Transportation Research Part E 72 (2014) 101–124

Contents lists available at ScienceDirect

Transportation Research Part E

journal homepage: www.elsevier.com/locate/tre

Multi-objective design of an organ transplant network under

uncertainty
B. Zahiri a, R. Tavakkoli-Moghaddam a, M. Mohammadi a, P. Jula b,⇑
a
School of Industrial Engineering, College of Engineering, University of Tehran, Tehran, Iran
b
Beedie School of Business, Simon Fraser University, 8888 University Drive, Burnaby, BC V5A 1S6, Canada

a r t i c l e i n f o a b s t r a c t

Article history: We propose a novel multi-period location–allocation model for the design of an organ
Received 10 January 2014 transplant transportation network under uncertainty. The model consists of a bi-objective
Received in revised form 31 July 2014 mathematical programming model that minimizes total cost and time, including waiting
Accepted 20 September 2014
time in the queue for the transplant operation, while considering organs’ priorities. A fuzzy
multi-objective programming based approach is presented to solve the small and medium
size problems to optimality. For larger problems, we propose two meta-heuristics based
Keywords:
algorithms. Lower bounds, and several numerical examples with managerial insights are
Healthcare network
Multi-period organ transplant network
discussed. A real case-study is provided, and the existing and the proposed optimal
Fuzzy multi-objective programming solutions are compared.
Queuing theory Ó 2014 Elsevier Ltd. All rights reserved.
Meta-heuristics

1. Introduction

Supply chain management (SCM) is the process of planning, implementing and controlling the operations of the supply
chain efficiently (Melo et al., 2009). The supply chain network design (SCND), as one of the most important subsets of stra-
tegic decision making of SCM, plays an important role in the overall economic performance of the chain. SCDN addresses the
number of facilities, their location, and the allocation of flows among these facilities (Pishvaee and Razmi, 2012). Location–
allocation models play critical roles in health service planning. They provide frameworks for investigating accessibility prob-
lems, comparing the quality (i.e., in terms of cost efficiency and public welfare), and suggesting alternative solutions to
improve existing systems (Shariff et al., 2012; Rahman and Smith, 2000).
Organ transplantation, which is one of the most vital subset of healthcare systems, has become a successful treatment for
many diseases that otherwise would have been fatal (Beliën et al., 2012). Despite all the advances and sophisticated tech-
nologies in operations and transportation methods in healthcare facilities, management on location–allocation of organ
transplant centers in some districts has remained far from being cost and time efficient on a consistent basis.
The transplant surgical process involves a donor (i.e., a volunteer or brain-death patient) and a recipient (i.e., the patient
who receives the organ). The concerned network system consists of donors, recipients’ region/zone, shipping agents, hospi-
tals and transplant centers (TCs). Donors are kept in hospitals, where some necessary analyses are carried out and organ is
removed (harvested). Recipient zones are populated areas, in which majority of applicants for receiving organs locate. Trans-
plant centers are facilities where the registration and surgical operations for transplanting the recipients are done. And
finally, shipping agents are in charge of transporting the organs and the required materials from hospitals to TCs. Therefore,

⇑ Corresponding author. Tel.: +1 778 782 3564; fax: +1 778 782 4920.
E-mail addresses: b.zahiri@ut.ac.ir (B. Zahiri), tavakoli@ut.ac.ir (R. Tavakkoli-Moghaddam), mehrdadmohamadi@ut.ac.ir (M. Mohammadi), pjula@sfu.ca
(P. Jula).

http://dx.doi.org/10.1016/j.tre.2014.09.007
1366-5545/Ó 2014 Elsevier Ltd. All rights reserved.
102 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124

there are several decisions that should be simultaneously made, including: (1) establishing hospitals among a potential set of
locations, (2) establishing TCs among a given set of candidate locations, (3) determining which hospitals or TCs are capable of
treating specific organs. Furthermore, the following decisions are dynamically made in each time period: (4) assigning dif-
ferent shipping agents to transfer donors’ organs and necessary materials between hospitals and TCs, (5) selecting transpor-
tation modes with different cost and time while the transplant operation is performed within specified perishable time for
each organ, and (6) allocating different regions of patients to TCs.
The complexity of the organ transplant network and the intense interaction between facilities require efficient manage-
ment to maximize the benefits of the stakeholders involved. In such networks, donors are usually transferred to hospitals, or
may refer directly to TCs for donation. As the demand zones and the TCs may be located in disperse geographical areas with
often long distances, servicing all demands via a single transportation mode may not be practical. On the other hand, organ
perishability is a critical issue in transplant supply chains. According to Uehlinger et al. (2010), each organ is constrained by
the cold ischemia time, defined as the maximum time that the organ can be kept outside the body.
Due to dynamic nature of this industry, the respective costs, supplies, demands, distances, travel times and other relevant
parameters may change due to uncertain circumstances. This will greatly affect the design of the network. Therefore, it is
important to address uncertainties in the problem.
This paper introduces a new bi-objective mixed-integer programming (BOMIP) model for the multi-period location–allo-
cation problem of the organ transplant supply chain that minimizes the total cost and travel time. The main contributions of
this paper, which differentiate our efforts from related studies, are as follows:

 Designing a novel multi-period transplant network that addresses fluctuations in demands and supplies, while consider-
ing different transportation modes, and possible integrated facilities (i.e., facilities which can simultaneously function as a
hospital and a TC).
 Developing a multi-priority queuing system to model the congestion of organs at each TC.
 Applying an efficient hybrid fuzzy multi-objective programming approach to cope with uncertainties.
 Developing two meta-heuristic based algorithms to efficiently solve large size instances of the problem.
 Proposing lower bounds for evaluating the performance of the proposed meta-heuristic algorithms in cases that the opti-
mal solutions are not available.
 Applying the proposed methods in a real-case study.

The rest of the paper is organized as follows: Section 2 represents a literature review. Problem definition and mathemat-
ical formulation are presented in Section 3. In Section 4, the proposed possibilistic programming is explained. Meta-heuristic
algorithms and the lower bound approach are presented in Section 5. Section 6 handles the computational experiments and
an industrial case-study. Finally, conclusion and future research are provided in Section 7.

2. Literature review

Focus on the current research can be classified into two main categories: first, designing location–allocation of healthcare
facilities (or healthcare supply chain network design), and second, optimizing organ transplant supply chains.

2.1. Location–allocation design of healthcare networks

In the area of healthcare facilities network design, Papageorgiou (1978) and Rais and Viana (2010) proposed a compre-
hensive survey on applying operations research on such systems. In this field, several researchers have presented different
models for location–allocation of healthcare facilities. Shariff et al. (2012) formulated a capacitated maximal covering
healthcare location–allocation problem and applied it to one of the districts of Malaysia. For determining the percentage
of coverage of the existing facilities, they proposed a new genetic algorithm (GA). Benneyan et al. (2012) presented a model
for multi-period location–allocation of Veterans Health Administration. They presented a multi-period mathematical integer
programming model considering the trade-offs between costs, coverage, service location and capacity. Syam and Côté (2012)
proposed a model for location–allocation of a treatment department related to traumatic brain injuries. They considered the
minimization of the total cost as an objective function. The data obtained from department of veterans affairs have been used
for testing the applicability of their model. They also examined the effects of five critical factors, such as the degree of service
centralization, service level mandates by acuity, lost admission cost by acuity, facility overload penalty cost by acuity and
target utilization level by acuity and treatment unit. Furthermore, Syam and Côté (2010) for location–allocation of special-
ized healthcare systems proposed a model based on three main factors; i.e., geographic density of the patient population,
degree of centralization of services, and the role of patient retention as a function of distance to a treatment unit. Zhang
and Jiang (2013) presented a bi-objective robust programming approach for an emergency medical services system under
uncertainty that minimizes the total costs and determines the assignment of demand points to EMS, location of EMS facil-
ities, and the number of EMS vehicles at each station. Shariff et al. (2012) focused on multi-period location–allocation of
healthcare systems (i.e., emergency blood facilities) for a case study in Beijing. They proposed a heuristic algorithm based
on the Lagrangian relaxation method.
 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124 103

2.2. Optimizing organ transplant supply chains

In the domain of optimizing organ transplant network systems, Bruni et al. (2006) proposed a mixed-integer linear pro-
gramming (MILP) model based on equity. In this approach, equity refers to equalizing the transplant priority among patients
with the same needs but belonging to different regions. To optimize their transplant system, they assumed special centers –
referred as OPO (Organ Procurement Organization) in the USA – that play critical rules in managing and procuring the organs.
Each OPO center covers one or more transplant centers, in which transplant operations are done. Balancing the mean waiting
time of a patient on the transplant waiting lists is another issue that is aimed in the paper. A real case study in Italy is also
applied. Kong et al. (2010) aimed at maximizing the efficiency of the liver allocation systems in the USA. For solving the prob-
lem, a branch-and-price approach was utilized. Clinical data were used for solving numerical examples. Beliën et al. (2012)
proposed a mixed-integer linear programming model with deterministic parameters to define the optimal location of trans-
plant centers. Their objective function minimizes the total weighted time, by assigning a weight coefficient to the time period
when a removed organ should be implanted to the patient body and the time required for the round-trip traveling of a unit
from a donor hospital to donor transplant centers. They took five different organs (i.e., Kidney, Liver, Lung, Heart and Pancreas)
into consideration and have linked them by a budget constraint. A case study in Belgian was also presented. Zahiri et al. (2014)
presented a mathematical model for multi-period location–allocation of organ transplant centers under uncertainty with
minimization of the total costs as the objective function. To cope with epistemic uncertainty of input data, they have utilized
a robust possibilistic programming approach. Numerical examples were inspired by a real case study in Iran.
There is a gap in existing literature addressing the integration of designing the healthcare facility locations, and the opti-
mization of organ supply chain networks. Several authors have considered these categories separately (Shariff et al., 2012;
Papageorgiou, 1978; Benneyan et al., 2012; Syam and Côté, 2012; Kong et al., 2010), while integrating them to make a more
realistic model is needed. Towards this goal, Zahiri et al. (2014) presented a mathematical model for multi-period location–
allocation of organ transplant centers under uncertainty and used robust possibilistic programming approach to solve the
problem. Minimizing the total cost was the sole objective of the model presented. Our study targets another gap in existing
literature by considering alternative transportation modes with different time and cost involved. Alternative transportation
modes should be considered when organs have different demands, and perishable times. For organs with fast perishable time,
a swift transportation mode may be more appropriate.
Furthermore, existing literature does not sufficiently address uncertainties in the system. In industrial cases, the demand
and supply parameters are affected by uncertain circumstances. In addition, since the demand and supply of organs as well
as the processing times are uncertain, and because of differences in perishable times among organs, there will be congestion
in TCs. To address the gap in existing literature on considering congestion in the system, we develop a multi-priority queuing
system in our approach. Finally, we contribute to existing literature by simultaneously considering the surgical and trans-
portation operations in our approach.

3. Modeling framework and mathematical programming

3.1. Modeling framework

Fig. 1 shows the schematic view of an organ transplant supply chain network and the interactions between its elements.
As soon as a volunteer or a brain-death patient donor (D) arrives (1), the lab tests and the necessary analyses will be con-
ducted in the hospital (H). In the case of an operation, a removal (harvesting) process will be done in the hospital and the
organ will be sent to the TC through the shipping agents (ShA) using one of alternative transportation modes. Meanwhile, the
allocated recipients’ zones (Z) are notified and the recipient patient should be present at the TC as soon as possible for the
operation (4). We assume each hospital has a contract with a single shipping agent in each time period (i.e., monthly, quar-
terly, or yearly) (2). Shipping agents deliver organs to the TCs by ground mode (3) or through air mode (30 ).
Furthermore, organ removal in the hospital needs several processes (e.g., the initial medical examinations, and the prep-
aration for surgery, and the surgery). The cumulative time spent in these operations is referred as ‘‘operation time’’ in this
article. Since operation time significantly varies for each case, organ recipients must wait in a queue before surgery. In this
study, based on the importance of the organs (i.e., based on their demand and perish time) each of the organs has a priority
associated with it, and is processed according to its priority. Since having surgery to transplant an organ requires a significant
operation time, waiting time of other patients should be taken into account. On the other hand, as the waiting time for a
transplant operation directly affects the perishability of the organ and survival chance of the recipients, controlling this time
can significantly elevate the efficiency of the proposed network. Finally, the proposed model minimizes total transportation
cost and time as a multi-objective optimization model.

Fig. 1. A typical organ transplant network.

104 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124

Another feature of our model is facility integration where hospitals and TCs are integrated in a same center, resulting in
savings in fixed and transportation costs and higher utilization of resources. By integrating the facilities, patients’ move-
ments will be less, resulting in lower risks of health complications and death.

3.2. M/M/c priority queuing system

A considerable body of research has shown that queuing theory can be applied in real-world healthcare situations
(McClain, 1976; Nosek and Wilson, 2001; Preater, 2002; Green, 2006; Wang, 2004). Many authors concluded that since there
is always a variation of arrivals around scheduled times, a Poisson process is a good representation of the arrival rates. They
also declared that Poisson process is applicable to other service systems where arrival rates and capacity levels vary signif-
icantly over time. In this paper, since the unit flows are transferred by shipping agents, the number and frequency of entering
organs entities to the transplant centers is low. There is always a variation in the inter-arrival times of transporters due to
factors such as environmental conditions, traffic jams, link disruptions, etc. Hence, an Exponential distribution is considered
for modeling the inter-arrival times of organs entering to the transplant centers. We also use the peak hours analysis, assum-
ing that during the peak hours, the average arrival rate and the service rate are both constant and the arrivals of patients to
the transplant centers follow Poisson distribution. This allows us to model the queue formed by organs entering to the trans-
plant centers as an M/M/c queuing system. Furthermore, Poisson is the most commonly used distribution for modeling
healthcare services when there are not many specified details on patient behavior and randomness is an essential part of
the behavior of the system. The transplant centers in our queuing system are multiple servers with infinite capacities, thus
no balking of arrivals is considered. In the proposed system, the priority of the organs is addressed based on their numbering
index. That is, organs with index o = 1 have the highest priority, and lower priorities are numbered in ascending order. The
following notation is used to present the priority queuing system:

Sets
k index of candidate transplant centers, k = {1, . . . ,K}
i index of candidate hospitals, i = {1, . . . , I}
o,o0 index of type of organs, o = {1, . . . , O}
t index of time periods, t = {1, . . . , T}

Parameters
Wk number of servers at TC k (e.g., available operating rooms, etc.)
ltko service rate of TC k for processing organ type o at period t

Decision variables
ktko arrival rate of organ type o at TC k in period t
X tiko number of organ type o transferred from hospital i to TC k in period t
Xatiko air flow of organ type o from hospital i to a TC k in period t
Wqtko average waiting time for organ type o at TC k in period t
W tko average total time spent by organ type o at TC k in period t

Based on above notations, the arrival rate of an organ to a TC is calculated by:

X
I
 t 
ktko ¼ X iko þ Xatiko ; 8k; o; t; ð1Þ
i¼1

Alternatively, the total service times, which is the sum of waiting time in the queue and operation time, is calculated by:
1
W tko ¼ Wqtko þ ; 8k; o; t; ð2Þ
ltko
In order to calculating the queue waiting time of arrival organs into the TCs, Eq. (3) is presented, in which Atko and Btko are
calculated as Eqs. (4) and (5), respectively (Gross et al., 2013).
1
Wqtko ¼ ; 8k; o; t; ð3Þ
Atko  Btko  Btk;o1
2 !Wk W 1" j # 3
  ltko Xk
ktko =lt
Atko ¼ 4Wk ! Wk l  t
ko ktko ko
þ Wk l t 5
ko ; 8k; o; t; ð4Þ
ktko j¼0
j!
Po t
o0 ¼1 kko0
Btko ¼ 1  ; 8k; o; t; ð5Þ
Wl k ko
t

where Btk0 ¼ 1. Furthermore, by substituting Eqs. (4) and (5) in Eq. (3), the queue waiting time of arrival organs into the trans-
plant centers is presented by:
 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124 105

1
Wqtko ¼  
Po t 
Po1 t  ; 8k; o; t; ð6Þ
 lt Wk PWk 1 ðktko =ltko Þj k 0 k
Wk ! Wk ltko  ktko ko
ktko j¼0 j!
þ Wk ltko 1  Wo0 ¼1
l
k
t
ko
1  o0 ¼1 ko0
W lt k
ko k;o1

Finally, the total service time, W tko , is computed by:

1 1
W tko ¼  
Po t 
Po1 t  þ t ; 8k; o; t: ð7Þ
 ltko Wk PWk 1 ðktko =ltko Þj k k l
Wk ! Wk ltko  ktko ktko j¼0 j!
þ Wk ltko 1  Wo0 ¼1
lt
ko0

k ko
1  W ol0 ¼1
t
ko 0

k k;o1
ko

3.3. Mathematical programming

This section proposes a new mathematical model for an organ transplant network utilizing the modeling framework of
Section 3.1 and the queuing priority system presented in Section 3.2.
Assumptions:

 Locations of recipient zones are known, while only candidate shipping agents, and potential locations for establishing
hospitals and TCs are given.
 Shipping agents are able to serve multiple hospitals at the same time.
 Shipping agents are located in close proximity of hospitals.
 TCs have dedicated capacity and resources for transplantation.

Sets
v index of candidate shipping agents, v = {1, . . . , V}
h index of recipients’ zones, h = {1, . . . , H}
e index of candidate locations for establishing integrated facilities (i.e., a facility serves as both a hospital and a TC),
e = {1, . . . , E}

Parameters
ci establishing cost (e.g., costs of design and construction) of a hospital in candidate location i
c0k establishing cost of a TC in candidate location k
rio cost of equipping hospital i for organ type o
r 0ko cost of equipping TC k for organ type o
0
fe Saving costs in a case of the facility integration in candidate location e
Sti number of patients arriving at hospital i in period t
M tio average number of organ type o obtained from each donor at hospital i in period t
uti ratio of elective patients to total patients (i.e., elective and non-elective) for organ donation at hospital i in
period t
Dtho total demand of recipients’ zone h for organ type o in period t
cvvi cost of having a contract of hospital i with shipping agent v in each period
ciik cost of moving a unit from hospital i to TC k by ground mode
caik cost of moving a unit from hospital ito TC k by air mode
chhk cost of transporting a patient from recipient zone h to TC k (e.g., Ambulance, medical team, etc.)
tiko traveling time of delivering organ type o from hospital i to TC k by ground mode
taiko traveling time of delivering organ type o from hospital i to TC k by air mode
thhk traveling time from recipient zone h to TC k
tpo defined threshold for traveling time indicating appropriate mode of transportation for organ type o
ISCo cold ischemia time for organ type o
t 0io operation time for the removal process of organ type o at hospital i
xho minimum target satisfaction level of demand for organ type o in recipients’ zone h

Decision variables
zi 1 if a hospital is established in candidate location i; and 0, otherwise
z0k 1 if a TC is established in candidate location k; and 0, otherwise
yio 1 if hospital i is equipped for organ type o; and 0, otherwise
y0ko 1 if TC k is equipped for organ type o; and 0, otherwise
wtv 1 if shipping agent v is selected in period t; and 0, otherwise
X v tv i 1 if hospital i is serviced by shipping agent v in period t; and 0, otherwise
X tiko ground flow of organ type o from hospital i to a TC k in period t

(continued on next page)

106 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124

(continued)

Sets
Xatiko air flow of organ type o from hospital i to a TC k in period t
t
Xhhko flow of recipients from region h to TC k for transplantation of organ type o in period t
Qeo 1 if both yeo and y0eo are 1; and 0, otherwise
NI total number of established hospitals
NK total number of established TCs

The multi-objective mathematical model is presented as follows.

X
I X
K X
O X
I X
O X
K X
T X
I X
V X
T X
O X
K X
I
Min OFV 1 ¼ c i zi þ c0k z0k þ r io yio þ r 0ko y0ko þ cv v i X v tv i þ ciik X tiko
i k o i o k t i v t o k i

X
T X
O X
K X
I X
T X
O X
K X
H
t
X
O X
E
0
þ caik Xatiko þ chhk Xhhko  f e Q eo ; ð8Þ
t o k i t o k h o e

X
T X
O X
I X
T X
O X
K X
I X
T X
O X
K X
I X
T X
O X
K
Min OFV 2 ¼ t 0io ðX tiko þ Xatiko Þ þ tiko X tiko þ taiko Xatiko þ W tko ktko ; ð9Þ
t o i t o k i t o k i t o k

s:t:
yio 6 zi ; 8i; o; ð10Þ
y0ko 6 z0k ; 8k; o; ð11Þ
X
I
yio P 1; 8o; ð12Þ
i
X
K
y0ko P 1; 8o; ð13Þ
k

X v tv i 6 wtv ; 8i; v ; t; ð14Þ

X
I
wtv 6 X v tv i ; 8v ; t; ð15Þ
i
X
V
yio 6 X v tv i ; 8i; o; t; ð16Þ
v
X
V
X v tv i 6 1; 8i; t; ð17Þ
v
X tiko ¼ 0; jmaxft iko ; thhk g > ISC o or t iko > tpo ; 8i; k; o; h; t; ð18Þ
Xatiko ¼ 0; jmaxftaiko ; thhk g > ISC o ; 8i; k; o; h; t; ð19Þ
X tiko 6 Sti M tio uti y0ko ; 8
i; k; o; t; ð20Þ
t t t t 0
Xaiko 6 Si Mio ui yko ; i; k; o; t;8 ð21Þ
X
K
X tiko 6 Sti M tio uti yio ; 8i; o; t; ð22Þ
k
X
K
Xatiko 6 Sti M tio uti yio ; 8i; o; t; ð23Þ
k

X
H
t
X
I X
I
Xhhko ¼ X tiko þ Xatiko ; 8k; o; t; ð24Þ
h i i
X
K
t
Xhhko P xho Dtho ; 8h; o; t; ð25Þ
k

Q eo 6 yeo ; 8e; o; ð26Þ

Q eo 6 y0eo ; 8e; o; ð27Þ
yeo þ y0eo  Q eo 6 1; 8e; o; ð28Þ
 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124 107

X
I
zi ¼ NI; ð29Þ
i
X
K
z0k ¼ NK; ð30Þ
k

zi ; z0k ; yio ; y0ko ; wtv ; X v tv i ; Q eo 2 f0; 1g; 8i; k; o; v ; t; e; ð31Þ

t
X tiko ; Xatiko ; Xhhko P 0; Integer; 8i; k; o; h; t: ð32Þ
It should be noted that in a location–allocation problem, location decisions are strategic decisions and do not change over
time. However, some allocation decisions (e.g., making contract with a specific shipping agent or allocating a hospital to a
shipping agent) are tactical decisions and needs to be changed over time. In this problem, some decisions (e.g., determining
the location and equipment of a hospital) are strategic decisions. In practice, when a hospital is established and equipped, it
is not rational or cost-efficient to be relocated. Therefore, the costs of these decisions are taken into account just one time
over the time horizon. However, other decisions (e.g., allocating a hospital to a specific shipping agent or the amount of flows
between each pair of nodes in the network) are changed periodically due to different policies, and are treated accordingly in
the model. For example, in a peak rate of demand and supply, it is reasonable to use air flow transportation, while in periods
with lower rates, using other transportation modes is recommended.
The first objective function (8) minimizes the total cost, including costs of establishing different centers, transportation
costs among facilities, while considering the saving costs of integrating facilities. The second objective function (9) mini-
mizes the total time, including operation time of surgery in hospitals, travelling time among facilities, and waiting time
in queues in TCs.
Constraints (10) and (11) guarantee that a hospital and a TC can only be equipped for a particular type of organ if the
associated facility is already established. Constraints (12) and (13) ensure that there exists at least one hospital and one
transplant center treating each organ. Constraint sets (14) and (15) guarantee that only if a shipping agent is selected,
the flow to the hospital can take place in that period, and vice versa. Constraints (16) show that the assignment of shipping
agent to a hospital can only be done if the hospital is established. Constraint (17) indicates that each established hospital
should be covered by only one shipping agent in each period. Constraint (18) ensures that the time period for delivering an
organ from the hospital to the TC, or for a recipient to arrive at the TC, should not exceed the cold ischemia time; otherwise,
the flow of delivering an organ from the hospital to the TC is set to be 0. In addition, if the travel time from the hospital to
the TC exceeds tpo for organ o(but remains less than the cold ischemia time), the flow between these facilities should be via
air. Similarly, constraints (19) ensures that the travel time from the hospital to the TC or arrival time of a recipient from its
zone should not exceed the cold ischemia time of organ type o. Constraints (20) and (21) ensure that only the ground and
air flows to TCs are feasible if the TC is equipped for that organ type. In addition, the upper bounds of the flows are
observed. Eqs. (22) and (23) ensure that the ground and air flows from hospitals to TCs are only feasible if the hospital
is equipped for that organ type, while the upper bound of the flow is observed. Constraint (24) is balancing equation which
indicate that the total flows from recipients’ zones to TCs are equal to the total flows from hospitals to TCs in period t. Con-
straint (25) ensures that in each period, at least xho percent of the total demand for each organ type at each zone should be
covered by the available supplies. Constraints (26)–(28) define the integrated facilities. Constraints (29) and (30) calculate
the number of hospital and TC locations in the network, respectively. Finally, constraint (31) and (32) are domain
constraints.

Remark 1. Constraint sets (18) and (19) can also be written as follows. Let y00t
iko be a binary variable equals to 1 if there is flow
X tiko ; and 0, otherwise. For that case, we should define constraint sets as follows:

X tiko 6 M 0 y00t
iko ; 8i; k; o; t; ð33Þ
t
y00t
iko 6 X iko ; 8i; k; o; t; ð34Þ
y00t
iko 2 f0; 1g; 8i; k; o; t: ð35Þ
where M0 is a sufficient large number. Separating constraint set (18) results to:

X tiko ¼ 0; jtiko > ISC o ; 8i; k; o; t; ð36Þ

X tiko ¼ 0; jthhk > ISC o ; 8i; k; o; h; t; ð37Þ
X tiko ¼ 0; jtiko > tpo ; 8i; k; o; t: ð38Þ
We will have:
ðt iko  ISC o Þy00t 00t
iko 6 eyiko ; 8i; k; o; t; ð39Þ
00t 00t
ðthhk  ISC o Þyiko 6 eyiko ; 8i; k; o; h; t; ð40Þ
ðt iko  tpo Þy00t 00t
iko 6 eyiko ; 8i; k; o; t: ð41Þ
108 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124

v=1

h=1
v=2
22 k=1 i=2

h=2 23
41
i=1,
40 i=3 v=3
k=2
45
h=3
v=4

Fig. 2. Optimal solution considering objective function 1.

v=1

h=1
v=2
k=1 45 i=2

h=2
i=1,
i=3 v=3
k=2

h=3
v=4

Fig. 3. Optimal solution considering objective function 2.

Table 1
Parameter ranges for small-scale problems.

Parameters Values Parameters Values

ci Uniform (4800, 5100) caik Uniform (0.27, 0.29)
c0k Uniform (28000, 31000) tiko Uniform (11, 44)
rio Uniform (100, 130) thhk Uniform (2, 7)
r0ko Uniform (300, 600) taiko Uniform (2, 4)
0
fe Uniform (2500, 3500) tpo Uniform (16, 18)
Sti Uniform (25, 60) ISCo Uniform (5, 13)
uti Uniform (0, 1) t0 io Uniform (5, 7)
Dtho Uniform (35, 80) ltko Uniform (350, 650)
cvvi Uniform (2, 4) xho Uniform (0.7, 0.9)
cik Uniform (0.14, 0.25)

where e is a very small positive number. If the difference of the left hand side (LHS) is positive, y00t
iko has to be 0, while if the
difference of the LHS is negative or zero, y00t
iko will be either 0 or 1. Constraint set (19) can also be similarly expanded.

3.4. Model validation

In order to validate the proposed model, a small-size test problem (i.e., jOj  jVj  jIj  jKj  jHj  jTj = 2  4  3 
2  3  3) is designed and the generated solution using different objective functions are depicted in Figs. 2 and 3. The param-
eters are listed in Table 1. It should be noted that deterministic values, generated from probability distributions listed in
Table 1, are used in this section.
In Figs. 2 and 3, the grey nodes indicate the selected facilities and the shaded nodes refer to the integrated ones. Also,
dotted lines indicate air flow and the number of transferred organs and patients has been shown on each corresponding link.
The associated results are demonstrated for a selected organ (i.e., o = 1), and t = 1. Since both flows and assigned facilities in
each solution are different, it can be concluded that the results from using different objective functions are not necessarily
consistent, and therefore objective functions should be considered separately.

4. Crisp counterpart formulation

In this section, the uncertain parameters of the proposed SCND are first presented. Then, a hybrid two-stage approach is
introduced to convert the model into its auxiliary crisp counterpart.
 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124 109

Table 2
Uncertain parameters.

Parameters Parameters
cei ¼ ðcið1Þ ; cið2Þ ; cið3Þ Þ cfv v i ¼ ðcv v ið1Þ ; cv v ið2Þ ; cv v ið3Þ Þ
~c0k ¼ ðc0kð1Þ ; c0kð2Þ ; c0kð3Þ Þ ~cik ¼ ðcikð1Þ ; cikð2Þ ; cikð3Þ Þ
rfio ¼ ðr ioð1Þ ; r ioð2Þ ; r ioð3Þ Þ f
ca ik ¼ ðcaikð1Þ ; caikð2Þ ; caikð3Þ Þ
~r0ko ¼ ðr 0koð1Þ ; r0koð2Þ ; r0koð3Þ Þ ~tiko ¼ ðtikoð1Þ ; tikoð2Þ ; tikoð3Þ Þ
~f 0 ¼ ðf 0 ; f 0 ; f 0 Þ f
th hk ¼ ðthhkð1Þ ; thhkð2Þ ; thhkð3Þ Þ
e eð1Þ eð2Þ eð3Þ
e
S ti ¼ ðStið1Þ ; Stið2Þ ; Stið3Þ Þ e iko ¼ ðtaikoð1Þ ; taikoð2Þ ; taikoð3Þ Þ
ta
De t ¼ ðDt
ho
t t
hoð1Þ ; Dhoð2Þ ; Dhoð3Þ Þ te0 io ¼ ðt0ioð1Þ ; t0ioð2Þ ; t 0ioð3Þ Þ

4.1. Uncertain parameters

In practice, due to the nature of data in SCND problems, some parameters are uncertain. Therefore, in this paper we con-
sider establishment costs, costs of equipping facilities, saving costs in integrated facilities, transportation costs and travel
time, supplies, demands, and required time for removal process to be possibilistic data in the form of triangular fuzzy num-
bers as shown in Table 2.

4.2. Proposed possibilistic programming

To solve the proposed mixed integer programming model, a new hybrid two-stage approach is presented in this section.
In the first stage, the model is converted into its crisp counterpart by applying an efficient solution methodology based on Me
measure proposed by Xu and Zhou (2013). In the second stage, to accommodate both objective functions, Torabi and Hassini
(TH) approach (Torabi and Hassini, 2008) is used.

4.2.1. Equivalent auxiliary crisp model

To cope with the uncertainty of the constraint sets, among the available possibilistic programming approaches, we have
utilized the chance constrained programming approach. It should be noted that in this industry, basic measures (i.e., Pos and
Nec) are set to be limited and extremely optimistic and pessimistic. Following Xu and Zhou (2013), we present measure Me
which is adjustable according to the varying attitudes of the decision maker. To convert the proposed chance constraints
with the Me measure into their crisp counterparts, we start with a compact form of the basic possibilistic chance-constraint
programming model as follows (see Xu and Zhou, 2013; Heilpern, 1992; Liu and Iwamura, 1998; Pishvaee et al., 2012):
Min E½Z ¼ E½~cy
s:t:
~ P a;
MefAx P dg ð42Þ
e P b;
MefSx 6 Nyg
y 2 f0; 1g; x P 0:
In the formulation (42), it is assumed that the chance constraints should be satisfied with the minimum confidence levels
of a and b.
The above model can be transformed into two models; namely, the lower approximation model (LAM) and the upper
approximation model (UAM).
8
> Min E½Z ¼ E½~cy
>
>
>
> s:t: n
>
> o
< ~ P a;
ðLAMÞ Pos Ax P d ð43Þ
>
> n o
> Pos Sx 6 Ny
> e P b;
>
>
>
:
y 2 f0; 1g; x P 0
and
8
> Min E½Z ¼ E½~cy
>
>
>
> s:t: n
>
> o
< ~ P a;
ðUAMÞ Nec Ax P d ð44Þ
>
> n o
>
> Nec Sx 6 Nye P b;
>
>
>
:
y 2 f0; 1g; x P 0:
The crisp counterpart formulation of the above model can be stated by:
110 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124

8  
>
> Min E½Z ¼ ð1nÞ cð1Þ þ 12 cð2Þ þ 2n cð3Þ y
>
> 2
>
>
>
< s:t:
ðLAMÞ Ax P dð2Þ  ð1  aÞðdð2Þ  dð1Þ Þ; ð45Þ
>
>   
>
>
>
> Sx 6 Nð2Þ þ ð1  bÞ Nð3Þ  Nð2Þ y;
>
:
y 2 f0; 1g; x P 0

and

8  
> ð1nÞ
cð1Þ þ 12 cð2Þ þ 2n cð3Þ y
> Min E½Z ¼
>
> 2
>
>
>
< s:t:
ðUAMÞ Ax P dð2Þ þ aðdð3Þ  dð2Þ Þ; ð46Þ
>
>  
>
>
>
> Sx 6 Nð2Þ  bðNð2Þ  Nð1Þ Þ y;
>
:
y 2 f0; 1g; x P 0;

where n (0 6 n 6 1) is the optimistic-pessimistic parameter to define the combined attitude of the decision maker (DM) (Xu
and Zhou, 2013). So, the crisp equivalent of the proposed model can be formulated by:

 LAM:

XI
 XI

ð1  nÞ 1 n ð1  nÞ 0 1 n
Min OFV 1 ¼ cið1Þ þ cið2Þ þ cið3Þ zi þ ckð1Þ þ c0kð2Þ þ c0kð3Þ z0k
i
2 2 2 i
2 2 2
X
O X I
 XO X K

ð1  nÞ 1 n ð1  nÞ 0 1 n
þ r ioð1Þ þ r ioð2Þ þ rioð3Þ yio þ r koð1Þ þ r0koð2Þ þ r 0koð3Þ y0ko
o i
2 2 2 o k
2 2 2
XXX ð1  nÞ
T I V

1 n
þ cv v ið1Þ þ cv v ið2Þ þ cv v ið3Þ X v tv i
t i v 2 2 2
X
T X O X K X I

ð1  nÞ 1 n
þ cikð1Þ þ cikð2Þ þ cikð3Þ X tiko
t o k i
2 2 2
X
T X O X K X I

ð1  nÞ 1 n
þ caikð1Þ þ caikð2Þ þ caikð3Þ Xatiko
t o k i
2 2 2
X
T X O X K X H

ð1  nÞ 1 n t
þ chhkð1Þ þ chhkð2Þ þ chhkð3Þ Xhhko
t o k h
2 2 2
X
O X E

ð1  nÞ 0 1 0 n 0
 f eð1Þ þ f eð2Þ þ f eð3Þ Q eo ð47Þ
o e
2 2 2

XT X O X I

ð1  nÞ 0 1 n
Min OFV 2 ¼ t ioð1Þ þ t 0ioð2Þ þ t0ioð3Þ ðX tiko þ Xatiko Þ
t o i
2 2 2
X
T X O X K X I

ð1  nÞ 1 n
þ t ikoð1Þ þ t ikoð2Þ þ t ikoð3Þ X tiko
t o k i
2 2 2
X
T X O X K X I
 X
T X
O X
K
ð1  nÞ 1 n
þ taikoð1Þ þ taikoð2Þ þ taikoð3Þ Xatiko þ W tko ktko ð48Þ
t o k i
2 2 2 t o k

s:t:
h  i
X tiko 6 Mtio uti y0ko Stið2Þ þ ð1  aÞ Stið3Þ  Stið2Þ ; 8i; k; o; t; ð49Þ
h  i
Xatiko 6 M tio uti y0ko Stið2Þ þ ð1  aÞ Stið3Þ  Stið2Þ ; 8i; k; o; t; ð50Þ
X
K h  i
X tiko 6 M tio uti yio Stið2Þ þ ð1  aÞ Stið3Þ  Stið2Þ ; 8i; o; t; ð51Þ
k
 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124 111

X
K h  i
Xatiko 6 Mtio uti yio Stið2Þ þ ð1  aÞ Stið3Þ  Stið2Þ ; 8i; o; t; ð52Þ
k
X
K h  i
t
Xhhko P xho Dthoð2Þ  ð1  aÞ Dthoð2Þ  Dthoð1Þ ; 8h; o; t; ð53Þ
k

Constraints (1)–(19), (24), (26)–(32).

and
 UAM:

Min E½OFV 1  ð47Þ

Min E½OFV 2  ð48Þ

s:t:
h i
X tiko 6 M tio uti y0ko Stið2Þ  aðStið2Þ  Stið1Þ Þ ; 8i; k; o; t; ð54Þ
h i
Xatiko 6 Mtio uti y0ko Stið2Þ  aðStið2Þ  Stið1Þ Þ ; 8i; k; o; t; ð55Þ

X
K h i
X tiko 6 M tio uti yio Stið2Þ  aðStið2Þ  Stið1Þ Þ ; 8i; k; o; t; ð56Þ
k

X
K h i
Xatiko 6 Mtio uti yio Stið2Þ  aðStið2Þ  Stið1Þ Þ ; 8i; k; o; t; ð57Þ
k

X
K
t
Xhhko P xho ½Dthoð2Þ þ aðDthoð3Þ  Dthoð2Þ Þ; 8h; o; t: ð58Þ
k

Constraints (1)–(19), (24), (26)–(32).

4.3. Proposed solution approach

Various approaches have been developed to solve the crisp multi-objective programming models in the literature includ-
ing priori, interactive and posteriori methods (Vahdani et al., 2012). Among these methods, fuzzy interactive methods are
one of the most attractive approaches because of their ability in measuring and adjusting the satisfaction level of each objec-
tive function based on the decision maker’s preferences in an interactive and progressive way (Torabi and Hassini, 2008). In
this paper, to solve the developed model, a hybrid solution approach is utilized by combining the presented possibilistic pro-
gramming in Section 4.1 with the TH method (Torabi and Hassini, 2008). Despite the classical multi-objective programming
methods such as the weighted sum method that may result in weakly efficient solutions, the TH method guarantees to find
efficient solutions. Steps of the proposed hybrid method can be summarized as follows:

Step 1: Determine the positive ideal solution (PIS) and negative ideal solution (NIS) for each objective function for both LAM
and UAM.
Step 2: Determine a linear membership function for each objective function in each LAM and UAM as:
8
>
> 1 if OFV j < OFV PIS
j
>
>
< OFV NIS OFV
if OFV NIS 6 OFV j 6 OFV PIS
j
lj ðxÞ ¼ j
OFV NIS OFV PIS j j ð59Þ
>
> j j
>
>
:
0 if OFV j > OFV NIS
j

where lj(x) denotes the satisfaction degree of the jth objective function.
Step 3: For each models (LAM and UAM), convert the equivalent crisp bi-objective models into a single-objective model
using the TH’s aggregation function (Torabi and Hassini, 2008), which is computed by:
X
Max kðxÞ ¼ uk0 þ ð1  uÞ hj lj ðxÞ
j

s:t: ð60Þ
k0 6 lj ðxÞ;
x 2 FðxÞ; k0 and k 2 ½0; 1:
112 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124

where F(x) and k0 denotes the feasible region involving some constraints of the equivalent crisp model and the min-
imum satisfaction degree of objectives, respectively. In addition, hj and u indicate the relative importance of the j  th
objective function and the coefficient of compensation, respectively, which are usually determined by decision mak-
P
ers’ preferences ( j hj ¼ 1; hj > 0Þ.
Step 4: Determine the values of the uncertainty levels (a) and the coefficient of compensation (u) and importance weight of
the objectives (hj), and solve the single-objective model for each model with possibility and necessity chance con-
straints to obtain an interval of the upper and lower bounds of both objectives. If the current solution satisfies the
decision maker, stop; otherwise change the values of parameters to provide another solution (Torabi and Hassini,
2008; Vahdani et al., 2012).

5. Lower bound and meta-heuristic algorithms

The model presented in Section 3 is a mixed-integer nonlinear programming (MINLP) optimization problem. It is nonlin-
ear because of nonlinear form of the objective function (9). In order to solve these problems, several optimization techniques
such as branch-and-bound, branch-and-cut and branch-and-price have been developed in literature. However, solving such
an MINLP problem is very computationally challenging especially for large-size instances.
The presented model was coded in GAMS 22.9 software. The reliability and validity of the proposed model was tested on
some randomly generated data sets of the problem on a Pentium 4 computer with 2.53 GHz CPU and 6 GB RAM. These prob-
lems were successfully solved to optimality using the nonlinear branch and reduce approach in BARON solver. However, the
computational time increases exponentially with the increase of problem size. For example, it may take several days to solve
instances up to four organs, eight nodes for hospitals and TCs, 4 periods and 6 shipping agents. Further attempts to solve
industrial cases were unsuccessful due to required long computational time. To address this limitation, we developed two
efficient meta-heuristics, i.e., a simulated annealing (SA) and a hybrid differential evolution (DE) algorithms to find accept-
able (i.e., near optimal) solutions in a reasonable amount of computational time. Furthermore, in order to evaluate the per-
formance of these meta-heuristics, a lower bound was developed following Ishfaq and Sox (2011) approach.

5.1. Lower bound procedure

In order to evaluate the effectiveness of the proposed meta-heuristic algorithms, a lower bound procedure is presented in
this section. Following Vahdani et al. (2012), a partial relaxation of a subproblem of the original problem P is used. In this
procedure, a subproblem of P can be created by dividing the set of hospitals and TCs distinctly, in which any two sets of TC
locations, Si and Sj, are distinct if there is at least one TC k 2 Si and k R Sj. It is also similar for hospital facilities. There will be
C jKj jIj
NK ¼ K!=ððK  NKÞ!NK!Þ and C NI ¼ I!=ððI  NIÞ!NI!Þ number of distinct TC and hospital sets in a network, respectively. There-
fore, the set of total sub-problems is SKI with size of C jKj jKj
NK  C NI . For a selected sub-problem from set SKI say Si where
0
jS i j ¼ NI  NK, corresponding location decision variables zi and zk are fixed at 1, and the P is then reduced to a mixed-integer,
non-linear assignment sub-problem (SPd) over the binary decision variables zi and z0k . Now let RP(SPd) be the partial relax-
ation of SPd that relaxes the integrality of some of other binary variables, i.e., we assume that: 0 < yio ; y0ko ; wtv < 1. Besides,
let OFVRP(SPd) be the optimal objective function value of RP(SPd). Since this value can be declared as a lower bound for the
optimal value of SPd, the minimum value among these values from the possible KI sets determines a lower bound for the
optimal value of the original problem P where OFVLB(P) = mindOFVRP(SPd).
In order to determine a valid lower bound for P, the lower bound procedure must find a global optimal solution for each
mixed-integer nonlinear optimization subproblem RP(SPd). Note that the objective function (8) and all constraints in RP(SPd)
are linear, but objective function (9) is nonlinear. Therefore, to guarantee that the MINLP solver will find a true global optimal
solution for each subproblem RP(SPd), the convexity of the objective function (9) needs to be verified.

Lemma 5.1.1. The waiting time expression in Eq. (7) is convex in k.

Proof. This is done by showing that the nonlinear equation (7) is convex in k and subsequently showing that it is also convex
in X, as k is linear function of X. When the other parameters are constant, Eq. (7) is of the form of Eq. (61) as follow:
1
W¼ ; ð61Þ
ðcl  kÞ  kA ð1  DkÞ2
where A and D are constant. It can be shown that for k < cl,
 
2 Aþ2
dW ðA þ 3ÞD k þ ð2clðA þ 1ÞD þ A þ 1ÞkA  clðA þ 2ÞD2 þ 2ðA þ 2ÞD kAþ1  clAkA1
¼ > 0: ð62Þ
dk ðcl  kÞ2  k2A ð1  DkÞ4
d2 W
In addition, for the second order derivative, it can be shown that dk2
> 0 for k < cl. Both of first and second derivatives
were numerically verified in MATLAB software. h
 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124 113

NFC ← 0
Set the parameters of SA (Mutation No., , )
Create initial solution
Terminate ← false
While (terminate = false) do
Mutate each initial solution ← Mutation No., Inversion, Reversion & Swap strategies
Find the best solution among new created solutions ← Smaller OFV
Update NFC
If (OFVnew solution < OFVcurrent solution) then
Accept the new solution
Else
Calculate
Apply probable acceptance function ← ← &T
Create random value r
If (r < P) then
Accept new solution
Else
Hold current solution
End if
End if
Update NFC
T←
If (NFC = predeined value) then
Terminate = true
End if
End while

Fig. 4. Pseudo code of the proposed SA.

Lemma 5.1.2. Eq. (7) is convex in X.

Proof. If a function F(Y) is convex in Y, and g(X) is an affine function of X, then F[g(X)] is convex in X (Bazaraa et al., 2013).
Note that Eq. (7) is convex in k. Since k is linear function of X in RP(SPd), it follows that Eq. (7) in RP(SPd) is convex in X. h

5.2. Meta-heuristic solution algorithms

In this section, two meta-heuristic algorithms (i.e., a SA and a hybrid DE) are developed for the presented model.

5.2.1. Simulated annealing (SA)

Simulated annealing (SA) is a path meta-heuristic based on annealing in solids. This method was first applied to optimi-
zation problems by Kirkpatrick et al. (1983). Indeed, it is a local search-based heuristic method designed to avoid being
trapped in poor local optima. It begins from a random initial solution. At each iteration, SA creates a new solution from
the predefined neighborhood of the current solution by specific mutation strategies. For more details about the implemen-
tation of different mutation strategies, the readers are referred to Mohammadi et al. (2013). Afterwards, the objective func-
tion value (OFV) of this new solution is subsequently compared with the current solution. If this OFV is better than the
current solution, the current solution will be replaced with the new one. Otherwise, a new solution
  may be accepted as
the new current solution, with a small probability determined by the Boltzmann function, exp kTDf ; where Df is the differ-
ence between OFVs, k is a predetermined constant and T is the current temperature (Metropolis et al., 1953; Kirkpatrick et al.,
1983; Mohammadi et al., 2014). Moreover, the number of function calls (NFC) is considered as stopping criterion. The pseudo
code of the proposed SA is presented in Fig. 4.

5.2.2. Differential evolution (DE)

In this section, a hybrid DE algorithm is developed to obtain solutions with higher quality than classical meta-heuristic
algorithms. First, a brief explanation of the DE algorithm is presented, and then the hybrid DE algorithm is described in
details.

5.2.2.1. DE algorithm in general. In order to find global optima for numerous unconstrained problems, Storn and Price (1997)
developed a simple and efficient heuristic, called differential evolution (DE) algorithm. Similar to other evolutionary compu-
tational algorithms, during DE, a population of solutions with size of PS (i.e., population size) is evolved using specific oper-
ators such as mutation and crossover (Kirkpatrick et al., 1983). The initial population is often randomly generated over the
variables domain. Each solution vector in the population has to be selected once as the target vector so thatPS competitions
take place in one generation. A new solution vector is generated by DE’s mutation operator, in which the weighted difference
114 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124

Fig. 5. Pseudo code of the SADE’s initialization phase.

Table 3
Mutation strategies.

Mutation strategy Operator

MS #1 Mi ¼ xbest þ Fðxv 2  xv 3 Þ
MS #2 Mi ¼ xv 1 þ Fðxv 2  xv 3 Þ
MS #3 Mi ¼ xbest þ Fðxv 1 þ xv 2  xv 3  xv 4 Þ
MS #4 Mi ¼ xi þ Fðxv 1 þ xv 2  xv 3  xv 4 Þ

between two population vectors is added to a third vector. Hence, this algorithm was named as differential evolution. Note
that these three vectors are randomly selected and must be different from the target vector; therefore, Pop must be at least 4.
Let xi, i = 1, . . . , Pop, be the target vector, a mutated vector is generated according to Eq. (63).

Mi ¼ xv 1 þ F ðxv 2  xv 3 Þ; ð63Þ

where v1, v2, and v3 are mutually different random indexes taking from {1, 2, . . . , Pop}, and are not equal to i. F in Eq. (63) is a
constant real value 2 [0, 2], which controls the amplification of the differential variation between the second and the third
randomly chosen population vectors. (i.e., xv2  xv3). Each mutated vector shares its information with target vector using
the crossover operation in order to create new solution si ¼ fsi1 ; . . . ; sij ; . . . ; siD g as Eq. (64).

M ij if randðjÞ 6 CR and j ¼ rnbrðiÞ;
sij ¼ ð64Þ
xij if randðjÞ > CR and j – rnbrðiÞ;
where rand(j) is the jth component of a D-dimensional uniform random number 2 [0, 1], CR corresponds to crossover rate,
and rnbr(i) is a randomly chosen index 2 {1, . . . , D} to ensure that at least one mutated dimensional value is used in the new
created solution. If the new created solution yields a lower cost function value than the target vector for a minimization
problem, then the new solution is replaced by the target vector in the following generation.

5.2.2.2. Proposed hybrid DE algorithm. In this paper, a new variant of DE algorithm, namely self-adaptive DE (SADE) algorithm,
is developed to obtain better solutions than classical DE. In the SADE, in addition to new crossover operator, a self-adaptive
mutation operator is also applied. This operator, called merging operator, is developed to function similar to assimilation
operator of imperialist competitive algorithm, discussed in Mohammadi et al. (2013). In the literature, several mutation
strategies have been introduced (Storn, 1996; Qin et al., 2009; Price et al., 2005; Iorio and Li, 2005). Simultaneous use of
 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124 115

Fig. 6. Merging operator with a random angle h.

all of them in an algorithm, however, significantly increases the computational time. Therefore, many articles have used only
one of them to search the solution space. On the contrary, in the proposed SADE algorithm, we simultaneously use several
mutation operators while keeping the computation time low. The proposed SADE includes an initialization phase in which
each mutation operator obtains a score if it could create a better solution at each iteration. At the end of the initialization
phase, which is limited by a predefined number of iterations, a selection probability (SP) metric is calculated for each oper-
ator by dividing the obtained score by the number of iterations. It should be noted that the sum of all selection probabilities
is equal to 1. When the initialization phase is finished, the main phase of SADE is started which includes searching the solu-
tion space by using the self-adapted mutation and merging operators. The Pseudo code of the initialization phase of the pro-
posed SADE algorithm is shown in Fig. 5.
In the SADE algorithm, a new solution vector is generated by a mutation strategy selected through the roulette wheel
competition based on SP metric among all strategies (see Table 3). In the first strategy, two randomly selected vectors
(i.e., solutions) are first deducted from each other (i.e., xv2  xv3) and then combined with the best solution in the popula-
tion (i.e., xbest). In the second strategy, two deducted vectors are combined with another randomly selected vector from the
population (i.e., xv1). Strategies number 3 and 4 are similar to strategies number 1 and 2, respectively. In the strategies num-
ber 3 and 4, not only four randomly selected vectors are first combined as (xv1 + xv2  xv3  xv4), but also they are com-
bined with the best and another randomly selected vectors, respectively. Afterward, the mutated vector is moved toward the
target vector. This movement is shown in Fig. 6, in which d is the distance between mutated and target vectors, X is the
amount of mutated vector’s movement toward the target vector that belongs to a uniform interval between 0 and b  d,
and b is a randomly generated number greater than 1. For mutated vector’s movement toward target vector in other direc-
tion, an h angle is applied with the uniform distribution between c and c. It has been estimated that p/4 is the most proper
value of c (Calégari et al., 1999). Notably, target vector can be a randomly selected vector or the best vector among the whole
population (i.e., xbest).

6. Computational experiments

In this section, to validate the proposed model and the possibilistic programming approach, several medium-size exam-
ples are presented. The results are evaluated for the deterministic model and the model under uncertainty. Distribution func-
tions of randomly generated samples follow Table 1. The performances of developed meta-heuristic algorithms, in terms of
both solution quality and computation time, are evaluated in a benchmarking study.

6.1. Sensitivity analyses and managerial insights

In this section, the lower and upper approximation models are developed under different uncertainty levels a (i.e.,
a = 0.1–0.9) and values of k (i.e., k = 0.1–0.9), for each objective function as shown in Tables 4 and 5, respectively. The objec-
tive functions of each model are then aggregated by the TH approach (Torabi and Hassini, 2008). The intervals for objective
function values (OFVs) are computed for different values of u (i.e. u = 0.1, 0.5, 0.9) and hj (i.e.hj = 0.1, 0.5, 0.9), and are pre-
sented in Tables 6 and 7.

Table 4
OFVs vs. alteration of a (k = 0.2).

Problem size Uncertainty level (a) Objective function values (LAM, UAM)
jOj  jVj  jI j  jKj  jHj  jTj (OFV1) (OFV2)
365333 0.1 (70,769.289, 71,385.258) (714.2700, 996.850)
0.2 (70,773.009, 71,389.049) (738.8400, 1038.40)
0.3 (70,918.929, 71,532.134) (763.410, 1082.355)
0.4 (71,062.649, 71,535.940) (787.980, 1129.640)
0.5 (71,066.369, 71,539.746) (815.325, 1176.925)
0.6 (71,077.589, 71,819.916) (843.780, 1222.960)
0.7 (71,228.810, 71,823.850) (872.235, 1267.920)
0.8 (71,232.530, 80,459.917) (900.690, 1319.100)
0.9 (71,239.247, 80,465.130) (928.420, 1367.750)
116 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124

Table 5
OFVs vs. alteration of k (a = 0.3).

Problem size Values of k Objective function values (LAM, UAM)

jOj  jVj  jI j  jKj  jHj  jTj (OFV1) (OFV2)
365333 0.1 (70,667.338, 71,260.549) (738.105, 1047.165)
0.2 (70,918.929, 71,532.134) (763.410, 1082.355)
0.3 (71,170.520, 71,803.720) (788.715, 1117.545)
0.4 (71,422.112, 72,075.306) (814.020, 1152.735)
0.5 (71,673.703, 72,346.892) (839.325, 1187.925)
0.6 (71,925.294, 72,618.478) (864.630, 1223.115)
0.7 (72,176.886, 72,890.064) (889.935, 1258.305)
0.8 (72,428.477, 73,161.650) (915.240, 1293.495)
0.9 (72,680.068, 73,433.235) (940.545, 1328.685)

Table 6
OFVs vs. different values of u based on (hj = 0.5, a = 0.3, k = 0.2).

Problem size (u) Objective function value under the TH approach

jOj  jVj  jIj  jKj  jHj  jTj (OFV1) (OFV2)
365333 0.1 (89,011.729, 89,342.099) (1206.685, 1573.725)
0.5 (89,005.729, 89,342.099) (1208.185, 1573.725)
0.9 (83,498.506, 83,986.639) (2275.570, 2817.720)

Table 7
OFVs vs. different values of hj based on (u = 0.5, a = 0.3, k = 0.2).

Problem size hj Objective function value under the TH approach

jOj  jVj  jI j  jKj  jHj  jTj (OFV1) (OFV2)
365333 0.1 (89,011.729, 89,342.099) (1206.685, 1573.725)
0.5 (89,005.729, 89,342.099) (1208.185, 1573.725)
0.9 (83,498.506, 83,980.739) (2275.570, 2817.820)

Table 8
OFVs vs. different numbers of NI and NK.

NI NK
1 2 3
1 (Infeasible, Infeasible) (Infeasible, Infeasible) (Infeasible, Infeasible)
2 (Infeasible, Infeasible) (46,367.69, 2976) (76,768.2, 2883)
3 (48,890.285, 4128) (51,328.69, 1043) (81,729.20, 950)
4 (53,852.005, 2149) (56,416.54, 1003) (86,817.05, 873)
5 (59,152.005, 2149) (61,516.54, 1003) (91,917.05, 873)

Fig. 7. OFV 1 vs. xho.

Two parameters which significantly affect the structure of the network and consequently the value of OFs, are NK and NI,
the total number of selected hospitals, and TCs, respectively. Therefore, the sensitivity analysis on these parameters can pro-
vide valuable managerial insights. Table 8 shows a sensitivity analysis on different values of NI and NK considering deter-
B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124 117

Fig. 8. OFV 2 vs. xho.

Fig. 9. OFV 1 vs. saving costs for the facility integration.

Fig. 10. OFV 1 vs. tpo.

Fig. 11. OFV 2 vs. tpo.

118 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124

Table 9
SA and SADE parameters settings.

Algorithm Parameter Value

SA T0 12
a 0.95
Mutation No. 8
NFC 40,000
SADE Pop 250
F 0.8
b 2
NFC 40,000
c p/4

ministic parameters. As shown in Table 8, although an increase in NI and NK, will increase the total costs (i.e., first objective),
the total time (i.e., second objective) will be decreased. In this case, there will be more facility to service the patients, and
consequently, the nearest facility will be utilized resulting in less time. Please note in case of NI 6 4, at least one organ is
allocated to each hospital. In this situation, each hospital should be served by a shipping agent. Therefore, by increasing
the number of NI, the total time is decreased. However, in case of NI = 5, the network structure does not change. It means
that no organ is allocated to a new located hospital. Hence, although the total cost is increased, the total time is not
decreased.
Figs. 7 and 8, show the sensitivity analysis on different values of xho (i.e., minimum target satisfaction level of demand)
for each objective function. Increasing the value of xho, forces the model to cover more demand points, which results in
establishing more facilities to satisfy the total demand. As can be seen in Fig. 7, increasing the value of xho between 0.6
and 0.8 does not have a significant effect on the slope of the total costs’ graph. It can be concluded that the main cost in this
interval is due to transportation costs which are negligible compare to the other costs (i.e., establishing costs of facilities).
While, by increasing the value of xho, to 0.9, the slope significantly increases, which is due to the costs for establishment
new facilities. For the second objective function value, by increasing xho, the waiting time in the queue increases and as
a consequence, the OFV 2 will be increased with an approximately constant rate.
Fig. 9 shows the changes in OFV 1 with changes in f0 e (i.e., saving costs in a case of the facility integration). According to this
figure, before a certain threshold (here, 2700 of saving costs), the facilities are not integrated. After this threshold, some facil-
ities are integrated and the OFV 1 is decreased.
Figs. 10 and 11 illustrate the changes of OFVs vs. changes of tpo. By increasing the value of tpo, the total cost is decreased,
because all shipments are transported by ground mode, which is cheaper than air transportation. Consequently, in this case
the total time will be increased. For OFV 1 , for the values of tpo between 25 and 26 h, a significant decrease in total cost indi-
cates that most of the travel times from hospital to TCs are less than 26 h, and therefore at this tpo most of flows are forced to
be transported by ground transportation. This behavior happens in the interval of [23, 26] for OFV 2 , in which a significant
increase in the total time value is observed. It should be noted that the slope of OFV 2 in the interval of [25, 26] is (almost
58%) greater than that of for the values the interval of [23, 25]. In the interval [25, 26], choosing the appropriate value of
tpo depends on the DM preference and the priority of each objective.

Fig. 12. Comparison between ZLB and TH of the optimal solution.

 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124 119

Table 10
APRG and CPU Time (in second) for small and medium size.

Data set SA SADE

Replications Gap Time Replications Gap Time
1 2 3 4 5 1 2 3 4 5
1 .05 .06 .06 .07 .05 .05 45 .03 .04 .04 .05 .03 .03 56
2 .06 .07 .05 .08 .08 .06 46 .04 .05 .03 .05 .05 .04 57
3 .08 .09 .11 .10 .09 .09 48 .05 .07 .05 .07 .06 .06 59
4 .09 1.0 1.1 1.0 1.1 .85 49 .07 .78 .77 .77 .84 .64 63
5 1.0 1.2 1.3 1.1 1.2 1.2 51 .76 .95 .92 .85 .94 .88 64
6 1.3 1.4 1.1 1.2 1.4 1.3 54 .92 1.1 .86 .87 1.1 .97 66
7 1.5 1.3 1.6 1.3 1.5 1.4 57 1.0 0.9 1.2 .91 1.1 1.0 69
8 1.6 1.5 1.6 1.7 1.7 1.6 59 1.2 1.1 1.1 1.3 1.2 1.1 75
9 1.8 1.9 1.8 1.7 1.8 1.8 63 1.4 1.5 1.4 1.2 1.2 1.3 76
10 1.9 2.0 2.1 1.9 2.0 1.9 68 1.5 1.5 1.4 1.3 1.4 1.4 79
11 2.1 2.2 2.2 1.9 2.1 2.1 72 1.5 1.7 1.6 1.4 1.6 1.5 89
12 2.3 2.4 2.3 2.4 2.4 2.4 74 1.8 1.8 1.6 1.9 1.7 1.7 94
13 2.5 2.2 2.3 2.6 2.6 2.4 78 1.9 1.6 1.4 1.9 2.0 1.7 96
14 2.7 2.7 2.9 2.9 2.8 2.8 84 1.9 2.0 2.0 2.1 2.0 2.0 112
15 2.9 2.9 3.1 3.2 3.1 3.0 89 2.2 2.0 2.2 2.1 2.1 2.1 121

Table 11
APRG and CPU Time (in second) for large size instances.

Data set SA SADE

Replications Gap Time Replications Gap Time
1 2 3 4 5 1 2 3 4 5
16 3.6 3.8 3.9 3.6 3.7 3.7 103 2.3 2.4 2.3 2.2 2.3 2.3 129
17 3.7 3.9 3.6 3.6 3.8 3.7 104 2.2 2.3 2.4 2.2 2.2 2.2 132
18 4.6 4.3 4.7 4.6 4.6 4.5 123 2.6 2.7 2.7 2.5 2.5 2.6 134
19 4.8 4.9 4.7 4.7 4.6 4.7 127 2.5 2.6 2.5 2.7 2.8 2.6 142
20 5.3 5.2 5.1 5.2 5.3 5.2 134 2.8 2.9 2.6 2.7 2.7 2.7 159
21 5.2 5.4 5.5 5.5 5.5 5.4 136 2.6 2.8 2.8 2.6 2.9 2.7 163
22 5.6 5.6 5.4 5.7 5.7 5.6 137 3.1 2.9 2.9 2.8 2.9 2.9 169
23 6.9 7.1 7.0 6.9 7.2 7.0 146 3.2 3.3 3.2 3.4 3.3 3.2 174
24 7.2 7.3 7.3 7.4 7.3 7.3 149 3.3 3.4 3.5 3.4 3.3 3.3 174
25 7.4 7.4 7.8 7.6 7.6 7.5 163 3.5 3.6 3.5 3.5 3.6 3.5 178
26 8.3 8.8 8.7 8.4 8.5 8.5 164 3.8 3.9 3.8 3.8 3.9 3.8 187
27 8.8 8.7 8.9 9.0 8.9 8.8 166 3.8 3.9 4.1 3.8 3.9 3.9 192
28 12.5 12.6 12.8 12.9 13.0 12.7 194 4.4 4.5 4.6 4.5 4.5 4.5 246
29 13.7 13.9 14.1 14.3 13.9 13.9 214 4.7 4.6 4.7 4.6 4.6 4.6 269
30 15.3 15.5 15.2 15.4 15.3 15.3 234 4.9 4.8 4.9 4.8 4.8 4.8 285

Fig. 13. CPU time vs. the number of data set.

120 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124

Fig. 14. Mean gap vs. the number of data set.

6.2. Benchmarking study

In this section, the performance of developed meta-heuristic algorithms in terms of both solution quality and computa-
tion time is evaluated in a benchmark study over a number of randomly generated data sets. These data sets contain 30 test
problems ranging from small-size to large-size instances. Besides, for each problem, five different intervals of parameters
were created for better comparison. Therefore, in total 150 problem instances were tested. The parameters of SA and SADE
algorithms were tuned using the response surface methodology (RSM) (Calégari et al., 1999) and are presented in Table 9.
These parameter settings were pre-tested over randomly selected problem instances, before being used in the comparative
study.
The percentage of relative gap (PRG) measure is used to compare the SA and SADE solutions with lower bound solutions.
The PRG is calculated as ½100  ðOFV LB  OFV Meta Þ=OFV Meta , where OFVLB and OFVMeta are the TH objective function value of
the obtained solution for the lower bound approach and the meta-heuristic algorithm, respectively. As TH is in the form
of maximizing, OFVLB is bigger than OFVMeta. For calculating the OFVLB, the lower bound of each objective function is obtained
according to Section 5.1, and then the value of OFVLB is calculated following Eq. (50).
In addition, a comparison between OFVLB and the optimal solution based on TH objective function is carried out in order to
show the tightness of the lower bound approach using some small test problems (see Fig. 12). The mean gap of TH objective
function between lower bound and optimal solution is 1.1%. Therefore, the tightness of the lower bound is about 1.1% in
comparison to the optimal solutions.
The percentage of relative gaps for the five replications of each test problem were averaged and recorded as the average
percentage of relative gap (APRG). The APRG obtained by solving the problem instances using SA and SADE and the corre-
sponding average computational times are reported in Tables 10 and 11.
The average of percentage of relative gaps over all test instances for SA and SADE are 4.58% and 2.23%, respectively. The
minimum percentage of relative gap for SA and SADE are 0.05% and 0.03%, respectively. Besides, the maximum percentage of
relative gap for SA and SADE are 15.5% and 4.9%, respectively.
It can be seen that SADE outperforms SA in the context of presented problem. Since OFVLB < OFVOptimal < OFVMeta, then the
average gap between OFVLB and OFVOptimal is 1.1% (MG_LB). Therefore, it can be concluded that the mean gap between SA and
SADE algorithms with optimal solution are 3.48% (i.e., (MG_SA  MG_LB)%) and 1.13% (i.e., (MG_SADE  MG_LB)%), respec-
tively. The required computation time for the SA and SADE is also given in the last column of their corresponding parts in
Tables 10 and 11. As shown, the computation time of proposed meta-heuristic algorithms increases with the increase of
the problem size, but stays low. These results show that both SA and SADE approaches can find near optimal solutions in
a reasonable time while SADE performs absolutely better than SA algorithm.
In addition, Figs. 13 and 14 show the sensitivity of computation time and relative percentage gap with respect to the
change in problem size for both meta-heuristic algorithms. Furthermore, Fig. 13 shows that the computation time of the
SA algorithm is less than SADE, while Fig. 14 depicts the lower deviation gap for SADE in comparison to SA algorithm.

6.3. Case study

Regional planning is an important problem in developing countries, where location decisions are commonly made by
local officials mostly based on political, economical, and cultural factors. These decisions may result in a placement of facil-
ities that are far from being optimal in terms of geographical accessibility (Cocking et al., 2006). Improving accessibility of
patients with maximum efficiency (i.e., minimum cost) and maximum responsiveness (i.e., minimum time) is becoming vital
issue in such countries.
Designing an optimal transplant network to improve accessibility of patients with minimum cost and time for Iran is con-
sidered as an application of the problem under study. Iran has 31 provinces, out of which 22 are the most populated, and are
 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124 121

Fig. 15. Iran’s map and the candidate locations of nodes.

Fig. 16. LAM optimum facilities and their interactions.

considered as recipient zones in our study. According to the experts’ opinion, candidate locations of each center have been
determined and marked on a geographic map shown in Fig. 15. Noteworthy, the current location of centers are included as
the candidate locations of facilities. The organ transplant supply chain in Iran does not include shipping agents, thus they are
122 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124

Fig. 17. UAM optimum facilities and their interactions.

Table 12
Interval of fixed and transportation costs vs. different values of n.

n Fixed costs Range Transportation costs Range

0.2 [175,697, 223,483] 47,786 [8855.306, 44,586.574] 35731
0.4 [176,499, 224,889] 48,390 [9334.594, 45,406.828] 36072
0.6 [177,301, 225,771] 48,470 [9813.882, 46,005.787] 36191
0.8 [178,103, 226,653] 48,550 [10,293.17, 46,604.746] 36311

not considered in this study. Although obtaining the exact values of all parameters used by presented model are challenging,
these values can be estimated by the analysis of aggregate data. For instance, we may not be able to exactly obtain the saving
costs of the integrated facilities, we can, however, estimate this value by the excess cost of establishment of a new hospital
(i.e. construction costs). It should be noted that cost of building TCs and hospitals can be estimated using past data. To
address the fluctuations and lack of detailed data, we have considered these costs under uncertainty. Traveling times with
different transportation modes can be estimated using Google maps. We have also considered these times under uncertainty.
Due to the large problem size of this case-study, the proposed SADE meta-heuristic algorithm is used to obtain near opti-
mal network. Parameters of the proposed SADE algorithm were set according Table 8 while the stopping criterion (NFC) was
set on 100000 to find better solutions.
After applying SADE, the derived LAM and UAM network are shown in Figs. 16 and 17, respectively. As shown in Fig. 17,
with xho = 0.9, the UAM network results in establishing more hospitals compared to the LAM network. Also, the number of
transplant centers in both UAM and LAM are equal, while the locations of them are different. It can be concluded that some
candidate locations of facilities (i.e., hospitals and TCs), regardless of whether it is LAM or UAM, should be selected. For
example, a TC at candidate location number 2 has been selected in both LAM and UAM networks while candidate location
number 1 is not selected. Although both problems cover total demands with xho = 0.9, the UAM model can be applied as a
more reliable network. As can be seen in Fig. 17, in UAM, due to establishing more facilities, more hospitals are accessible to
the TCs. Hence, both the ground and the air flows as well as alternative routes are more than the LAM problem.
Table 12 shows the intervals of the cost objective functions, which are derived from the TH approach. The interval of the
fixed and the transportation costs have been calculated separately for different values of n. The results may help the DM
select the best value of n. Moreover, the range of each interval has been included. As the value of n increases, the interval
of the objective function value widens and offers more sets of possible solutions for the DM.
 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124 123

Fig. 18. Prioritized regions of Iran derived from LAM and UAM models.

Table 13
Interval of fixed costs and comparing to the current network.

Optimum interval Current network in Iran

Fixed costs [97,700, 107,100] 146,700

Considering the results presented in Figs. 16 and 17, the districts of Iran have been prioritized as shown in Fig. 18. In the
high priority districts No. 1 (illustrated by cross-hatched pattern), both models (i.e., LAM and UAM) establish the same facil-
ities. It can be concluded that these districts have the most impact on the total demand coverage. Decision districts No. 2
(diagonal lines pattern) have been derived through only one model (either LAM or UAM). For example, the district No. 2
which includes the candidate location TC #4, has been derived only from the LAM problem, while the other districts are
derived from the UAM. Since regions No. 3 (solid light pattern) have not been assigned in any of LAM or UAM models, it
can be concluded that establishing facilities in these districts is not efficient. District 4 (solid dark pattern) represent depop-
ulated areas where no facilities is needed.
To compare the result obtained from Figs. 16 and 17, the values of first objective function are shown in Table 13. As
shown, not only fixed cost of establishing facilities is not in the proposed interval, but also it has a significant difference with
the upper bound of the optimum solution interval. The current inefficient network for locating facilities in Iran has imposed
almost an additional 27% of fixed costs (regardless of transportation costs) to the country.

7. Conclusions

The contributions of this paper to the literature are: (1) presenting a novel bi-objective mathematical programming
model for the organ transplant supply chain network design considering the multi-priority queuing system for different
organs under uncertainty, (2) applying an efficient approach proposed by Xu and Zhou (2013) hybridizing with the fuzzy
multi-objective programming approach given by Torabi and Hassini (2008) to cope with the uncertainty, and (3) developing
two meta-heuristics approaches based on simulated annealing (SA), and self-adaptive differential evolution (SADE) algo-
rithms with an efficient lower bound procedure, to address large-size industrial problems. Besides, several sensitivity anal-
yses were conducted to validate the effectiveness and efficiency of the proposed model, and managerial insights were
provided. Finally, a real case was studied in Iran and concluded that the current transplant network is not as efficient as
the optimum network obtained from the proposed model. For future research, considering other challenges (e.g. disruption
and reliability) in designing the network, and developing efficient and effective meta-heuristics to address these challenges
are recommended.
124 B. Zahiri et al. / Transportation Research Part E 72 (2014) 101–124

References

Bazaraa, M.S., Sherali, H.D., Shetty, C.M., 2013. Nonlinear Programming: Theory and Algorithms. John Wiley & Sons.
Beliën, J., De Boeck, L., Colpaert, J., Devesse, S., Van den Bossche, F., 2012. Optimizing the facility location design of organ transplant centers. Decis. Support
Syst. 54, 1568–1579.
Benneyan, J.C., Musdal, H., Ceyhan, M.E., Shiner, B., Watts, B.V., 2012. Specialty care single and multi-period location–allocation models within the veterans
health administration. Socio-Econ. Plann Sci. 46, 136–148.
Bruni, M.E., Conforti, D., Sicilia, N., Trotta, S., 2006. A new organ transplantation location–allocation policy: a case study of Italy. Health care Manage. Sci. 9,
125–142.
Calégari, P., Coray, G., Hertz, A., Kobler, D., Kuonen, P., 1999. A taxonomy of evolutionary algorithms in combinatorial optimization. J. Heuristics 5, 145–158.
Cocking, C., Flessa, S., Reinelt, G., 2006. Locating health facilities in Nouna district, Burkina Faso. In: Operations Research Proceedings 2005. Springer, pp.
431–436.
Green, L., 2006. Queueing analysis in healthcare. In: Hall, R.W. (Ed.), Patient Flow: Reducing Delay in Healthcare Delivery. Springer, New York, pp. 281–308.
Gross, D., Shortle, J.F., Thompson, J.M., Harris, C.M., 2013. Fundamentals of Queueing Theory. Wiley.
Heilpern, S., 1992. The expected value of a fuzzy number. Fuzzy Sets Syst. 47, 81–86.
Iorio, A.W., Li, X., 2005. Solving rotated multi-objective optimization problems using differential evolution. In: AI 2004: Advances in Artificial Intelligence.
Springer, pp. 861–872.
Ishfaq, R., Sox, C.R., 2011. Hub location–allocation in intermodal logistic networks. Eur. J. Oper. Res. 210, 213–230.
Kirkpatrick, S., Gelatt Jr., C.D., Vecchi, M.P., 1983. Optimization by simulated annealing. Science 220, 671–680.
Kong, N., Schaefer, A.J., Hunsaker, B., Roberts, M.S., 2010. Maximizing the efficiency of the US liver allocation system through region design. Manage. Sci. 56,
2111–2122.
Liu, B., Iwamura, K., 1998. Chance constrained programming with fuzzy parameters. Fuzzy Sets Syst. 94, 227–237.
McClain, J.O., 1976. Bed planning using queuing theory models of hospital occupancy: a sensitivity analysis. Inquiry 13, 167–176.
Melo, M.T., Nickel, S., Saldanha-Da-Gama, F., 2009. Facility location and supply chain management –a review. Eur. J. Oper. Res. 196, 401–412.
Metropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H., Teller, E., 1953. Equation of state calculations by fast computing machines. J. Chem. Phys. 21
(vol. 1087).
Mohammadi, M., Jolai, F., Tavakkoli-Moghaddam, R., 2013. Solving a new stochastic multi-mode p-hub covering location problem considering risk by a
novel multi-objective algorithm. Appl. Math. Modell. 37, 10053–10073.
Mohammadi, M., Torabi, S.A., Tavakkoli-Moghaddam, R., 2014. Sustainable hub location under mixed uncertainty. Transp. Res. Part E 62, 89–115.
Nosek Jr., Wilson, J.P., 2001. Queuing theory and customer satisfaction: a review of terminology, trends, and applications to pharmacy practice. Hosp.
Pharm. 36, 275–279.
Papageorgiou, J.C., 1978. Some operations research applications to problems of health care systems (a survey). Int. J. Bio-med. Comput. 9, 101–114.
Pishvaee, M.S., Razmi, J., 2012. Environmental supply chain network design using multi-objective fuzzy mathematical programming. Appl. Math. Modell. 36,
3433–3446.
Pishvaee, M., Razmi, J., Torabi, S., 2012. Robust possibilistic programming for socially responsible supply chain network design: a new approach. Fuzzy Sets
Syst. 206, 1–20.
Preater, J., 2002. Queues in health. Health Care Manage. Sci. 5, 283.
Price, K.V., Storn, R.M., Lampinen, J.A., 2005. Differential Evolution – A Practical Approach to Global Optimization. Springer.
Qin, A.K., Huang, V.L., Suganthan, P.N., 2009. Differential evolution algorithm with strategy adaptation for global numerical optimization. IEEE Trans. Evol.
Comput. 13, 398–417.
Rahman, S.-U., Smith, D.K., 2000. Use of location–allocation models in health service development planning in developing nations. Eur. J. Oper. Res. 123,
437–452.
Rais, A., Viana, A., 2010. Operations research in healthcare: a survey. Int. Trans. Oper. Res. 18, 1–31.
Shariff, S., Moin, N.H., Omar, M., 2012. Location allocation modeling for healthcare facility planning in Malaysia. Comput. Ind. Eng. 62, 1000–1010.
Storn, R., 1996. On the usage of differential evolution for function optimization. In: Fuzzy Information Processing Society, 1996. NAFIPS., 1996 Biennial
Conference of the North American, pp. 519–523.
Storn, R., Price, K., 1997. Differential evolution – a simple and efficient heuristic for global optimization over continuous spaces. J. Global Optim. 11, 341–
359.
Syam, S.S., Côté, M.J., 2010. A location–allocation model for service providers with application to not-for-profit health care organizations. Omega 38, 157–
166.
Syam, S.S., Côté, M.J., 2012. A comprehensive location–allocation method for specialized healthcare services. Oper. Res. Health Care 1, 73–83.
Torabi, S., Hassini, E., 2008. An interactive possibilistic programming approach for multiple objective supply chain master planning. Fuzzy Sets Syst. 159,
193–214.
Uehlinger, N., Beyeler, F., Marti, H., Immer, F., 2010. Organ transplantation in Switzerland: impact of the new transplant law on cold ischaemia time and
organ transports. Swiss Med. Wkly. 140, 15–16.
Vahdani, B., Tavakkoli-Moghaddam, R., Modarres, M., Baboli, A., 2012. Reliable design of a forward/reverse logistics network under uncertainty: a robust-M/
M/c queuing model. Transp. Res. Part E: Logist. Transp. Rev. 48, 1152–1168.
Wang, Q., 2004. Modeling and analysis of high risk patient queues. Eur. J. Oper. Res. 155, 502–515.
Xu, J., Zhou, X., 2013. Approximation based fuzzy multi-objective models with expected objectives & chance constraints: application to earth-rock work
allocation. Inf. Sci. 238, 75–95.
Zhang, Z.-H., Jiang, H., 2013. A robust counterpart approach to the bi-objective emergency medical service design problem. Appl. Math. Modell. 38, 1033–
1040.
Zahiri, B., Tavakkoli-Moghaddam, R., Pishvaee, M.S., 2014. A robust possibilistic programming approach to multi–period location–allocation of organ
transplant centers under uncertainty. Comput. Ind. Eng. 74, 139–148.