilling (1982). By the late 1990s, a number of and Daskin (1998) distinguish between probabi-
authors had picked up this theme, applying it to listic approaches and scenario approaches, the
competitive location as well as the p-median, see latter having been discussed above. Probabilis-
Owen and Daskin (1998) and Serra and Marianov tic methods may employ tools from mathemati-
(1999). At present, the scenario approach appears cal programming such as chance-constrained
to dominate other methods that deal with time- programming. Constraints on the probability of
evolving demands, and new applications of this servers availability within a time standard were
methodology are expected in the future. offered by ReVelle and Hogan, 1989 and are re-
In contrast to the location models cited above viewed by Marianov and ReVelle, 1995. Another
that focus on demands that grow or shrink in a approach derives from queuing, where customers
maturing pattern through time, another line of will patronize facilities not only based on their
location research has taken up the issue of de- proximity to them, but also on the (expected)
12 C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19

congestion at the facilities. Larson’s (1974) A medi-center (see, e.g., Handler, 1985) minimizes
hypercube model and Brandeau and Chiu’s (1992) the average distance, subject to a constraint that
stochastic queue center problem belong to this all customer-facility distances do not exceed a
category as do models due to Batta (1989) and prespecified limit. Both can easily be parameter-
others. It is fair to say that stochastic zero–one ized by varying the weights and bounds, respec-
programming is a challenging area and that the tively. The resulting output––the tradeoff between
rate of new articles in this area will likely mirror minimum average distance and maximum dis-
the relatively slow progress in stochastic pro- tance––gives invaluable information regarding the
gramming itself. robustness of the solution.
Despite our focus on a multiplicity of problem
settings, there is one assumption common to vir- 6.2. Undesirable facility locations
tually all but a few location models. Here, we refer
to the assumption that demand occurs only at The concept of undesirable facility locations has
specified points in the plane or nodes in a network. already been alluded to above. An immediate
It is apparent that this assumption is a convenience observation is that, as opposed to minisum and
for modeling purposes. In both, public and private minimax problems, the corresponding maxisum
facility location problems many users may have and maximin problems require a bounded set, in
been aggregated and modeled as a single demand which the location of the facility is to be chosen,
point. Such a point may represent a number of since otherwise the optimal location moves an
potential emergency calls, customers with an infinite distance from the nodes to be avoided.
anticipated demand, or clusters of voters in polit- Another important consideration is that the im-
ical models. Beginning with Current and Schilling pact of the facility or facilities on the population at
(1990), a steady stream of articles has considered a node is a cumulative one, reflecting the burden
the impact of aggregation. The most recent article placed on the node by all of the facilities, including
at the time of writing is a review by Francis et al. the nearest facility. In problems of customer access
(2002). Location scientist are well-advised to and of goods movement, the transportation of
ponder the issue of aggregation because many goods usually takes place on the least expensive
users, themselves typically not location scientists, route, making the shortest path between customer
frequently question the impact of aggregation. It is and facility a natural choice (except in the case of
mandatory that analysts are aware of the errors urban transportation when transportation may be
introduced by aggregation, if they want to see the embedded in a traffic assignment model). In con-
solutions of their models applied in practice. trast, when dealing with polluting facilities, a
customer is not just affected by the closest facility,
but typically by all of them. This has serious
6. Extensions of the basic problem structures algorithmic consequences.

In this section, we discuss some of the more 6.3. Balancing objectives


interesting ‘‘other’’ location models, focusing on
novel structures, unusual settings, and clever There are very few actual (as opposed to po-
objectives that have been developed. tential) applications of location models with bal-
ancing objectives, objectives that attempt to
6.1. Cent-Dians and Medi-Centers equalize the treatment of all customers. The main
reason appears to be the problem of measurement.
We begin with two obvious extensions, both It is not surprising that the definition of equity
designed to reconcile medians and centers. A cent- does, of course, lend itself to socio-economic and
dian (see, e.g., Halpern, 1976) is a point that even philosophical interpretations. Typical equity
minimizes a linear convex combination of minisum objectives use the variance of access distances,
(i.e., median) and minimax (i.e., center) objectives. Lorenz curve (that is typically used to measure
C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19 13

income inequality), or the related Gini index as an ing, this is known as a supply chain. Slats et al.
expression that measures the equity or inequity of (1995) and similar work focuses on the operation
a set of sites. A simple example of location on the of the supply chain as the primary consideration.
plane demonstrates that equity objectives should In contrast, location analysts are expected to focus
be used in conjunction with an efficiency objective on the design at first and the operation second.
rather than as stand-alone objectives. Consider The supply chain idea may be inverted as well.
three customers, whose locations form a triangle, Solid wastes are collected at the local level, mod-
in which one angle approaches 180. The point eled at the postal code level, perhaps, of an urban
that optimizes the equality of distances is the area, transported on small trucks to a transfer
center of the circle on whose circumference the station (perhaps to be sited) and then loaded onto
three points are located. This point can be very large trucks for transport to landfills or other final
distant from each of the three points. In moving disposal facilities, also to be sited. The system is
the distant facility towards the three customers, very similar to the industrial system cited above;
each customer gains in that the customer-facility see, i.e., Kulcar (1996). A review of solid wastes
distance decreases. However, this gain is achieved management from an OR perspective is provided
at the ‘‘expense’’ of making the three distances by Liebman (1997) who includes in his review the
more unequal. A discussion of equity objectives models which site transfer stations. As Fleisch-
can be found in Erkut (1993), Marsh and Schilling mann et al. (1997) point out, the idea of reverse
(1994) and Eiselt and Laporte (1995). An inter- logistics fits in this schema as well.
esting implementation is provided by Mandell An early work on the siting of plants and
(1991). warehouses is Kaufman et al. (1977), who de-
scribed an implicit enumeration algorithm for the
6.4. Hierarchical siting problems problem. The problem structure led to an excessive
number of variables making solution of reason-
Hierarchical location problems occur in many ably sized problems problematic. Ratick et al.
health care contexts. There are different levels of (1987) and Osleeb and Ratick (1990) focus on
health care, e.g., a doctor’s office (or maybe just a siting the facilities in a supply chain through time.
nursing station) at the lowest level, a small clinic Similarly, Bhaskaran (1992) also emphasizes the
on the next-higher level, and a full-service hospital supply chain operation in the location of inter-
at the highest level. Each facility on some level mediate transfer facilities. More recently, Bender
generally provides all the services of a facility on et al. (2002) have described the implementation of
the next-lower level. The question is then how to location software along with a GIS interface to
serve or cover as many people as possible within elucidate location decisions in supply chain man-
feasibility (e.g., cost) constraints. Two of the early agement. The research on location in a supply
references in this context are Narula (1984) and chain environment is challenging because of the
Moore and ReVelle (1982) with recent work by potential size, especially the number of variables,
Serra and ReVelle (1993, 1994a,b). A classical of the problems and the number of zero–one siting
reference of hierarchies in descriptive geography is variables.
L€osch (1944), who established locational patterns
of central places in Southern Germany which 6.5. Hub location problems
resemble honeycomb patterns typically found in
Voronoi diagrams. Hub location problems arise in contexts in which
Another type of hierarchy naturally occurs in customers require transport between a large
the industrial context, as goods move from the number of origin–destination pairs. This area of
point of extraction to manufacturing plants that research was introduced to the location literature
convert the resource into a product, followed by by O’Kelly (1986, 1987). Frequently, it is cheaper
distribution through warehouses to points of de- for the carrier, to avoid direct routing between all
mand. In the parlance of industrial decision mak- pairs between which flow occurs. Instead, hub
14 C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19

networks are assembled that require customers to (Serra and ReVelle, 1993, 1994b). The idea is
travel from their origin to a central hub, from based on the work of the German economist von
there either to their destination or to another hub, Stackelberg (1943), who was the first to consider
and if to a second hub, then to continue on to their duopolies in which one firm acts as leader of the
final destination. Hub systems are very popular in game, while the other is the follower. The earliest
the airline industry; a special issue of Location analytical location contribution in this area is by
Science (volume 4, issue 3, 1996) is devoted to such Hakimi (1983), who termed the locations of the
problems. The reviews by Bryan and O’Kelly follower medianoids. Medianoids result from con-
(1999) and Campbell et al. (2002) are compre- ditional location problems, in the sense that the
hensive. It is worth pointing out that the solution follower takes the locations of the leader as given.
of such problems does not exclusively benefit the Hakimi referred to the leader’s location as cent-
carrier: typically, traffic between most origin–des- roids in the sense that the leader of the game will
tination pairs does not justify frequent service, take the optimized reaction of the follower into
whereas the hub system allows customers to fly consideration and protect against it by using a
much more frequently to their destination, at the minimax strategy. ReVelle (1986) formulated and
expense of one or two transfers. solved the follower’s problem. He referred to it as
a maximum capture problem, and recognized cap-
6.6. Competitive location problems ture as a form of covering.
Competitive location problems may be ex-
Competitive location problems may incorporate tended to include free entry, in which additional
not only location decisions, but also pricing deci- facilities can enter the market, as long as it is
sions as variables in the model, see, for instance, profitable to do so. In these problems, each entrant
Serra and ReVelle (1999). In its simplest form, a into the market must act as leader by protecting
competitive location problem seeks the positions himself against new entrants. The complexity of
and prices which maximize the market captured these models usually requires computational sim-
(as measured in customers or profit) from previ- ulations.
ously positioned competitors. The field emerged
with Hotelling’s (1929) contribution of duopolists 6.7. Combined siting and routing
competing in a linear market––Hotelling’s famous
‘‘main street’’. The possibility of price undercut- Location–routing problems incorporate the
ting results in a large degree of instability, as wit- issues of location and routing in a single model. As
nessed by the many contributions in the literature; an example, Bouliane and Laporte (1992) consider
for surveys see, Eiselt et al. (1993), Serra and the simultaneous location of postal relay boxes
ReVelle (1995) and Plastria (2001). The contribu- and the routing of postal trucks for deliveries and
tions in the economics literature typically deal with pickups. Perl and Daskin (1985) make an early
the existence of Nash equilibria, as derived using contribution to this area. Clearly, as just about all
the tools of game theory. Such an equilibrium can known routing problems are NP-hard, the addi-
loosely be stated as a solution, in which neither of tion of a location component does not make them
the competitors has an incentive to unilaterally any easier to solve.
change the current solution. Research in the area We also need to mention p-dispersion problems
has revealed that Hotelling-models are inherently which do not include any customers, but attempt
unstable, in the sense that even seemingly minor to locate facilities, so as to maximize the shortest
changes of location rules or parameters can not distance between any facility and its nearest
only change the solution, but can also destroy or neighbor. Typical applications for this type of
establish Nash equilibria. problem are found in the military context, where
Another approach is sequential competitive the decision maker attempts to locate, say,
location, in which a leader in the (location or ammunition dumps, so as to ensure that the dis-
pricing) game is followed by a follower in the game covery and subsequent destruction of one such
C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19 15

facility has as little impact as possible on the other perspective, when it comes to applications, there
facilities. An early reference is Shier (1977). Similar appears to be a significant deficit, at least as
problems are p-defense problems which maximize compared to other, similar, fields. One reason may
the sum of distances between facilities closest to be the fact that many of the practical location
each other. For details, readers are referred to problems faced by executives and public sector
Moon and Chaudry (1984). decision makers are strategic problems, having
multiple objectives and structures that do not fall
6.8. Capture problems into the relatively easy categories outlined above.
In other words, the main obstacles to the appli-
So far, we have assumed that demand in loca- cation of the many models that are available to-
tion–allocation models occurs at points of a net- day, could be problems of measurement, finding
work, and that we serve this demand through consensus among groups of decision makers, and
point-to-point travel. For many types of service, ‘‘selling’’ agreed-upon location choices to the
service stations, fast food outlets, and automatic general public. In their insightful contribution
teller machines, customer demand does not occur regarding the location of hazardous materials,
entirely at points (e.g., the residence of the cus- Kleindorfer and Kunreuther (1994) describe ways
tomer), but along a path, e.g., a customer’s daily of dealing with the planner–public interface.
route to work. The idea is now to capture or One possibility to overcome this applications
intercept as much potential demand as possible. deficit is to build more comprehensive integrated
For that purpose, Hodgson (1990) introduced the models, at the expense of an exponentially in-
flow capturing location-allocation model (FCLM) creased complexity of the models and required
which serves such demands by locating facilities to algorithms. Other, possibly more promising, ap-
intercept them en route. He indicated that the proaches might employ fairly simple subproblems
model has the structure of the MCLP, and stressed with structures that fit the above categories within
the need to avoid flow cannibalization, or double the context of decision support systems. Coupled
counting. The model has been adapted to vehicle with extensive sensitivity analyses, this approach
inspection stations (Hodgson et al., 1996) and appears to have potential. However, falling back
billboards (Hodgson and Berman, 1997). Berman onto a one-size-fits-all general mixed integer pro-
et al. (1995) review these models. From a technical gramming problem may be required in the case of
point of view, one of the important issues in more complex scenarios.
flow capturing (also known as flow interception) Another recently developed tool may help
models is the potential for double counting of the overcome the dearth of applications of location
capture. In order to avoid this, it is necessary to models. Here, we are referring to spatial display
define additional zero–one variables that indicate tools that are, in part, designed to overcome the
whether or not an individual who has taken a communications barrier between location analysts
particular route, has or has not been exposed to at and locators. Geographic information systems
least one of the facilities. (GIS), especially when coupled with location
models and efficient algorithms, have the ability to
vividly display the results of the user input and,
7. Conclusion and future prospects most importantly, the changes that result from
input modifications. This display tool, coupled
In this paper we have surveyed a number of the with techniques designed to deal with fuzzy input
important problems in facility location. The field is from multiple decision makers may, after all, result
very active with many interesting problems still in many more location models being applied by
being investigated, both from a problem state- practitioners.
ment/formulation standpoint and from an algo- Communication barriers aside, we also know
rithmic point of view; for details, see ReVelle that university professors are usually given
(1997). Although the field is active from a research little incentive to apply their methodologies.
16 C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19

Publications in journals are the currency of pro- Brandeau, M.L., Chiu, S.S., 1992. A center location problem
motion cases, and applications achieve little cred- with congestion. Annals of Operations Research 40, 17–32.
Brimberg, J., Love, R.F., 1995. Estimating distances. In:
ibility unless they include some theoretical content. Drezner, Z. (Ed.), Facility Location: A Survey of Applica-
The reversal of this longstanding pattern seems tions and Methods. Springer-Verlag, Berlin, pp. 9–32.
difficult unless some level of government provides Brimberg, J., ReVelle, C., 2000. The maximum return on
financial incentives for transferring the technology investment plant location problem. Journal of the Opera-
of location or unless industrial sponsorship is tional Research Society 51, 729–735.
Bryan, D., O’Kelly, M., 1999. Hub-and-spoke networks in air
offered; see ReVelle (1997). transportation: An analytical review. Journal of Regional
Science 39, 275–295.
Campbell, J.F., Ernst, A.T., Krishnamoorthy, M., 2002. Hub
location problems. In: Drezner, Z., Hamacher, H.W. (Eds.),
Acknowledgements Facility Location: Applications and Theory. Springer-Ver-
lag, Berlin, pp. 373–407.
This research was, in part, supported by a grant Chrissis, J., Davis, R., Miller, D., 1982. The dynamic set
from the Natural Sciences and Engineering covering problem. Applied Mathematical Modelling 6, 2–6.
Cooper, L., 1963. Location–allocation problems. Operations
Council of Canada. This support is gratefully
Research 11, 331–343.
acknowledged. The authors wish to express their Cornuejols, G., Nemhauser, G.L., Wolsey, L.A., 1990. The
appreciation to an anonymous referee for a num- uncapacitated facility location problem. In: Mirchandani,
ber of valuable comments that helped to clarify a P.B., Francis, R.L. (Eds.), Discrete Location Theory. Wiley-
variety of issues. Interscience, New York, pp. 119–171.
Church, R.L., ReVelle, C., 1974. The maximal covering
location problem. Papers of the Regional Science Associa-
tion 32, 101–118.
References Current, J., Daskin, M., Schilling, D., 2002. Discrete network
location models. In: Drezner, Z., Hamacher, H. (Eds.),
Balinski, M.L., 1965. Integer programming: Methods, uses, Facility Location: Applications and Theory. Springer-Ver-
computation. Management Science 12, 253–313. lag, Berlin, pp. 81–118.
Ballou, R., 1968. Dynamic warehouse location analysis. Jour- Current, J., Schilling, D., 1990. Analysis of errors due to
nal of Marketing Research 5, 271–276. demand data aggregation in the set covering and maximal
Batta, R., 1989. A queueing-location model with expected covering location-problems. Geographical Analysis 22, 116–
service time-dependent queueing disciplines. European 126.
Journal of Operational Research 39, 192–205. Daskin, M.L., 1983. Maximal expected covering location
Bender, T., Hennes, H., Kacsics, J., Melo, M., Nickel, S., 2002. model: Formulation, properties, and heuristic solution.
Location software and interface with GIS and supply chain Transportation Science 17, 48–70.
management. In: Drezner, Z., Hamacher, H. (Eds.), Facility Daskin, M.S., 1995. Network and Discrete Location: Models,
Location: Applications and Theory. Springer-Verlag, Ber- Algorithms, and Applications. Wiley & Sons, New York.
lin. Densham, P., Rushton, G., 1992. A more efficient heuristic for
Berman, O., Hodgson, M.J., Krass, D., 1995. Flow intercepting solving large p-median problems. Papers in Regional
models. In: Drezner, Z. (Ed.), Facility Location: A Survey Science 71, 307–329.
of Applications and Methods. Springer-Verlag, pp. 427– Dijkstra, E.W., 1959. A note on two problems in connection
452. with graphs. Numerische Mathematik 1, 269–271.
Bhaskaran, S., 1992. Identification of transshipment center Domschke, W., Drexl, A., 1985. In: Location and Layout
locations. European Journal of Operational Research 63, Planning: An International Bibliography. Volume 238 of
141–150. Lecture Notes in Economics and Mathematical Systems.
Bilde, O., Krarup, J., 1977. Sharp lower bounds and efficient Springer-Verlag, Berlin.
algorithms for the simple plant location problem. Annals of Drezner, Z., 1992. A note on the Weber location problem.
Discrete Mathematics 1, 79–97. Annals of Operations Research 40, 153–161.
Bouliane, J., Laporte, G., 1992. Locating postal relay boxes Drezner, Z., Simchi-Levi, D., 1992. Asymptotic behavior of the
using a set covering algorithm. American Journal of Weber location problem on the plane. Annals of Operations
Mathematical and Management Sciences 12, 65–74. Research 40, 163–172.
Bozkaya, B., Zhang, J., Erkut, E., 2002. An efficient genetic Drezner, Z., Klamroth, K., Sch€ obel, A., Wesolowsky, G.O.,
algorithm for the p-median problem. In: Drezner, Z., 2002. The Weber problem. In: Drezner, Z., Hamacher, H.
Hamacher, H. (Eds.), Facility Location: Applications and (Eds.), Facility Location: Applications and Theory.
Theory. Springer-Verlag, Berlin, pp. 179–205. Springer-Verlag, Berlin, pp. 1–36.
C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19 17

Eiselt, H.A., 1992. Location Modeling in Practice. American Hakimi, S.L., 1965. Optimal distribution of switching centers in
Journal of Mathematical and Management Sciences 12, 3– a communication network and some related theoretic graph
18. theoretic problems. Operations Research 13, 462–475.
Eiselt, H.A., Laporte, G., 1995. Objectives in location prob- Hakimi, S.L., 1983. On locating new facilities in a competitive
lems. In: Drezner, Z. (Ed.), Facility Location: A Survey of environment. European Journal of Operational Research
Applications and Methods. Springer-Verlag, Berlin, pp. 12, 29–35.
151–180. Handler, G.Y., 1985. Medi-centers of a tree. Transportation
Eiselt, H.A., Sandblom, C.-L., 2004. Decision Analysis, Loca- Science 19, 246–260.
tion Models, and Scheduling Problems. Springer-Verlag, Harsanyi, J.C., 1975. Can the maximin principle serve as a basis
Berlin. for morality? A critique of John Rawls’s theory. American
Eiselt, H.A., Laporte, G., Thisse, J.-F., 1993. Competitive Political Science Review 69, 594–606.
location models: A framework and bibliography. Transpor- Heller, M., 1985. Location optimization and simulation for the
tation Science 27, 44–54. analysis of emergency medical service systems. Doctoral
Elzinga, J., Hearn, D.W., 1972. Geometrical solutions for some Dissertation, Johns Hopkins University, Baltimore, MD, p.
minimax location problems. Transportation Science 6, 379– 132.
394. Hodgson, M.J., 1990. A flow-capturing location–allocation
Erkut, E., 1993. Inequality measures for location problems. model. Geographical Analysis 22, 270–279.
Location Science 1, 199–217. Hodgson, M.J., Berman, O., 1997. A billboard location
Erkut, E., Newman, S., 1989. Analytical models for locating model. Geographical and Environmental Modelling 1, 25–
undesirable facilities. European Journal of Operational 45.
Research 40, 275–279. Hodgson, M.J., Rosing, K.E., Zhang, J., 1996. Locating vehicle
Erlenkotter, D., 1978. A dual-based procedure for uncapaci- inspection stations to protect a transportation network.
tated facility location. Operations Research 26, 992–1009. Geographical Analysis 28, 299–314.
Eyster, J.W., White, J.A., Wierwille, W.W., 1973. On solving Hogan, K., 1990. Reducing errors in rainfall estimates through
multifacility location problems using a hyperboloid approx- rain gauge location. Geographical Analysis 22, 33–49.
imation procedure. AIIE Transactions 5, 1–6. Hopmans, A.C.M., 1986. A spatial interaction model for
Fleischmann, M., Bloemhof-Ruwaard, J.M., Dekker, R., van branch bank accounts. European Journal of Operational
der Laan, E., van Nunen, J.A.E.E., Van Wassenhove, Research 27, 242–250.
L.N., 1997. Quantitative models for reverse logistics: A Hotelling, H., 1929. Stability in Competition. Economics
review. European Journal of Operational Research 103, 1– Journal 39, 41–57.
17. Hua, L.-K. et al., 1962. Application of mathematical methods
Francis, R., Lowe, T., Tamir, A., 2002. Demand point to wheat harvesting. Chinese Mathematics 2, 77–91.
aggregation for location models. In: Drezner, Z., Ham- Huxley, S.J., 1982. Finding the right spot for a church camp in
acher, H. (Eds.), Facility Location; Applications and Spain. Interfaces 12, 108–114.
Theory, pp. 207–232. Jacobsen, S.K., Madsen, O.B.G., 1980. A comparative study
Frank, H., 1966. Optimum locations on a graph with proba- of heuristics for a two-level routing–location problem.
bilistic demand. Operations Research 14, 409–421. European Journal of Operational Research 5, 278–
Gabzewicz, J.J., Thisse, J.-F., 1989. Spatial competition and 287.
location of firms. Fundamentals of Pure and Applied Kariv, O., Hakimi, S.L., 1979a. An algorithmic approach to
Economics 5, 2–37. network location problems. Part I: The p-centers. SIAM
Galv~ao, R., 1989. A method for solving optimally uncapaci- Journal of Applied Mathematics 37, 513–538.
tated plant location problems. Annals of Operations Kariv, O., Hakimi, S.L., 1979b. An algorithmic approach to
Research 18, 225–244. network location problems. Part II: The p-median. SIAM
Galv~ao, R., 1993. The use of Lagrangean relaxation in the Journal of Applied Mathematics 37, 539–560.
solution of uncapacitated facility location problems. Loca- Kaufman, L., Eede, M., Hansen, P., 1977. A plant and
tion Science 1, 57–79. warehouse location problem. Operations Research Quar-
Goldman, A.J., 1971. Optimal center location in simple terly 28, 547–554.
networks. Transportation Science 5, 212–221. Kimes, S.E., Fitzsimmons, J.A., 1990. Selecting profitable hotel
Gunawardane, G., 1982. Dynamic versions of set covering type sites at La Quinta Inns. Interfaces 20, 12–20.
public facility location problems. European Journal of Kleindorfer, P.R., Kunreuther, H.C., 1994. Siting of hazardous
Operational Research 10, 190–195. facilities. In: Pollock, S.M. et al. (Eds.), Handbooks in OR
Halpern, J., 1976. The location of a center–median convex and MS. Elsevier Science, pp. 403–440.
combination on an undirected tree. Journal of Regional K€orkel, M., 1989. On the exact solution of large-scale simple
Science 16, 237–245. plant location problems. European Journal of Operational
Hakimi, S.L., 1964. Optimal locations of switching centers and Research 39, 157–173.
the absolute centers and medians of a graph. Operations Kuhn, H., 1973. A note on Fermat’s problem. Mathematical
Research 12, 450–459. Programming 4, 98–107.
18 C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19

Kuhn, K., Kuenne, R., 1962. An efficient algorithm for the O’Kelly, M., 1987. A quadratic integer program for the location
numerical solution of the generalized Weber problem in of interacting hub facilities. European Journal of Opera-
spatial economics. Journal of Regional Science 4, 21–33. tional Research 32, 393–404.
Kulcar, T., 1996. Optimizing solid waste collection in Brussels. Osleeb, J., Ratick, S., 1990. A dynamic location–allocation
European Journal of Operational Research 90, 71–77. model for evaluating the spatial impacts for just-in-time
Larson, R.C., 1974. A hypercube queuing model for facility planning. Geographical Analysis 22, 50–69.
location and redistricting in urban emergency services. Owen, S., Daskin, M., 1998. Strategic facility location: A
Computers and Operations Research 1, 67–95. review. European Journal of Operational Research 111,
Liebman, J., 1997. Solid waste management. In: Revelle, C., 423–447.
McGavity, A. (Eds.), Design and Operation of Civil and Perl, J., Daskin, M., 1985. A warehouse location-routing
Environmental Engineering Systems, Chapter 5. Wiley, problem. Transportation Research B––Methods 19, 381–
New York, pp. 205–232. 396.
L€
osch, A., 1944. Die r€aumliche Ordnung der Wirtschaft, G. Plastria, F., 1995. Continuous location problems. In: Drezner,
Fischer Verlag, Jena, Germany (English translation: The Z. (Ed.), Facility Location: A Survey of Applications and
Economics of Location, translated by Woglom, W.H. and Methods. Springer-Verlag, Berlin, pp. 225–262.
Stolper, W.F., Yale University Press, New Haven, CT, Plastria, F., 2001. Static competitive facility location: An
1954). overview of optimization approaches. European Journal
Love, R.F., Morris, J.G., Wesolowsky, G.O., 1988. Facilities of Operational Research 129, 461–470.
Location: Models and Methods. North Holland, New Ratick, S., Osleeb, J., Kuby, M., Lee, K., 1987. Interperiod
York. network storage location–allocation (INSLA) models. In:
Mandell, M., 1991. Modeling effectiveness–equity tradeoffs in Ghosh, A., Rushton, G. (Eds.), Spatial Analysis and
public service delivery systems. Management Science 37, Location–Allocation Models. Van Nostrand Rheinhold,
467–482. New York, pp. 269–301.
Maranzana, F., 1964. On the location of supply points to Rawls, J., 1971. A Theory of Justice. Harvard University Press,
minimize transport costs. Operations Research Quarterly Cambridge, MA.
15, 261–270. Repede, J., Bernardo, J., 1994. Developing and validating a
Marianov, V., ReVelle, C., 1995. Siting emergency services. In: decision-support system for locating emergency medical
Zvi Drezner, (Ed.), Facility Location. Springer-Verlag, pp. vehicles in Louisville, Kentucky. European Journal of
203–227 (Chapter 10). Operational Research 75, 567–581.
Marks, D., Liebman, J., 1971. Location models: Solid wastes ReVelle, C., 1986. The maximum capture or sphere of influence
collection examples, Journal of the Urban Planning and location problem––Hotelling revisited on a network. Jour-
Development Division, Proceedings of the American Soci- nal of Regional Science 26, 343–358.
ety of Civil Engineers, vol. 97, No. UP1, April, pp. 15–30. ReVelle, C., 1993. Facility siting and integer friendly program-
Marsh, M.T., Schilling, D.A., 1994. Equity measurement in ming. European Journal of Operational Research 65, 147–
facility location analysis: A review and framework. Euro- 158.
pean Journal of Operational Research 74, 1–17. ReVelle, C., 1997. A perspective on location science, invited
Miehle, W., 1958. Link-length minimization in networks. review for Location Science 5, 3–13.
Operations Research 6, 232–243. ReVelle, C., Hogan, K., 1989. The maximum availability
Minieka, E., 1970. The m-center problem. Siam Review 12, location problem. Transportation Science 23, 192–
138–141. 200.
Moon, I.D., Chaudry, S.S., 1984. An analysis of network ReVelle, C., Laporte, G., 1996. The plant location problem:
location problems with distance constraints. Management New models and research prospects. Operations Research
Science 30, 290–307. 44, 864–874.
Moore, G., ReVelle, C., 1982. The hierarchical service location ReVelle, C.S., Swain, R.W., 1970. Central facilities location.
problem. Management Science 28, 775–780. Geographical Analysis 2, 30–42.
Morris, J.G., 1978. On the extent to which certain fixed-charge ReVelle, C., Marks, D., Liebman, J., 1970. Analysis of private
depot location problems can be solved by LP. Journal of the and public sector location problems. Management Science
Operational Research Society 29, 71–76. 16, 692–707.
Narula, S.C., 1984. Hierarchical location–allocation problems: ReVelle, C., Toregas, C., Falkson, L., 1976. Applications of the
A classification scheme. European Journal of Operational location set covering problem. Geographical Analysis 8, 67–
Research 15, 183–189. 76.
Niemi, R.G., Weisberg, H.F. (Eds.), 1976. Controversies in Rey, M., Soriano, P., Stampfli, E., 1995. La localization d’une
American Voting Behavior. W.H. Freeman & Sons, San installation de stockage de dechets stabilizes: Le cas de la
Francisco, CA. Suisse romande. INFOR 33, 50–62.
O’Kelly, M., 1986. The location of interacting hub facilities. Roodman, G., Schwarz, L., 1975. Optimal and heuristic facility
Transportation Science 20, 92–106. phase out strategies. AIIE Transactions 7, 177–184.
C.S. ReVelle, H.A. Eiselt / European Journal of Operational Research 165 (2005) 1–19 19

Rosing, K.E., 1992. An optimal method for solving the Shier, D.R., 1977. A min–max theorem for p-center problems
(generalized) multi-Weber problem. European Journal of on a tree. Transportation Science 11, 243–252.
Operational Research 58, 414–426. Slats, P., Bhola, B., Evers, J., Dijkhuizen, G., 1995. Logistic
Rosing, K., ReVelle, C., 1997. Heuristic concentration: Two chain modeling. European Journal of Operational Research
stage solution construction. European Journal of Opera- 87, 1–20.
tional Research 97, 75–86. Teitz, M., Bart, P., 1968. Heuristic methods for estimating the
Rosing, K., ReVelle, C., Rolland, E., Schilling, D., Current, J., generalized vertex median of a weighted graph. Operations
1998. Heuristic concentration and Tabu search: A head to Research 16, 955–961.
head comparison. European Journal of Operational Re- Toregas, C., Swain, R., ReVelle, C., Bergman, L., 1971. The
search 104, 93–99. Location of emergency service facilities. Operations Re-
Rosing, K., ReVelle, C., Schilling, D., 1999. A gamma heuristic search 19, 1363–1373.
for the p-median problem. European Journal of Operational Toregas, C., ReVelle, C., 1973. Binary logic solutions to a class
Research 117, 522–532. of location problem. Geographical Analysis 5, 145–155.
Schilling, D., 1982. Strategic facility planning: The analysis of Tryfos, P., 1986. A integer programming approach to the
options. Decision Sciences 13, 714–724. apparel sizing problem. Journal of the Operational Re-
Schilling, D., Jayaraman, V., Barkhi, R., 1993. A review of search Society 37, 1001–1006.
covering problems in facility location. Location Science 1, Vasko, F., Wolf, F., Stott, K., 1987. Optimal selection of ingot
22–55. sizes via set covering. Operations Research 35, 346–
Schilling, D., Rosing, K., ReVelle, C., 2000. Network distance 353.
characteristics that affect computational effort in the p- Vergin, R.C., Rogers, J.D., 1967. An algorithm and computa-
median problem. European Journal of Operational Re- tional procedure for locating economic facilities. Manage-
search 127, 525–536. ment Science 13, B240–B254.
Serra, D., Marianov, V., 1999. The p-median problem in a von Stackelberg, H., 1943. Grundlagen der theoretischen
changing network. Location Science 4, 383–394. Volkswirtschaftslehre., W. Hodge & Co. Ltd, London
Serra, D., ReVelle, C., 1995. Competitive location in discrete (translated as The Theory of the Market Economy).
space. In: Zvi Drezner, (Ed.), Facility Location. Springer- €
Weber, A., 1909. Uber den Standort der Industrien, T€ubingen
Verlag, pp. 337–356 (Chapter 15). (translated by C.J. Friedrich as Theory of the Location of
Serra, D., ReVelle, C., 1993. The PQ-median problem: Loca- Industries, University of Chicago Press, Chicago, IL).
tion and districting of hierarchical facilities: Part I: An exact Weiszfeld, E., 1937. Sur le point pour lequel la somme des
solution method. Location Science 1, 299–312. distances de n points donnees est minimum. Tohoku
Serra, D., ReVelle, C., 1994a. The PQ-median problem: Mathematical Journal 43, 355–386.
Location and districting of hierarchical facilities: Part II, Wesolowsky, G.O., 1993. The Weber problem: History and
heuristic solution methods. Location Science 2, 63–82. procedures. Location Science 1, 5–23.
Serra, D., ReVelle, C., 1994b. Market Capture by 2 compet- Wesolowsky, G., Truscott, W., 1975. The multi-period loca-
itors––the preemptive location problem. Journal of Regio- tion–allocation problem with relocation of facilities. Man-
nal Science 34, 549–561. agement Science 22, 57–65.
Serra, D., ReVelle, C., 1999. Competitive location and pricing Wirasinghe, S.C., Waters, N.M., 1983. An approximate proce-
on networks. Geographical Analysis 31, 109–129. dure for determining the number, capacities, and locations
Shamos, M.I., Hoey, D., 1975. Closest point problems. In: of solid waste transfer stations in an urban region. Euro-
Proceedings of the 16th Annual Symposium on Founda- pean Journal of Operational Research 12, 105–
tions of Computer Science, pp. 151–162. 111.

You might also like