Computers & Industrial Engineering 129 (2019) 192–209

Contents lists available at ScienceDirect

Computers & Industrial Engineering

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

Sustainable supplier selection and order allocation: A novel multi-objective T

programming model with a hybrid solution approach
Hadi Moheb-Alizadeha, Robert Handfieldb,
⁎

a
Graduate Program in Operations Research, North Carolina State University, Raleigh, NC 27695-7913, USA
b
Department of Business Management, College of Management, North Carolina State University, Raleigh, NC 27695-7229, USA

ARTICLE INFO ABSTRACT

Keywords: Sustainable supplier selection is the process of identifying the appropriate supply partners of an organization
Supplier selection with the most beneficial monetary value, while diminishing the various effects of its operations on society and
Order allocation environment. Therefore, it plays a significant role in moving an organization toward sustainable development.
Sustainability Alongside sustainable supplier selection, assigning proper quantity of orders to suppliers satisfying sustainability
-constraint method
criteria is another activity that should be simultaneously carried out in order to develop a comprehensive tool in
Benders decomposition algorithm
Data envelopment analysis
sustainable supplier management. Meanwhile, the issue of sustainable supplier selection and order allocation has
attracted limited attention in the literature. The present paper aims at developing an inclusive multi-objective
mixed integer linear programming model, which accounts for multiple periods, multiple products and multi-
modal transportation, to evaluate suppliers and allocate order quantities. Furthermore, the developed model that
includes all sustainability aspects also takes both shortage and discount conditions into account. A hybrid three-
step solution methodology is then presented, using which the original multi-objective problem is firstly con-
verted to a single objective model by -constraint method. In the second step, the current single objective
programming model is solved using the Benders decomposition algorithm that in turn is accelerated by a variety
of algorithmic enhancements. Finally, among all the Pareto optimal solutions of the original multi-objective
programming problem, the preferable solution is intelligently selected based on the DEA super efficiency score of
all purchasing firms as a decision support tool. The applicability of the proposed approach is illustrated by a real-
world case study in automotive industry.

1. Introduction multi-criteria problem, and that it is important to consider trade-offs

between tangible and intangible factors in selecting suppliers. Supplier
The supplier selection decision process is an important determinant selection can be made more complicated by the fact that different and
of improved supply chain performance. This decision involves evalu- sometimes conflicting criteria must be factored into the decision.
ating and selecting the best suppliers for a sourcing requirement, and As the result of increased customer knowledge and ecological
assigning production volumes to selected suppliers. The decision results pressures from markets and various stakeholders, organizations must
in reduced purchasing cost and improved organizational competitive- consider environmental and social impacts of their operations, in ad-
ness. For these reasons, supplier selection is considered a strategic de- dition to traditional economic profits. The concern around environ-
cision in the field of supply chain management (Moheb-Alizadeh, mental and social impacts of business activities has led to a new
Mahmoudi, & Bagheri, 2017; Willis, Huston, & Pohlkamp, 1993). paradigm called sustainable development. Sustainable development is
Multiple factors must be considered in evaluating suppliers during the defined as an organizational development that meets the needs of the
selection decision. Dickson (1966) identified 23 criteria determined by present without compromising the ability of future generations to meet
purchasing managers for different types of supplier selection problems. their own needs (WCED (1987)). While diverse interpretations of sus-
A review of methods and criteria of supplier selection by Weber, tainability exist, a central concept employed in operationalizing sus-
Current, and Benton (1991) revealed that 47 out of 76 reviewed papers tainability is the triple bottom line approach, which stipulates that
considered multiple criteria during the evaluation process. This re- performance must be considered in the three dimensions of economic,
search ascertains that the supplier selection problem is intrinsically a environmental and social performance (Elkington (1998)). Dyllick and

⁎
Corresponding author at: 2806-A Hillsborough St., Upper Level, Campus Box 7229, Raleigh, NC 27695-7229, USA.
E-mail addresses: [email protected] (H. Moheb-Alizadeh), [email protected] (R. Handfield).

https://doi.org/10.1016/j.cie.2019.01.011
Received 10 August 2018; Received in revised form 26 November 2018; Accepted 4 January 2019
Available online 09 January 2019
0360-8352/ © 2019 Elsevier Ltd. All rights reserved.
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

Hockerts (2002) framed the three dimensions of sustainability as the Therefore, the preferable solution is chosen among all Pareto optimal
business case (economic), the natural case (environmental), and the solutions in such a way that the purchasing firms are as efficient as
societal case (social). Accordingly, practitioners and academia alike possible. In this case, the decision maker is not required to select from
have emphasized the field of sustainable supply chain management numerous Pareto optimal solutions for the preferable solution based on
(Brandenburg, Govindan, Sarkis, & Seuring, 2014; Moheb-Alizadeh, individual (and thus subjective) criteria. Instead, an analytical decision
2018; Seuring & Muller, 2008). support tool is developed to assist decision makers in selecting the
In the context of increased focus on sustainability, it is therefore preferable solution intelligently. The proposed methodology in this step
critical that organizations source from economically, environmentally presents several unique advantages over previous studies: (i) our lit-
and socially responsible suppliers (Buyukozkan & Cifci, 2011). How- erature review suggests that the majority of former studies have applied
ever, there are multiple challenges that arise in simultaneously seeking DEA only in supplier selection problem, and (ii) in all reviewed papers,
to optimize economic, environmental and social dimensions of perfor- the primary resources and performance measures were inputs and
mance within supplier selection processes (Govindan, Rajendran, outputs of DEA model, respectively. This standpoint was initially pro-
Sarkis, & Murugesan, 2015; Zimmer, Frohling, & Schultmann, 2016). posed by Boussofiane, Dyson, and Thanassoulis (1991). In these studies,
This has given rise to the sustainable supplier selection problem, which a supplier evaluation criterion, e.g. purchasing cost, whose smaller
considers both environmental and social elements of suppliers perfor- value is desirable was the primary input while an evaluation criterion
mance (Zimmer et al., 2016). Prior research has included economic and such as quality whose greater value is satisfactory was the primary
environmental aspects of sustainability in supplier selection (Govindan output of the DEA model. These criteria are not in seeking to under-
et al., 2015), however the social responsibility dimension is often stand the overall efficiency of purchasing firms relative to other forms
considered separately or overlooked altogether (Bai & Sarkis, 2010; of sustainable value defined earlier.
Govindan, Khodaverdi, & Jafarian, 2013; Kuo, Wang, & Tien, 2010; Therefore, the challenges tackled in the developed solution meth-
Moheb-Alizadeh & Handfield, 2018; Punniyamoorthy, Mathiyalagan, & odology are threefold: (i) the multiple objective functions dealt with by
Parthiban, 2011). In addition, there is a dearth of research that con- -constraint method, (ii) the large-scale property of the proposed pro-
siders all three dimensions of sustainable supplier selection and order blem tackled by BDA, and (iii) choosing the final preferable solution
allocation simultaneously. In sum, research that simultaneously con- resolved by DEA super efficiency. The proposed framework in this
siders both sustainable supplier selection criterion and order allocation paper to sustainable supplier selection and order allocation is depicted
is nascent (Azadnia, Saman, & Wong, 2015). in Fig. 1.
In addressing this gap, a novel and comprehensive multi-objective In summary, the contributions of the present paper are as follows:
mixed integer linear programming (MILP) model is proposed in this
paper, which considers all sustainability aspects (economic, environ- • A novel and comprehensive multi-objective MILP model to sus-
mental and social responsibility) aligned with the triple bottom line tainable supplier selection and order allocation problem is devel-
approach. The proposed model models the decision across multiple oped, which accounts for multiple periods, multiple products and
periods, multiple products and multimodal transportation options in multimodal transportation. The developed model, which includes all
the presence of shortage and discount conditions. sustainability aspects, also takes both shortage and discount condi-
As the first step of the solution procedure, the -constraint method is tions into account.
used in this paper in order to deal with the multiple objective functions • The original multi-objective MILP model is converted to a single
and derive a single objective MILP model. In this model, the objective objective programming model using -constraint method.
function considers the economic aspect of sustainability as the primary • A Benders decomposition algorithm that is computationally ac-
objective function of the model. Accordingly, the objective functions celerated is developed to solve the single objective programming
associated with the environmental and social responsibility aspects are model and derive a Pareto optimal solution.
treated as appropriately defined constraints. • Using physical inputs and outputs, DEA super-efficiency score of
The resulting single objective MILP model is a very complex large- purchasing firms is obtained to assist decision maker with in-
scale sustainable supplier selection and order allocation model, due to telligently choosing the preferable solution among all Pareto op-
the numerous raw materials considered, a long list of suppliers for each timal solutions. Furthermore, the efficiency score of purchasing
raw material and various types of transportation modes. Therefore, in firms themselves are attained.
the second step of the solution procedure, an effective solution meth- • The impact of suppliers’ performance on the DEA super-efficiency
odology based on Benders decomposition algorithm (BDA) is developed score of purchasing firms can be clearly explored.
to solve the single objective MILP model and consequently derive a
Pareto optimal solution for the original multi-objective MILP problem. To the best of the authors’ knowledge, our approach to the sus-
In addition, a variety of enhancements are deployed to accelerate the tainable supplier selection and order allocation problem is unique as
Benders decomposition algorithm, including: (i) adding valid inequal- prior studies have not addressed this problem.
ities to the master problem to reduce the number of feasibility cuts, (ii) The rest of this paper is organized as follows: Section 2 reviews
generating enhanced Pareto optimal cuts to exclude a larger space of previous studies in the areas of sustainable supplier selection and order
the master problem, (iii) specializing the local branching search to the allocation, and also DEA application to this problem. Section 3 in-
derived single objective MILP model to concurrently improve both troduces a new multi-objective MILP model to sustainable supplier se-
lower and upper bounds during the execution of Benders decomposi- lection and order allocation. The proposed solution procedure is ela-
tion, and (iv) generating knapsack cuts for master problem. borated in Section 4. Section 5 explains how to implement the
Once a set of Pareto optimal solutions is derived, the preferred so- developed model and its solution procedure using a real-world case
lution is selected using data envelopment analysis (DEA) super effi- study. Finally, Section 6 is devoted to conclusions and future works.
ciency in the final step of the developed solution methodology in an
innovative fashion, where physical inputs and outputs are derived to
acquire DEA super efficiency of purchasing firms. The physical inputs 2. Literature review
(raw materials purchased from suppliers) are attained from Pareto
optimal solutions. On the other hand, physical outputs (end products) The papers reviewed in this section are classified into two cate-
manufactured by purchasing firms are estimated using properly defined gories: (i) the studies examining sustainable supplier selection and/or
production functions. Afterwards, the DEA super efficiency of pur- order allocation, and (ii) the papers that apply DEA model to evaluate
chasing firms is derived using the obtained physical inputs and outputs. suppliers and/or to allocate orders.

193
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

Construct a new multi-objective Convert the original multi-

programming model for objective programming problem
sustainable supplier selection and to a single objective model using
order allocation problem. ε-constraint method.

Solve the single objective

programming model using the
accelerated BDA to derive a set of
Pareto optimal solutions.

Calculate the value of each raw

Derive the proper production material (physical input) shipped
function using historical data. to all purchasing firms in each
Pareto optimal solution

Estimate the value of each end- Derive the DEA super efficiency
product (physical output) of all score of all purchasing firms.
purchasing firms.

Select suppliers and assigning Choose the Pareto optimal

order quantities according to the solution with the maximum DEA
selected preferable solution. super efficiency score as the
preferable solution.

Fig. 1. The proposed methodology in this paper.

2.1. Sustainable supplier selection and/or order allocation makers’ assessments, Amindoust, Ahmed, Saghafinia, and
Bahreininejad (2012) applied fuzzy logic and proposed a ranking
Table 1 lists papers studying sustainable supplier selection and/or method on the basis of fuzzy inference system for supplier selection
order allocation problem, in which all sustainability aspects are in- problem. Identifying 22 sustainable supplier selection criteria and three
cluded. In this table, SS and SSOA stand for supplier selection and dimensions of criteria, i.e. economic, environmental, and social, Luthra,
supplier selection and order allocation, respectively. Govindan, Kannan, Mangla, and Garg (2017) used an integrated ana-
In particular, Bai and Sarkis (2010) utilized grey system and rough lytical tool to evaluate sustainable suppliers. Orji and Wei (2015) pre-
set theory to sustainable supplier selection problem by explicitly con- sented a modeling approach of integrating information on supplier
sidering sustainability attributes. To handle the subjectivity of decision behavior in fuzzy environment with system dynamics simulation

Table 1
Review of previously published literature for sustainable supplier selection and/or order allocation.
Article Problem Methodology Sustainability Multimodal Multiple Multiple Shortage Discount

Eco Env Soc Transportation Periods Products Condition Condition

Bai and Sarkis (2010) SS Rough set theory and Grey system • • •
Amindoust et al. (2012) SS Fuzzy inference • • •
Luthra et al. (2017) SS AHP, VIKOR, multi-criteria optimization, compromise • • •
solution
Orji and Wei (2015) SS Fuzzy logic, Systems dynamics Bayesian framework, • • •
Sarkis and Dhavale (2015) SS Monte Carlo Markov Chain • • •
Buyukozkan and Cifci SS Fuzzy ANP, Incomplete preference relations • • •
(2011)
Trapp and Sarkis (2016) SS MILP • • • •
Song et al. (2017) SS DEMATEL, pairwise comparison, rough sets • • •
Kumar et al. (2017) SSOA fuzzy AHP and fuzzy multi-objective linear • • • •
programming
Gupta et al. (2016) SSOA AHP, fuzzy multi-objective integer linear • • • • •
programming
Aktin and Gergin (2016) SSOA MILP • • • •
Azadnia et al. (2015) SSOA Rule-based weighted fuzzy method, fuzzy AHP, • • • • •
MOPP
Ghadimi et al. (2018) SSOA Multi-Agent system, MODM • • • •
Shalke et al. (2017) SSOA MODM • • • • • •
The present paper SSOA MODM and DEA super efficiency • • • • • • • •

194
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

modeling technique. Taking a triple bottom line approach and con- criteria were transformed into the corresponding quantitative ones.
sidering business operations as well as environmental impacts and so- Saen (2008) proposed an algorithm for ranking suppliers in the pre-
cial responsibilities of the suppliers, Sarkis and Dhavale (2015) devel- sence of volume discount based on super efficiency analysis. Saen
oped an approach based on a Bayesian framework and Monte Carlo (2010) applied an imprecise DEA model considering weight restriction
Markov chain simulation. Buyukozkan and Cifci (2011) developed an and imprecise data simultaneously to evaluate suppliers. Jafari-
approach based on fuzzy analytic network process within multi-person Songhori, Tavana, Azadeh, and Khakbaz (2011) presented a DEA model
decision-making schema to evaluate sustainable suppliers. Integrating and a multi-objective MILP model to supplier selection and order al-
supplier sustainability in the form of ratings and sustainable supplier location problem, where the former (latter) model was used to supplier
development through investment and training budgets, Trapp and selection (order allocation). Kuo and Lin (2012) considered supplier
Sarkis (2016) developed an optimization model that simultaneously selection problem using analytical network process and DEA, where the
addressed supplier selection, supplier development, and sustainability latter approach was used to consider the interdependency among cri-
considerations. Song, Xu, and Liu (2017) developed a hybrid approach teria. Falagario, Sciancalepore, Costantino, and Pietroforte (2012) ad-
to explore interrelationship of sustainability aspects in supplier selec- dressed the decision problem of supplier selection by applying DEA
tion problem. Kumar, Rahman, and Chan (2017) used integrated fuzzy cross-efficiency. Dotoli and Falagario (2012) addressed supplier selec-
analytical hierarchy process and fuzzy multi-objective linear program- tion and order allocation problem using DEA, TOSIS and linear pro-
ming approach for order allocation among suppliers. By integrating gramming, where DEA and TOPSIS were used to identify appropriate
fuzzy multi-objective integer linear programming and analytic hier- suppliers and linear programming to allocate orders. Karsak and Dursun
archy process techniques, Gupta, Govindan, Mehlawat, and Kumar (2014) implemented an imprecise DEA for supplier selection, which
(2016) developed an optimization model for supplier selection and employed the weights of supplier assessment criteria computed by
order allocation. Aktin and Gergin (2016) introduced a questionnaire fuzzy weighted average. Zhou, Pedrycz, Kuang, and Zhang (2016) de-
for measuring the sustainability scores of a company’s potential sup- veloped a multi-objective DEA model in a setting of type-2 fuzzy
pliers and developed mixed integer linear programming models for modeling to evaluate and select the most sustainable suppliers. Yousefi,
distributing demand to the most sustainable suppliers of the company. Soltani, Saen, and Pishvaee (2017) proposed a network goal program-
Azadnia et al. (2015) proposed an integrated approach for sustainable ming and DEA model for sustainable supplier selection problem, where
supplier selection and order allocation combined with multi-period and inputs and outputs of DEA were the goals determined by managers and
multi-product lot-sizing problem. Ghadimi, Ghassemi-Toosi, and experts.
Heavey (2018) developed a multi-agent systems approach for sustain-
able supplier selection and order allocation that resulted in a more co- 3. Model development
operative partnership. Shalke, Paydar, and Hajiaghaei-Keshteli (2017)
developed a multi-objective mathematical programming model to A novel multi-objective MILP model for sustainable supplier selec-
consider sustainable supplier selection and order allocation simulta- tion and order allocation problem with multiple periods, multiple
neously in a multi-period, multi-item and multi-supplier supply chain products and multimodal transportation in the presence of price dis-
with quantity discount. count and shortage conditions, which henceforth is named as SSSOA
The papers reviewed in this subsection have examined all sustain- model, is developed in this section using the following indexes, para-
ability aspects, i.e. economic, environmental and social responsibility, meters and decision variables.
in their proposed frameworks. However, there are studies in the lit-
erature in which not all sustainability pillars are explored (Govindan Indexes
et al., 2015). An immediate contribution of the present paper over the i: Index of suppliers; i = 1, …, m
current set of studies is the inclusion of all sustainability aspects. j: Index of raw materials; j = 1, …, n
k: Index of purchasing firms; k = 1, …, K
2.2. DEA application to supplier selection and/or order allocation r: Index of end products; r = 1, …, s
l: Index of price levels by suppliers; l = 1, …, Li
Table 2 categorizes the related research papers, where a variant of p: the index of transportation modes; p = 1, …, P
DEA model is applied to evaluate suppliers and/or to allocate orders. t: Index of period; t = 1, …, T
In this context, Weber (1996) applied DEA model to identify the
best suppliers based on their operational criteria. Liu, Ding, and Lall Parameters
(2000) exploited DEA model proposed by Banker and Moray (1986) to demtjk : required demand of raw material j in purchasing firm k in
evaluate the total efficiency of suppliers and decreased the number of period t.
suppliers based on their efficiency scores. Developing a multi-objective repijt : rejection percentage of raw material j provided by supplier i in
programming model for supplier selection problem, Weber, Current, period t.
and Dessai (2000) exploited the values of objective functions corre- pprijlt : unit purchasing price of raw material j at price level l provided
sponding to the Pareto optimal solutions as the inputs of DEA model by supplier i in period t.
and specified the number of suppliers. Talluri, Narasimhan, and Nair t
fscijkp : fixed shipment cost of raw material j shipped from supplier i
(2006) used chance constrained DEA to take intrinsic variation in to purchasing firm k by transportation mode p in period t.
characteristics of suppliers performance into account and subsequently t
vscijkp : variable shipment cost of unit raw material j shipped from the
to evaluate suppliers risk. When suppliers do not use the same inputs supplier i to purchasing firm k by transportation mode p in period t.
and outputs, Saen (2007) proposed an approach using analytical hier- mshtjk : maximum amount of shortage for raw material j at purchasing
archy process and interpolation to estimate missed data and then ap- firm k in period t.
plied chance constrained DEA to acquire relative efficiency of each ststjk : maximum available storing space for raw material j in pur-
supplier. Celebi and Bayraktar (2008) developed a hybrid approach chasing firm k in period t.
including neural network and DEA with missed data for supplier se- nsutjk : maximum number of suppliers hired by purchasing firm k to
lection problem. Wu (2009) presented a hybrid approach using DEA, provide raw material j in period t.
decision tree and neural network, where DEA model was applied to otdijt : on-time delivery percentage of raw material j provided by
classify suppliers into efficient and inefficient groups. Ha and Krishnan supplier i in period t.
(2008) applied a hybrid approach including analytical hierarchy pro-
apdjk t
: minimum acceptable percentage for on-time delivery of raw
cess, DEA and neural networks to evaluate suppliers, where qualitative material j to the purchasing firm k in period t.

195
 H. Moheb-Alizadeh, R. Handfield

Table 2
A taxonomy of papers studying supplier selection and/or order allocation by DEA.
Article Problem Methodology DEA inputs DEA output

Weber (1996) SS Basic DEA model Unite price and % rejects % on-time delivery
Liu et al. (2000) SS A variant of DEA model Unite price, delivery performance and distance factor Supply variety and quality
Weber et al. (2000) SSOA MODM and DEA Unite price, % rejectsand % late delivery A constant value as output
Talluri et al. (2006) SS Chance constrained DEA Unite price %On-time delivery and % acceptance
Saen (2007) SS AHP and chance constrained General inputs (2 unspecified inputs) General inputs (2 unspecified outputs)
DEA
Celebi and Bayraktar SS Neural network and DEA A fixed input On time delivery, price, quality and serving
(2008)
Wu (2009) SS Decision trees, neural Quality management practices, self-audit, process capability, firm Quality,price,delivery,cost reduction performance and other
networks and DEA management, design capabilities and cost reduction capability

196
Saen (2008) SS DEA super efficiency Price and number of shipments Number of bills without error,number of on time shipment
Saen (2010) SS Imprecise DEA Total shipments costs, distance and supplier reputation Number of bills received from the supplier without error
Jafari-Songhori et al. SSOA DEA and MODM Supplier price, ordering costs, transportation costs Supplier Quality Score, lead time, variance lead time
(2011)
Kuo and Lin (2012) SSOA Analysis network process and delivery schedule, price, manufacturing process improvement, Delivery quality, cooperation, environmental planning, green purchasing, environmental
DEA inspection situation of qualitive control and implementation situation systems, environmental administration systems, technological capability,manufacturing
of quality control capability,financial status and supply quality assurance
Falagario et al. (2012) SS DEA cross efficiency Price, execution time Post-delivery maintenance and enhancement plans
Dotoli and Falagario SSOA DEA, TOPSIS and linear Price, ICT integration cost, geographical distance On time delivery, quality, lead time
(2012) programming
Karsak and Dursun SS Imprecise DEA and quality A dummy input with a constant value of 1 Product volume, delivery, payment method, supply variety,reliability, experience, early
(2014) function deployment relationship management, location
Zhou et al. (2016) SS Fuzzy DEA Technology and financial capabilities, environmental and safety costs Number of on-time shipments, number of bills without errors
Yousefi et al. (2017) SS DEA, and neural network Goals determined by managers/experts Goals determined by managers/experts
The present paper SSOA MODM and DEA super Physical raw materials Physical end products
efficiency
Computers & Industrial Engineering 129 (2019) 192–209
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

apr jk t
: maximum acceptable percentage for rejecting raw material j Emission objective function: Objective function (2) minimizes total
in purchasing firm k in period t. CO2 emission, which consists of transportation emission, processing
capijt : capacity of supplier i to provide raw material j in period t. emission of suppliers and holding emission of extra raw materials over
t
fseijkp : fixed shipment CO2 emission of raw material j shipped from all time periods. This objective function is related to the environmental
supplier i to purchasing firm k by transportation mode p in period t. aspect of sustainability.
vseijkpt
: variable shipment CO2 emission of raw material j shipped
T P m n Li K
from supplier i to purchasing firm k by transportation mode p in t t t
Min TE = fseijkp Xijlk / tcaijkp +
period t. t=1 p=1 i=1 j=1 l=1 k=1
hoetjk : holding CO2 emission of each unit of raw material j in pur- T P m v K T m v Li K
chasing firm k in period t. t
Xijkp t
vseijkp + t
Xijlk preij +
conijk : fixed contract cost of supplier i with purchasing firm k for raw t=1 p=1 i =1 j=1 k=1 t=1 i=1 j=1 l=1 k=1
material j. T K v
z jp : 1 if raw material j can be transported by transportation mode p; ILP tjk hoetjk
otherwise 0. t=1 k=1 j=1

lbrjk : lower bound of rating value on raw material j for purchasing (2)
firm k.
Social responsibility objective function: This objective function that is
rweij : rating weight of supplier i for providing raw material j.
associated with social responsibility aspect of sustainability contains
tcaijkpt
: capacity of transportation mode p for raw material j from
supplier i to the purchasing firm k at time period t. two components including job opportunities created and lost days
hoc tjk : holding cost of each unit of raw material j in purchasing firm k caused from works damages. These components are formulated as the
in period t. first and second terms in objective function (3), respectively, where 1
sphj : required space for holding a unit of raw material j. and 2 are normalizing weights. The definition given by Devika,
Jafarian, and Nourbakhsh (2014) is adapted in formulating the current
ijl : upper limit of purchase volume for raw material i in supplier j at
t
terms.
price level l in period t; tij1 t
ij2 … tijLi .
shc jk : shortage cost of each unit of raw material j in purchasing firm
t T m v Li K
k in period t. Max SR = 1
t
vark Xijlk / pcakt
pcakt : production capacity of purchasing firm k in period t. t=1 i =1 j=1 l=1 k=1

Li : number of price levels offered by supplier i. T m v Li K

preij : processing CO2 emission of material j by supplier i. 2
t
losk Xijlk /pcakt
losk : lost days caused by works damages in purchasing firm k. t=1 i=1 j =1 l =1 k=1

vark : variable number of created jobs at purchasing firm k. T m v Li K

t
M: a big positive number. = 1 vark 2 losk Xijlk / pcakt
t=1 i=1 j=1 l=1 k=1 (3)
Decision variables
t
Xijlk : order quantity of raw material j at price level l provided by
supplier i for purchasing firm k in period t. 3.2. Constraints
t
Yijlk : 1 if supplier i provides raw material j at price level l to pur-
chasing firm k in period t; Otherwise 0. Constraint (4) determines the amount of each raw material held by
t
Qijkp : amount of raw material j transported from supplier i to pur- each purchasing firm at the end of each time period, while constraint
chasing firm k by using the transportation mode p at time period t. (5) derives the shortage amount for each raw material faced by each
H tjk : if raw material j is held by purchasing firm k in period t; purchasing firm at the end of each period, where ILP jk
0 0
= ILP jk =0:
Otherwise, i.e. shortage condition, 0.
t
Uijkp : 1 if raw material j is transported from supplier i to purchasing m P

firm k by transportation mode p at time period t; Otherwise 0. ILP tjk = ILP tjk 1 + t
Qijkp demtjk ILP tjk 1 ; k, j, t
i=1 p =1 (4)
ILP tjk : inventory level of material j at the end of period t in pur-
chasing firm k. m P
ILP tjk : shortage level of material j at the end of period t in purchasing ILP tjk = ILP tjk 1 + demtjk t
Qijkp ILP tjk 1 ; k, j, t
firm k. i=1 p =1 (5)

3.1. Objective functions Constraint (6) requires the total rejected amount of each raw ma-
terial in each purchasing firm be less than or equal to its respective
Cost objective function: Objective function (1) that is associated with predetermined quantity at each time period.
the economic aspect of sustainability minimizes total cost over the time m Li
horizon. It includes the purchasing cost, fixed and variable transpor- t
Xijlk repijt demtjk apr jk
t
; k, j, t
tation costs, holding cost of extra raw materials, shortage cost and fixed i =1 l=1 (6)
contract cost over all time periods.
That the total amount of each raw material delivered on-time to
T m v Li K T P m n K each purchasing firm at each time period should be greater than or
t ppr t + t vsc t +
Min TC = Xijlk ijl Qijkp ijkp equal to its corresponding desired quantity is satisfied by constraint (7).
t= 1 i =1 j= 1 l =1 k = 1 t= 1 p = 1 i = 1 j = 1 k = 1
T P m v Li K m Li
t
fscijkp t / tca t
Xijlk ijkp + t
Xijlk otdijt demtjk apdjk
t
; k, j, t
t= 1 p =1 i =1 j =1 l =1 k = 1
i =1 l=1 (7)
T K v T K v
ILP tjk hoc tjk + ILP tjk shc tjk+
t= 1 k =1 j =1 t = 1 k =1 j= 1 Constraint (8) implies the total weighted amount of each raw ma-
T m n Li K terial purchased from all suppliers to each purchasing firm should ex-
t con t
Yijlk ijk ceed the respective pre-specified lower bound of rating value at each
t= 1 i =1 j= 1 l =1 k = 1 (1)
time period.

197
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

m Li Li
t
Xijlk rweij demtjk lbrjk ; k , j, t t
Yijlk 1; i, j , k, t
i =1 l=1 (8) l=1 (18)
Any purchasing firm has a limited space for holding any raw ma- Finally, constraint (19) requires Yijlk t
, H tjk and Uijkp
t
be binary, and
terial. Therefore, constraint (9) satisfies that the area occupied by each t
Xijlk t
, Qijkp , ILP tjk and ILP tjk be nonnegative variables.
raw material supplied by all suppliers to each purchasing firm at each t
Yijlk , H tjk , Uijkp
t t
{0, 1}, Xijlk t
, Qijkp , ILP tjk , ILP tjk 0; i, j , l, k, p , t
time period should be less than or equal to the respective total available
space. (19)
m P For the above multi-objective programming model, is called a
ILP tjk 1 + t
Qijkp demtjk ILP tjk sphj H tjk sts tjk ; k , j, t Pareto optimal solution if it is impossible to improve all objective
i=1 p=1 functions simultaneously by any other solution. Among all Pareto op-
(9) timal solutions, the one selected by a decision maker is called the
preferable solution (Branke, Deb, Miettinen, & Slowinski, 2008).
Constraint (10) implies that the shortage amount of each raw ma-
terial should be less than or equal to a predetermined quantity at each
4. Solution methodology
period.
m P In this section, a solution procedure for the SSSOA model developed
ILP tjk + demtjk t
Qijkp ILP tjk 1 1 H tjk mshtjk ; k , j, t in the preceding section is presented. In the first step of the proposed
i=1 p=1
solution methodology, the multiple objective functions are transformed
(10) into one single objective function using the -constraint method. The
The amount of raw material j provided by supplier i in all price obtained single objective programming model is then solved using the
levels for all purchasing firms should be less than or equal to the ca- Benders decomposition algorithm, which itself is accelerated by several
pacity of that supplier for raw material j, which is guaranteed by con- computational enhancements, to attain a set of Pareto optimal solu-
straint (11). tions. Having Pareto optimal solutions derived, the preferable solution
is selected using DEA super efficiency of purchasing firms in the final
Li K
t step of the proposed solution procedure based on physical inputs (raw
Xijlk capijt ; i, j , t
l=1 k=1 (11) materials) and outputs (end products). In other words, among all the
Pareto optimal solutions, the preferable solution is the one maximizing
In order to help a purchasing firm manage its supply network more DEA super efficiency of purchasing firms.
appropriately and control the associated risks, constraint (12) satisfies
that, for each raw material, a particular number of suppliers is chosen
4.1. The -constraint method
for each purchasing firm at each time period.
m Li
t
In the developed SSSOA multi-objective programming model, the
Yijlk nsutjk ; j, k, t goal is to provide a solution compromising among economic, environ-
i =1 l=1 (12)
mental and social responsibility aspects of sustainability. Among the
Constraint (13) indicates, at each time period, all purchased raw available methods to deal with a multi-objective programming pro-
materials from suppliers should be transported to purchasing firms, i.e. blem, the -constraint method is selected in the present paper due to the
no purchased raw materials are left in suppliers. simplicity and applicability of its implementation. It has been also ap-
Li P plied to various multi-objective problems in the literature (e.g.
t
Xijlk = t
Qijkp ; i , j, k, t Fahimnia, Jabbarzadeh, Ghavamifar, & Bell, 2017; Mohammed &
l=1 p=1 (13) Wang, 2017; Paydar, Babaveisi, & Safaei, 2017; Ustun & Anagun,
2015).
Constraint (14) requires that the amount of each raw material that is
The proposed SSSOA model, in which total cost and CO2 emission
transported by a transportation mode from a supplier to a purchasing
are minimized and social responsibility is maximized, can be sum-
firm should be less than or equal to the respective transportation ca-
marized as follows:
pacity at each time period.
t t t Min {TC , TE }
Qijkp Uijkp tcaijkp ; i, j , k, p , t (14)
Max SR
Constraint (15) satisfies that if a raw material cannot be transported s. t . x S (20)
by a transportation mode, then no supplier should use that transpor- where x is the vector of decision variables, TC, TE and SR are total cost,
tation mode to send that type of raw material to any purchasing firm. CO2 emission and social responsibility objective functions, respectively,
T m K and S denotes the space of feasible solutions. In the -constraint
t
Uijkp zjp TmK ; j, p method, the original multi-objective problem is converted into a single
(15)
t=1 i =1 k=1
objective programming model by retaining one of the objective func-
Constraints (16) and (17) imply that if = 1, then should be t
Yijlk t
Xijlk tions as the primary objective function in the model. Accordingly, the
within the respective purchase volume of price level l at each time remaining objective functions are expressed as the properly defined
period. Conversely, if Yijlk
t
= 0 , then they guarantee that no raw material constraints with enforcing upper and/or lower bounds. Therefore, if
j at price level l is purchased from supplier i for purchasing firm k at total cost objective function (1) is chosen as the primary objective
time period t, i.e. Xijlk
t
= 0 . In addition, constraint (18) guarantees that function, the following single objective programming model is attained
at most one discount interval can be selected for each raw material where total CO2 emission and social responsibility objective functions
provided by each supplier for a purchasing firm in a particular period of are treated as model constraints:
time.
Min TC
t t t t t
Yijlk ijl Xijlk Yijlk ij, l + 1 ; i , j, l = 1, …, Li 1, k, t (16) s. t . TE TEmin + v TE
SR SRmin + v SR
t t t t
YijLik ijLi XijL ik
YijLik
M; i , j, k, t (17) x S (21)

198
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

where v = 0, 1, …, , TE = and SR =
TEmax
. In
TEmin SRmax SRmin 4.2. Benders decomposition algorithm
order to derive the maximum and minimum values of total CO2 emis-
sion and social responsibility objective functions, the following steps The single objective programming problem (22)–(25) is a MILP
are taken: model, which typically needs to be solved in practice with numerous
suppliers, raw materials and transportation modes in order to achieve
1. Obtain the global optimum solution of each objective function over one Pareto optimal solution associated with a particular value of v. The
S. Then, let = {XTC , XTE , XSR } , where XTC , XTE and XSR are the large-scale property of the developed model in real world yields sub-
obtained global optimum solutions corresponding to TC, TE and SR stantially computational difficulty that cannot be alleviated by utilizing
objective functions. commercial optimization software such as CPLEX, Xpress and LINGO.
2. Find the values of the objective functions TE and SR at each point of Such a difficulty is intensified once one notes that the aforementioned
. model should be solved + 1 times to obtain a set of + 1 Pareto
3. Derive the minimum value of TE and SR objective functions as optimal solutions. Hence, in order to deal with such a computational
TE min = min {TE (x ), x } and SRmin = min {SR (x ), x } , and complexity, a Benders decomposition algorithm (BDA), which is com-
their maximum value as TE max = max {TE (x ), x } and putationally accelerated, is developed in this section. The first en-
SRmax = max {SR (x ), x }. hancement is to add a set of valid inequalities to the master problem in
order to reduce the number of feasibility cuts. The second acceleration
However, to guarantee an optimum solution of (21) is a Pareto is to exclude a larger space of the master problem by generating the
optimal solution of the original multi-objective problem (20), the enhanced Pareto-optimal cuts. To concurrently improve both lower and
modification proposed by Mavrotas (2009) is applied in this paper, upper bounds during the execution of Benders decomposition, the local
where the constraints associated with the added objective functions are branching search is then specialized to the derived single objective
transformed to equality by explicitly incorporating the appropriate MILP model as the third enhancement. Finally, knapsack cuts are
slack or surplus variables (S1 and S2 ) and then penalizing the current generated for the master problem.
new variables at the single objective function. In this case, model (21) BDA initially proposed by Benders (1962) is currently known as an
becomes: efficient algorithm to solve large-scale mixed integer programming
problems. In this algorithm, the original complicated MILP problem is
T m v Li K
t
not directly solved and is instead decomposed into a pure integer pro-
Min TC = Xijlk pprijlt
gramming problem called the master problem and a linear program-
t=1 i=1 j=1 l=1 k=1
T P m n K ming problem called the subproblem. The optimum solutions of the
+ t
Qijkp t
vscijkp+ latter problems are used in an iterative fashion towards acquiring the
t=1 p =1 i=1 j =1 k=1 optimum solution of the original MILP problem. BDA has significant
T P m v Li K advantages compared to other solution methods, e.g. metaheuristic
t
fscijkp t
Xijlk t
/tcaijkp + algorithms, such as: (1) it relies on strong algebra concepts, (2) the
t=1 p=1 i =1 j=1 l=1 k=1 convergence of this algorithm and achievement of optimum solution is
T K v T K v analytically proven, (3) the decision maker can adjust the optimality
ILP tjk hoc tjk + ILP tjk shc tjk+ gap precisely when it is needed, and (4) other efficient solution
t=1 k=1 j=1 t =1 k=1 j=1 methods can be employed while solving the decomposed problems, i.e.
T m n Li K master problem and subproblem (Pishvaee, Razmi, & Torabi, 2014).
t t
Yijlk conijk + S1 + S2 BDA has been successfully exploited in various contexts such as energy
t=1 i=1 j =1 l =1 k=1 (22) management, supply chain management, etc. (Rahmaniani, Crainic,
Gendreau, & Rei, 2017).
s.t. (4)–(18)
To develop BDA for the proposed single objective model (22)–(25),
T P m n Li K dual subproblem (DSP) and master problem (MP) should be firstly
t
fseijkp t
Xijlk t
/ tcaijkp + formulated. Fixing the binary variables to particular given values, i.e.
t
t=1 p =1 i =1 j=1 l=1 k=1
t
Yijlk t
= y¯ijlk , H tjk = h¯jk and Uijkp
t t
= ūijkp , Benders primal subproblem (PSP)
T P m v K T m v Li K is formulated as follows:
t t t
Xijkp vseijkp + Xijlk preij +
t=1 p =1 i =1 j=1 k=1 t=1 i=1 j=1 l=1 k=1 Min TC (26)
T K v
ILP tjk hoetjk + S1 = TE min + v s. t . (4)–(8), (11), (13), (23), (24)
TE
t=1 k =1 j=1
m P
(23) ILP tjk 1 + t
Qijkp demtjk ILP tjk sphj
t
h¯jk sts tjk ; k, j, t
i=1 p=1
T m v Li K
t (27)
1 vark 2 losk Xijlk / pcakt S2 = SRmin + v SR
t=1 i =1 j=1 l=1 k=1 m P
t
(24) ILP tjk + demtjk t
Qijkp ILP tjk 1 1 h¯ jk mshtjk ; k, j, t
i=1 p=1
t
Yijlk , H tjk , Uijkp
t t
{0, 1}, Xijlk t
, Qijkp , ILP tjk , ILP tjk , S1, S2 (28)
0; i , j, l, k, p, t (25) t
Qijkp t
u¯ ijkp t
tcaijkp ; i, j , k , p , t (29)
where is an adequately small value (usually between 10 and 10 3 6) t
Xijlk t
y¯ijlk t
; i, j , l = 1, …, Li 1, k, t
ij, l + 1 (30)
that does not affect the objective function. For a particular value of v, a
Pareto optimal solution for the original multi-objective problem (20) is t
Xijlk t
y¯ijlk t
; i, j , l = 1, …, Li 1, k, t
attained by solving the current single objective programming model.
ijl (31)
Hence, a set of + 1 Pareto optimal solutions is acquired that defines a t t
XijL y¯ijL M ; i, j , k, t (32)
discretized Pareto front. ik ik

199
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

t
XijL ik
t
y¯ijL k i
t
ijLi ; i, j , k , t (33) iteration is attained as follows:
T m v Li K
t t
Xijlk , Qijkp , ILP tjk , ILP tjk , S1, S2 0; i, j , l, k, p , t (34) Min ZMP = + t
Yijlk t
conijk
t=1 i=1 j=1 l=1 k=1 (46)
Let 1 … 5 , 6 , 7 , 8 , 9 , 10 … 16 be vectors of dual variables associated
with constraints (4)–(8), (11), (13), (23), (24), (28)–(33), respectively. s. t . (12), (15), (18)
Then, the DSP that provides an upper bound for the single objective
problem (22)–(25) can be derived as follows:
T n K T n K T n K
demtjk ¯1, t
jk + demtjk ¯2,
demtjk apr tjk ¯3,
jk
t t
jk +
T n K T n K t =1 j= 1 k =1 t= 1 j =1 k= 1 t = 1 j =1 k= 1
Max Z DSP = demtjk 1, t
jk + demtjk 2, t
jk
T n K T n K T n m
6, t
4, t
t =1 j =1 k= 1 t =1 j =1 k= 1 demtjk apdtjk ¯jk + demtjk lbr tjk ¯5,
jk
t
capijt ¯ij +
T n K T n K t =1 j= 1 k =1 t=1 j =1 k= 1 t =1 j= 1 i =1
demtjk apr tjk 3,
jk +
t
demtjk apdtjk 4, t
jk + (TEmin + v TE )
¯8 + (SRmin + v
SR )
¯9
t =1 j= 1 k =1 t= 1 j = 1 k =1 T n K
T n K T n m
H tjk sts tjk + demtjk sphtj ¯10,
jk +
t
demtjk lbr tjk 5,
jk
t
capijt ij6, t+ t =1 j= 1 k =1
t =1 j= 1 k =1 t =1 j =1 i=1
T n K T m n K P
(TEmin + v TE ) 8 + (SRmin + v SR ) 9
demtjk 1 H tjk mshtjk ¯11,
jk
t t ¯ 12, t
U tjk tcaijkp ijkp
T n K t =1 j= 1 k =1 t= 1 i = 1 j= 1 k = 1 p = 1
t
h¯jk ststjk + demtjk sphtj 10, t
jk + T m n Li 1 K T m n Li 1 K
t =1 j= 1 k =1 t
Yijlk t ¯ 13, t t t ¯ 14, t
ij, l + 1 ijlk + Yijlk ijl ijlk
T n K t=1 i=1 j =1 l=1 k =1 t = 1 i = 1 j = 1 l= 1 k = 1
t 11, t
demtjk 1 h¯jk mshtjk jk T m n K T m n K
t =1 j= 1 k =1 t
YijL 15, t
M ¯ijk + t
YijL t ¯ 16, t ; ¯1, …, ¯16 p
ik i k ijLi ijk
T m n K P T m n Li 1 K t=1 i=1 j =1 k= 1 t=1 i= 1 j =1 k= 1
12, t t 13, t
u¯ tjk tcaijkp
t y¯ijlk t
ij, l + 1 ijlk +
(47)
ijkp
t =1 i=1 j =1 k= 1 p= 1 t = 1 i = 1 j = 1 l= 1 k = 1
T m n Li 1 K T m n K
t t 14, t t 15, t T n K T n K T n K
y¯ijlk y¯ijL M ijk + 1, t 2, t 3, t
ijl ijlk ik 0 demtjk jk + demtjk jk demtjk apr tjk jk +
t =1 i=1 j =1 l =1 k =1 t=1 i= 1 j =1 k= 1
T m n K t = 1 j =1 k= 1 t =1 j= 1 k =1 t =1 j= 1 k =1
t t 16, t T n K T n K T n m
y¯ijL 4, t 5, t 6, t
i k ijLi ijk demtjk apdtjk jk + demtjk lbr tjk jk capijt ij +
t =1 i=1 j =1 k= 1 (35)
t = 1 j =1 k= 1 t= 1 j = 1 k =1 t = 1 j =1 i= 1

s. t . TEmin + v TE
8 + SRmin + v
SR
9

repijt jk
3, t
+ otdijt jk
4, t
+ rweij 5,
jk
t 6, t
ij + 7, t
ijk + T n K
10, t
H tjk sts tjk + demtjk sphtj jk +
P t = 1 j =1 k= 1
t t
preij + fseijkp /tcaijkp 8 + 1 vark 2 losk /pcakt 9
T n K T m n K P
11, t 12, t
p=1 demtjk 1 H tjk mshtjk jk U tjk tcaijkp
t
ijkp
t = 1 j =1 k= 1 t =1 i= 1 j =1 k= 1 p =1
P
13, t 14, t 15, t 16, t T m n Li 1 K T m n Li 1 K
+ + pprijlt + t
fscijkp t
/ tcaijkp ; i , j, l, k , t t
Yijlk t
13, t
t t
14, t
ijlk ijlk ijk ijk ij, l + 1 ijlk + Yijlk ijl ijlk
p=1 t = 1 i = 1 j = 1 l= 1 k = 1 t = 1 i = 1 j = 1 l= 1 k = 1

(36) T m n K
t
YijL M ijk +
15, t
T m n K
t
YijL t
16, t
;
1
, …,
16
r
ik i k ijLi ijk
t = 1 i= 1 j = 1 k =1 t =1 i= 1 j =1 k= 1
1, t
jk + 1, t + 1
jk + 2, t + 1
jk + hoe tjk 8 sphj 10, t + 1
jk + 11, t + 1
jk
(48)
hoc tjk ; j, k, t < T (37) t
Yijlk , H tjk , Uijkp
t
{0, 1} ; i, j , l, k, p (49)
1, t
jk + hoetjk 8 hoc tjk ; j, k, t = T (38)
(50)
1, t + 1
jk + 2, t
jk + sphj 10, t + 1
jk
11, t + 1
jk shc tjk ; j, k , t < T (39) 1 16
In the MP, constraint (47) is an optimality cut in which ¯jk , …, ¯ijk
are the optimum values of the dual variables in the bounded DSP. In the
2, t
jk shc tjk ; j, k, t = T (40) latter constraint, is a free continuous variable that is used to under-
estimate total cost. In addition, constraint (48) represents a feasibility
1, t 2, t 7, t t 8 10, t 11, t 12, t
+ + vseijkp sphj + 1 16
jk jk ijk jk jk ijkp cut in which jk , …, ijk are the extreme rays of the unbounded DSP.
t
vscijkp ; i, j , k, p , t (41) Algorithm 1. The classical BDA
8 (42)
1: Y, H, U initial feasible solutions, z upper , z lower
9 (43) 2: while (z upper z lower ) > do
3: Solve DSP
1, t 2, t 3, t 4, t 5, t 6, t 10, t 11, t 12, t 13, t 14, t 15, t 16, t 4: if the DSP is unbounded then
jk , jk , jk , jk , jk , ij , jk , jk , ijkp , ijlk , ijlk , ijk , ijk 5: Add the feasibility cut (48) to the MP
0; i , j, l, k, p, t (44) 6: else
7: Add the optimality cut (47) to the MP
7, t 8, 9 8: z upper z DSP
ijk , ; i, j , k , t (45)
9: end if
Based on the theory of duality, if the DSP is unbounded, then PSP is 10: Solve the MP
11: z lower z MP
infeasible. In such a circumstance, r denotes the set of extreme rays of
12: end while
the polyhedron formed by constraints (36)–(45). Otherwise, if the DSP
has a bounded optimum solution, then p represents the set of extreme
points of such a polyhedron. The classical BDA described in Algorithm 1 is known to have quite
Now according to the solution of the DSP, the MP that provides a slow convergence (Geoffrion & Graves, 1974). Therefore, several al-
lower bound for the single objective problem (22)–(25) at each gorithmic enhancements are presented in the following subsections in

200
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

order to accelerate the classical BDA. for all Y, H and U with strict inequality for at least one point. An op-
timality cut that cannot be dominated by any other cut is called a
4.2.1. Valid inequalities Pareto-optimal cut. Adding Pareto-optimal cuts accelerates con-
One of the reasons for slow convergence of the classical BDA is the vergence of the BDA. Let denote the polyhedron defined by the
m Li
weak lower bounds provided by the MP at primary iterations of the t
Yijlk k
nsujk ; j, k , t
algorithm (Saharidis, Boile, & Theofanis, 2011). To avoid this in- i =1 l=1
efficiency, some valid inequalities are derived and added to the MP in T m K
order to restrict the feasible region and produce high quality solutions.
t
Uijkp zjp TmK ; j, p
Therefore, it is expected that the defined valid inequalities reduce the following constraints:
t=1 i=1 k =1
Li
number of feasibility cuts (48) during the execution of solution proce- t
Yijlk 1; i , j, k, t
dure. l=1
t
Definition 1. Consider constraint (12) for the number of suppliers 0 Yijlk 1; i, j , l, k, t ,
selected. Let N = {(i, l) i {1, …, m}, l {1, …, Li }} . A set N N is a 0 H tjk 1 ; j, k, t
cover if nsutjk < N ; j, k, t . and 0 Uijkp t
1; i, j, k, p , t . In addition, let ri ( ) indicate the re-
lative interior of . In this case, a Pareto-optimal cut can be obtained by
Proposition 1. If N N is a cover, the cover inequality
solving the following auxiliary dual problem:
Yt
i, l N ijlk
N 1; j, k, t is valid for constraint (12).
T n K
Proof. The proof is straightforward based on Wolsey (1998); page: 147. Max Z DSP _ AUX = 1, … , 9 0, t t
h¯jk sts jk + demtjk sphtj 10, t
jk +
t= 1 j = 1 k =1
Proposition 2. If N N is a cover, then the valid inequality
T n K T m n K P
0, t 11, t
demtjk 1 h¯jk mshtjk jk u¯ 0, t t 12, t
jk tcaijkp ijkp
t t t t =1 j= 1 k =1 t=1 i=1 j =1 k= 1 p= 1
Yijlk + ijlk Yijlk N 1; j, k , t T m n Li 1 K T m n Li 1 K
i, l N i, l N (51) 0, t t
y¯ijlk 13, t
ij, l + 1 ijlk +
0, t t 14, t
y¯ijlk ijl ijlk
t =1 i=1 j =1 l =1 k =1 t= 1 i =1 j= 1 l =1 k = 1
is non-redundant (facet-defining) and thus as strong as possible, where T m n K T m n K
0, t 15, t 0, t t 16, t
y¯ijL k M ijk + y¯ijL
i i k ijLi ijk
t
ijlk = N nsutjk ; j, k, t & i , l N t =1 i=1 j =1 k= 1 t= 1 i =1 j= 1 k =1

(53)
Proof. Following the procedure to obtain a lifted cover inequality by
Wolsey (1998); page: 149, the proof is straightforward. s. t . (36)–(45)

In addition, constraint (52) implies that the selected suppliers pro- where ȳ 0 , h̄ 0 , ū 0 ri ( ) . The optimality cut generated by the optimum
vide sufficient capacity to all raw materials in purchasing firms. solution of the above auxiliary dual problem is a Pareto-optimal cut in
m Li
ri ( ) . It is worth mentioning that Pareto-optimal cuts, which are non-
demtjk capijt Yijlk
t
; j, k, t dominated, would be generated using ȳ 0 , h̄ 0 , ū 0 ri ( ) because the
i =1 l=1 (52) description of the convex hull formed by the MPs constraints is not
already known and that it is quite difficult to find out ȳ 0 , h̄ 0 and ū 0 in
ri ( ) .
4.2.2. Pareto-optimal cuts
The PSP is typically degenerate because of its customary network 4.2.3. Local branching
structure. Accordingly, the DSP might have multiple optimum solutions As stated before, one of the main reasons for slow convergence of
(Magnanti & Wong, 1981). It implies that different Benders optimality the classical BDA is computational complexity of the MP, which is an
cuts can be generated from a set of alternative optimum solutions. Let integer programming problem with new optimality and feasibility cuts
(.) denote the parts of the objective function in the DSP that are in- appended in each iteration. The newly added cuts make the MP even
dependent of the binary decision variables. In this case, according to more complicated, whereas the DSP is a linear programming model that
the dominating cut definition (Magnanti & Wong, 1981), the optimality can be typically solved immediately. Therefore, it is required to spend a
T
cut generated by the dual solution vector dominates that build by the tremendous amount of time to solve the MP particularly at the middle
T
dual solution vector if and only if and terminating iterations of BDA. To overcome this pitfall, a near
T n K optimum solution of the MP is attained at each iteration of BDA in lieu
of acquiring its optimum solution (Geoffrion & Graves, 1974). To im-
1 9 10, t
, …, H tjk sts tjk + demtjk sphtj jk +
t = 1 j =1 k= 1
plement such an idea, a local branching approach, which was initially
developed by Fischetti and Lodi (2003) to deal with complex integer
T n K T m n K P
11, t 12, t
demtjk 1 H tjk mshtjk jk U tjk tcaijkp
t
ijkp
t= 1 j = 1 k =1 t= 1 i= 1 j = 1 k =1 p=1 programming problems, was proposed by Rei, Cordeau, Gendreau, and
T m n Li 1 K
t t 13, t
T m n Li 1 K
t t 14, t
Soriano (2009) to efficiently solve complex MP in BDA. In local
Yijlk ij, l + 1 ijlk + Yijlk
t= 1 i =1 j= 1 l =1 k = 1 t= 1 i =1 j= 1 l =1 k = 1
ijl ijlk branching methodology, a reasonable neighborhood around the pre-
T m n K T m n K vious solution is specified to confine the region of feasible solutions in
15, t 16, t
the MP. The current restricted MP is then solved to either optimality or
t t t 1 9
YijL M ijk + YijL , …,
ik i k ijLi ijk
t= 1 i =1 j= 1 k =1 t= 1 i= 1 j = 1 k =1
a good feasible solution. Consequently, a significant amount of time is
potentially saved under this methodology because the original com-
T n K T n K
10, t 11, t
H tjk ststjk + demtjk sphtj jk + demtjk 1 H tjk mshtjk jk
t= 1 j = 1 k =1 t= 1 j = 1 k =1 plicated MP is now solved within a restricted region of feasible solu-
T m n K P
12, t
T m n Li 1 K
13, t tions.
U tjk tcaijkp
t t
Yijlk t
ij, l + 1 ijlk +
t= 1 i =1 j= 1 k =1 p = 1
ijkp
t =1 i=1 j =1 l =1 k =1
Let 1 = (y¯ 1 , h¯ 1, u¯ 1) be the MPs optimum solution at the current
T m n Li 1 K T m n K iteration. Then, the region of its feasible solutions can be divided into
t 14, t 15, t
two sub-regions in the next iteration by introducing the following dis-
t
Yijlk t
YijL M ijk +
ijl ijlk ik
t= 1 i =1 j= 1 l =1 k = 1 t=1 i=1 j =1 k= 1
T m n K junction:
t t 16, t
YijL ijk
i k ijLi
t= 1 i =1 j= 1 k =1 (y , h, u, y¯ 1 , h¯ 1, u¯ 1) (y , h, u, y¯ 1 , h¯ 1, u¯ 1) +1 (54)

201
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

where denotes a proper distance function and is a reasonably-va- should be noted that the MP now includes the expressions (46)–(52),
lued positive integer number. In this case, the reduced MP that is re- (56), (12), (15) and (18). Furthermore, c denotes a non-negative
stricted to a sub-region by adding the left branching cut can efficiently parameter defined for a convex combination that updates the core point
be solved by a commercial solver. To implement local branching pro- in generating the Pareto-optimal cuts. The current non-negative para-
cedure, Hamming distance metric is used in the present paper. Let meter is usually set to 0.5, which has led to satisfactory results in most
t
Y = {i , j, l, k, t: Yijlk = 1}, H = {j, k, t: H tjk = 1} and U = {i , cases.
j, k, p , t : Uijkp = 1} . Then, the extended representation of the left
t
Algorithm 2. En-BDA
branching cut is given as follows:
1: Y, H, U initial feasible solutions, z upper , z lower
t t
1 Yijlk + Yijlk + 1 H tjk + t
Hijlk 2: y¯ 0 , h¯ 0, u¯ 0 initial points in ri ( ), c = 0.5
i, j , l , k , t Y i, j , l , k , t Y j, k , t H j, k , t H 3: while (z upper z lower ) > do
4: Solve the auxiliary DSP
+ 1 t
Uijkp + t
Uijkp 5: Add the Pareto-optimal cut (47) to the MP
6: Solve the MP
i, j, k, p, t U i , j, k , p , t U (55)
7: z lower z MP
If = h¯ 2, u¯ 2 ) is assumed to be the restricted MP’s solution on
2 (y¯ 2 , 8: Solve the DSP
9: if the DSP is unbounded then
the left sub-region, then one of the following situations might happen in
10: Add the feasibility cut (48) to the MP
practice (Jeihoonian, Zanjani, & Gendreau, 2016): ¯0 ¯ 0 + , u¯ 0
11: y¯ 0 cy¯ 0 + , h ch cu
¯0 +
12: else
1. if 2 is an optimum solution of the current subproblem obtained 13: Add the optimality cut (47) to the MP
within the predetermined time limit, then the left branching cut is 14: z upper z DSP
substituted with the right branching one, i.e. 15: y¯ 0 + (1 ¯0
c ) y¯ 0 , h
cy¯ 0 ¯ 0 + (1
ch ¯0,
c) h u¯ 0 cu
¯0 + (1 c)u
¯0
(y , h, u, y¯ 1 , h¯ 1, u¯ 1) + 1, and 2 becomes the new reference 16: 1 (y¯ , h¯ , u¯ )
point for local branching procedure. In this case, local branching 17: Implement the local branching procedure
proceeds with solving the new local branching subproblem. 17: end if
18: end while
2. If the current subproblem is infeasible, i.e. no 2 can be acquired,
then the left branching cut is replaced by the right branching one
and the region’s size of feasible solutions in the current subproblem
is increased by /2 , i.e. + /2 . Afterwards, local branching
4.3. DEA super efficiency to choose the preferable solution
continues with solving the new local branching subproblem.
3. If 2 is a feasible solution improving the objective function within
For each particular value of v, one Pareto optimal solution for the
the prespecified time limit, it is regarded as the new reference point
original SSSOA problem is attained by solving the single objective
for local branching procedure. In addition, 2 is eliminated from the
programming model (22)–(25) using En-BDA algorithm. Therefore, a
region of feasible solutions by adding (y , h, u, y¯ 2 , h¯ 2, u¯ 2) and
set of + 1 Pareto optimal solutions is given to the decision maker in
(y , h, u, y¯ 2 , h¯ 2, u¯ 2) 1 to the MP.
order to select the final preferable solution. In this section, a decision
4. if 2 is a feasible solution that does not improve the objective
support tool based on DEA super efficiency is elucidated, which enables
function within the predetermined time limit, then it is removed
to choose the preferable solution with respect to the super efficiency of
from the region of feasible solutions and is augmented by unit. In
all purchasing firms. In other words, the preferable solution is a Pareto
other words, the branching cuts (y , h, u, y¯ 2 , h¯ 2, u¯ 2) + 1 and
optimal solution for which DEA super efficiency of all purchasing firms
(y , h, u, y¯ 2 , h¯ 2, u¯ 2) 1 are added to the MP.
is maximum.
DEA initially proposed by Charnes, Cooper, and Rhodes (1978) is a
4.2.4. Knapsack cut nonparametric multi-factor productivity analysis model, which evalu-
If a reasonably tight upper bound is available from the DSP, then the ates the relative efficiencies of a homogeneous set of decision-making
knapsack cut initially developed by Santoso, Ahmed, Goetschalckx, and units (DMUs). Although DEA is a beneficial tool for measuring the re-
Shapiro (2005) can be derived, which has a substantial impact on ob- lative efficiency, it often recognizes too many DMUs as efficient. This
taining the MPs solution. Since the optimum values of objective func- deficiency known as low discrimination power is due to self-evaluation
tions in the DSP and MP provide the upper and lower bounds respec- nature of DEA method and becomes more intense when the number of
tively at each iteration, it is concluded that z upper ZMP . Therefore, DMUs is relatively small in comparison with the total number of inputs
based on the current inequality, objective function (46) and constraint and outputs. There are numerous approaches in the literature to cope
(47), the following knapsack cut is formulated to be added to the MP: with low discrimination power (Adler, Friedman, & Sinuany-Stern,
T n K
2002). In the present paper, the super efficiency ranking method pro-
z upper 1, …, 9 H tjk sts tjk + demtjk sphtj ¯10,
jk +
t posed by Li, Jahanshahloo, and Khodabakhshi (2007) is utilized, which
t =1 j= 1 k =1 is always a feasible model and conquers some drawbacks of other ap-
T n K T m n K P
proaches, as follows:
demtjk 1 H tjk mshtjk ¯11,
jk
t t ¯ 12, t
U tjk tcaijkp ijkp
t= 1 j = 1 k =1 t= 1 i= 1 j = 1 k =1 p=1 n
1 sj+2
T m n Li 1 K
t t ¯ 13, t
T m n Li 1 K
t t ¯ 14, t
Min h 0s = 1 +
Yijlk ij, l + 1 ijlk + Yijlk ijl ijlk n j =1
Rj (57)
t= 1 i =1 j= 1 l =1 k = 1 t =1 i= 1 j =1 l = 1 k=1
T m n K T m n K
t
YijL 15, t
M ¯ijk + t
YijL t ¯ 16, t s.t.
ik i k ijLi ijk
t= 1 i =1 j= 1 k =1 t=1 i=1 j =1 k= 1
+1
T m v Li K
t con t
Yijlk v x¯jv + s j1 s+j2 = x¯j0 ; j
t=1 i= 1 j =1 l= 1 k =1
ijk
v=1 (58)
(56) +1

With respect to all the presented enhancements, Algorithm 2 out- v o¯rv sr+ = o¯r 0 ; r
lines the enhanced BDA, i.e. En-BDA. In the current accelerated BDA, it v=1 (59)

202
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

v, s j1, s+j2, sr+ 0; v , j, r (60) Remark 1. That the inputs and outputs are aggregated using Eqs. (61)
and (63) in this study is inspired from the aggregate model proposed by
where ōrv is the amount of output r from DMU v, while x̄ jv denotes the Kao and Liu (2014) and Park and Park (2009).
amount of input j to DMU v. In this model, h0s is DEA super efficiency
score of DMU0 and Rj is the maximum amount of all inputs j, i.e. Remark 2. n inputs and s outputs for the individual DMU are attained
Rj = max v (x¯jv ) . Constraint (58) allows input j of DMU0 to increase by using Eqs. (61) and (63), respectively, in + 1 different conditions. To
s+j2 or decrease by s j1. Constraint (59) implies that output r of DMU0 is implement a DEA model, it is recommended that the number of DMUs
permitted to increase by at most sr+. DMU0 is super efficient if optimum be at least 3 times the total number of inputs and outputs (Charnes
value of the objective function in this model is greater than one. On the et al., 1978). Hence, the magnitude of in problem (21) is determined
other hand, a super efficiency score that is smaller than one denotes the so as to have 3(n + s ) 1 satisfied.
respective DMU is not super efficient. Consequently, the above DEA
super efficiency model can be solely solved to rank all DMUs (Li et al.,
5. Empirical study
2007). If there are + 1 DMUs to be evaluated, this model needs to be
solved + 1 times in order to determine the super efficiency score of all
A real case study is presented in this section to elucidate how the
DMUs under study.
SSSOA model and its solution procedure are implemented in a practical
In order to implement DEA super efficiency model (57)–(60) to
context. In this regard, the proposed approach is employed for a man-
choose the preferable solution among all previously attained Pareto
ufacturing and engineering company in automotive industry, which is
optimal solutions, it is essential to define proper DMUs and their inputs
the sole manufacturer of three types of automobile transmission sys-
and outputs. In the following, it is explained how to specify an in-
tems in Iran, i.e. K = 1 and s = 3. The purpose of this empirical study is
dividual DMU and derive its inputs and outputs in a unique way.
to evaluate four potential suppliers for two required raw materials in a
DMU determination – In this study, the set of all purchasing firms is
time horizon of two years, i.e. m = 4, n = 2 and T = 2 , in such a way
regarded as one individual DMU that is evaluated in different condi-
that not only the company’s cost, CO2 emission and social responsibility
tions. A condition is indeed one Pareto optimal solution derived. In
are optimized, but also it operates in its most efficient condition.
other words, instead of evaluating different DMUs, one individual DMU
Because of the economic troubles encountered recently and lack of the
that is the set of all combined purchasing firms is evaluated in different
required manufacturing resources, the issue of efficiency is of great
+ 1 conditions. Contrary to previous studies in the literature, this
importance for this company (Moheb-Alizadeh & Faez, 2009). The lo-
helps derive DEA super efficiency of aggregated purchasing firms.
cations of this manufacturing company and its suppliers are illustrated
Inputs characterization – The inputs are defined as physical raw
in Fig. 3.
materials transported from all suppliers to the individual DMU. Let
The required raw materials are in fact the forged parts of two shafts,
v ; v = 0, …, denote the v-th Pareto optimal solution. Obviously, each i.e. primary shaft (PS) and secondary shaft (SS). Due to the complex
Pareto optimal solution v contains an amount of raw material j in price
chemical combinations and high-tech manufacturing processes required
level l provided by supplier i to purchasing firm k at time period t, i.e.
for these materials, the company succeeded in identifying only four
t
v = {(Xijlk ) v ; i , j, l, k, t }; v = 0, …, . In this case, the following ex-
potential domestic suppliers across the country, each of which has two
pression
price levels for each shaft, Li = 2 ; i . Furthermore, three types of
T m Li K transportation modes including flatbed truck, box truck and tractor unit
t
x¯jv = (Xijlk )v ; j, v are utilized to ship forged parts from suppliers to the manufacturer, i.e.
t =1 i=1 l=1 k=1 (61)
P = 3. This manufacturing company produces three types of transmis-
aggregates the order quantities of each raw material j transported from sion systems; automatic transmission (AT), manual transmission (MT)
all suppliers in all price levels during all time periods to the individual and continuous variable transmission (CVT). In the manufacturing
DMU in each Pareto optimal solution. In other words, x̄ jv characterizes process, one PS and one SS are assembled together in each type of
input j (physical raw material j) of the individual DMU with respect to transmission systems. All parameter values are defined as presented in
the Pareto optimal solution v in DEA super efficiency model (57)–(60). supplementary material. These data are derived from historical records,
Output characterization – The outputs are defined as the end products which are mainly available in production planning and control de-
manufactured by purchasing firms. Based on the available historical partment, purchasing department and quality control department of
data, a properly defined production function is utilized to estimate the this manufacturing company.
production level of end products for each purchasing firm. For each Moreover, it suffices to have 14 in the present case study to
Pareto optimal solution v , the Cobb-Douglas production function gives satisfy 3(n + s ) 1. Therefore, is arbitrarily set to 14 and sub-
the production level of end product r in purchasing firm k, i.e. orkv , as sequently the individual DMU, which is indeed the single manufacturer,
follows: is evaluated in 15 different conditions.
jk
5.1. Results
n T m Li
t
orkv = rk (Xijlk )v ; r, k, v
j =1 t =1 i=1 l=1 (62) Using the procedure described in Section 4.1, the minimum and
where rk and jk are estimated based on available historical data. In maximum values of TE and SR objective functions are calculated as
this case, the amount of end product r of the individual DMU with re- presented in Table 3.
spect to the Pareto optimal solution v in DEA super efficiency model A strong valid inequality for constraint (12) was derived as in-
(57)–(60) is derived as: equality (51). In this regard, N is defined in this paper as a subset of N
with N = 6 . Although 28 of such subsets, i.e. C68 , can be characterized,
K
only 4 arbitrary cover inequalities associated with constraint (12) are
o¯rv = orkv ; r, v
k=1 (63) defined as follows:

Therefore, + 1 DMUs (one individual DMU in + 1 different Y1tj1k + Y1tj2k + Y2tj1k + Y2tj2k + Y3tj1k + Y3tj2k 5; j, k , t
conditions) are characterized, each of which has n inputs and s outputs.
Fig. 2 illustrates the defined single DMU and its inputs and outputs for Y1tj1k + Y1tj2k + Y2tj1k + Y2tj2k + Y3tj1k + Y4tj1k 5; j, k, t
some Pareto optimal solution v . It also depicts inputs and outputs of
Y1tj1k + Y1tj2k + Y2tj1k + Y2tj2k + Y3tj1k + Y4tj2k 5; j, k, t
each purchasing firm for the same Pareto optimal solution.

203
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

purchasing firms

Inputs DMU Outputs

Fig. 2. The single DMU of all purchasing firms with its inputs and otputs.

Y1tj1k + Y1tj2k + Y2tj1k + Y2tj2k + Y3tj2k + Y4t j1k 5; j, k, t Table 3

Minimum and maximum values of total emission (TE) and social responsibility
Therefore, strong valid inequalities for constraint (12) are derived (SR) objective functions.
as follows: Objective function Minimum Maximum

Y1tj1k + Y1tj2k + Y2tj1k + Y2tj2k + Y3tj1k + Y3tj2k + 3(Y4t j1k + Y4t j2k ) TE 110,155.1 218,892.2
SR 9.15 10.27
5; j, k, t

Fig. 3. The structure of the supply network in the case study.

204
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

Table 4
Optimal order quantities and the associated values of objective functions for each Pareto solution.

Y1tj1k + Y1tj2k + Y2tj1k + Y2tj2k + Y3tj1k + Y4t j1k + 3(Y3tj2k + Y4t j2k )
Table 5
5; j, k, t The physical inputs and outputs of the manufacturing company and its super
efficiency score in each Pareto optimal solution.
Y1tj1k + Y1tj2k + Y2tj1k + Y2tj2k + Y3tj1k + Y4t j2k + 3(Y3tj2k + Y4tj1k ) v Available forged shafts Transmission systems (Outputs) Super
(Inputs) efficiency
5; j, k, t
PS SS AT MT CVT
Y1tj1k + Y1tj2k + Y2tj1k + Y2tj2k + Y3tj2k + Y4t j1k + 3(Y3tj1k + Y4t j2k )
0 173,400.0 198,542.1 245,247.9 238,488.0 219,681.3 1.2284
5; j, k, t 1 175,793.1 198,542.1 246,312.4 239,451.0 220,435.4 1.1234
2 179,023.1 198,542.1 247,733.7 240,736.2 221,441.1 1.1167
The En-BDA is implemented in Python 3.6 with GUROBI 8.0 opti-
3 182,252.9 198,542.1 249,137.4 242,005.0 222,433.1 1.1100
mizer on a computer with CPU of 2.5 GHz and 6.0 GB RAM, where the 4 186,060.1 197,964.8 250,468.2 243,169.4 223,267.0 1.0033
relative optimality gap is set to 0.01, i.e. = 0.01. For v = 0, …, 14 in the 5 189,272.6 197,982.2 251,836.0 244,405.8 224,234.2 1.0144
single objective model (22)–(25), Table 4 presents the optimum order 6 192,582.5 197,902.2 253,177.0 245,611.1 225,163.6 1.1720
7 195,826.6 197,888.0 254,509.4 246,812.5 226,097.9 1.2299
quantities of 15 different Pareto optimal solutions. As stated before, the
8∗ 199,056.6 197,887.9 255,828.5 248,002.4 227,024.5 1.3124
individual manufacturer is evaluated in these 15 Pareto optimal solu- 9 199,400.0 200,774.4 257,507.1 249,710.6 228,739.7 1.1285
tions as though 15 different DMUs with distinct inputs and outputs are 10 199,400.0 204,004.3 259,214.1 251,466.2 230,535.4 1.0339
considered. For each Pareto optimal solution derived, the value of the 11 199,400.0 207,234.1 260,905.3 253,206.3 232,316.7 1.0000
corresponding objective function is also given in this table, where the 12 199,400.0 210,464.0 262,581.2 254,931.2 234,083.9 1.0000
13 199,400.0 213,693.9 264,242.0 256,641.4 235,837.3 1.0000
trade-offs among objective functions are observed in the sense that
14 199,400.0 216,923.6 265,888.1 258,337.0 237,577.2 1.0000
improving TC and SR objective functions result in worsening TE ob-
jective function.
Table 5 shows the available amounts of forged PS and SS (physical Table 6
inputs) for this manufacturing company calculated using Eq. (61). As Dimensions of the generated test problems.
stated before, the Cobb-Douglass production function helps derive the
Instance m n s K Li T P # of # of total # of
production level of end transmission systems (physical outputs). Ap- integer variables constraints
propriate production functions are characterized based on a set of variables
historical records for forged PS and SS, and the respective production
1 15 8 6 4 2 6 5 20,352 40,908 33,981
levels of manufactured transmission systems in this company as follows:
2 25 12 6 4 2 12 5 101,376 203,344 166,673
AT = 34.734PS 0.316SS 0.414 (64) 3 35 15 8 4 2 12 8 252,720 506,179 364,985
4 45 22 12 6 2 18 8 1,071,576 2,145,554 1,533,889
MT = 32.466PS 0.294SS 0.439 (65) 5 50 30 18 8 2 24 10 3,461,760 6,929,314 4,690,385

205
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

Table 7 4500
Classical BDA
Characterization of various parameters in test problems. 4000

Average number of iterations

En-BDA
Parameter Value Parameter Value 3500
3000
demtjk (ton) U (145, 200) pcakt (ton) U (380, 500)
pprijlt ("$") U (210, 270) t
tcaijkp (ton) U (38, 50) 2500
U (1.8, 2.2) mshtjk (kg) U (55, 85) 2000
t
fscijk ×102"$"
1500
t ("$")
vscijk U (5, 10) lbrjk U (5, 8) 1000
hoc tjk ("$") U (210, 230) rweij U (0.05, 0.50)
500
shc tjk ("$") U (660, 750) t
fseijk (kg) U (160, 430)
0
ststjk (m3 ) U (210, 320) t (kg)
vseijk U (0.07, 0.4) 1 2 3 4 5
sphj (m3 ) U (0.001, 0.002) U (31, 41) Instance
hoetjk × 10 5 kg

Fig. 4. Average number of iterations for each instance using the classical BDA
repijt (%) U (4, 9) preij (kg) U (0.16, 0.2) and En-BDA.
otdijt (%) U (92, 100) conijk ( ×103"$") U (10, 18)
apr tjk (%) U (8, 12.5) losk U (0.4, 1)

apdtjk (%) U (58, 60) var k U (0.5, 1.1)

capijt (ton) U (420, 820)

efficient in state 8. Referring to the column 8 of Table 4, the set of
suppliers selected at each time period to provide each forged shaft can
be determined. In addition, the amounts of forged PS and SS assigned to
each selected supplier can be also determined in this efficient state.
CVT = 27.295PS 0.250SS 0.490 (66)
5.2. Evaluating BDA performance
If the current production functions are applied to the amounts of
forged PS and SS presented in Table 5, then the corresponding pro- As explicated before, the En-BDA developed in this paper is used to
duction levels for three transmission systems in each Pareto optimal cope with the large-scale property of the proposed single objective
solution are computed as also given in Table 5. The current physical programming model (22)–(25). However, the impact of En-BDA on
inputs and outputs in Table 5 are applied to evaluate DEA super effi- solution time of the empirical study in this paper might not be com-
ciency of this manufacturing company based on model (57)–(60). DEA pletely observed because the problem explored does not have sig-
super efficiency scores of this company in 15 different conditions are nificantly high dimension. Hence, in this section, five substantially
also presented in Table 5. larger instances of the single objective programming model (22)–(25)
It is observed in Table 5 that DEA super efficiency score of this are generated, where the precise quantity of all the parameters is
manufacturing company fluctuates significantly, while all 15 different characterized by Monte Carlo sampling method implemented on par-
conditions presented are Pareto optimal solutions of SSSOA model for ticular uniform distribution functions with predetermined intervals.
which the company is absolutely indifferent. Contrary to all previous Each instance is solved three times in order to examine the performance
studies, it signifies that the manufacturing company should not be of En-BDA more accurately. The performance of En-BDA is compared
neutral against different Pareto optimal solutions of the SSSOA model with that of the classical BDA and GUROBI 8.0 within a time limit of
because it has different performance in each Pareto optimal solution. 6 h. The dimensions of generated instances with respect to the number
Moreover, it turns out that the super efficiency model (57)–(60) has of integer variables, the number of total variables and the number of
aptly discriminated different states of the company. Therefore, it can be constraints are given in Table 6, which present how large the developed
declared by considering Table 5 that the manufacturing company is single objective programming model can be in practice for

Table 8
Comparison results of the classical BDA, En-BDA and Gurobi (optimum values in bold).
Inst. Run Classical BDA En-BDA Gurobi

Time (h) # of iterations Cost GAP (%) Time (h) # of iterations Cost Time (h) Cost

1 1 2.198 1658 95,368,889 1 2.045 1369 95,368,889 2.267 95,368,889

2 2.307 1246 102,201,008 1 2.419 913 102,201,008 2.718 102,201,008
3 2.493 1599 88,911,297 1 2.127 1052 88,911,297 3.083 88,911,297
2 1 2.811 1983 111,584,669 1 3.011 1283 111,584,669 3.632 111,584,669
2 2.749 2028 114,498,706 1 2.651 1390 114,498,706 3.719 114,498,706
3 2.800 2095 99,100,611 1 2.612 1326 99,100,611 3.512 99,100,611
3 1 4.567 2796 115,218,044 1 3.159 1417 115,218,044 6 116,823,719
2 4.218 3014 116,074,183 1 2.917 1864 116,074,183 6 117,041,621
3 4.739 3154 116,523,330 1 2.648 1809 116,523,330 6 117,910,465
4 1 6 4019 124,321,083 4.95 3.364 2127 123,085,284 6 124,581,340
2 6 3729 125,473,180 11.63 3.415 1953 124,381,928 6 125,908,443
3 6 3843 125,049,775 5.11 3.291 2026 123,592,068 6 124,893,071
5 1 6 3559 130,811,079 8.17 3.718 1973 128,419,734 6 131,098,157
2 6 4035 130,227,164 4.26 3.822 2119 127,645,129 6 130,853,294
3 6 3384 130,649,382 9.54 3.904 2214 127,273,844 6 130,419,063

206
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

commonplace supply networks. which consisted of -constraint method, enhanced BDA and DEA, was
The uniform distribution functions are defined in such a way that devised. In particular, the original multi-objective programming model
the proposed model retains its feasibility and the obtained results sound was converted to a single objective programming model using the
reasonable. In this regard, if U (a, b) denotes a uniform random -constraint method in the first step of the solution procedure.
variable in the interval (a, b) , then Table 7 presents the uniform dis- Afterwards, the current single objective model was solved using BDA
tribution functions that are used to generate the quantity of the cor- accelerated by a set of computational enhancements in order to attain
responding parameters. Furthermore, nsutjk = m/2 + one Pareto optimal solution. Since a Pareto optimal solution char-
1; j, k, t , = (m + K + T )/2 + 1 for local branching, ij1 = 80 tons; acterizes the physical raw materials transported from suppliers to
i, j, t and ij2 = 190 tons; i, j, t , in all instances. In each instance, purchasing firms, the production level of physical end products was
five valid inequalities are also derived with N = nsutjk + 2. derived by Cobb-Douglas production function in the third step of the
Table 8 summarizes the computational results from solving each proposed solution methodology. In the latter step, using the obtained
instance by En-BDA, classical BDA and GUROBI, where the elapsed time physical inputs and outputs, the DEA super efficiency of purchasing
in hours, the number of iterations and the value of total cost objective firms was then evaluated for all Pareto optimal solutions generated. In
function are also reported. The column entitled GAP in classical BDA this case, the final preferable solution, which indicated the set of se-
indicates the relative difference between lower and upper bounds that lected suppliers and allocated orders, was a Pareto optimal solution
are derived within the predetermined time limit. associated with the greatest DEA super efficiency score.
In addition, the bold numbers represent the optimum value of the The developed BDA that was enhanced in this paper by various
total cost objective function obtained within the time limit. According computational enhancements had its own merits to emphasize when
to Table 8, the following insightful findings can be drawn: applying to large-scale instances. In particular, valid inequalities,
Pareto-optimal cuts, local branching procedure and knapsack cut were
• During the time limit of 6 h, GUROBI is only capable of solving in- incorporated towards devising En-BDA. The evaluation of its perfor-
stances 1 and 2 to optimality. However, it fails to find the optimum mance in comparison with the classical BDA and GUROBI revealed that
solution of the remaining instances within the dedicated time limit. En-BDA is significantly capable of solving large-scale instances of the
Therefore, for such instances, the best value of total cost objective developed programming model in this paper within the predetermined
function acquired by GUROBI after 6 h is presented. time limit. It was observed that En-BDA could improve the average total
• For instances 1, 2 and 3, the classical BDA outperforms GUROBI cost by $2,022,279 and $1,923,959 in instances that the classical BDA
with respect to the solution time. In particular, for instances 1 and 2 and GUROBI were not able to solve, respectively. On the other hand,
for which both the classical BDA and GUROBI are able to find the with respect to the number of iterations, En-BDA could find the optimal
optimum solutions, the solution time of GUROBI is greater than that solution in less number of iterations on average in comparison with the
of the classical BDA by 23.3% on average. In addition, the classical classical BDA.
BDA finds the optimum solution of instance 3 in 4.508 h on average, DEA model is a frontier-based methodology and consequently is
whereas GUROBI fails to do so. However, it is observed that the sensitive to the quality of input and output data. Therefore, any noise
classical BDA is unable to solve instances 4 and 5 within the time existing in measured data can lead to inauthentic results when evalu-
limit of 6 h. That is why the gap between the lower and upper ating DMUs efficiency. In other words, precise data of inputs and out-
bounds obtained in the latter instances is greater than the optimality puts should be gathered to apply any DEA model successfully. In the
gap of 1%. present paper, the production level of end products (physical outputs)
• All instances can be successfully solved by En-BDA in the time limit was derived by Cobb-Douglas production function based on formerly
of 6 h. Particularly, the solution time of En-BDA for instances 1 and recorded data, which may not be as precise as desired. In such a cir-
2 is on average 3.3% (27.4%) less than that of the classical BDA cumstance, the inherent uncertainty is needed to be somehow taken
(GUROBI). In addition, the solution time of En-BDA for instance 3 is into account. One way to present the uncertainty in data is to apply the
on average 55.0% less than that of classical BDA. While neither the membership function in fuzzy set theory proposed by Zadeh (1965). In
classical BDA nor GUROBI are able to solve instances 4 and 5 to this case, a fuzzy production function would be firstly built using re-
optimality in the specified time limit, En-BDA solves them in 3.357 h corded crisp inputs and fuzzy outputs, and then a fuzzy DEA super-
and 3.814 h on average, respectively. It is also worth mentioning efficiency model would be developed to derive the super-efficiency
that the average relative gap between total cost reported by En-BDA scores.
and the classical BDA (GUROBI) for instances 4 and 5 (3, 4 and 5), Furthermore, the proposed solution approach provided DEA supper-
for which the classical BDA (GUROBI) fails to find the optimum efficiency score of all purchasing firms over the whole time horizon. An
solution, is equal to 1.61% (1.57%) that values $2,022,279 interesting development of the present study is to propose a framework
($1,923,959). This implies that the best feasible solutions obtained by which DEA super efficiency score of each purchasing firm in each
by the classical BDA and GUROBI within 6 h are far from the op- period of time is obtained. In this case, the impact of suppliers’ per-
timum solutions derived by En-BDA. formance on each purchasing firm with respect to sustainability criteria
• With respect to the number of iterations, Fig. 4 demonstrates the would be monitored.
average number of iterations when solving each instance using the
classical BDA and En-BDA. In summary, it is concluded that the Appendix A. Supplementary material
developed computational enhancements are significantly able to
improve the solution time and the number of iterations of En-BDA in Supplementary data associated with this article can be found, in the
comparison with the classical BDA and GUROBI. online version, at https://doi.org/10.1016/j.cie.2019.01.011.

6. Conclusion and future works References

In this paper, the sustainable supplier selection and order allocation Adler, N., Friedman, L., & Sinuany-Stern, Z. (2002). Review of ranking methods in the
data envelopment analysis context. European Journal of Operational Research, 140(2),
problem was addressed through development of a comprehensive
249–265.
multi-period, multi-product multi-objective MILP model that also took Aktin, T., & Gergin, Z. (2016). Mathematical modeling of sustainable procurement stra-
price discount, multimodal transportation and shortage condition into tegies: Three case studies. Journal of Cleaner Production, 113, 767–780.
Amindoust, A., Ahmed, S., Saghafinia, A., & Bahreininejad, A. (2012). Sustainable sup-
account, in which all sustainability aspects were included based on the plier selection: A ranking model based on fuzzy inference system. Applied Soft
triple bottom line approach. In addition, a hybrid solution procedure,

207
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

Computing, 12, 1668–1677. suppliers for supplier selection and performance improvement. Supply Chain
Azadnia, A. H., Saman, M. Z. M., & Wong, K. Y. (2015). Sustainable supplier selection and Management, 5(3), 143–150.
order lot-sizing: An integrated multi-objective decision-making process. International Luthra, S., Govindan, K., Kannan, D., Mangla, S. K., & Garg, C. P. (2017). An integrated
Journal of Production Research, 53(2), 383–408. framework for sustainable supplier selection and evaluation in supply chains. Journal
Bai, C., & Sarkis, J. (2010). Integrating sustainability into supplier selection with grey of Cleaner Production, 140(3), 1686–1698.
system and rough set methodologies. International Journal of Production Economics, Magnanti, T. L., & Wong, R. T. (1981). Accelerating Benders decomposition: Algorithmic
124, 252–264. enhancement and model selection criteria. Operations Research, 29(3), 464–484.
Banker, R. D., & Moray, R. C. (1986). Efficiency analysis for exogenously fixed input and Mavrotas, G. (2009). Effective implementation of the -constraint method in multi-ob-
output. Operation Research, 34(4), 513–518. jective mathematical programming problems. Applied Mathematics and Computation,
Benders, J. F. (1962). Partitioning procedures for solving mixed-variables programming 213, 455–465.
problems. Numerische Mathematik, 4, 238–252. Mohammed, A., & Wang, Q. (2017). The fuzzy multi-objective distribution planner for a
Boussofiane, A., Dyson, R. G., & Thanassoulis, E. (1991). Applied data envelopment green meat supply chain. International Journal of Production Economics, 184, 47–58.
analysis. European Journal of Operational Research, 52, 1–15. Moheb-Alizadeh, H. (2018). Efficient design of sustainable supply chain networks: de-
Brandenburg, M., Govindan, K., Sarkis, J., & Seuring, S. (2014). Quantitative models for velopment and extensions. PhD Dissertation, North Carolina State University,
sustainable supply chain management: Developments and directions. European Raleigh, USA.
Journal of Operational Research, 233, 299–312. Moheb-Alizadeh, H., & Faez, F. (2009). A multi-objective approach to supplier evaluation
Branke, J., Deb, K., Miettinen, K., & Slowinski, R. (2008). Multi-objective optimization: using multiple criteria data envelopment analysis (MCDEA). Journal of Industrial
Interactive and evolutionary approaches. Berlin: Springer-Verlag. Engineering (Journal of Faculty of Engineering), 43(1), 67–82.
Buyukozkan, G., & Cifci, G. (2011). A novel fuzzy multi-criteria decision framework for Moheb-Alizadeh, H., & Handfield, R. (2018). An integrated chance-constrained stochastic
sustainable supplier selection with incomplete information. Computers in Industry, 62, model for efficient and sustainable supplier selection and order allocation.
164–174. International Journal of Production Research, 56(21), 6890–6916.
Celebi, D. C., & Bayraktar, D. (2008). An integrated neural network and data envelopment Moheb-Alizadeh, H., Mahmoudi, M., & Bagheri, R. (2017). Supplier selection and order
analysis for supplier evaluation under incomplete information. Expert Systems with allocation using a stochastic multi-objective programming model and genetic algo-
Applications, 35, 1698–1710. rithm. International Journal of Integrated Supply Management, 11(4), 291–315.
Charnes, A., Cooper, W. W., & Rhodes, E. (1978). Measuring efficiency of decision making Orji, I. J., & Wei, S. (2015). An innovative integration of fuzzy-logic and systems dy-
units. European Journal of Operational Research, 2, 429–444. namics in sustainable supplier selection: A case on manufacturing industry. Computers
Devika, K., Jafarian, A., & Nourbakhsh, V. (2014). Designing a sustainable closed-loop & Industrial Engineering, 88, 1–12.
supply chain network based on triple bottom line approach: A comparison of meta- Park, K. S., & Park, K. (2009). Measurement of multi-period aggregative efficiency.
heuristics hybridization techniques. European Journal of Operational Research, 235, European Journal of Operational Research, 193, 567–580.
594–615. Paydar, M. M., Babaveisi, V., & Safaei, A. S. (2017). An engine oil closed-loop supply
Dickson, G. W. (1966). An analysis of vendor selection systems and decisions. Journal of chain design considering collection risk. Computers and Chemical Engineering, 104,
Purchasing, 2(1), 5–17. 38–55.
Dotoli, M., & Falagario, M. (2012). A hierarchical model for optimal supplier selection in Pishvaee, M. S., Razmi, J., & Torabi, S. A. (2014). An accelerated Benders decomposition
multiple sourcing contexts. International Journal of Production Research, 50(11), algorithm for sustainable supply chain network design under uncertainty: A case
2953–2967. study of medical needle and syringe supply chain. Transportation Research Part E, 67,
Dyllick, T., & Hockerts, K. (2002). Beyond the business case for corporate sustainability. 14–38.
Business Strategy and the Environment, 11(2), 130–141. Punniyamoorthy, M., Mathiyalagan, P., & Parthiban, P. (2011). A strategic model using
Elkington, J. (1998). Cannibals with forks: The triple bottom line of 21st century business. structural equation modeling and fuzzy logic in supplier selection. Expert Systems with
Oxford: Capstone. Applications, 38(1), 458–474.
Fahimnia, B., Jabbarzadeh, A., Ghavamifar, A., & Bell, M. (2017). Supply chain design for Rahmaniani, R., Crainic, T. G., Gendreau, M., & Rei, W. (2017). The Benders decom-
efficient and effective blood supply in disasters. International Journal of Production position algorithm: A literature review. European Journal of Operational Research, 259,
Economics, 183, 700–709. 801–817.
Falagario, M., Sciancalepore, F., Costantino, N., & Pietroforte, R. (2012). Using a DEA- Rei, W., Cordeau, J.-F., Gendreau, M., & Soriano, P. (2009). Accelerating Benders de-
cross efficiency approach in public procurement tenders. European Journal of composition by local branching. INFORMS Journal on Computing, 21(2), 333–345.
Operational Research, 218, 523–529. Saen, R. F. (2007). A new mathematical approach for suppliers selection: Accounting for
Fischetti, M., & Lodi, A. (2003). Local branching. Mathematical Programming, 98(1-3), non-homogeneity is important. Applied Mathematics and Computation, 185(1), 84–95.
23–47. Saen, R. F. (2008). Using super-efficiency analysis for ranking suppliers in the presence of
Geoffrion, A. M., & Graves, G. W. (1974). Multicommodity distribution system design by volume discount offers. International Journal of Physical Distribution & Logistics
benders decomposition. Management Science, 20(5), 822–844. Management, 38(8), 637–651.
Ghadimi, P., Ghassemi-Toosi, F., & Heavey, C. (2018). A multi-agent systems approach for Saen, R. F. (2010). Developing a new data envelopment analysis methodology for supplier
sustainable supplier selection and order allocation in a partnership supply chain. selection in the presence of both undesirable outputs and imprecise data. International
European Journal of Operational Research, 269(1), 286–301. Journal of Advanced Manufacturing Technology, 51, 1243–1250.
Govindan, K., Khodaverdi, R., & Jafarian, A. (2013). A fuzzy multi criteria approach for Saharidis, G. K., Boile, M., & Theofanis, S. (2011). Initialization of the Benders master
measuring sustainability performance of a supplier based on triple bottom line ap- problem using valid inequalities applied to fixed-charge network problems. Expert
proach. Journal of Cleaner Production, 47, 345–354. Systems with Applications, 38, 6627–6636.
Govindan, K., Rajendran, S., Sarkis, J., & Murugesan, P. (2015). Multi criteria decision Santoso, T., Ahmed, S., Goetschalckx, M., & Shapiro, A. (2005). A stochastic programming
making approaches for green supplier evaluation and selection: A literature review. approach for supply chain network design under uncertainty. European Journal of
Journal of Cleaner Production, 98, 66–83. Operational Research, 167, 96–115.
Gupta, P., Govindan, K., Mehlawat, M. K., & Kumar, S. (2016). A weighted possibilistic Sarkis, J., & Dhavale, D. G. (2015). Supplier selection for sustainable operations: A triple-
programming approach for sustainable vendor selection and order allocation in fuzzy bottom-line approach using a Bayesian framework. International Journal of Production
environment. International Journal of Advanced Manufacturing Technology, 86, Economics, 166, 177–191.
1785–1804. Seuring, S., & Muller, M. (2008). From a literature review to a conceptual framework for
Ha, S. H., & Krishnan, R. (2008). A hybrid approach to supplier selection for the main- sustainable supply chain management. Journal of Cleaner Production, 16(15),
tenance of a competitive supply chain. Expert Systems with Applications, 34, 1699–1710.
1303–1311. Shalke, P. N., Paydar, M. M., & Hajiaghaei-Keshteli, M. (2017). Sustainable supplier se-
Jafari-Songhori, M., Tavana, M., Azadeh, A., & Khakbaz, M. H. (2011). A supplier se- lection and order allocation through quantity discounts. International Journal of
lection and order allocation model with multiple transportation alternatives. Management Science and Engineering Management, 12, 1–13.
International Journal of Advanced Manufacturing Technology, 52, 365–376. Song, W., Xu, Z., & Liu, H-. C. (2017). Developing sustainable supplier selection criteria
Jeihoonian, M., Zanjani, M. K., & Gendreau, M. (2016). Accelerating Benders decom- for solar air-conditioner manufacturer: An integrated approach. Renewable and
position for closed-loop supply chain network design: Case of used durable products Sustainable Energy Reviews, 79, 1461–1471.
with different quality levels. European Journal of Operational Research, 251(3), Talluri, S., Narasimhan, R., & Nair, A. (2006). Vendor performance with supply risk: A
830–845. chance-constrained DEA approach. International Journal of Production Economics, 100,
Kao, C., & Liu, S-. T. (2014). Multi-period efficiency measurement in data envelopment 212–222.
analysis: The case of Taiwanese commercial banks. Omega, 47, 90–98. Trapp, A. C., & Sarkis, J. (2016). Identifying Robust portfolios of suppliers: A sustain-
Karsak, E. E., & Dursun, M. (2014). An integrated supplier selection methodology in- ability selection and development perspective. Journal of Cleaner Production, 112,
corporating QFD and DEA with imprecise data. Expert Systems with Applications, 41, 2088–2100.
6995–7004. Ustun, A. K., & Anagun, A. S. (2015). Multi-objective mitigation budget allocation pro-
Kumar, D., Rahman, Z., & Chan, F. T. S. (2017). A fuzzy AHP and fuzzy multi-objective blem and solution approaches: The case of stanbul. Computers & Industrial Engineering,
linear programming model for order allocation in a sustainable supply chain: A case 81, 118–129.
study. International Journal of Computer Integrated Manufacturing, 30(6), 535–551. WCED (World Commission on Environment and Development). (1987). Our common fu-
Kuo, R. J., & Lin, Y. J. (2012). Supplier selection using analytic network process and data ture. Oxford: Oxford University Press.
envelopment analysis. International Journal of Production Research, 50(11), Weber, C. A. (1996). A data envelopment analysis approaches to measuring vendor
2852–2863. performance. Supply Chain Management, 1(1), 28–39.
Kuo, R. J., Wang, Y. C., & Tien, F. C. (2010). Integration of artificial neural network and Weber, C. A., Current, J. R., & Benton, W. C. (1991). Vendor selection criteria and
MADA methods for green supplier selection. Journal of Cleaner Production, 18(12), methods. European Journal of Operational Research, 50, 2–18.
1161–1170. Weber, C. A., Current, J. R., & Dessai, A. (2000). An optimization approach to de-
Li, S., Jahanshahloo, G. R., & Khodabakhshi, M. (2007). A super-efficiency model for termining the number of vendors to employ. Supply Chain Management, 5(2), 90–98.
ranking efficient units in data envelopment analysis. Applied Mathematics and Willis, H. T., Huston, R. C., & Pohlkamp, F. (1993). Evaluation measures of just in time
Computation, 184, 638–648. supplier performance. Production and Inventory Management Journal, 34(2), 1–5.
Liu, F., Ding, F. Y., & Lall, V. (2000). Using data envelopment analysis to compare Wolsey, L. A. (1998). Integer programming. New York: Wiley.

208
H. Moheb-Alizadeh, R. Handfield Computers & Industrial Engineering 129 (2019) 192–209

Wu, D. (2009). Supplier selection: A hybrid model using DEA, decision tree and neural Zhou, X., Pedrycz, W., Kuang, Y., & Zhang, Z. (2016). Type-2 fuzzy multi-objective DEA
network. Expert Systems with Applications, 36, 9105–9112. model: An application to sustainable supplier evaluation. Applied Soft Computing, 46,
Yousefi, S., Soltani, R., Saen, R. F., & Pishvaee, M. S. (2017). A robust fuzzy possibilistic 424–440.
programming for a new network GP-DEA model to evaluate sustainable supply Zimmer, K., Frohling, M., & Schultmann, F. (2016). Sustainable supplier managementa
chains. Journal of Cleaner Production, 166, 537–549. review of models supporting sustainable supplier selection, monitoring and devel-
Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8, 338–353. opment. International Journal of Production Research, 54(5), 1412–1442.

209