You are on page 1of 29

Computational and Applied Mathematics (2019) 38:194

https://doi.org/10.1007/s40314-019-0966-6

A fuzzy-budgeted robust optimization model for joint


network design-pricing problem in a forward−reverse
supply chain: the viewpoint of third-party logistics

Alireza Ghafarimoghadam1 · Sima Ghayebloo2 · Mir Saman Pishvaee3

Received: 10 December 2018 / Revised: 25 June 2019 / Accepted: 24 September 2019


© SBMAC - Sociedade Brasileira de Matemática Aplicada e Computacional 2019

Abstract
The increasing opportunities for cost savings and customer satisfaction have propelled third-
party logistics providers (3PLs) to get involved in the forward and reverse logistics operations.
The forward−reverse supply chain network design (FRSCND) for 3PLs have been somehow
studied under various conditions, so far. However, some very common business configurations
and real-world concerns such as pricing and uncertainty are less investigated in the literature.
Accordingly, this paper proposes a novel robust model for the design of a 3PL’s logistics
network and pricing decisions. Since the value of uncertainty budget parameter in the robust
model is an epistemic uncertain one, the fuzzy-robust model regarding the uncertainty budget
parameter as a fuzzy number is also developed. The results of numerical examples show
that the proposed models outperform the deterministic model regarding solution robustness
and computational time. Considering the robust sensitivity analysis, the effects of uncertain
parameters on total cost, based on their conservatism levels are analyzed. In addition, the
conducted sensitivity analysis over penalty values for constraints violation reveal that for
the medium- and large-size problems, the proposed models are more cost-effective for high
penalty values, while the models’ performance is related to the problem size, for their low
amounts.

Keywords Third-party logistics providers · Network design · Forward−reverse supply


chain · Robust optimization · Fuzzy uncertainty budget · Pricing

Mathematics Subject Classification 90B06

Communicated by Rosana Sueli da Motta Jafelice.

B Mir Saman Pishvaee


pishvaee@iust.ac.ir; ms-pishvaee@aut.ac.ir
1 Department of Industrial Engineering, Iran University of Science and Technology, University
Ave., Narmak, Tehran 16846-13114, Islamic Republic of Iran
2 Department of Mechanic, Faculty of Mechanic, University of Zanjan, University
Blvd, Zanjan 45371-38791, Islamic Republic of Iran
3 School of Industrial Engineering, Iran University of Science and Technology, Narmak,
Tehran 16846-13114, Islamic Republic of Iran

123
194 Page 2 of 29 A. Ghafarimoghadam et al.

1 Introduction

The management of end-of-life products, also called reverse logistics (RL) (Nukala and
Gupta 2007; Fattahi and Govindan 2017), becomes more important in today’s environmen-
tally conscious manufacturing sector. As pointed out by (Aras et al. 2008), environmental
protection and remaining advantages value in the used products that can be captured by
recovery processes is the corporation’s principal motivation for the collection operation. In
most of the past researches, the design of forward and reverse supply chain networks was
considered separately; however, the configuration of the RL network has a strong influence
on the forward logistics network and vice versa. Separating the design may result in sub-
optimality; therefore, the design of the FRSCND should be integrated (Lee and Dong 2008;
Lee et al. 2013; Ghayebloo et al. 2015). Fleischmann et al. (2000) stated that contrary to
forward logistics systems, the reversal of flows in the reverse supply chain implies numerous
supply sources and few demand points. The design of the network has also been severely
complicated by a high degree of uncertainty in collection and distribution. More specifi-
cally, one of the main difficulties associated with implementing reverse logistics activities is
the degree of uncertainty in terms of the timing and quantity of products. Thus, managing
return flow usually requires a specialized infrastructure and relatively high handling cost and
time (Ko and Evans 2007). In addition, many manufacturers have understood that their core
competences do not lie in the logistics field, and have, therefore, progressively sought to buy
logistics services and functions from 3PLs (Bottani and Rizzi 2006; Lee et al. 2013; Sen et al.
2017). Therefore, outsourcing the non-core processes and activities makes it possible to focus
on core manufacturing activities, while, at the same time, 3PLs have specific logistics core
competencies, and consequently, they can manage logistics processes more efficiently than
their customers. In addition, outsourcing logistics functions to third-party reverse logistics
(3PRLs) have been a resource of competitive benefit for most companies. Most companies
quote greater flexibility, operational efficiency, enhanced customer service levels, and a better
focus on their core businesses as part of the advantages of appealing the services of 3PLs
(Azadi and Saen 2011).
As described above, 3PLs play a significant role in today’s supply chain management
context, because business organizations are increasingly looking for effective logistics and
supply chain management that can provide competitive advantages for them (Giri and Sarker
2017; Kayvanfar et al. 2018). Therefore, 3PLs gain a growing research interest because of
its increasing importance in practice, and the 3PLs selection problem has been extensively
studied in the literature (Aguezzoul 2014; Ecer 2018). Since outsourcing logistics activities
to 3PLs is a relatively new context, few works have been accomplished regarding logistics
network design. In 3PRLs design, for the first time, Ko et al. (2005) developed a distribution
network, considering integrated forward and reverse flows for 3PLs. The network for 3PLs
may consist of client’s facilities, warehouses or distribution centers (DCs), collection centers,
and market places. Since the type and time of 3PLs’ contracts are variable, some demands
are not satisfied in the planning horizon. Therefore, the possibility of increasing capacity at
warehouses and repair centers is included in conducted researches by Ko et al. (2005) and
Min and Ko (2008).
Some works only considered forward supply chain for 3PLs. Regarding this issue, Mallidis
et al. (2012) presented a green supply chain network design (SCND) model including input
ports, DCs and transportation modes along with decisions about dedicated or shared usage
of storehouses of 3PLs and transportation vehicles considering environmental effects of
transportation. Kumar et al. (2013) developed a three-stage inventory model to address the

123
A fuzzy-budgeted robust optimization model for joint… Page 3 of 29 194

outsourcing of logistics activities to 3PLs with different shipment policies between manu-
facturer, exporter and assembly point for a manufacturing industry. Arabzad et al. (2015)
proposed a multi-objective optimization algorithm for solving the new multi-objective loca-
tion–inventory problem in a DC network considering different transportation modes and
3PLs. Their model was limited to forward logistics. By the way, there are few works which
considered forward and reverse supply chains, simultaneously. For example, Li et al. (2018)
studied location-inventory decisions jointly in a closed-loop system with 3PLs. First, a mathe-
matical model was proposed to develop the mixed-integer non-linear programming (MINLP)
model for the location-inventory problem under study. Then, a novel heuristics based on dif-
ferential evolution and genetic algorithm (GA) was designed to solve the MINLP models,
efficiently.
The dynamic and complex nature of forward supply chain especially reverse supply chain
imposes a high degree of uncertainty in supply chain planning decisions and significantly
influences the overall performance of the supply chain network. Uncertainty is included in
some proposed models for 3PLs in the literature. Ko and Evans (2007) presented a net-
work design model for 3PLs. They investigated forward and reverse logistics movements
using dynamic parameters. They used simulation technique to consider uncertainty into their
model, and then tried to solve the model using hybrid GA. Zhang et al. (2007) presented a
fuzzy model to design a network for remanufacturing logistics, from 3PLs’ point of view.
This model was under reverse logistics process. They also used fuzzy chance constrained
programming model in network design problem, where transportation costs and parameters
related to backward demand of products were considered as triangular fuzzy numbers. Du
and Evans (2008) minimized tardiness and total costs for location and capacity decisions
in a closed-loop logistics network operated by 3PLs. To solve their proposed bi-objective
mixed-integer linear programming (MILP) model, a hybrid scatter search method was devel-
oped. Xanthopoulos and Iakovou (2009) proposed a simulation-based solution approach for
capturing the uncertainties in reverse logistics. Suyabatmaz et al. (2014) considered the RL
network design of the 3PLs including finding the number and places of the test centers
under supply uncertainty associated with the quantity of the returns. They applied hybrid
simulation−analytical modeling, which iteratively uses mixed-integer programming (MIP)
models and simulation, as a suitable framework for handling the uncertainties in the stochastic
RL network design problem. Ghaffari-Nasab et al. (2016) proposed a hub logistics network
design from the perspective of third-party companies considering inventory costs and the
existing uncertainty in demand regarding production and distribution. Daghigh et al. (2016)
proposed a multi-objective model to design logistics network for 3PLs considering sustain-
able objectives under uncertainty.
As summarized above, the number of existing works in developing FRSCND for 3PLs
is very limited and new, as well. Therefore, there are many issues and subject areas could
be considered in this field. One of the main challenges of network design models is the
inclusion of uncertainty. Robust optimization (RO) as a more recent and preeminent approach
to deal with uncertainty is applied in this paper. Unlike the proposed approaches in the
literature to deal with the uncertainty, RO only needs the upper and lower bounds of uncertain
data and does not cause computational intractability. In addition, its sensitivity analysis
enables decision maker (DM) to analyze the effects of changes in input data on system’s
output based on DM’s reliability level. There are several well-known RO approaches that
can be classified into two categories: conservative robust programming (e.g., Soyster 1973,
Mulvey et al. 1995, Li et al. 2011) and flexible robust programming (e.g., Bertsimas and
Sim 2004). Some applications of RO approach in the FRSCND are presented in recent
years such as Hasani et al. (2012), Ramezani et al. (2014), Ghafarimoghadam et al. (2016),

123
194 Page 4 of 29 A. Ghafarimoghadam et al.

Talaei et al. (2016), Prakash et al. (2018). The approach used in this research has been
developed by Bertsimas and Sim (2004) following related developments from the robust
control area. In this case, data are assumed to belong to some sets with no specific probabilistic
information. One of the parameters added in the robust counterpart model based on the
approach of Bertsimas and Sim (2004) is DM’s reliability level called ‘uncertainty budget
level’. It is determined by the DM and is tainted by a high degree of epistemic uncertainty. To
cope with this challenging issue, a credibility-based chance constrained programming model
proposed first by Liu (2004) and Liu and Liu (2002) is employed in this work. In general,
this technique is a computationally efficient fuzzy mathematical programming approach that
relies on strong mathematical concepts (i.e., the expected value of a fuzzy number and the
credibility measure). It can support different kinds of fuzzy numbers such as triangular and
trapezoidal forms and enables the DM to satisfy some chance constraints in at least some
given confidence levels. In addition, contrary to the possibility measure that has no self-
duality property, the credibility measure is a self-dual measure (Li and Liu 2006). That is, if
the credibility value of a fuzzy event attains 1, DM believes that the fuzzy event will surely
be happened; however, when the corresponding possibility value achieves 1, the fuzzy event
may fail to happen. In other words, a fuzzy event may fail, even though its possibility achieves
1, and hold even though its necessity is 0. However, the fuzzy event must hold if its credibility
is 1, and fail if its credibility is 0.
Another main challenge of the companies involved in product recovery is used product
(Pokharel and Mutha 2009). Therefore, considering pricing decisions besides location, ware-
housing, transportation, and other related decisions will make the model more practical and
useful. As mentioned before, there are considerable uncertainties in the quantity, timing,
and quality of returned products. Because of high percentage of returned products owing
to have a poor quality, adopting a proactive approach and implementing a used product
acquisition strategy thorough offering the appropriate acquisition price is necessary for a
company engaged in product recovery. In this issue, some methods are proposed to calculate
the optimal acquisition price for returned products (Choi et al. 2004; Yalabik et al. 2005; Aras
et al. 2008; Liang et al. 2009). In the context of joint network design and pricing decisions,
Keyvanshokooh et al. (2013) developed a model determining the acquisition price for used
products and network design decisions at the same time. To the best of the authors’ knowl-
edge, there is no outstanding work which handles the mentioned concerns simultaneously
for a 3PLs’s network. Taleizadeh et al. (2019) developed a mixed-integer optimization model
to plan a multi-period multi-echelon sustainable closed-loop supply chain. The social and
environmental effects of the supply chain decisions were taken into account in their research
which has been measured by GRI guideline indicators. A discount offer on the returned
products was also considered as an incentive policy to increase the customers’ willingness to
return back their used products more. In this offer, the amount of the proposed discount was
depended on the quality of the returned products. Fattahi et al. (2018) considered a multi-
period supply chain network redesign problem in which the customer zones’ demands were
stochastic and price-dependent for multiple products. A novel multi-stage stochastic program
was also developed in their research to simultaneously make tactical decisions including the
products’ prices and strategic redesign decisions. Moreover, they developed an MILP model
for their problem, and then, Benders decomposition (BD) approach was employed to solve
it.
Motivated by the mentioned drawbacks in the proposed models, we developed a joint
network design-pricing model which supports strategic decisions regarding forward and
reverse flows of products in the 3PLs’ supply chain as well as the optimal unit acquisition
price. The proposed model is an MINLP one and since the non-linear models are inherently

123
A fuzzy-budgeted robust optimization model for joint… Page 5 of 29 194

much more difficult to be optimized and, on the other hand, it is also hard to distinguish
a local optimum from a global optimum (Chinneck 2006), the developed model is then
linearized using the approach proposed by Chang and Chang (2000). The proposed approach
is a concise linearization method for solving mixed 0–1 polynomial (M01P) programming
problems. The method is both more compact and also uses fewer additional binary variables,
extra continuous variables, and auxiliary constraints to linearize the M01P problem than the
previous methods such as the ones proposed by Glover (1975), Oral and Kettani (1992), Li
(1994). Since the basic model in the proposed approach is to some extent different from
the proposed model in this paper, the linearization is somehow more different. Therefore,
some more transformations are applied to make the model linear. Some parameters in reverse
logistics such as the quality of returned products have a high degree of uncertainty. In addition,
in the last decade, the fluctuations in fuel prices have significantly affected shipping costs
and it seems that this uncertain condition on fuel prices will be continued with respect to new
economic crises. Thus, he robust counterpart of the proposed model is developed to cope
with the uncertainties in recoverable rate of returned products and shipping costs. One of the
main issues which should be determined, while applying the RO model is the determination
of uncertainty budget level which has epistemic uncertainty type. Therefore, the fuzzy-robust
model is developed at the end step of the modeling. Numerical tests (see Sect. 5) show the
power of the proposed fuzzy-robust model in handling the uncertainty of parameters and
generating robust and fuzzy optimal solutions.
The main contributions of this paper that differentiate it from the existing researches in
the literature can be summarized as follows:
• Developing an MILP model for 3PLs’s FRSCND problem which determines both network
design and pricing decisions of returned products, simultaneously.
• Proposing a robust optimization approach based on uncertainty budget which makes a
trade-off between robustness and its cost, appealing to risk-seeking DMs who prefer an
optimal solution in most of the cases not all, in order to incur less cost.
• Considering the uncertainty budget parameter of the proposed robust optimization model
to cope with its epistemic nature.
• Linearization of non-linearity caused by pricing decision so as to capture the global opti-
mum solution.
• Conducting fuzzy and robust sensitivity analyses and presenting some useful managerial
insights.
The rest of this paper is organized as follows. In Sect. 2, the problem on-hand is defined and
the proposed FRSCND model for 3PLs as well as incentive pricing and linearization issues
presented. The robust and fuzzy-robust models are developed in sect. 3. Section 4 presents
numerical examples and discusses the main computational results. Finally, the conclusions
from this work are drawn in Sect. 5.

2 Problem definition

In this paper, an FRSCND model for a 3PL with production centers (as 3PL’s customers),
3PL’s storages, as well as its collection centers and customer zones is modeled and analyzed.
The studied supply chain network is illustrated in Fig. 1. In the forward flow, new products
are procured by 3PL which is shipped from production center to storages site of the 3PL. The
procurement price is included in the shipping cost. Then, the products are shipped to customer
zones. In the reverse flow, each customer zone sells the second hand (returned) products to

123
194 Page 6 of 29 A. Ghafarimoghadam et al.

3PL

Storages

Production centers
(3PL’s customers)

Customer zones

Collection centers

Fig. 1 Forward–reverse supply network for the 3PL under study

3PL. The procured returned products are shipped to collection centers, sorted and categorized
into two groups of qualified and unqualified ones. Then, the qualified ones are shipped to
production centers, while the unqualified ones are disposed (at disposal centers), where the
latter category (i.e., disposal centers) is not considered in the model of this research. In the
proposed FRSCND model for 3PL, some prices are proposed by 3PL (called acquisition
price) for each customer zone and the model then determines the optimum price. It must
be emphasized that a higher acquisition price reduces the unit cost saving from a return;
however, increases the customers’ willingness to return their used products at the same time.

The proposed model is developed based on the following assumptions.


• There are some customers at each customer zone with different demands amount.
• The number of customers at each customer zone is predefined.
• Storages and collections centers’ capacities are limited and determined.
• The all demand of customer zones for new products must be satisfied.
• The all demand of production centers for returned products must be satisfied in reverse
flow.
• The proposed price of 3PL to customer zones is discrete and ranges over 1 and an upper
limit.
• The proposed price of each customer to 3PL has discrete uniform distribution.
• Storing and sorting costs in storages and collection centers are included in the shipping
costs.
Considering the above-mentioned assumptions, one should choose the location, determine
the number of storages and collection centers, as well as specify the quantity of flows between
each pair of network facilities, as the main objectives which should be addressed in this study.
Besides, the accepted proposed price to each customer zones will be determined.
The methodology of the paper can be summarized as follows:
1. Developing a deterministic non-linear joint network design-pricing model for the defined
problem.

123
A fuzzy-budgeted robust optimization model for joint… Page 7 of 29 194

2. Linearization of the developed model.


3. Incorporating uncertainty in some parameters and developing the robust counterpart.
4. Incorporating deep uncertainty of the uncertainty budget level in the robust model and
developing the fuzzy model.
5. Verifying and validating the proposed models by numerical experiments and extracting
some useful managerial insights.

2.1 Incentive pricing

In the proposed network, the 3PL should procure returned products from customer zones
and ship them to collection centers. The main challenging issue for the 3PL is that what
price should be proposed to satisfy the production centers’ demand as well as optimize the
procurement costs, simultaneously. To deal with this issue, the developed model by Aras
et al. (2008) is used.
It is assumed that the quantity of the second hand (returned) products of each customer
zone is different and their proposed price, denoted by R∗k , has a discrete uniform distribution.
The discrete uniform distribution is assumed for the proposed price in this research according
to Aras et al. (2008), since the main contents of both models are in the same field. In this
regard, Aras et al. (2008) stated that the reason for choosing the discrete uniform distribution
is that it not only provides analytical tractability in the formulation of the proposed model,
but also it helps to incorporate a large degree of variability among the proposed prices. With
this choice, R∗k takes the values over the interval [1,uk], where uk ≥ 1 and represents the
maximum proposed sale price of the returned products by customers in customer zone k.
As a straight result, the proposed procurement price by the 3PL in customer zone k must
be less or equal than uk . If the proposed price level by customer zone k is denoted by cp p
( p  1, 2, . . . , u k ), the supremum and infimum values of cp p are u k and 1, respectively.
To deal with the 3PL’s challenge regarding the proposed price, a binary variable is defined
named YPkp . It is 1, if the price level p (i.e., cp p ) in zone, k is proposed by the 3PL. Since
only one price level can be proposed by 3PL, the following constraints should be applied:

YPkp ≤ 1 ∀k (1)
p

cp p YPkp ≤ u k ∀k. (2)
p

In Eqs. (1) and (2), the term p cp p YPkp represents the proposed price by 3PL in customer
zone k which could be equal to one of the proposed prices by customers, i.e.,cp p .
Since R∗k ∼ U [1, u k ], the probability density function (PDF) and cumulative distribution
functions (CFD) of R*k can be specified by f (R∗k )  u1k and F(R∗k )  Ru∗kk , respectively.
The returned product will be sold by a customer, if the proposed price by 3PL be greater or
equal than the proposed price by customer, where its probability can be computed as Eq. (3):
   cp YPkp
p p
p cp p YPkp − R∗k ≥ 0  ∀k. (3)
p uk
As a result, if the proposed price by 3PL is equal to supremum value, the all customers in
customer zone k are willing to sell their products. In reverse, if it is equal to infimum value,
no customer wants to sell his/her products. The total number of customers  who are willing to
cp YPkp
sell products to 3PLat the 3PL’s proposed price is determined by ndk p u kp , where ndk
is the total number of customers in customer zone k. Considering rk as the quantity of each
customer’s products ready to sale at customer zone k (a predefined percentage of demand

123
194 Page 8 of 29 A. Ghafarimoghadam et al.

of each customer zone is assumed to result in return products), the total number of products
ready 
for sale according to 3PL’s proposed price in customer zone k can be expressed as
cp YPkp
rk ndk p u kp . In addition, total number of procured products in customer zone k which

shipped to all collection centers is denoted by l XRCkl . Constraint (4) assures that it must
be less or equal than the total number  of products ready for sale according to the 3PL’s
cp YPkp
proposed price in zone k, i.e., rk ndk p u kp :

 p cp p YPkp
XRCkl ≤ rk ndk ∀k. (4)
l uk
In addition, procurement cost of returned products from customer zone k which is non-
linear can be formulated as Eq. (5):
  
cp p YPkp XRCkl ∀k. (5)
p l

2.2 Model formulation

In this section, indices, input parameters, decision variables as well as the mathematical
model are presented.
Indices:
i Index of production centers i  1, . . . , I .
j Index of 3PL’s storage j  1, . . . , J .
l Index of 3PL’s collection centers l  1, . . . , L.
k Index of customer zones k  1, . . . , K .
p Index of suggested price levels to customer zones p  1, . . . , u k .
Parameters:

fw j Fixed cost of opening storage center j.


frl Fixed cost of opening collection center l.
cfi j Shipping cost for per unit of new product from production center i to storage center j.
cc jk Shipping cost for per unit of new product from storage center j to customer zone k.
cikl Shipping cost for per unit of returned product from customer zone k to collection center
l.
crli Shipping cost for per unit of returned product from collection center l to production
center i.
pfi Quantity of new products for sale at production center i.
pri Quantity of returned products needed at production center i (demand of production
center i for returned products).
bw j Capacity of storage center j.
bcl Capacity of collection center l.
τl Recoverable rate of returned products at collection center l.
ndk Number of customers at customer zone k.
dk Demand of each customer at customer zone k.
rk Quantity of returned product ready to sale for each customer at customer zone k.
cp p Suggested price at level p.
uk Supremum (upper limit) of suggested price level at customer zone k.

123
A fuzzy-budgeted robust optimization model for joint… Page 9 of 29 194

Decision variables:
XFWi j Quantity of new products shipped from production center ito storage center j.
XFC jk Quantity of new products shipped from storage center j to customer zone k.
XRCkl Quantity of returned products shipped from customer zone k to collection center l.
XRPli Quantity of returned products shipped from collection center l to production center
i.


1 if storage center j is opened
YW j 
0 otherwise

1 if collection center l is opened
YCl 
0 otherwise

1 if proposed price level p is accepted at customer zone k
YPkp  .
0 otherwise
The mathematical formulation of the model is described by an objective function followed
by a set of constraints:
    
Min z  fw j YW j + cfi j XFWi j + cc jk XFC jk
j i j j k
    
+ frl YCl + cikl XRCkl + crli XRPli
l k l
     l i

+ cp p Y Pkp XRCkl . (6)


k p l

Subject to:

XFC jk ≥ ndk dk ∀k (7)
j

XRPli ≥ pri ∀i (8)
l
 
XFWi j  XFC jk ∀ j (9)
i k
 
XRPli  τl XRC K l ∀l (10)
i K

XFWi j ≤ pfi ∀i (11)
j

XFWi j ≤ bw j YW j ∀ j (12)
i

XRC K l ≤ bcl YCl ∀l (13)
K
 
XRC K l ≤ nd K r K (cp p Y PK p /u K ) ∀k (14)
l p

YP K p ≤ 1 ∀k (15)
p
X F Wi j , XFC jk , XRC K l , XRPli ≥ 0 ∀i, j, l, k (16)

YW j , YCl , YP K p ∈ {0, 1} ∀ j, l, k, p. (17)


Objective function, Eq. (6), minimizes the total cost including fixed opening costs, ship-
ping, and procurement costs. Constraints (7) and (8) assure that demand of new products
for all customers and demand of returned products for production centers are fully satisfied,
respectively. Constraint (9) establishes the flow balance of new products at storages centers.
Constraint (10) assures that all qualified (or recoverable) returned products are shipped from

123
194 Page 10 of 29 A. Ghafarimoghadam et al.

collection centers to production centers. Equations (11) to (13) are capacity constraints of
facilities. Constraint (14) assures that the total returned products shipped to all collection
centers are less than the total procured returned products by the 3PL at each customer. Con-
straint (15) indicates that only one price level is accepted at each customer zone. For more
explanation about constraints (14) and (15), see Sect. 2.1. Finally, Constraints (16) and (17)
enforce the non-negativity and binary restrictions on corresponding decision variables.

2.3 The model linearization


   
The last term in the objective function, i.e., k p cp p YPkp l XRCkl is non-linear,
because it involves the multiplication of two binary and continuous variables. To handle the
complexity of such MINLP model, one can linearize it using the procedure proposed by
Chang and Chang (2000) as follows.
Consider X ∗ Y is a non-linear term in a MINLP model, where Y is a binary variable and
X is a continuous variable with upper bound U (already determined).This non-linearity can
be written as Eq. (18):
X ∗Y
0≤ X ≤U (18)
Y ∈ [0, 1]

According to the procedure proposed by Chang and Chang (2000), the correspondent
linear term using some auxiliary constraints can be defined as Eq. (19):

UZ
X
U −Z ≥0
(19)
X
U − Z ≤1−Y
0≤Z ≤Y

In the developed model, the non-linear term which should be linearized, is as Eq. (20):
  
k p cp p Y Pkp l X RCkl
 
0≤ l X RCkl ≤ p ndk rk (cp p Y Pkp /u k ) ∀k (20)
Y Pkp ∈ [0, 1] ∀k, p

According to the above descriptions, the linear form of the non-linear term in the objective
function can be stated as Term (21) as well as Eqs. (22) to (24):
   
cp p ZPkp ndk rk (cp p YPkp /u k ) (21)
k p p



l XRCkl
 − ZPkp ≥ 0 ∀k (22)
p nd r
k k (cp p YP /u
kp k ) p


XRCkl  
 l
− ZPkp ≤ 1 − YPkp ∀k (23)
p ndk rk (cp p YPkp /u k )
p p

ZPkp ≤ YPkp ∀k, p. (24)

123
A fuzzy-budgeted robust optimization model for joint… Page 11 of 29 194

The Term (21) is replaced by the non-linear term in the objective function and Constraints
(22) to (24) should be added to the model. As it can be implied, the Term (21) and Constraints
(22) and (23) are not linear yet. It originates from the upper bound of the continuous variable
which is not a parameter. Accordingly, some more transformations are made in the following
to make them linear. Constraints (22) and (23) can be rewritten as Inequalities (25) and (26):
    
XRCkl ≥ ZPkp ∗ ndk rk (cp p YPkp /u k ) ∀k (25)
 l  p   p
 
XRCkl ≤ ndk rk (cp p YPkp /u k ) ∗ ZPkp + (1 − YPkp ) ∀k. (26)
l p p p
   
The term which makes the constraints non-linear is p ZP kp × p YP kp which
should be linearized. According to Constraints (15) and (24), the value of the term ZPkp ∗
YPkp at each customer zone (i.e. k) when, p , p ∈ p is equal to zero. Therefore, the
non-linear term can be written as Eq. (27):
    
ZPkp ∗ YPkp  ZPkp YPkp ∀k. (27)
p p p

Since the result of ZPkp YPkp is always equal to ZPkp according to Constraint (24), Eq. (28)
can be then resulted:
 
ZPkp YPkp  ZPkp ∀k. (28)
p p

Regarding the all transformations, Term (21) is replaced by Term (29) and Constraints
(30) and (31) should be replaced by (25) and (26), respectively. Constraints (30) and (31)
can also be written as Constraint (32):
 
cp p ndk rk (cp p ZPkp /u k ) (29)
k p
 
XRCkl ≥ ndk rk (cp p ZPkp /u k ) ∀k (30)
l p
XRCkl ≤ ndk rk (cp p ZPkp /u k ) ∀k (31)
p 
l 
XRCkl  ndk rk cp p ZPkp /u k ∀k. (32)
l p

The proposed model regarding more simplifications of transformed terms and constraints
can be presented as follows:
    
Min z  fw j YW j + cfi j XFWi j + cc jk XFC jk
j i j j k
    
+ frl YRl + cikl XRCkl + crli XRPli
l k l l i
  
+ (rk ndk /u k ) cp2p ZPkp (33)
k p
subject to:
  
XRC K l  nd K r K cp p ZP K p /u K ∀k (34)
l p
Z PK P ≤ YP K p ∀k, p (35)

Z Pkp ≥ ∀k, p. (36)


Constraints (7–17) except (14).
It should be pointed out that the error of linearization must be calculated and analyzed. In
the proposed model, the non-linearity is caused by multiplying two real and binary variables
and since its result is zero or a real number, no error is occurred.

123
194 Page 12 of 29 A. Ghafarimoghadam et al.

3 Robust optimization model

In order to present the robust optimization framework, consider the following optimization
model of the convex polyhedron X as Eq. set (37):
Min cx + d
s.t. Ax ≥ b,
(37)
x ∈ X,
c, d, A, b ∈ U ,
where the parameters c, d, A, b are determined according to uncertainty and vary in a convex
uncertainty set U. According to Ben-Tal and Nemirovski (1999), the robust counterpart of
the optimization Model (37) can be written as Model (38) where robust value Ĉ(x) is the
worst value of the objective function cx + d for candidate solution x:

Min  Ĉ(x)  sup [cx + d] : Ax ≥ b, ∀(c, d, A, b) ∈ U . (38)


x c,d,A,b∈U

The optimization problem, Eq. (38) is a semi-infinite optimization problem suffering from
computational intractability. As a demonstration of equivalent form, consider the uncertain
parameters in the ith constraint and let Ji be the set of the uncertain coefficients. Each
uncertain ãi j with symmetric distribution in interval [ai j − âi j , ai j + âi j ] can be demonstrated
as Eq. (39), where ai j is the nominal value, âi j is the variation amplitude and ξi j is an
independent random variable taking values over range [− 1, 1]. According to this definition,
Ben-Tal and Nemirovski (1999) showed that the equivalent robust counterpart of the original
ith constraint can be described as Inequality (40):
ãi j  ai j + ξi j âi j ∀ j ∈ Ji (39)
⎡ ⎤
 
ai j x j − ⎣max ξi j âi j x j ⎦ ≥ bi ∀i. (40)
ξ ∈U
j j∈ ji

As it is evident, the equivalent form is associated with the uncertainty set U which is
defined using an arbitrary norm of vector spaces. We first employ a box and polyhedral
uncertainty set defined by L1-norm and L − ∞ norm of the uncertain data, obtaining the
counterpart formulation as Constraint (41) see Li et al. (2011) for the proof:
 
ai j x j + Ψ wi j + Γ z i ≤ bi
j j∈Ji
 
z i + wi j ≥ âi j x j ∀ j ∈ Ji
z i ≥ 0, wi j ≥ 0. (41)
The box uncertainty set is described using the ∞-norm of the uncertain data vector as
Eq. (42):
      
U∞  ξ  ξ ∞ ≤ ψ  ξ ξ j  ≤ ψ, ∀ j ∈ Ji , (42)
where ψ is the adjustable parameter controlling the size of the uncertainty set. In addition,
the polyhedral uncertainty set is described using the 1-norm of the uncertain data vector as
Eq. (43):
⎧  ⎫

   ⎨    ⎬
U1  ξ  ξ 1 ≤ Γ  ξ  ξ j  ≤ Γ , (43)
⎩  ⎭
j∈Ji

123
A fuzzy-budgeted robust optimization model for joint… Page 13 of 29 194

where Γ is the adjustable parameter controlling the size of the uncertainty set. When ψ 
1, i.e., the set U is defined as the “interval + polyhedral” uncertainty set, the corresponding
robust counterpart optimization formulation is as Eq. set (44):
⎧ 

⎪ ai j x j + wi j + z i ≤ bi

⎨ j j∈Ji
z i + wi j ≥ âi j u j , ∀ j ∈ Ji (44)

⎪ −u
⎪ j ≤ x j ≤ u j , ∀ j ∈ Ji

z i ≥ 0, wi j ≥ 0
As it can be implied, it is exactly the robust counterpart formulation proposed by Bertsimas
and Sim (2004). To describe the Bertsimas and Sim’s approach, consider the following linear
programming model as Eq. set (45):
Min c x

s.t : âi j x j ≥ bi ∀i,
j

x∈X (45)
where ãi j is uncertain data that fluctuates based on Eq. (39). For each constraint i, a parameter
Γi is introduced as uncertainty budget taking values over interval [0, | ji |], where ji has the
same definition as before. Utilization of parameter Γi is intended to reach Γi of the coef-
ficients ai j to their worst-case values and change another coefficient ait by (Γi − Γi )âit .
Therefore, parameter Γi offers a mechanism to control the deterioration effect of objective
function value against the portability of constraints violation. Bertsimas and Sim (2004)
proved that non-linear formulation of uncertain problem can be expressed as Eq. set (46):
Minc x
⎧ ⎫
 ⎨  ⎬
s.t : ai j x j − max âi j x j + Γi − Γi âiti x j ≥ bi ∀i,
{Si ∪{ti }|Si ⊆Ji ,|Si | Γi ,ti ∈Ji \Si }⎩ ⎭
j j∈Si

x∈X (46)
where Si represents the coefficients getting their worst values and ti represents a coefficient
changing by (Γi − Γi )âit . Given the optimal solution x ∗ , Bertsimas and Sim (2004) showed
that the protection part of ith constraint is equivalent to the following linear optimization
problem as Eq. set (47):
  
Max âi j x ∗j ηi j
j∈Ji

s.t : ηi j ≤ Γi
j∈Ji

0 ≤ ηi j ≤ 1 ∀ j ∈ ji . (47)
According to the strong duality theorem, since the problem (47) is feasible and bounded
for all Γi ∈ [0, | ji |], its dual pair is also feasible and bounded with the same objective function
value, expressed as Eq. set (48):

Min Γi pi + wi j
j∈Ji
 
 
s.t : pi + wi j ≥ âi j x ∗j  ∀i, j ∈ ji

123
194 Page 14 of 29 A. Ghafarimoghadam et al.

wi j ≥ 0 ∀ j ∈ ji
pi ≥ 0 ∀i, (48)
where wi j and pi are dual variables. After reinjecting the dual form (48) in (46), the following
robust counterpart is obtained as Eq. set (49):
Min c x
 
s.t : ai j x j − Γi pi − wi j ≥ bi ∀i
j j∈Ji

pi + wi j ≥ âi j x j ∀i, j ∈ ji
wi j , pi , x j ≥ 0. (49)
As expressed before, the role of parameters Γ and ψ is to control the size of uncertainty
space. Hence, the value of these parameters should be determined based on decision maker’s
conservatism level before solving robust counterpart problem.
 The upper
 bound for the prob-
ability of constraint violation can be calculated as exp − γi2 /2| ji | , where γ is the tunable
parameter i.e., ψ or  (Li and Floudas 2012). Finally, it is worth mentioning that when
Γ  Ψ ≤ 1, the uncertainty sets defined by L1 and D-norms cover the uncertainty space
similarly, otherwise, they behave differently.
The objective mincx can be equivalently transformed as Eq. (50):
minz
s.t.z − cx ≥ 0. (50)
Thus, the uncertainty in the objective function c̃ can be treated as uncertainty in the
following type of constraints in terms of Eq. (51):
z − c̃x ≥ 0. (51)
As described before, the uncertain parameters of the proposed model are shipping costs
and recoverable rate of returned products. Taking uncertain parameters into account in the
objective function (i.e., shipping costs) and according to the provided descriptions in Sect. 3
(i.e., Bertsimas and Sim (2004)’s approach, the robust counterpart of the proposed network
design for the 3PL under D-norm is equivalent to Eqs. (52) to (56). In addition, parameters
Γ and Γ t control the uncertainty space for shipping costs and recoverable rates which take
values in intervals [0, i × j + j × k + k × l + l × i] and [0, 1], respectively. Note that in
order to obtain the robust counterpart formulation of the objective function, it is equivalently
transformed as a constraint:
Min Z
   
fw j YW j + frl YCl + (rk ndk cp2p ZP K p )/u K
j l K P
     
cfi j XFWi j + cc jk XFC jk + cikl XRCkl
i j j k l K
       
+ crli XRPli + λΓ + wcfi j + wcc jk + wcikl
i l i j j k l K
 
+ wcrli ≤ Z (52)
i l

subject to:
λ + wcfi j ≥ c fˆi j XFWi j ∀i, j

123
A fuzzy-budgeted robust optimization model for joint… Page 15 of 29 194

λ + wcc jk ≥ cĉ jk XFC jk ∀ j, k


λ + wcikl ≥ clˆkl XRCkl ∀k, l
λ + wcrli ≥ cr̂li XRPli ∀l, i (53)
 
τl XRCkl − λtl Γ tl − Wl ≥ XRPli ∀l (54)
K  i
λtl + Wl ≥ τ̂l XRCkl ∀l (55)
K
λtl , λ, Wl , wc f i j , wcc jk , wci kl , wcrli ≥ 0 ∀i, j, k, l. (56)
Constraints (34–36).
Constraints (7–17) except (10) and (14).

3.1 Fuzzy-robust optimization model

The control parameters of the robust model, i.e., uncertainty budget level (i.e., Γ and Γ t),
proposed in the previous section are assumed to be deterministic and are set by DMs. Since,
in most situations DMs have insufficient or imperfect knowledge of them, it is better to
be considered fuzzy numbers instead of deterministic ones. In this section the fuzzy-robust
model using credibility-based chance constraint programming model is developed.
Let ξ̃ be a fuzzy variable with membership function μ(x), and let r be a real number.
According to Liu and Liu (2002), the credibility measure is defined as Eq. (57):
  1 
Cr ξ̃ ≤ r  sup μ(x) + 1 − sup μ(x) . (57)
2 x≤r x>r
   
Noteworthy, since Pos ξ̃ ≤ r  supx≤r μ(x) and Nec ξ̃ ≤ r  1 − supx>r μ(x), the
credibility measure can also be defined as Eq. (58):
  1    
Cr ξ̃ ≤ r  (Pos ξ̃ ≤ r + Nec ξ̃ ≤ r ). (58)
2
Accordingly, the credibility measure could be defined as the average of the possibility
(Pos) and necessity (Nec) measures.
In addition, the expected value of ξ̃ can be determined based on the credibility measure
as Eq. (59)(Liu & Liu, 2002):
   ∞    0  
E ξ̃  Cr ξ̃ ≥ r dr − ξ̃ ≤ r dr . (59)
0 −∞

Now, assume that ξ̃ is a trapezoidal


 fuzzy number denoted by four prominent points
as ξ̃  ξ̃(1) , ξ̃(2) , ξ̃(3) , ξ̃(4) . According to Eq. (59), the expected value of ξ̃ is
 
ξ̃(1) + ξ̃(2) + ξ̃(3) + ξ̃(4) /4 and the corresponding credibility measures are as follows:


⎪ 0, r ∈ (−∞, ξ(1) ],



⎪ r −ξ(1)

⎪ , r ∈ (ξ(1) , ξ(2) ],
  ⎪
⎨ 2(ξ(2) −ξ(1) )
Cr ξ̃ ≤ r  1
r ∈ (ξ(2) , ξ(3) ], (60)

⎪ 2

⎪ −2ξ(3) +ξ(4)
⎪ r2(ξ
⎪ , r ∈ (ξ(3) , ξ(4) ],

⎪ (4) −ξ(3) )

⎩ 1, r ∈ (ξ(4) , +∞],

123
194 Page 16 of 29 A. Ghafarimoghadam et al.



⎪ ∈ (−∞, ξ(1) ],
⎪ 1,

r

⎪ 2ξ(2) −ξ(1) −r

⎪ ,r ∈ (ξ(1) , ξ(2) ],
  ⎪
⎨ 2(ξ(2) −ξ(1) )
Cr ξ̃ ≥ r  1
r ∈ (ξ(2) , ξ(3) ], (61)

⎪ 2

⎪ ξ(4) −r
⎪ 2(ξ −ξ ) , r
⎪ ∈ (ξ(3) , ξ(4) ],

⎪ (4) (3)

⎩0 r ∈ (ξ(4) , +∞],

According to Eqs. (60) and (61), it can be proven (see Zhu and Zhang 2009) that if ξ̃ is a
trapezoidal fuzzy number and α > 0.5 then:
 
Cr ξ̃ ≤ r ≥ α ⇔ r ≥ (2 − 2α)ξ(3) + (2α − 1)ξ(4) , (62)
 
Cr ξ̃ ≥ r ≥ α ⇔ r ≤ (2α − 1)ξ(1) + (2 − 2α)ξ(2) . (63)

Equations (62) and (63) can be applied directly and more conveniently when compared
to α-critical values proposed by Liu (2004), so as to convert fuzzy chance constraints into
their equivalent crisp ones.
Since there is no parameter having epistemic uncertainty in the objective function the
credibility-based fuzzy-robust mathematical programming using Eqs. (62) and (63) can be
formulated as follows:

Min Z
   
Z− fw j YW j − frl YCl − (rk ndk cp2p ZP K p )/u K
j l K P
     
− cfi j XFWi j − cc jk XFC jk − cikl XRCkl
i j j k l K
       
− crli XRPli − wcfi j − wcc jk − wcikl
i l i j j k l K
  !
− wcrli ≥ λ (2 − 2α)Γ(3) + (2α − 1)Γ(4) (64)
i l
  !
XRPli − τl X RCkl + Wl ≤ −λtl (2θl − 1)Γ tl(1) + (2 − 2θl )Γ tl(2) ∀l. (65)
i K

Constraints (53, 55 and 56).


Constraints (34–36).
Constraints (7–17) except (10)! and (14). !
Γ˜  Γ(1) , Γ(2) , Γ(3) , Γ(4) and Γ˜ tl  Γ tl(1) , Γ tl(2) , Γ tl(3) , Γ tl(4) are independent
trapezoidal fuzzy numbers and α and θl are the confidence (minimum credibility) levels
of constraints (52) and (54), respectively.

4 Computational results

In this section, several numerical experiments and sensitivity analyses are implemented and
the obtained results are reported to illustrate the validity of the proposed models. To this
end, three test problems are designed including small, medium, and large sizes, as shown in
Table 1.
Nominal data are randomly generated using different random distributions reported in
Table 2. It should be mentioned that all prices are counted in dollars.

123
A fuzzy-budgeted robust optimization model for joint… Page 17 of 29 194

Table 1 Size of test problems Problem no. i j l k p

1 8 7 7 12 3
2 20 15 15 30 10
3 35 30 30 50 15

Table 2 Sources of random Parameter Corresponding random distribution


generation of the nominal data
fw j , frl ∼ Uniform (100, 000, 150, 000)
cfi j , cc jk , cikl , crli ∼ Uniform (100, 300)
pfi ∼ Uniform (2000, 3000)
pri ∼ Uniform (500, 1000)
bw j , bcl ∼ Uniform (5000, 7000)
ndk ∼ Uniform (10, 20)
dk ∼ Uniform (200, 500)
rk ∼ Uniform (0.4, 0.8)
uk ∼ Uniform (80, 100)
τl 0.7

To compare the deterministic, robust and fuzzy-robust models, the all the mathematical
models are coded in GAMS and solved via CPLEX solver on a PC with corei5 and 4G RAM.

4.1 Economic and computational performance of the models under nominal data

In this subsection, the effects of uncertainty levels on total supply chain cost are investigated
through altering the conservatism level of the robust model, the minimum credibility level
of the fuzzy-robust model and the perturbation amplitude of all uncertain parameters (i.e.,
shipping costs and recoverable rate). To have a better understanding regarding the effects of
conservatism degree, data perturbation and confidence level on supply chain costs, different
values for these parameters are considered which can almost cover optimistic, realistic and
pessimistic decision making attitudes. Perturbation range (âi j ) is assumed to be 30%, 50%,
and 80%, the conservatism level is set to 70%, and 90%, and the minimum credibility level
is set to 60%, 75%, and 90%. The results are exhibited in Table 3. As expected, the objec-
tive function values of robust and fuzzy-robust models are higher than their corresponding
deterministic models due to additional costs of coping with uncertainty.
To investigate the magnitude of effects on total cost, the normalized deviation between
deterministic and robust models as well as between deterministic and fuzzy-robust models
are calculated and presented in Table 4. As it can be seen in Table 4, the increased dif-
ference becomes more tangible when the levels of conservatism and perturbation increase.
For instance, a highly conservative and risk-averse version of the robust model in problem
#3 (with conservatism level  90% and perturbation amplitude  80%) gives us decisions
which lead to 97.52% higher total cost. However, lower uncertainty levels make a trade-off
between robustness and its cost, and enhance the reliability of the solution at a more reason-
able cost. For example, for the conservatism and perturbation values equal to 70% and 30%,
respectively, the difference is 29.75% versus 97.52%.

123
194

123
Page 18 of 29

Table 3 Performance of the proposed models under nominal data (total supply chain cost-Thousands)
Problem no. Perturbation level Deterministic Robust model Fuzzy-robust model
(%)
Conservatism level Conservatism level Conservatism level

70% 90% 70% 90%

Minimum credibility level Minimum credibility level

60% 75% 90% 60% 75% 90%

1 30 5691 7327 7358 7329 7329 7330 7360 7361 7361


50 5691 8471 8568 8476 8477 8478 8574 8576 8578
80 5691 10,468 11,011 10,489 10,495 10,500 11,064 11,078 11,092
2 30 14,080 18,174 18,253 18,178 18,179 18,180 18,258 18,259 18,261
50 14,080 21,037 21,385 21,050 21,053 21,057 21,403 21,408 21,412
80 14,080 25,898 27,493 25,952 25,967 25,980 27,630 27,667 27,704
3 30 23,422 30,391 30,531 30,399 30,401 30,403 30,541 30,543 30,545
50 23,422 35,309 35,749 35,332 35,338 35,344 35,733 35,740 35,748
80 23,422 43,462 46,263 43,558 43,583 43,608 46,508 46,575 46,642
A. Ghafarimoghadam et al.
A fuzzy-budgeted robust optimization model for joint… Page 19 of 29 194

Table 4 Change in total supply chain cost (thousands)


Problem no. Perturbation Deterministic Robust model Fuzzy-robust model
level (%)
Conservatism Conservatism level Conservatism level
level

70% 90% 70% 90%

Minimum credibility Minimum credibility


level level

60% 75% 90% 60% 75% 90%

1 30 – 28.75 29.30 28.78 28.79 28.80 29.33 29.34 29.35


50 – 48.85 50.55 48.94 48.96 48.98 50.67 50.70 50.73
80 – 83.95 93.48 84.32 84.41 84.51 94.42 94.66 94.91
2 30 – 29.07 29.63 29.10 29.11 29.12 29.67 29.68 29.69
50 – 49.41 51.88 49.50 49.53 49.55 52.01 52.04 52.07
80 – 83.93 95.26 84.32 84.42 84.52 96.24 96.50 96.76
3 30 – 29.75 30.35 29.79 29.79 29.80 30.39 30.40 30.41
50 – 50.75 52.63 50.85 50.87 50.90 52.56 52.59 52.62
80 – 85.56 97.52 85.97 86.08 86.18 98.57 98.85 99.13

The trend of differences in both of robust and fuzzy-robust models are the same; however
such difference is slightly higher in fuzzy-robust model, i.e., 99.13% and 29.79% compared
to 97.52% and 29.75% for the robust model.
Table 5 presents the computational time of deterministic, robust and fuzzy-robust models.
As it can be implied, though the robust and also fuzzy-robust approaches increase the prob-
lems size through introducing new variables and constraints, they preserve the computational
tractability of the original problem.

4.1.1 The models’ performances when shipping costs and recoverable rates are subject
to uncertainty individually

In Sect. 5.1, the simultaneous effect of all uncertain parameters on the supply chain cost
is investigated. In this subsection, for problem #1, the cost difference of the deterministic
and robust models is apportioned between various uncertain input parameters (i.e., shipping
or recoverable rate) to clarify individual role of each uncertain parameter in cost increment
under different levels of perturbation and conservatism. So, it is presumed that only one
parameter (for example shipping cost or recoverable rate) is uncertain, while the other one
is deterministic. Afterwards, conservatism and perturbation levels are gradually altered and
the objective function response is assessed by normalized deviation (Z r − Z d )/Z d , where
Z r and Z d are the optimal values of robust and deterministic solutions, respectively. The
outcomes of this experiment for two uncertain parameters of the supply chain are depicted
in Figs. 1 and 2.
The primary insight is attained by comparing the magnitude of the objective function
responses to perturbations in uncertain parameters. Setting conservatism level to 100% and
perturbation level to 80%, variations in shipping costs and recoverable rate parameters indi-
vidually lead to increases in objective function by 73% and 19.96%, respectively. These

123
194 Page 20 of 29 A. Ghafarimoghadam et al.

Table 5 Performance of the proposed models under nominal data [CPU time(s)]
Problem no. Perturbation Deterministic Robust model Fuzzy-robust model
level (%)
Conservatism Conservatism level Conservatism level
level

70% 90% 70% 90%

Minimum credibility Minimum credibility


level level

60% 75% 90% 60% 75% 90%

1 30 1.94 1.99 2.25 2.56 1.32 1.45 1.91 1.32 1.90


50 1.94 1.05 1.42 2.45 1.99 1.05 6.05 2.36 1.06
80 1.94 4.65 2.54 2.14 1.50 2.51 1.25 5.66 2.59
2 30 1.92 3.45 2.88 1.49 2.07 2.54 2.35 1.79 2.45
50 1.92 2.26 2.45 3.40 2.01 3.12 2.35 2.19 2.14
80 1.92 2.89 3.55 7.12 6.54 1.65 2.02 10.25 5.89
3 30 8.69 1.96 3.08 5.98 8.54 4.76 4.02 3.65 3.22
50 8.69 5.19 5.33 4.80 4.50 4.93 6.47 5.06 4.55
80 8.69 14.80 14.24 10.04 18.14 3.35 24.48 5.19 25.36

90 30% perturbation
change in total cost (%)

50% perturbation
70 80% perturbation 73 73 73

62

50
46 46 46
39 40
30
26 26 27 27 27
18
10
20 40 60 80 100
conservatism level (%)
Fig. 2 Response of supply chain to variations in shipping cost (robust model)

outcomes indicate that shipping data is more sensitive to uncertainty than recoverable rate
data. Therefore, more endeavors should be made to estimate shipping costs, so that the
destructive effects of perturbation in these data are prohibited.
Another engaging outcome is derived by comparing the magnitude of responses to different
levels of conservatism, especially for values upper and lower than 60%. In shipping cost, as
an uncertain parameter, the change in total cost swiftly increases and reaches at a maximum
value equal to 73% when perturbation and conservatism levels equal to 80% and 60%,
respectively. After this value, it fixes at a constant level 73% (see Fig. 2), while in the
recoverable rate, for those conservatism level values below 60%, the change in total cost
does not increase significantly and sustainable change can be observed after this point. For

123
A fuzzy-budgeted robust optimization model for joint… Page 21 of 29 194

25

30% perturbation
change in total cost (%)

20 50% perturbation 19.96


80% perturbation
15

10
9.00

5
4.06
2.70
1.28 1.74
0
20 40 60 80 100
conservatism level (%)
Fig. 3 Response of supply chain to variations in recoverable rate (robust model)

Table 6 Values of the model’s variables under different perturbation levels for recoverable rate
Perturbation level (%) Total quantity of Total number of opened The proposed price
returned products collection centers (YC) levels accepted at
shipped from customer customer zones (related
zones to collection to YP)
centers (XRC)

30 3008 1 1
50 3810 1 1
80 6349 2 1, 2
100 11,429 3 2, 3

example, for perturbation levels equal to 50% and 80% and conservatism level equal to 100%,
it rises from 4.06 to 19.96 (see Fig. 3). Having a look at the values of the model’s variables
presented in Table 6, these trends can be more clarified. The variables whose values are not
changed for each perturbation level is not presented in Table 6.
As it can be seen in Table 6, when perturbation level for recoverable rate goes up, the
network gathers more returned products (see the value of XRC in Table 6), with high quality
(high price levels) to immune the demand satisfaction of production center for returned
products. As a consequence, the number of opened collection centers is increased to handle
the retuned products. Therefore, a sharp rising in the value of objective function can be seen
while the perturbation level for returned products increases. Since the level of perturbation for
shipping cost does not have any effect on demand satisfaction for returned products, the value
of the model’s variables are constant and not that much change is seen in high perturbation
levels. In this case, only the value of control variables in the robust model changes. It should
be mentioned that similar trends can be seen for fuzzy-robust model.

4.2 Performance evaluation of robust and fuzzy-robust models

Monte Carlo simulation is used for performance evaluation of the models. For this purpose,
under each perturbation level, 100 random realizations are uniformly generated between

123
194 Page 22 of 29 A. Ghafarimoghadam et al.

(a) Problem #1-small size (b) Problem # 2 - medium size


5,000 20,000
CV (Billions)

4,000 deterministic deterministi

CV (Billions)
15,000 c
robust robust
3,000
Fuzzy robust
10,000
2,000

1,000 5,000

0
0
10 20 30 40 50 80 10 20 30 40 50 80
perturbation level (%) perturbation level (%)

(c) Problem # 3 - large size


50,000
deterministic
40,000 robust
CV(Billions)

Fuzzy robust
30,000

20,000

10,000

0
10 20 30 40 50 80
perturbation level (%)
Fig. 4 Performance of the models under realizations for conservatism level  70% and different size problems

the extreme points of the corresponding perturbation range of uncertain parameters and then
substituted in Eqs. (7) to (17) except (14) and (33) to (36), which would result in a realization-
based model that its compact form is as Eq. (66):

min Creal x ∗ + Freal y ∗ + πi Ri
i
s.t. Areal,i y ∗ + Breal,i x ∗ + Ri ≥ Dreal,i ∀i. (66)

In this model, the upper case letters denote the realized values and (x ∗ , y ∗ ) are the optimal
values of continuous and integer decision variables, respectively, introduced from the deter-
ministic, robust and fuzzy-robust models to be evaluated. In addition, R is the only decision
variable measuring the constraint violation, when the solution (x ∗ , y ∗ ) is infeasible and πi is a
penalty for possible violation (see Pishvaee et al. 2011). After inserting the solution obtained
from deterministic, robust and fuzzy-robust models under different levels of conservatism
and perturbation into the model (Eq. set 66), the mean and standard deviation of the objective
values obtained for different random realizations are calculated to compare the results. Since
the robust and fuzzy-robust models outperform the deterministic model in one of the criteria
(for example mean), while they are dominated in the other criterion, so the multiplication
of these two criteria (i.e. mean and standard deviation) is considered in the comparisons
denoted by CV (Coefficient variable). The obtained results of three problems with minimum
credibility level 60%, the penalty for possible violation 1000, and conservatism levels 70%,
are presented in Fig. 4.
The CV values demonstrate that the fuzzy-robust model outperforms robust and determin-
istic models in problem # 1 (see Fig. 4a). In addition, for large -size problems (see Fig. 4b, c),

123
A fuzzy-budgeted robust optimization model for joint… Page 23 of 29 194

(a) Problem #1-small size (b) Problem #1-medium size


5,000 20,000
deterministic
4,000 determinist
robust

CV(Billions)
CV(Billions)

ic 15,000
robust Fuzzy robust
3,000
10,000
2,000
5,000
1,000

0 0
10 20 30 40 50 80 10 20 30 40 50 80

perturbation level (%) perturbation level (%)

(c) Problem # 3 - large size


50,000

40,000 deterministic
CV(Billions)

robust
30,000 Fuzzy robust

20,000

10,000

0
10 20 30 40 50 80

perturbation level (%)


Fig. 5 Performance of the models under realizations for conservatism level  90% and different size problems

the robust model’s performance gets better, as good as the fuzzy-robust model. The interesting
point is that the superiority of robust and fuzzy-robust models is amplified when perturbation
level and problem size increase. Figure 4b, c illustrate this matter clearly. As pointed out in
previous section, the proposed model for 3PL is very sensitive to the uncertain parameters.
Hence, the proposed robust models with higher perturbation levels match better with the
proposed models with very sensitive parameters to the uncertainty.
Some more experiments are run to show the effect of different conservatism levels on the
proposed models’ performances, depicted in Fig. 5. As it can be seen, there is no substantial
difference between the performances for conservatism level 70% (see Fig. 4) and 90% (see
Fig. 5). In other words, the behavior of curves with conservatism level 90% is similar to the
curve with conservatism level 70%.
Another experiment is conducted so as to investigate the effects of changing the minimum
credibility level on the objective function value in the realization model. Three levels of
minimum credibility levels, i.e., 60%, 75% and 90% are considered. The results for all
problems are shown in Tables 7, 8, 9. As the minimum credibility level increases, although
there is no a clear trend in CV values, the mean also increases, due to costs incurred for
improving the supply chain stability. Moreover, in many runs, the standard deviation (Std.)
decreases by increasing the minimum credibility values which indicate higher performance
of fuzzy-robust model compared to robust and deterministic ones.

123
194 Page 24 of 29 A. Ghafarimoghadam et al.

Table 7 Performance of fuzzy-robust model under realization for different levels of minimum credibility (prob-
lem # 1, mean and Std. are in Thousands and CV is in million)
Conservatism Minimum credibility Minimum credibility Minimum credibility
level level  60% level  75% level  90%

Mean Std. CV Mean Std. CV Mean Std. CV

70% 5702 65 372,909 5702 65 372,776 5702 65 372,645


5716 132 752,332 5716 132 752,051 5716 132 751,775
5727 170 973,614 5727 170 972,927 5727 170 972,256
5760 266 1,534,964 5760 266 1,534,288 5761 266 1,533,624
5859 314 1,838,804 5859 314 1,836,957 5859 313 1,835,133
6052 529 3,199,521 6112 499 3,046,879 6047 524 3,165,616
90% 5704 63 360,197 5704 63 360,262 5704 63 360,327
5723 126 721,978 5723 126 722,133 5723 126 722,290
5732 165 946,423 5732 165 946,237 5733 165 946,062
5786 258 1,491,975 5783 252 1,458,984 5788 258 1,491,262
5876 309 1,817,232 5881 308 1,808,558 5885 300 1,766,241
6407 526 3,371,495 6394 503 3,217,961 6403 498 3,187,010

Table 8 Performance of fuzzy-robust model under realization for different levels of minimum credibility (prob-
lem # 2, mean and Std. are in Thousands and CV is in million)
Conservatism Minimum credibility Minimum credibility Minimum credibility
level level  60% level  75% level  90%

Mean Std. CV Mean Std. CV Mean Std. CV

70% 14,145 97 1,367,256 14,145 97 1,365,887 14,145 97 1,370,167


14,224 194 2,760,497 14,224 193 2,750,105 14,224 194 2,760,134
14,264 322 4,591,558 14,264 324 4,617,177 14,264 298 4,254,366
14,360 392 5,635,644 14,356 396 5,679,874 14,360 403 5,793,734
14,460 436 6,306,928 14,443 411 5,933,510 14,450 446 6,446,446
15,180 761 11,554,722 15,171 820 12,434,967 15,241 832 12,681,209
90% 14,148 94 1,335,448 14,146 93 1,312,618 14,149 94 1,328,998
14,234 194 2,761,918 14,204 215 3,060,990 14,239 195 2,782,523
14,294 293 4,183,855 14,276 277 3,952,113 14,276 278 3,975,112
14,402 395 5,692,008 14,419 388 5,594,716 14,409 396 5,702,643
14,658 492 7,209,410 14,669 488 7,158,229 14,679 490 7,193,167
15,981 831 13,276,810 15,936 715 11,392,432 16,016 869 13,911,760

4.2.1 Sensitivity analysis over penalty of violation

To provide some managerial insights, a sensitivity analysis is conducted over penalty values
of constraints violation i.e., π. The deterministic, robust and fuzzy-robust models with per-
turbation level 30%, conservatism and minimum level of credibility values respectively equal
to 70% and 60% are evaluated under the realization model. The results are demonstrated in
Figs. 6 and 7.

123
A fuzzy-budgeted robust optimization model for joint… Page 25 of 29 194

Table 9 Performance of fuzzy-robust model under realization for different levels of minimum credibility (prob-
lem # 2, mean and Std. are in Thousands and CV is in million)
Conservatism Minimum credibility Minimum credibility Minimum credibility
level level  60% level  75% level  90%

Mean Std. CV Mean Std. CV Mean Std. CV

70% 23,622 110 2,592,369 23,622 108 2,550,889 23,622 109 2,574,207
23,732 216 5,123,746 23,739 217 5,148,979 23,734 218 5,176,281
23,870 326 7,775,861 23,870 330 7,878,242 23,868 328 7,832,358
24,035 453 10,879,978 24,036 449 10,787,300 24,029 431 10,363,075
24,279 648 15,738,579 24,373 623 15,172,678 24,296 656 15,931,478
25,648 1,015 26,043,913 25,653 1,111 28,495,084 25,671 1,065 27,336,226
90% 23,658 122 2,886,700 23,657 121 2,868,002 23,657 122 2,893,444
23,775 212 5,048,431 23,773 216 5,136,906 23,773 214 5,088,722
23,896 388 9,261,800 23,986 343 8,230,426 23,905 389 9,296,890
24,344 473 11,511,144 24,303 512 12,446,964 24,225 498 12,069,528
24,633 621 15,287,219 24,517 619 15,180,771 24,499 616 15,091,544
26,996 1089 29,399,987 27,031 1058 28,585,413 27,115 1034 28,048,371

16,000
14,000
12,000 Deterministic
CV (Billions)

Robust
10,000 Fuzzy robust
8,000
6,000
4,000
2,000
10 50 100 1000 3000
Values of π
Fig. 6 CV values of models under different levels of π (Problem # 2)

As it can be observed, in the medium-size problem (i.e., Problem #2), for high amounts
of π, the robust and fuzzy-robust models are more cost-effective, while for the low amounts
of penalty, the deterministic model has a lower value of objective function (see Fig. 6). The
same sensitivity analysis is conducted over the large-size problem (i.e., Problem #3) and its
result is presented in Fig. 7. As it can be seen, the robust and fuzzy-robust models outperform
deterministic one for all penalty values.

5 Conclusions and future streams

Because of the globalization and advantages of outsourcing such as operational efficiency,


enhanced customer service levels, and a better focus on the core businesses, more com-

123
194 Page 26 of 29 A. Ghafarimoghadam et al.

16,000
15,000
14,000 Deterministic
Robust
13,000 Fuzzy robust
CV (Billions)

12,000
11,000
10,000
9,000
8,000
7,000
6,000
1 10 50 100 1000
Values of π
Fig. 7 CV values of models under different levels of π (Problem # 3)

panies outsource their logistics. Accordingly, designing an effective network as a strategic


decision has become important to both manufacturers and 3PLs. To get this goal, a joint
network design-pricing model including forward and reverse supply chains is presented in
this research for 3PLs. To do so, a mixed-integer non-linear programming (MINLP) model
was first developed and then linearized to overcome the difficulties of non-linearity. Since
the uncertainty is unavoidable in the network design supply chain problems, especially in
reverse supply chain, the robust approach was applied. Afterward, the fuzzy counterpart of
the robust model was presented to deal with the imprecise nature of the uncertainty budget
of the robust model.
Based on the experiments, we reached the following conclusions for the example network:
• The robust and also fuzzy-robust optimization models in most experiments significantly
improved the immunity and robustness of the solutions and rectified the difficulties of
scenario-based stochastic programming.
• When the shipping cost and recoverable rate individually are subject to uncertainty, the
results indicate that although shipping cost is more sensitive than recoverable rate under
different perturbation levels, the change in total cost for recoverable rate is way more than
shipping cost under different conservatism levels, especially for values more than 60%.
• As another result, conducted sensitivity analysis over different values of violation penalty in
medium-size problems (i.e., Problem # 2) showed that the robust and fuzzy-robust models
have better performance under higher values of violation penalty, while the deterministic
model is more applicable under lower values of penalties.
• In addition, in the large-sized problem (i.e., Problem # 3) the robust and fuzzy-robust
models outperformed deterministic model for all values of violation penalty.
To the best of the authors’ knowledge, this study is one of the primary researches apply-
ing joint network design-pricing model and robust approach with fuzzy uncertainty budget
value for 3PLs, simultaneously. However, it can be enhanced in some ways in the future: (1)
optimizing the network by considering multiple criteria, such as environmental and social
parameters; (2) extending the existing robust programming model into the fuzzy-robust pro-
gramming one which can address the uncertainty in the lower and upper bound of the uncertain
parameters using the fuzzy boundary interval (Nie et al. 2007); (3) and using meta-heuristic
algorithms to solve the addressed problem for very-large-size ones.

123
A fuzzy-budgeted robust optimization model for joint… Page 27 of 29 194

References
Aguezzoul A (2014) Third-party logistics selection problem: a literature review on criteria and methods.
Omega 49:69–78
Arabzad SM, Ghorbani M, Tavakkoli-Moghaddam R (2015) An evolutionary algorithm for a new multi-
objective location-inventory model in a distribution network with transportation modes and third-party
logistics providers. Int J Prod Res 53(4):1038–1050
Aras N, Aksen D, Tanuğur AG (2008) Locating collection centers for incentive-dependent returns under a
pick-up policy with capacitated vehicles. Eur J Oper Res 191(3):1223–1240
Azadi M, Saen RF (2011) A new chance-constrained data envelopment analysis for selecting third-party
reverse logistics providers in the existence of dual-role factors. Expert Syst Appl 38(10):12231–12236
Ben-Tal A, Nemirovski A (1999) Robust solutions of uncertain linear programs. Oper Res Lett 25(1):1–13
Bertsimas D, Sim M (2004) The price of robustness. Oper Res 52(1):35–53
Bottani E, Rizzi A (2006) A fuzzy TOPSIS methodology to support outsourcing of logistics services. Supply
Chain Manag Int J 11(4):294–308
Chang C-T, Chang C-C (2000) A linearization method for mixed 0–1 polynomial programs. Comput Oper
Res 27(10):1005–1016
Chinneck JW (2006) Practical optimization: a gentle introduction. Systems and Computer Engineering. Car-
leton University, Ottawa. http://www.sce.carleton.ca/faculty/chinneck/po.html
Choi T-M, Li D, Yan H (2004) Optimal returns policy for supply chain with e-marketplace. Int J Prod Econ
88(2):205–227
Daghigh R, Jabalameli M, Amiri A, Pishvaee M (2016) A multi-objective location-inventory model for 3PL
providers with sustainable considerations under uncertainty. Int J Ind Eng Comput 7(4):615–634
Du F, Evans GW (2008) A bi-objective reverse logistics network analysis for post-sale service. Comput Oper
Res 35(8):2617–2634
Ecer F (2018) Third-party logistics (3PLs) provider selection via Fuzzy AHP and EDAS integrated model.
Technol Econ Dev Econ 24(2):615–634
Fattahi M, Govindan K (2017) Integrated forward/reverse logistics network design under uncertainty with
pricing for collection of used products. Ann Oper Res 253(1):193–225
Fattahi M, Govindan K, Keyvanshokooh E (2018) A multi-stage stochastic program for supply chain network
redesign problem with price-dependent uncertain demands. Comput Oper Res 100:314–332
Fleischmann M, Krikke HR, Dekker R, Flapper SDP (2000) A characterisation of logistics networks for
product recovery. Omega 28(6):653–666
Ghafarimoghadam A, Karimi A, Mousazadeh M, Pishvaee MS (2016) A robust optimisation model for reman-
ufacturing network design problem with one-way substitution. Int J Serv Oper Manag 24(4):484–503
Ghaffari-Nasab N, Ghazanfari M, Teimoury E (2016) Hub-and-spoke logistics network design for third party
logistics service providers. Int J Manag Sci Eng Manag 11(1):49–61
Ghayebloo S, Tarokh MJ, Venkatadri U, Diallo C (2015) Developing a bi-objective model of the closed-loop
supply chain network with green supplier selection and disassembly of products: the impact of parts
reliability and product greenness on the recovery network. J Manuf Syst 36:76–86
Giri B, Sarker BR (2017) Improving performance by coordinating a supply chain with third party logistics
outsourcing under production disruption. Comput Ind Eng 103:168–177
Glover F (1975) Improved linear integer programming formulations of nonlinear integer problems. Manag
Sci 22(4):455–460
Hasani A, Zegordi SH, Nikbakhsh E (2012) Robust closed-loop supply chain network design for perishable
goods in agile manufacturing under uncertainty. Int J Prod Res 50(16):4649–4669
Kayvanfar V, Husseini SM, Sajadieh MS, Karimi B (2018) A multi-echelon multi-product stochastic model
to supply chain of small-and-medium enterprises in industrial clusters. Comput Ind Eng 115:69–79
Keyvanshokooh E, Fattahi M, Seyed-Hosseini S, Tavakkoli-Moghaddam R (2013) A dynamic pricing
approach for returned products in integrated forward/reverse logistics network design. Appl Math Model
37(24):10182–10202
Ko HJ, Evans GW (2007) A genetic algorithm-based heuristic for the dynamic integrated forward/reverse
logistics network for 3PLs. Comput Oper Res 34(2):346–366
Ko HJ, Park BI (2005) The impacts of an integrated solution on the logistics networks of third-party logistics
providers. J Eastern Asia Soc Transp Stud 6:2762–2777
Kumar S, Muddada RR, Pandey M, Mahanty B, Tiwari M (2013) Logistics planning and inventory optimization
using swarm intelligence: a third party perspective. Int J Adv Manuf Technol 65(9–12):1535–1551

123
194 Page 28 of 29 A. Ghafarimoghadam et al.

Lee D-H, Dong M (2008) A heuristic approach to logistics network design for end-of-lease computer products
recovery. Transp Res Part E Logist Transp Rev 44(3):455–474
Lee H, Zhang T, Boile M, Theofanis S, Choo S (2013) Designing an integrated logistics network in a supply
chain system. KSCE J Civ Eng 17(4):806–814
Li H-L (1994) Global optimization for mixed 0–1 programs with convex or separable continuous functions. J
Oper Res Soc 45(9):1068–1076
Li Z, Floudas CA (2012) Robust counterpart optimization: Uncertainty sets, formulations and probabilistic
guarantees. In: Proceedings of the 6th conference on foundations of computer-aided process operations,
Savannah (Georgia)
Li X, Liu B (2006) A sufficient and necessary condition for credibility measures. Int J Uncertain Fuzziness
Knowl Based Syst 14(05):527–535
Li Z, Ding R, Floudas CA (2011) A comparative theoretical and computational study on robust counterpart
optimization: I. Robust linear optimization and robust mixed integer linear optimization. Ind Eng Chem
Res 50(18):10567–10603
Li Y, Guo H, Zhang Y (2018) An integrated location-inventory problem in a closed-loop supply chain with
third-party logistics. Int J Prod Res 56(10):3462–3481
Liang Y, Pokharel S, Lim GH (2009) Pricing used products for remanufacturing. Eur J Oper Res
193(2):390–395
Liu B (2004) Uncertainty theory: an introduction to its axiomatic foundations. Springer, Berlin
Liu B, Liu Y-K (2002) Expected value of fuzzy variable and fuzzy expected value models. IEEE Trans Fuzzy
Syst 10(4):445–450
Mallidis I, Dekker R, Vlachos D (2012) The impact of greening on supply chain design and cost: a case for a
developing region. J Transp Geogr 22:118–128
Min H, Ko H-J (2008) The dynamic design of a reverse logistics network from the perspective of third-party
logistics service providers. Int J Prod Econ 113(1):176–192
Mulvey JM, Vanderbei RJ, Zenios SA (1995) Robust optimization of large-scale systems. Oper Res
43(2):264–281
Nie XH, Huang GH, Li YP, Liu L (2007) IFRP: A hybrid interval-parameter fuzzy robust programming
approach for waste management planning under uncertainty. J Environ Manag 84(1):1–11
Nukala S, Gupta SM (2007) A fuzzy mathematical programming approach for supplier selection in a closed-
loop supply chain network. In: Proceedings of the 2007 POMS-Dallas meeting
Oral M, Kettani O (1992) A linearization procedure for quadratic and cubic mixed-integer problems. Oper
Res 40(1):S109–S116
Pishvaee MS, Rabbani M, Torabi SA (2011) A robust optimization approach to closed-loop supply chain
network design under uncertainty. Appl Math Model 35(2):637–649
Pokharel S, Mutha A (2009) Perspectives in reverse logistics: a review. Resour Conserv Recycl 53(4):175–182
Prakash S, Kumar S, Soni G, Jain V, Rathore APS (2018) Closed-loop supply chain network design and
modelling under risks and demand uncertainty: an integrated robust optimization approach. Ann Oper
Res 1–28
Ramezani M, Kimiagari AM, Karimi B, Hejazi TH (2014) Closed-loop supply chain network design under a
fuzzy environment. Knowl Based Syst 59:108–120
Sen DK, Datta S, Mahapatra SS (2017) Decision support framework for selection of 3PL service providers:
dominance-based approach in combination with grey set theory. Int J Inf Technol Decis Mak 16(01):25–57
Soyster AL (1973) Convex programming with set-inclusive constraints and applications to inexact linear
programming. Oper Res 21(5):1154–1157
Suyabatmaz AÇ, Altekin FT, Şahin G (2014) Hybrid simulation-analytical modeling approaches for the reverse
logistics network design of a third-party logistics provider. Comput Ind Eng 70:74–89
Talaei M, Moghaddam BF, Pishvaee MS, Bozorgi-Amiri A, Gholamnejad S (2016) A robust fuzzy optimization
model for carbon-efficient closed-loop supply chain network design problem: a numerical illustration in
electronics industry. J Clean Prod 113:662–673
Taleizadeh AA, Haghighi F, Niaki STA (2019) Modeling and solving a sustainable closed loop supply chain
problem with pricing decisions and discounts on returned products. J Clean Prod 207:163–181
Xanthopoulos A, Iakovou E (2009) On the optimal design of the disassembly and recovery processes. Waste
Manag 29(5):1702–1711
Yalabik B, Petruzzi NC, Chhajed D (2005) An integrated product returns model with logistics and marketing
coordination. Eur J Oper Res 161(1):162–182
Zhang Y, Xie L, Hang W, Cui X (2007) A robust model for 3PLS to design a remanufacturing logistics network
under the uncertain environment. Automation and Logistics, 2007 IEEE International Conference on
IEEE

123
A fuzzy-budgeted robust optimization model for joint… Page 29 of 29 194

Zhu H, Zhang J (2009) A credibility-based fuzzy programming model for APP problem. Artificial Intelligence
and Computational Intelligence, 2009. AICI’09. International Conference on IEEE

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

123

You might also like