A Fuzzy Multi-Objective Covering-Based Vehicle Location Model

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 22

Computers & Operations Research 34 (2007) 705 – 726

www.elsevier.com/locate/cor

A fuzzy multi-objective covering-based vehicle location model


for emergency services
Ceyhun Araz∗ , Hasan Selim, Irem Ozkarahan
Department of Industrial Engineering, Dokuz Eylul University, Izmir, Turkey
Available online 27 April 2005

Abstract
Timeliness is one of the most important objectives that reflect the quality of emergency services such as ambulance
and firefighting systems. To provide timeliness, system administrators may increase the number of service vehicles
available. Unfortunately, increasing the number of vehicles is generally impossible due to capital constraints. In
such a case, the efficient deployment of emergency service vehicles becomes a crucial issue. In this paper, a
multi-objective covering-based emergency vehicle location model is proposed. The objectives considered in the
model are maximization of the population covered by one vehicle, maximization of the population with backup
coverage and increasing the service level by minimizing the total travel distance from locations at a distance bigger
than a prespecified distance standard for all zones. Model applications with different solution approaches such as
lexicographic linear programming and fuzzy goal programming (FGP) are provided through numerical illustrations
to demonstrate the applicability of the model. Numerical results indicate that the model generates satisfactory
solutions at an acceptable achievement level of desired goals.
Scope and purpose This paper considers the emergency service vehicles location problem. A multi-objective max-
imal covering location model is proposed in this paper. The model addresses the issue of determining the best
base locations for a limited number of vehicles so that the service level objectives are optimized. Three of the
important surrogates that reflect the quality of emergency service systems are considered as objectives in the model:
maximization of the population covered by one vehicle, maximization of the population with backup coverage and
minimization of the total travel distance from locations at a distance bigger than a prespecified distance standard
for all zones. The proposed model allows the incorporation of decision maker’s imprecise aspiration levels for the
goals by means of FGP approach. Thus, the solution efficiency inherent in the FGP approaches is also included into

∗ Corresponding author. Tel.: +90 232 3881047; fax: +90 232 3887869.
E-mail addresses: ceyhun.araz@deu.edu.tr (C. Araz), hasan.selim@deu.edu.tr (H. Selim), irem.ozkarahan@deu.edu.tr
(I. Ozkarahan).

0305-0548/$ - see front matter 䉷 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cor.2005.03.021
706 C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726

the model. To demonstrate the applicability of the model, numerical examples are provided using different solution
approaches.
䉷 2005 Elsevier Ltd. All rights reserved.

Keywords: Emergency vehicle location; Fuzzy goal programming; Multi-objective decision making

1. Introduction

Emergency services such as ambulance systems and fire departments must provide high service level
to ensure public safety. These services are typically provided by vehicles based at fixed locations. The
number and placement of vehicles have a strong influence on the quality of service. Increasing the
number of vehicles is generally limited due to capital constraints. In this case, the efficient deployment
of emergency service vehicles becomes a crucial issue.
The emergency service vehicle location problem is to determine the best base locations for vehicles so
that some service level objectives are optimized. Timeliness is one of the most important objectives that
reflect the quality of emergency service systems. Therefore, emergency service vehicles must be located
in such a way that they may reach to any demand point within a reasonable response time.
Timeliness can be measured in many ways such as [1]:

• minimization of the total or average time to serve all emergency calls;


• minimization of the maximum travel time to any single call;
• maximization of area coverage (ensures that as many zones in the area as possible is covered within
S minutes of travel);
• maximization of call coverage (ensures that as many calls in the area as possible is covered within S
minutes of travel).

These measures are actually surrogates for the true objective of reducing as much morbidity and
mortality as possible. Besides these objectives, cost minimization, maximization of coverage equity,
labor equity, back-up coverage and service level to the uncovered zones can be used as objectives of
emergency vehicle deployment systems.
Mathematical programming is one of the most frequently used approaches in order to solve the emer-
gency service vehicle location problems. More specifically, the maximal covering location problem
(MCLP) [2] has proved to be one of the most useful facility location models from both theoretical
and practical points of view (see e.g. [3–6]). Many of these models include multiple objectives usually in
conflict. These multi-objective models can be solved by using mathematical approaches such as weighting
method, goal programming (GP) and compromise programming.
GP is a well-known and widely used approach to solve various multi-objective problems. The GP
approach allows the decision maker (DM) to define satisfying levels of the value of each objective (the
goals) and then to find a solution which optimizes unfavorable deviations from those goals.
In real life, high degree of fuzziness and uncertainties exist in decision-making processes. Fuzzy set
theory (FST), introduced by Zadeh [7], provides an appropriate framework to describe and treat these
uncertainties. Additionally, FST helps to improve crisp models and provides more robust and flexible
models for real-world complex decision problems, especially those involving human aspects. In the
C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726 707

last decade, many fuzzy programming techniques such as fuzzy goal programming (FGP), interactive
fuzzy multi-objective decision making have been developed for solving multi-objective decision-making
problems. Applying FST into GP has the advantage of better representation of real situations in which
DM is not able to exactly establish the target value associated with each objective.
In this paper, a multi-objective covering-based vehicle location model for emergency services is pro-
posed. The objectives considered in the model are maximization of the population covered by one vehicle,
maximization of the population with backup coverage and increasing the service level by minimizing
the total travel distance from locations at a distance bigger than a prespecified distance standard for all
zones. In order to incorporate the DMs imprecise aspiration levels for the goals FGP approaches are used.
The proposed model inherits its formulation from the previous researches in maximal covering models.
The model is formed by using Hogan and ReVelle’s [3] maximal backup coverage model and Pirkul and
Schilling’s [4] capacitated maximal covering model as the base.
The paper is further organized as follows. Section 2 presents a brief literature review of the existing
approaches and the models related to the emergency service vehicle location. In Section 3, FGP approach
is described. Section 4 is devoted to the explanation of previous works of the proposed model. In Section
5, the proposed model is presented in detail. Section 6 includes an illustration of the proposed model to
demonstrate the effectiveness of the solution approaches used. In Section 7, computational results are
discussed. Finally, concluding remarks are presented in the last section.

2. Literature review

A vast modeling literature has been developed to assist planners in making location decisions of
emergency service vehicles. The models in the literature may be classified into two broad groups [8]:
those concerned with covering demands within a specified response time standard which are called
covering-based, and those aiming at minimizing the system-wide average response time which are called
p-median based.
The historical starting point of deterministic location models is the p-median problem. The aim in this
problem is to locate p facilities on a network of demand zones so that the average travel time to the zones
can be minimized. In solving p-median problems, it was seen that response times were unacceptably long
for some zones of the region. To get around this problem the notion of demand covering was developed.
Here, coverage means the presence of one vehicle at least within the predetermined distance or time
standard.
The basic coverage models began with the formulation of Toregas et al. [9] who developed the set
covering location problem (SCLP). Their model sought to minimize the number of vehicles needed to
cover all demand zones. Church and ReVelle [2] extended the SCLP to deal with the situation in which
the number of vehicles available is less than the number needed to cover all demand zones. This model
sought to cover the maximum population within the specified distance standard given a limited number of
vehicles available. The model is referred to as the MCLP and can result in some zones that are not covered.
In both SCLP and MCLP, demand is assumed to be generated at centers of zones only and it is said
that demand generated at zone i is covered by a facility at zone j if dij  S, where dij is the distance or
time from zone i to j, and S is the prespecified distance or time standard.
The MCLP has been a subject of considerable interest in the literature since its debut. The original
formulation of the MCLP has been modified and extended many times to meet the specific requirements
708 C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726

of various location problems. Schilling et al. [10] extended the maximal covering model by considering
two types of demand with a different priority in each zone. Maximizing the coverage of the highest
priority calls and next lower priority calls are the objectives in their formulation.
Hogan and ReVelle [3] suggested that backup coverage can be used as a decision criterion for location
of emergency service vehicles on a network. They describe backup coverage as the second coverage of
a demand zone, and express that it may be required in high-demand areas to maintain a uniform level of
service when vehicles can respond to only one call at a time. They applied backup coverage criterion to
SCLP and MCLP models. First coverage was traded off against backup coverage in their models. It was
shown that backup coverage can be provided without substantial first coverage loss.
Daskin et al. [8] examined the integration of different covering models such as multiple, excess,
backup and expected covering models. Pirkul and Schilling [11] modeled the objective of maximizing
covered calls subject to limits on the number of calls that each vehicle can answer and a
requirement for backup coverage. Pirkul and Schilling [4] were concerned with the addition of work-
load limits on the facilities and the quality of service delivered to the uncovered demand zones. They
suggested that the quality of service is controlled by the distance between facility locations and de-
mand zones. When uncovered demand must be assigned to facilities workload, their model assigns
those zones to the nearest facility locations with regard to the capacity constraints. Narasimhan
et al. [12] extended the model to include the objective of maximizing covered calls of different service
levels.
ReVelle et al. [5] considered several extensions of maximal conditional covering problem. In these
models, the facility locations are supposed to be covered by other facilities and may not be used to
cover their own zones. They discussed possible applications in a variety of emergency systems such as
emergency medical services and fire equipment services.
Berman and Krass [13] presented the generalized maximal cover location problem which allows for
partial coverage. They defined the degree of coverage as a decreasing step function of the distance to
the closest facility. They showed that their problem is equivalent to the uncapacitated facility location
problem.
Karasakal and Karasakal [14] introduced a notion, namely partial coverage, which is defined as a
function of distance of the demand point to the facility. They formulated the MCLP in the presence of
partial coverage. A solution algorithm for large problems was developed by using Lagrangean relaxation.
The reader is referred to Goldberg [1] for a review of literature related to deployment and planning
analysis pertaining to emergency medical services and fire departments, and to Marianov and ReVelle
[15] for a discussion on the problem and applications.

3. Fuzzy goal programming

Goal programming is one of the most powerful, multi-objective decision-making approaches in practi-
cal decision making. GP was originally proposed by Charnes et al. [16]. This method requires the DM to
set goals for each objective that he/she wishes to attain. GP solution technique focuses on the minimiza-
tion of the deviations from each goal, subject to the goal constraints and system constraints. In a standard
GP formulation, goals and constraints are defined precisely. However, one of the major drawbacks for a
DM in using GP is to determine precisely the goal value of each objective function.
C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726 709

Applying FST into GP has the advantage of allowing for the vague aspirations of a DM, which can
then be qualified by some natural language terms. When vague information related to the objectives are
present then the problem can be formulated as a FGP problem.
The FST in GP was first considered by Narasimhan [17]. Narasimhan and Rubin [18], Hannan [19]
and Tiwari et al. [20,21] extended the fuzzy theory to the field of GP. The works of, Ramik [22], Rao et
al. [23], Wang and Fu [24], Mohamed [25], Ohta and Yamaguchi [26], El-Wahed and Abo-Sinna [27],
Mohammed [28] have investigated various aspects of decision problems using FGP theoretically.
Zadeh [7] introduced FST that is a generalization of conventional set theory as a mathematical way
to represent vagueness in everyday life. A fuzzy set A can be characterized by a membership function,
usually denoted by , which assigns to each object of a domain its grade of membership in A. The nearer
the value of membership function to unity, the higher the grade of membership of element or object in a
fuzzy set A. Various types of membership functions can be used to represent the fuzzy set.
A typical FGP problem formulation can be stated as follows:

Find xi i = 1, 2, . . . , n,
to satisfy

Zm (xi ) ≺ Z̄m , m = 1, 2, . . . , M,
Zk (xi )  Z̄k , k = M + 1, M + 2, . . . , K,
gj (xi )  bj , j = 1, 2, . . . , J ,
xi  0, i = 1, 2, . . . , n, (1)

where Zm (xi ) is the mth goal constraint, Zk (xi ) is the kth goal constraint, Z̄m (xi ) is the target value of
the mth goal, Z̄k (xi ) is the target value of the kth goal, gj (xi ) is the j th inequality constraint, and bj is
the available resource of inequality constraint j .
In formulation (1), the symbols “≺ and ” denote the fuzzified versions of “  and  ” and can be
read as “approximately less (greater) than or equal to”. These two types of linguistic terms have different
meanings. Under “approximately less than or equal to” situation, the goal m is allowed to be spread
to the right-hand side of Z̄m (Z̄m = lm where lm denote the lower bound for the mth objective) with a
certain range of rm (Z̄m + rm = um , where um denote the upper bound for the mth objective). Similarly,
with “approximately greater than or equal to”, pk is the allowed left-hand side of Z̄k (Z̄k − pk = lk , and
Z̄k = uk ) [24].
As can be seen, GP and FGP have some similarities. Both of them need an aspiration level for each
objective. These aspiration levels are determined by the DM. In addition to the aspiration levels of the
goals, FGP needs max–min limits (uk , lk ) for each goal [25].
After constructing fuzzified aspiration levels with respect to the linguistic terms of “approximately less
than or equal to”, and “approximately greater than or equal to”, the fuzzy membership functions can be
developed for each goals as follows:
For “approximately less than or equal to”:

1 if Zm (x)  lm ,
Zm (x)−lm
zm (x) = 1− um −lm if lm  Zm (x)  um ,
0 if Zm (x)  um .
710 C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726

For “approximately greater than or equal to”:



1 if Zk (x)  uk ,
uk −Zk (x)
zk (x) = 1 − u −l if lk  Zk (x)  uk ,
k k
0 if Zk (x)  lk .
Using Belman and Zadeh’s [29] approach, the feasible fuzzy solution set is obtained by the intersection
of the all the membership functions representing the fuzzy goals. This feasible fuzzy solution set is then
characterized by its membership F (x) which is

F (x) = Z1 (x) ∩ Z2 (x) ∩ · · · ∩ Zk (x) = min[Z1 (x), Z2 (x), . . . , Zk (x)].

Then the optimum decision can be determined to be the maximum degree of membership for the fuzzy
decision:

max F (x) = max min[Z1 (x), Z2 (x), . . . , Zk (x)].


x∈F x∈F

Zimmermann [30] first used the maximin-operator of Bellman and Zadeh [29]. By introducing the aux-
iliary variable , which is the overall satisfactory level of compromise, formulation (1) can be equivalently
transformed as
Max Z = 
s.t.

Z1 ,
Z2 ,
....
..
Zk ,
gj (xi )  bj , j = 1, 2, . . . , J ,
xi  0, i = 1, 2, . . . , n,
0  1. (2)

In general, we often classify the goals by their levels of importance in a GP problem. Consideration of
different relative importance and priority of the goals in the FGP problem is also important because some
goals are more important than others [31]. In preemptive structure, the problem is solved by their orders
of levels. The rule of preemption states that the lower levels of priority cannot be considered unless the
higher levels of priority are satisfied [24]. This is equivalent to saying that if P1 P2  · · ·PK then 1 must
be maximized before the others (2 , . . . , k ).
These goals with the same kth level of priority have the same value of k . Once a level appears to
be unable to reach its goal, then maybe the after levels will force the model to provide an infeasible
solution. Even if the model gives a feasible solution that satisfies the DMs preemptive priority structure,
the computational efficiency will decrease with the number of priority levels.
The weighted additive model is widely used in GP problems to reflect the relative importance of
the goals. Tiwari et al. [21] have proposed a weighted additive model to aggregate priorities of the
C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726 711

fuzzy goals:

Max Z = wk  k
k
s.t. k (x) ∈ [0, 1], ∀k, x  0,

where wk denotes the weight of the kth goal.


Similar to weighted GP, the model may produce different, sometimes undesirable, solutions when the
weights are changed.
In order to deal with the FGP problem with a preemptive priority structure, Chen and Tsai [31] have
proposed an additive formulation. They incorporate the preemptive priority structure into the formulation
to find a set of solutions that optimize the sum of each fuzzy goal’s achievement degree.
To illustrate the formulation, an illustrative example can be useful: there are four fuzzy goals and the
priority levels of the goals are as follows:
Priority level 1: goal 1 and goal 4, Priority level 2: goal 3, Priority level 3: goal 2.
According to the above preemptive priority structure, the following relationship for the respective
achievement degrees for the goals can be written as

1 3 , 4 3 , 3 2 .

After adding the above relationship to the model, the FGP can be formulated as


4
Max Z = k
k=1
s.t.
k k ,
gj (xi )  bj , j = 1, 2, . . . , J ,
1 3 , 4 3 , 3 2 ,
xi  0, i = 1, 2, . . . , n,
0 k  1, k = 1, . . . , 4.

Although Chen and Tsai’s [31] approach have some computational superiority over other approaches,
it does not exactly represent a preemptive priority structure. Because, in exact preemption case, the lower
levels of priority cannot be considered unless the higher levels of priority are satisfied.
Chen and Tsai [31] also suggest an alternative way to incorporate preemptive priority structure into
the problem. This method allows the DM to determine a desirable achievement degree for each fuzzy
goal based on the preemptive priority structure, such as 1  0.9, 2  0.8 as the importance of the fuzzy
goals. Although this approach can ensure that the more important a fuzzy goal, the higher achievement
degree it has, it could make the model infeasible when the desirable achievement degree for each goal is
not chosen appropriately.
There can be different alternatives to deal with preemptive priority structure in literature. These alter-
natives can be used, when the number of priority levels increases or any feasible solution cannot be found
because of priority levels.
712 C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726

4. Previous works of the proposed model

4.1. Hogan and ReVelle’ s maximal backup coverage model

Congestion in service systems, as seen in a frequent lack of availability of a server within the distance
or time standard, motivated the creation of a new set of models which emphasized additional coverage
beyond the first.
Here, additional coverage means the presence within the distance standard of more than one facility, or
more precisely, the number of times a demand zone is covered. For a partially congested system, multiple
coverage may provide the capability of responding to demand within the distance standard even when
the closest vehicle is busy.
Hogan and ReVelle [3] suggested two models: maximal backup coverage problem one (BACOP1) and
maximal backup coverage problem two (BACOP2). The preceding backup coverage formulations require
coverage at each demand zone and this level of service may not be economically feasible while BACOP2
allows the simultaneous optimization of both backup coverage and first coverage. Formulation of Hogan
and ReVelle’s [3] BACOP2 model is presented below.
I and J are sets of demand zones and potential facility sites, respectively,

1 if demand i is covered at least once,
Yi =
0 otherwise,

1 if demand zone i is covered at least twice,
Ui =
0 otherwise,
where Xj is the integer number of facilities located at potential site j, and hi is the population in demand
zone i,

1 if dij  S,
aij =
0 otherwise,
where S is the distance or time standard, dij is the travel distance or time from j to i, and C is the number
of vehicles to be sited.
The model is as follows:

Max Z1 = hi Yi , (3)
i

Max Z2 = hi Yi , (4)
i
s.t.

aij Xj − Yi − Ui  0 ∀ i ∈ I, (5)
j

Ui − Yi  0 ∀ i ∈ I , (6)

Xj = C. (7)
j
C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726 713

The first two constraints of this model work in tandem to determine which zones receive backup
coverage. The first constraint (5) determines the number of facilities within the distance standard of a
zone while the second constraint (6) ensures that backup coverage can only be provided if first coverage
is already in place. Available number of vehicles to be located is limited by the last constraint (7).

4.2. Pirkul and Schilling’s capacitated maximal covering location problem

Most facilities, particularly those that are emergency related, have a limited capacity for meeting
demand. These kinds of limitations prevent service systems from providing effective service. To assure
that facilities are not overloaded and the quality of service is not degraded, such limitations must be
considered [32].
As Pirkul and Schilling [11] noted, when designing a system, capacity limits can serve as a means of
smoothing the workload among different facilities. Current and Storbeck [32] dealt with the capacitated
maximal covering location problem (CMCLP), which consider a facility’s workload to consist of all
demand assigned to the facility that lies within the coverage distance. The demand which lies beyond the
maximum service distance, while assigned to a facility, is not counted as part of that facility’s workload.
With such a problem structure, a given facility’s responsibility would appear to be only to provide service
to those points that it covers. According to Pirkul and Schilling [4], this is pathology. To correct this, they
modified the CMCLP and then extended it.
Pirkul and Schilling’s [4] CMCLP formulation is as follows:

1 if a facility located in zone j,
Xj =
0 otherwise,

1 if zone i is served by a facility in zone j,
pij =
0 otherwise,

where kj is the workload capacity for a facility in zone j,


All other sets, parameters, and variables are defined as before. We redefined the original notations of
Pirkul and Schilling’s [4] model to provide readability.

Max Z = aij hi pij (8)
i j
s.t.

Xj  C, (9)
j

pij = 1 ∀i ∈ I , (10)
j

pij  Xj ∀i ∈ I, j ∈ J , (11)

hi pij  kj ∀j ∈ J , (12)
i
714 C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726

Xj = [0, 1] ∀j ∈ J , (13)
pij = [0, 1] ∀i ∈ I, j ∈ J . (14)
The objective function maximizes the population assigned to a facility within the coverage distance S.
Constraint (9) limits the total number of facilities to no more than C, while (10) insure that all demand
points are assigned to a facility. Whenever a demand zone is assigned to a facility site, then a facility
must be constructed there. Constraint (11) requires that this occurs. The workload limits are imposed by
constraint set (12). Sets (13) and (14) are the integrality restrictions.
The level of service provided to the demand covered is controlled by S in this model. But there is no
force in the model to control the service level to the uncovered zones. Pirkul and Schilling [4] proposed a
new model which simultaneously considers coverage and the level of service delivered to the uncovered
zones to address this situation. Here, the service quality provided to uncovered zones is measured by a
traditional measure, minimizing average travel distance.
Pirkul and Schilling’s [4] extended CMCLP is presented below. The constraints of this model are same
as those of the previous model. Only the objective function was extended as follows:

Max Z1 = aij hi pij ,
i j

Min Z2 = eij hi dij pij ,
i j

where
 
1 if dij  S, 1 if dij > S,
aij = eij =
0 otherwise, 0 otherwise.

5. Proposed model

A new model, formed by considering Hogan and ReVelle’s [3] maximal backup coverage model and
Pirkul and Schilling’s [4] capacitated maximal covering model as the base, is introduced in this section.
The proposed model formulation is stated to allocate a fixed number of emergency service vehicles
to previously defined locations so that three important service level objectives can be achieved. These
objectives are maximization of the population covered by one vehicle, maximization of the population with
backup coverage and increasing the service level by minimizing the total travel distance from locations
at a distance bigger than a prespecified distance standard for all zones.
The proposed model can be distinguished from the base models in several ways. First, the model includes
all objectives considered in the base models. In this way, besides increasing the capability of responding
to demand within the distance or time standard through backup coverage, the level of service delivered
to uncovered zones is considered. Second, we use the continuous form of the assignment variable, pij ,
as Current and Storbeck [32] do, so as to provide a more realistic structure. Furthermore, the number of
zero–one variables is reduced by using the continuous form of pij , and as a result, the computation time
is improved. Third, the backup coverage variable, Ui , is also considered as a continuous variable in the
range of [0,1] to allow fractional backup coverage for a zone. Here, fractional backup coverage means
that some portion of population in a zone may have backup coverage, while the rest of it may not. Lastly,
C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726 715

contrary to Pirkul and Shilling’s [4] model, locating more than one vehicle in a zone is allowed in the
proposed model as in Hogan and Revelle’s [3]. This change eliminates the requirements that the capacity
of a vehicle must be greater than the largest population of all zones.
Furthermore, the proposed model distinguishes itself from previous maximal covering emergency
service vehicle location models in the solution approaches used. Considering that the target values of
the objectives in question are imposed by the DM, some degree of subjective imprecision is included. In
order to incorporate the DMs imprecise aspiration levels for the goals, FGP approaches are used.
The notational structure of the proposed model is as follows:
 
1 if dij  S, 1 if dij > S,
aij = eij =
0 if dij > S, 0 if dij  S,

1 if demand node i is covered once,
Yi =
0 otherwise,
where Ui is the fraction of population in zone i covered by more than one vehicle, hi is the population
in zone i, pij is the fraction of population in zone i served by a facility/facilities located in zone j, dij is
the travel distance or time from j to i, S is the distance or time standard, C is the number of vehicles to
be located, and kj is the workload capacity for a vehicle located in zone j.
The mathematical statement of the proposed model is as follows:

Max Z1 = hi Yi (15)
i

Max Z2 = hi Ui (16)
i

Min Z3 = eij hi dij pij (17)
i j

s.t.

aij pij − Yi − Ui  0, ∀i ∈ I (18)
j

Ui − Yi  0 ∀i ∈ I , (19)

pij  1 ∀i ∈ I , (20)
j

hi pij  kj Xj ∀j ∈ J , (21)
i

Xj  C, (22)
j

Ui  1 ∀i ∈ I , (23)
Xj  0 and integer ∀j ∈ J , (24)
716 C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726

Yi = [0, 1] ∀i ∈ I , (25)
pij  1 for all i, j pairs. (26)
In this model, first and second objectives maximize the population covered by at least one vehicle
and the population with backup coverage, respectively. Third objective provides the maximization of the
service level by minimizing the total travel distance from locations at a distance bigger than S for all
zones. Populations of the zones are included as weights in the third objective.
The first two constraints; determine which zones receive backup coverage. The first constraint set
determine the number of facilities within the coverage distance of a zone. The second constraint ensures
that backup coverage can only be provided if first coverage already exists.
Constraint set (20) provides that the entire population at each zone will be assigned to some facility.
Constraint set (21) ensures that the total population assigned to a facility does not exceed the capacity of
that facility. This constraint also ensures that population will only be assigned to facility sites which have
facilities actually located at them. Constraint (22) limits the total number of facilities to be located.
In constraint set (23), Ui , is considered as a continuous variable to allow fractional backup coverage
for a zone. We restrict the value of Ui in the range of [0,1] so that as much zone as possible can take the
advantage of having backup coverage.
As can be seen, locating more than one vehicle in a zone is allowed with constraint set (24). In contrast
to the basic location covering models it is no longer assumed that the entire population in a zone will be
served by its nearest facility. In this formulation it is also assumed that the entire population in a zone
does not have to be assigned to the same vehicle and that some portion of the population in a zone may
be covered and the rest of it not covered. It is provided by using the continuous form of the assignment
variable, pij .

6. Implementation of the model

We use four solution approaches, namely lexicographic multi-objective linear programming (LMOLP),
FGP, Chen and Tsai’s [31] additive FGP (A_FGP_C) and Tiwari et al.’s [21] weighted additive FGP
(WA_FGP_T), for solving the proposed model. In our experiments, two problem instances with different
number of zones (30- and 50-zone) are generated.
Firstly, LMOLP technique is used to solve the model. The solution method operates in a way that
addresses lower-priority objectives only after higher priority objectives have been optimized. In this
approach, an ordinal ranking of the objectives are specified as follows:
 
Priority level 1: Max Z1 = hi Yi , Priority level 2: Max Z2 = hi Ui ,
i i

Priority level 3: Min Z3 = eij hi dij pij .
i j

As discussed in Section 3, target values obtained from the DM may not be crisp and described precisely.
Here, the problem is modeled by using three different FGP approaches in order to incorporate the vague
aspirations of DMs into the formulation. Aspiration levels specified by DM to reflect relative flexibility
are presented in Table 1.
C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726 717

Table 1
Aspiration levels of the fuzzy goals

Objective Problem instance

30-zone 50-zone

Uk Lk Uk Lk

i hi Yi 240,000 220,000 490,000 480,000
i h
i Ui 50,000 40,000 60,000 45,000
i j eij hi dij pij 200,000 150,000 1,000,000 500,000

1 1

220000 240000
Σi hiYi 40000 50000
Σi hiU i

150000 200000
Σ Σ e hd
i j
ij i ij pij

Fig. 1. Membership functions of the fuzzy goals of the 30-zone problem.

After constructing aspiration levels with respect to the linguistic terms of “approximately less than or
equal to”, and “approximately greater than or equal to”, the fuzzy membership functions can be defined
for each goal. For instance, the membership function of the first goal of the 30-zone problem is given
below:
⎧ 

⎪ 1 if hi Yi  240, 000,



⎪  i


⎨ 240,000− hi Yi

1 (x) = 1 − i
if 220, 000  hi Yi  240, 000, (27)

⎪ 240,000−220,000



⎪  i

⎪ hi Yi  220, 000.
⎩0 if
i

Fig. 1 illustrates the membership functions of the fuzzy goals of the 30-zone problem.
According to the objective values in the interval [Lk , Uk ], the linear and continuous membership
functions can be defined in the same manner to quantify the fuzzy aspiration levels for the 50-zone
problem.
718 C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726

In the first FGP model, which will be called FGP1 in remaining sections, all objectives are wanted
to be satisfied simultaneously, and there are no priorities and also no relative importance assigned to
objectives. Then, for the 30-zone problem, the resulting linear programming formulation which uses
maximin-operator for aggregating goals to determine the fuzzy decision can be written as follows:

Max Z = 
s.t.
   
hi Yi − 220, 000 hi Ui − 40, 000 200, 000 − eij hi dij pij
i i i j
 ,  ,  ,
20, 000 10, 000 50, 000
0  1,

and constraints (18), (19),…, (26).


In the second fuzzy solution approach, we assume that the goals have different levels of importance.
In order to reflect the relative importance of the goals, Tiwari’s et al. [21] weighted additive model is
used. The relative weights of the three fuzzy goals are assumed as 0.7, 0.2 and 0.1, respectively. After
formulating the fuzzy goals, WA_FGP_T can be defined as follows:


3
Max Z = w k k
k=1
s.t.
   
hi Yi − 220, 000 hi Ui − 40, 000 200, 000 − eij hi dij pij
i i i j
1  , 2  , 3  ,
20, 000 10, 000 50, 000
0 k  1, k = 1, 2, 3

and constraints (18), (19), . . . , (26),

where wk denotes the weight of the kth fuzzy goal.


As the last fuzzy solution approach, the preemptive priority structure is incorporated into Chen’s [31]
additive formulation. According to the preemptive priority structure used in LMOLP model, the following
relationship for the respective achievement degrees for the goals can be written:

1 2 , 2 3 .

After adding the above relationship into the model as constraints, A_FGP_C model can be formulated
as follows:


3
Max Z = k
k=1
s.t.
C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726 719

1 2 ,
2 3 ,
0 k  1, k = 1, 2, 3

and constraints (18), (19), . . . , (26),


where the k ’s are defined as before.

7. Computational results

To illustrate the applicability of the proposed model, a set of computational experiments are performed
using LMOLP, FGP, additive FGP and weighted additive FGP approaches. As mentioned earlier, two
problem instances with different number of zones (30- and 50-zone) are randomly generated in our
experiments. The coordinates of the zones are drawn from the uniform distribution within the ranges
[0,45] and [0,85] for first and second instances, respectively. The Euclidean distances between zones are
used to define the coverage. We assume that a zone is covered, if it is served within the range of 15. Total
number of vehicles to be located is obtained by using set covering model for each instance. The workload
capacity which is the maximum number of people to be served is assumed as 75,000 for each vehicle.
To obtain solutions, the proposed model is solved by CPLEX 8.0 optimization software. The results of
the proposed model obtained by using the solution approaches mentioned are presented in Tables 2 and
3. Additionally, both problem instances are solved for BACOP2 and CMCLP models, and these results
are shown in Table 4.
In the LMOLP model, maximizing the first coverage is considered first since it has the highest priority.
Hence, solutions of the problems are 241,669 and 504,000 for 30- and 50-zone problems, respectively. It
is interesting to note in solutions of 30-zone problem that LMOLP model provides higher first coverage
solution than the CMCLP model, although the first coverage has the highest priority in both models.
This is probably because the proposed model allows the population at a zone can be served by more than
one vehicle located at different zones, whereas the CMCLP model considers pij as a binary variable.
Furthermore, the improvements in the first coverage objective results in better service level. The service
levels, which have the lowest priority, are found as 91,793 and 413,835.
It should also be noted that the BACOP2 model always gives better first coverage solutions. This solution
is clear and as expected since the capacity constraints are not considered in this model. Therefore, it should
not be compared with other models in terms of the first and backup coverage objectives. Fig. 2 illustrates
the level of first coverage objective for each solution approach, except BACOP2.
Besides the LMOLP method is better in the first coverage objective, it also provides 15,627 and 13,000
for the backup coverage objective in each instance, respectively. Since the backup coverage objective has
lower priority than the first coverage, the lexicographic solution approach causes unsatisfactory solutions
for backup coverage.
In the FGP approaches, the backup coverage is substantially improved by sacrificing the first coverage
in a small amount. According to results of FGP1 in 30-zone problem, the percentage of population with
backup coverage is 18.18%, while the percent of first coverage is 95.89%. In other words, deterioration
in the first coverage by 2.31% (from 98.2% to 95.89%) results in improvement in the backup coverage
by 192% (from 6.34% to 18.18%). Total satisfaction of objectives is 48%.
720

Table 2
Results of the proposed model with LMOLP approach
Number of zones Number Capacity The zones repre- Number of un- Number of peo- Number of peo- Service Total
of vehi- of a senting vehicle covered people ple covered by ple with backup level CPU time
cles vehicle locations one vehicle coverage (s)

30 4 75,000 1, 6, 17, 27 4637 (1.88%) 241,669 (98.12%) 15,627 (6.34%) 93,793 2.78
50 8 75,000 2, 8, 15, 22, 32, 35, 39, 49 14,000 (2.7%) 504,000 (97.30%) 13,000 (2.51%) 413,835 4.9
C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726
Table 3
Results of the proposed model with FGP approaches

Solution Number Number Capacity The zones repre- Satisfaction Number of un- Number of people Number of people Service Total
Approach of zones of vehi- of a senting vehicle levels covered people covered by one with backup cover- level CPU time
cles vehicle locations vehicle age (s)

FGP 30 4 75,000 1, 6, 9, 17  = 0.48 10,116 (4.11%) 236,190 (95.89%) 44,799 (18.18%) 176,003 2.94
Max  50 8 75,000 1, 2, 9, 18, 20,  = 0.70 31,000 (5.98%) 487,000 (94.02%) 55,487 (10.71%) 650,421 14.30
28, 35, 49

A-FGP-C 30 4 75,000 1, 6, 9, 17 1 = 0.81, 10,116 (4.11%) 236,190 (95.89%) 47,369 (19.23%) 180,788 2.94
2 = 0.74,
3 = 0.38
Max 1 + 50 8 75,000 1, 2, 16, 20, 24, 1 = 0.70, 31,000 (5.98%) 487,000 (94.02%) 55,500 (10.71%) 650,691 15.58
2 + 3 28, 35, 49 2 = 0.70,
3 = 0.69

WA- 30 4 75,000 1, 6, 9, 17 1 = 0.81, 10,116 (4.11%) 236,190 (95.89%) 48,734 (19.79%) 193,073 2
FGP-T 2 = 0.87,
3 = 0.14
Max 0.7 ∗ 50 8 75,000 2, 15, 16, 20, 23, 1 =1.0, 2 = 260,00 (5.02%) 492,000 (94.98%) 58,918 (11.37%) 1,000,000 9.06
C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726

1 + 0.2 ∗ 28, 35, 49 0.92, 3 = 0


2 + 0.1 ∗
3
721
722

Table 4
Results of BACOP2 and CMCLP models
Model Number Number Capacity The zones repre- Number of un- Number of people Number of people Service Total
of zones of vehi- of a senting vehicle covered people covered by one with backup cover- level CPU time
cles vehicle locations vehicle age (s)

BACOP2 30 4 Not con- 6, 17, 21, 27 0 246,306 (100%) 52,112 (21.16%) Not con- 0.2
sidered in sidered in
the model the model
50 8 Not con- 2, 22, 28, 30, 34, 0 518,000 (100%) 62,000 (11.96%) Not con- 0.2
sidered in 35, 39, 46 sidered in
the model the model

CMCLP 30 4 75,000 1, 15, 19, 21 8389 (3.5%) 237,917 (96.5%) Not considered in 224,823 1.1
the model
50 8 75,000 3, 15, 21, 22, 32, 14,000 (2.7%) 504,000 (97.30%) Not considered in 320,585 110.52
35, 39, 49 the model
C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726
C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726 723

245000 510000
The number of people with first

The number of people with first


240000 500000
235000 490000
230000
480000
coverage

coverage
225000
470000
220000
460000
215000
450000
210000
205000 440000

200000 430000
CMCLP LMOLP FGP WA-FGP-T A-FGP-C CMCLP LMOLP FGP WA-FGP-T A-FGP-C
50-Zone Problem
30-Zone Problem

Fig. 2. First coverage results.

Same results can be drawn for the 50-zone problem instance. The lower bound of aspiration levels
for the first coverage objective in FGP1 is reflected relative flexibility with respect to LMOLP model.
Therefore, FGP1 generates lower-first coverage and higher-backup coverage plan compared to LMOLP
model. Deterioration in the first coverage by 3.37% results in substantial improvement in the backup
coverage by 327%. In this case, total satisfaction level is 70%. These improvements are due to the fact
that FGP approaches find the satisfactory solutions by considering the aspiration level of the objectives
and allowing the tradeoffs among the goals. Fig. 3 illustrates the results of the solution approaches in
terms of the percentage levels of first and backup coverage.
As explained earlier, WA-FGP-T approach uses different weights in order to reflect the relative impor-
tance of the goals. It is important to note that this approach provides the best first and backup coverage
solutions among FGP approaches, while it gives worst solutions for the third goal for each problem
instance. These results show that the model reflects the relative importance of the goals to the solu-
tions. Additionally, the resulting achievement degrees for the three fuzzy goals are (0.81, 0.87, 0.14) and
(1, 0.92, 0) for the 30- and 50-zone problems, respectively.
A-FGP-C approach presents a preemptive structure in order to solve the proposed model. Instead of the
classical preemptive structure, the approach optimizes the sum of each fuzzy goal’s achievement degree.
The resulting achievement degrees for the three fuzzy goals are (0.81, 0.74, and 0.38) and (0.70, 0.70,
0.69) for the 30- and 50-zone problems, respectively. As can be seen, the resulting achievement degrees
show that this approach can ensure that the more important a fuzzy goal, the higher achievement degree
it has. However, it does not exactly represent a preemptive priority structure. A-FGP-C gives nearly the
same results as FGP1 in terms of all objectives for each instance.
In the light of the above results, it can be seen that backup coverage can be provided without sub-
stantial first coverage loss by using FGP approaches. It should also be noted that the proposed model
has computational superiority in terms of CPU times given in Tables 2 and 3, and computational bur-
den compared to other models. This superiority may be the consequence of the structure of the model
and efficiency of the fuzzy solution approaches. Additionally, when the number of zones is increased to
60, the fuzzy approaches solve the proposed model in at most 45.27 s while the lexicographic approach
reaches the optimum in 746 s. On the other hand, CMCLP model finds the optimum solution in more
than 1 h.
724 C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726

First Coverage Backup Coverage

100 12

95

percentage of backup coverage


Percentage of first coverage

10
90

85
8
80

75 6

70
4
65

60
2
55

50 0
CMCLP LMOLP FGP WA-FGP-T A-FGP-C

30-Zone Problem

First Coverage Backup Coverage

100 25

95

percentage of backup coverage


Percentage of first coverage

90 20

85

80 15

75

70 10

65

60 5

55

50 0
CMCLP LMOLP FGP WA-FGP-T A-FGP-C

50-Zone Problem

Fig. 3. First and backup coverage levels.

8. Concluding remarks

In this paper, a fuzzy multi-objective covering-based location model for emergency service vehicles is
proposed. In addition to the first coverage objective, backup coverage objective and service level that is
defined as the total travel distance from locations at a distance bigger than a prespecified distance standard
for all zones were both considered in the proposed model.
Model applications with different solution approaches such as LMOLP and FGP are provided through
numerical illustrations to demonstrate the applicability of the proposed model. The analysis has verified
that the proposed FGP model is an effective tool for generating a set of more realistic and flexible optimal
solution in solving emergency service vehicle location problems.
Because of the combinatorial complexity of the proposed model, the development of efficient heuristics
for large problems is an important area for future research. Herein, the proposed model and the membership
C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726 725

functions are assumed to be linear. However, in real life situations, nonlinear models and also nonlinear
membership functions may be more appropriate. As another direction for future research, besides the
DMs aspiration levels assigned to each goal, the model parameters, such as capacity of a vehicle, can be
modeled using FST.

References

[1] Goldberg JB. Operations research models for the deployment of emergency services vehicles. EMS Management Journal
2004;1(1):20–39.
[2] Church RL, ReVelle C. The maximal covering location problem. Papers of Regional Science Association 1974;32:
101–18.
[3] Hogan K, ReVelle C. Concepts and applications of backup coverage. Management Science 1986;32:1434–44.
[4] Pirkul H, Schilling DA. The maximal covering location problem with capacities on total workload. Management Science
1991;37:233–48.
[5] ReVelle C, Schweitzer J, Snyder S. The maximal conditional covering problem. INFOR 1996;34(2):77–91.
[6] Badri M, Mortagy A, Alsayed C. A multi-objective model for locating fire stations. European Journal of Operational
Research 1998;110(2):243–60.
[7] Zadeh LA. Fuzzy sets. Information and Control 1965;8:338–53.
[8] Daskin MS, Hogan K, ReVelle C. Integration of multiple, excess, backup, and expected covering models. Environment
and Planning 1988;15:15–35.
[9] Toregas C, Swain R, ReVelle C, Bergman L. The location of emergency service facilities. Operations Research 1971;19:
1363–73.
[10] Schilling D, ReVelle C, Cohen J, Elzinga D. Some models for fire protection locational decisions. European Journal of
Operational Research 1980;5(1):1–7.
[11] Pirkul H, Schilling DA. The siting of emergency service facilities with workload capacities and backup service. Management
Science 1988;34:896–908.
[12] Narasimhan S, Pirkul H, Schilling D. Capacitated emergency facility siting with multiple levels of backup. Annals of
Operations Research 1992;40(1):323–37.
[13] Berman O, Krass D. The generalized maximal covering location problem. Computers & Operations Research 2002;29:
563–81.
[14] Karasakal O, Karasakal EK. A maximal covering location model in the presence of partial coverage. Computers &
Operations Research 2004;3:1515–26.
[15] Marianov V, ReVelle C. Siting emergency services. In: Drezner Z, editor. Facility location. Berlin: Springer; 1995.
p. 199–223.
[16] CharnesA, Cooper WW, Ferguson RO. Optimal estimation of executive compensation by linear programming. Management
Science 1955;1:138–51.
[17] Narasimhan R. Goal programming in a fuzzy environment. Decision Science 1980;11:325–36.
[18] Narasimhan R, Rubin PA. Fuzzy goal programming with nested priorities. Fuzzy Sets and Systems 1984;14:115–29.
[19] Hannan EL. Some further comments on fuzzy priorities. Decision Science 1981;13:337–9.
[20] Tiwari RN, Dharmar S, Rao JR. Priority structure in fuzzy goal programming. Fuzzy Sets and Systems 1986;19:251–9.
[21] Tiwari RN, Dharmar S, Rao JR. Fuzzy goal programming—an additive method. Fuzzy Sets and Systems 1987;24:27–34.
[22] Ramik J. Fuzzy goals and fuzzy alternatives in goal programming problems. Fuzzy Sets and Systems 2000;111:81–6.
[23] Rao SS, Tiwari RN, Mohanty BK. A preference structure on aspiration levels in a goal programming problem—a fuzzy
approach. Fuzzy Sets and Systems 1988;25:175–82.
[24] Wang HF, Fu CC. A generalization of fuzzy goal programming with preemptive structure. Computers & Operations
Research 1997;24:819–28.
[25] Mohamed RH. The relationship between goal programming and fuzzy programming. Fuzzy Sets and Systems 1997;89:
215–22.
[26] Ohta H, Yamaguchi T. Linear fractional goal programming in consideration of fuzzy solution. European Journal of
Operational Research 1996;92:157–65.
726 C. Araz et al. / Computers & Operations Research 34 (2007) 705 – 726

[27] El-Wahed WF, Abo-Sinna MA. A hybrid fuzzy-goal programming approach to multiple objective decision making
problems. Fuzzy Sets and Systems 2001;119:71–85.
[28] Mohammed W. Chance constrained fuzzy goal programming with right-hand side uniform random variable coefficients.
Fuzzy Sets and Systems 2000;109:107–10.
[29] Bellman RE, Zadeh LA. Decision-making in a fuzzy environment. Management Science 1970;17:141–64.
[30] Zimmermann H-J. Fuzzy programming and linear programming with several objective function. Fuzzy Sets and System
1978;1:45–55.
[31] Chen LH, Tsai FC. Fuzzy goal programming with different importance and priorities. European Journal of Operation
Research 2001;133:548–56.
[32] Current JR, Storbeck JE. Capacitated covering models. Environment and Planning B: Planning and Design 1988;15:
153–63.

Ceyhun Araz is a research assistant and a Ph.D. student in Industrial Engineering Department at Dokuz Eylul University, Izmir,
Turkey. He obtained his M.S. in Industrial Engineering from Institute of Natural and Applied Sciences, Dokuz Eylul University
in 2001. His research interests are multi-criteria decision making, fuzzy set theory and its applications on multiple objective
decision making.

Hasan Selim is a research assistant and a Ph.D. student in Industrial Engineering Department at Dokuz Eylul University, Izmir,
Turkey. He obtained his M.S. in Industrial Engineering from Institute of Natural and Applied Sciences, Dokuz Eylul University
in 1999. His research interests are supply chain management, multiple objective decision making and OR applications on health
care systems.

Irem Ozkarahan is a Professor of Industrial Engineering at Dokuz Eylul University, Izmir, Turkey. She has a Ph.D. in Industrial
and Management Systems Engineering from Arizona State University, USA. M.A.Sc. in Industrial Engineering form University
of Toronto, Canada and B.Sc. in Industrial Engineering from Middle East Technical University in Turkey. Among the Journals
she has published are Naval Research Logistics Quarterly, IEEE Transactions, IEEE Journal of the Society for Health Systems,
Journal of Medical Systems, and Computers and Industrial Engineering. Her research interests include manpower scheduling
and its application on health care, heuristic scheduling, multiple objective decision making and supply chain management.
http://people.deu.edu.tr/irem.ozkarahan.

You might also like