Computers & Industrial Engineering 110 (2017) 527–537

Contents lists available at ScienceDirect

Computers & Industrial Engineering

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

A bi-level programming approach for production-distribution supply

chain problem
Omid Amirtaheri, Mostafa Zandieh ⇑, Behrouz Dorri, A.R. Motameni
Department of Industrial Management, Management and Accounting Faculty, Shahid Beheshti University, G.C., Tehran, Iran

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

Article history: This paper investigates a decentralized production-distribution supply chain consisting of one manufac-
Received 20 November 2016 turer and one distributor where the demand is jointly influenced by pricing and advertising policies. The
Received in revised form 16 May 2017 control of decision variables is partitioned amongst the members. The manufacturer decides about
Accepted 20 June 2017
wholesale price, production interval, expenditure of global advertising, and participation rate in the dis-
Available online 22 June 2017
tributor’s local advertising expenditure while the decisions on selling price, local advertising expenditure,
and allocation of demands to depots are made by the distributor. We propose a Stackelberg game frame-
Keywords:
work and develop two nonlinear bi-level programming models by switching the leader and follower roles
Bi-level programming
Supply chain coordination
between the manufacturer and the distributor. Four hierarchical solution algorithms are proposed by
Production-distribution supply chain combining genetic algorithm and particle swarm optimization to tackle the bi-level programming mod-
Stackelberg game els. Finally, computational experiments are carried out to analyze and compare the efficiency and effec-
Genetic algorithms tiveness of the proposed solution procedures.
Particle swarm optimization Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction the follower will take, as a response to the leader’s decisions.

Therefore, the actions of one influence the choices and payoffs
The last two decades have witnessed a considerable interest by available for the other.
researchers in supply chain management (SCM) and a new empha- The general form of a bi-level programming problem (BLPP) is
sis on the interactions among supply chain members. There are stated as follows:
two approaches towards decision making in supply chain: central-
min Fðx; yÞ
ized and decentralized. In a centralized supply chain, an individual x2X
member with dominant position delineates the policies for other subject to Gðx; yÞ 6 0
members. In such a case, members are willing to cooperate based min f ðx; yÞ
on the declared policies. In decentralized supply chain, however, y2Y

firms do not tend to cooperate and seek to optimize their own
objectives rather than those of the entire system.
Bi-level programming is an appropriate technique for modeling
decentralized management problems with two non-cooperative
decision makers in a hierarchical structure. From a historical point
of view, bi-level programming as a special case of multi-level opti-
mization is closely related to the economic problem of Stackelberg
(Von Stackelberg, 1952) in the field of game theory.
The decision maker at the upper level of the hierarchy is known
as leader, while the lower level decision maker is termed as fol-
lower. Each decision maker controls a set of decision variables sub-
ject to a set of constraints and aims to optimize his own objective
function. The actions taken by the leader affect the reactions that
⇑ Corresponding author.
E-mail address: m_zandieh@sbu.ac.ir (M. Zandieh).
subject to gðx; yÞ 6 0
where x 2 X # R is the set of leader’s decision variables and
n
y 2 Y # Rm is the set of follower’s decision variables. Similarly, the
functions F : Rnm ! R and f : Rnm ! R are objective functions
of leader and follower, while G : Rnm ! Rp and g : Rnm ! Rq
are upper and lower level constraints, respectively.
The relaxed feasible region (or constraint region) of BLPP is:
S ¼ fðx; yÞ : x 2 X; y 2 Y; Gðx; yÞ 6 0 and gðx; yÞ 6 0g
For a fixed x 2 X, the lower-level feasible set is defined by:
SðxÞ ¼ fy 2 Y : gðx; yÞ 6 0g
While the lower-level reaction set (or rational reaction set) is:
~Þ; y
RðxÞ ¼ fy 2 Y : y 2 arg minðf ðx; y ~ 2 SðxÞÞg
Every y 2 RðxÞ is a rational response.

http://dx.doi.org/10.1016/j.cie.2017.06.030
0360-8352/Ó 2017 Elsevier Ltd. All rights reserved.
528 O. Amirtaheri et al. / Computers & Industrial Engineering 110 (2017) 527–537
Finally, the induced region (or inducible region) is defined as:
IR ¼ fðx; yÞ : ðx; yÞ 2 S; y 2 RðxÞg
This set is formed by the feasible points of the BLPP, corresponds to
the feasible set of the leader. It is usually non-convex and can even
be disconnected or empty in presence of upper-level constraints.
The BLPP is equivalent to optimizing the leader’s objective F over
the inducible region IR. A feasible point ðx ; y Þ 2 IR is a global opti-
mum of BLPP (or Stackelberg equilibrium) if Fðx ; y Þ 6 Fðx; yÞ for all
ðx; yÞ 2 IR.
Bi-level programming techniques have been widely applied to
handle realistic decentralized decision problems. Gao and et al.
(2011) develop two nonlinear bi-level programming models for
pricing problem in a vendor-buyer supply chain. Sadigh,
Mozafari, and Karimi (2012) address a multi-product
manufacturer-retailer supply chain in which the demand of each
product is jointly influenced by price and advertising expenditure.
They apply bi-level programming approach and present several
solution procedures to find optimal policies in terms of pricing,
advertising and inventory management. Mokhlesian and Zegordi
(2014) develop a nonlinear bi-level programming model for a
multi-product two-echelon supply chain consisting of one manu-
facturer and multiple retailers. Zhang, Lu, and Gao (2015) establish
a bi-level pricing and replenishment strategy optimization model
in high-tech industry where the buyer and the vendor are respec-
tively designated as the leader and the follower. Ma, Wang, and
Zhu (2014) consider joint pricing and lot sizing problem in a
two-echelon supply chain system with one manufacturer and
one retailer. They establish two nonlinear bi-level programming
models by switching the leader and the follower roles between
the manufacturer and the retailer.
Although production-distribution supply chain problems have
been topics of wide interest in SCM research, little attention has
been given to analyzing the problems by means of bi-level pro-
gramming approach. Ryu, Dua, and Pistikopoulos (2004) present a
bi-level programming model to a production and distribution plan-
ning problem where the follower’s problem can be modeled by lin-
ear programs. A similar production and distribution problem with
probabilistic parameters is formulated as a bi-level linear multi-
objective programming problem by Roghanian (Roghanian,
Sadjadi, & Aryanezhad, 2007). Calvete, Galé, and Oliveros (2011)
study a hierarchical production-distribution planning in which
the distributor controls the design of the routes and the manufac-
turer controls the production process. In another paper (Calvete,
Galé, & Iranzo, 2014), they address a mixed integer bi-level opti-
mization model for the planning of a decentralized distribution net-
work consisting of manufacturing plants, depots and customers.
This paper addresses a decentralized production-distribution
supply chain consisting of one manufacturer and one distributor
and develop two nonlinear bi-level programming models, one bi-
level model considers the manufacturer as the leader
(Manufacturer-Stackelberg), who has priority in deciding, and the
other takes the distributor as the leader (Distributor-Stackelberg).
Both the manufacturer and distributor aim to maximize their prof-
its, but their decisions are related to each other in a hierarchical
way. The distributor owns several depots to supply customers in
different markets. The markets are also different in terms of size,
geographic location, price elasticity of demand, and response to
advertisement.
In order to boost the markets’ demands, the distributor under-
takes a given expenditure to conduct local advertising in each mar-
ket. On the other hand, the manufacturer may in turn, set a budget
for global advertising within the target markets. We assume that
the manufacturer participates in a cooperative advertising pro-
gram and undertakes a portion of distributor’s local advertising
expenditure. Cooperative advertising is seen to be among those
topics intensely investigated in the literatures during the recent
years (Aust & Buscher, 2014). This concept represents a financial
agreement between the manufacturer and the distributor to share
advertising expenditure based on which the manufacturer under-
takes part of local advertising expenditure (participation rate)
incurred by the distributor (Koh, 2013).
In the proposed BLPP, the control of decision variables is
partitioned amongst the manufacturer and the distributor. The
manufacturer decides about wholesale price, production
(replenishment) interval, global advertising expenditure, and par-
ticipation rate in the distributor’s local advertising expenditure.
The distributor, however, makes decisions on selling price, local
advertising expenditure in each market, and allocation of markets’
demands to depots in a way that total transportation cost is mini-
mized. In fact, neither the manufacturer nor the distributor has
direct control over the decision-making process of the other, but
their decisions affect the other member’s subsequent reactions.
With the assumption of perfect information to be available about
the follower, in both models, the leader could anticipate the reac-
tion of the follower and acts in a way that is optimal for him.
The first contribution of this paper is to model a decentralized
production-distribution supply chain problem by using of bi-level
programming, which allows us to take into account how decisions
made at the distribution segment of the supply chain can affect
and be affected by decisions made at the manufacturing segment.
Bi-level programming models are non-convex and strongly NP-
hard, even in their simplest case in which all the functions involved
are linear (Ben-Ayed & Blair, 1990; Hansen, Jaumard, & Savard,
1992). Thus, it is quite difficult to deal with nonlinear version of
these models. Traditional approaches for solving BLPP can be
roughly divided into the following categories (Dempe, 2002): (1)
vertex enumeration methods, (2) decent algorithm, (3) methods
based on Karush-Kuhn-Tucker conditions, and (4) penalty func-
tions, etc. The properties such as continuity, convexity and differ-
entiability are necessary when proposing these methods. Hence,
they have limitations on solving BLPP. meta-heuristic algorithms
including stochastic optimization techniques are recognized as
useful tools for solving BLPP which do not necessarily satisfy the
classical optimization assumptions. This has motivated researchers
to apply meta-heuristic methods to tackle BLPP (Gao & et al., 2011;
Khilwani & et al., 2008; Koh, 2007; Marinakis & Marinaki, 2008;
Mokhlesian & Zegordi, 2014; Rajesh & et al., 2003; Sadigh et al.,
2012; Sahin & Ciric, 1998).
Genetic Algorithm (GA) and Particle Swarm Optimization (PSO)
are the generic computational techniques espoused from the pro-
gression of biological life in the natural humanity (Khilwani &
et al., 2008). The second contribution of this study is to develop
four hierarchical algorithms by combining GA and PSO to tackle
the proposed bi-level programming models. In addition to these
algorithms, an exhaustive grid search is adopted for validation
purposes.
The rest of this paper is organized as follows. In Section 2, two
bi-level programming models (Manufacturer-Stackelberg &
Distributor-Stackelberg) are developed by formulating the manu-
facturer and the distributor problems. Section 3 proposes four hier-
archical procedures for solving the Stackelberg models. Section 4 is
devoted to experimental results and finally Section 5 discusses the
concluding remarks.
2. Bi-level programming models for production-distribution
supply chain
In this section, we describe general characteristics of the
production-distribution supply chain and formulate the bi-level
programming models.
 O. Amirtaheri et al. / Computers & Industrial Engineering 110 (2017) 527–537
Consider a two-echelon supply chain including one manufac-
turer and one distributor. The manufacturer wholesales the pro-
duct to his exclusive distributor at a wholesale unit price (w).
The distributor, in turn, sells the goods to customers in different
markets at unit retail price (p), where p ¼ w:ð1 þ rÞ and r is markup
percentage. It is assumed that there are K different markets in
terms of size, price elasticity of demand and response to global
and local advertising efforts.
A local advertising expenditure is incurred by the distributor in
each market (ak ; k ¼ 1; 2; . . . ; K) to stimulate the demand in short-
term. Comparatively, the manufacturer undertakes an expenditure
to conduct global advertising (A) so as to motivate the demand in
long-term. The manufacturer also compensates a portion of local
advertising expenditure (h), incurred by the distributor. Bearing
in mind the prevalent assumption in the literature (Xie & Neyret,
2009; Yue & et al., 2006), it can be assumed that the demand func-
tion of each market to have the following form:
Dk ðp; A; ak Þ ¼ zk :g k ðpÞ:hk ðA; ak Þ; k ¼ 1; 2; . . . ; K
where zk is the base demand while g k ðpÞ and hk ðA; ak Þ denote the
effects of retail price and advertising expenditures on k-th market’s
demand, respectively. Unlike a common assumption in many stud-
ies, where demand is a linear function of retail price (Swami & Shah,
2011; Xie & Wei, 2009), in this research a nonlinear function is con-
sidered to model the effect of retail price on demand. Inspired by
SeyedEsfahani, Biazaran, and Gharakhani (2011), we have:
g k ðpÞ ¼ ð1  ak :pÞ1=v k ; k ¼ 1; 2; . . . :; K
where ak is a positive constant and v k is the shape function of
demand curve for k-th market. By adjusting the parameter v k , we
incorporate specific demand curve for each market into the pro-
posed models (v k < 1, v k ¼ 1, and v k > 1 yields convex, linear,
and concave demand functions, respectively). It is obvious that
the condition p 6 mink¼1;2;...;K f1=ak g should be verified to avoid neg-
ative demand.
We use the function proposed by Xie and Wei (2009) to model
the effects of advertising expenditures on demand:
Fig. 1. Inventory levels for the manufacturer and the distributor’s depots.
529
pffiffiffi pffiffiffiffiffi
hk ðA; ak Þ ¼ n1k : A þ n2k : ak ; k ¼ 1; 2; . . . ; K ð3Þ
where n1k and n2k are positive constants reflecting the effects of glo-
bal and local advertising expenditures on k-th market’s demand,
respectively. This approach has been followed by Zhang, Xie, and
Chen (2013) in the literature. Combining Eq. (1) to Eq. (3), the
demand function of k-th market would be the following:
pffiffiffi pffiffiffiffiffi
Dk ðp; A; ak Þ ¼ zk :ð1  ak :pÞ1=v k :ðn1k : A þ n2k : ak Þ; k ¼ 1; 2; . . . ; K
ð4Þ
Therefore, total demand induced by pricing and advertising
P
policies is D0 ¼ Kk¼1 Dk . The distributor owns several depots with
known capacities (bj ; j ¼ 1; 2; . . . ; J where J is the number of depots)
and serves the markets geographically dispersed with the demand
of Dk (k ¼ 1; 2; . . . ; K). Furthermore, the distributor decides which
markets are assigned to each depot, while minimizing transporta-
ð1Þ tion costs. It is assumed that the demand of each market is sup-
plied by the depots in a unit-by-unit order. Therefore, we take cjk
as unit transportation cost from j-th depot to k-th market. In order
to meet the demands assigned to each depot, sufficient supply of
product has to be available at that depot. The manufacturer sup-
plies the product equal to the total demand (D0 ) within lots of
Q 0 units and delivers each lot to the depots based on their demand
shares. Denoting the production (replenishment) interval as
T ð0 < T 6 1Þ; the manufacturer delivers Q 0 ¼ D0 :T units to the
ð2Þ depots in each interval. The demand share of j-th depot is calcu-
lated as Q j ¼ dj =D0 where dj is the total demand of j-th depot.
Assuming two depots, the inventory levels for the manufacturer
as well as the distributor’s depots are shown in Fig. 1.
It is assumed that product is delivered to the distributor at the
manufacturer’s premises. Therefore, the transportation cost from
the manufacturer to the depots is incurred by the distributor. In
each replenishment cycle, the manufacturer delivers Q j units,
simultaneously within a lot, to j-th depot. As the capacity of trans-
portation vehicle is directly affected by the lot size, we can assume
that transportation cost from the manufacturer to each depot is a
530 O. Amirtaheri et al. / Computers & Industrial Engineering 110 (2017) 527–537

staircase function of Q j (see Fig. 2). Without loss of generality, it is Table 1

Notations for bi-level programming models.
possible to use a common staircase function for all depots.
It is assumed that distributor will not be interested in working Indices
with manufacturer if he does not obtain a minimum markup per- i Level of transportation cost (i ¼ 1; 2; . . . ; I)
j Depot (j ¼ 1; 2; . . . ; J)
centage (dD ). Similarly, manufacturer requires a minimum margin
k Market (k ¼ 1; 2; . . . ; K)
on wholesale price (dM ) to collaborate with distributor. dD and dM
Parameters (Manufacturer)
can be regulated to appropriate levels so as to avoid negative
f Production cost ($/Unit)
profits. uM Setup cost ($/Setup)
In order to formulate the bi-level programming models, we use qM Percentage of inventory holding cost ($/Unit)
the notations listed in Table 1. We assume that wholesale price (w), G Production capacity (Unit)
global advertising expenditure (A), participation rate in local Parameters (Distributor)
advertising expenditure (h), and production (replenishment) inter- uj D
Ordering cost for the j-th depot ($/Order)
val (T) are decided by the manufacturer and the distributor con- qDj Percentage of inventory holding cost for the j-th depot ($/Unit)
trols the markup percentage (r), local advertising expenditure in bj Capacity of the j-th depot (Unit)
each market (ak ), the assignment of markets demands to the cjk Unit transportation cost from j-th depot to k-th market ($/Unit)
depots (xjk ), and the level of transportation cost used to replenish qi Maximum capacity of i-th level of transportation cost (Unit)
si Transportation cost of i-th level ($)
each depot (yij ). M Large positive constant
In the following, we model the interactions between the manu-
Parameters (Markets)
facturer and the distributor through bi-level programming. zk Demand base of k-th market (Unit)
ak Effectiveness coefficient of retail price on demand of k-th market (1/$)
2.1. Manufacturer-Stackelberg (MS) model v k Shape factor of demand curve in k-th market
n1k Effectiveness coefficient of global advertising expenditure on demand of

Here we assume that the manufacturer holds the manipulative k-th market ($1=2 )
n2k Effectiveness coefficient of local advertising expenditure on demand of
power and acts as leader. This model can be expressed as follows:
k-th market ($1=2 )
X
K
1=v
pffiffiffi pffiffiffiffiffi Decision variables (Manufacturer)
max pM ¼ ðw  f Þ: zk :ð1  ak :w:ð1 þ rÞÞ k :ðn1k : A þ n2k : ak Þ  1=T:uM
w;A;T;h w Wholesale price ($/Unit)
k¼1
A Global advertising expenditure ($)
X
K pffiffiffi pffiffiffiffiffi
zk :ð1  ak :w:ð1 þ rÞÞ1=v k :ðn1k : A þ n2k : ak Þ
T Production (replenishment) interval
 T=2:qM :f : h Participation rate in local advertising expenditure
k¼1
Decision variables (Distributor)
X
K
 A  h: ak ð5aÞ r Markup percentage
k¼1 ak Local advertising expenditure in k-th market ($)
xjk Units of product sent from j-th depot to k-th market (Unit)
yij Indicator variable equals to 1 if i-th level of transportation
St:ðr; ak ; xjk ; yji Þ 2 argmaxpD cost is used by j-th depot and 0 otherwise
X
K  pffiffiffi pffiffiffiffiffi
¼ r:w: zk :ð1  ak :w:ð1 þ rÞÞ1=v k : n1k : A þ n2k : ak
Auxiliary variables
Dk Demand of k-th market (Unit)
k¼1 D0 Total demand (Unit)
pM Profit of the manufacturer ($)
pD Profit of the distributor ($)
X
J X
J X
K X
K
1=T: uDj  w:T=2 xjk :qDj  ð1  hÞ: ak
j¼1 j¼1 k¼1 k¼1 X
K
xjk 6 bj ; j ¼ 1; 2; . . . ; J ð5dÞ
XJ X
I X
J X
K k¼1
1=T: si :yij  cjk :xjk ð5bÞ
j¼1 i¼1 j¼1 k¼1 X
J pffiffiffi pffiffiffiffiffi
xjk ¼ zk :ð1  ak :w:ð1 þ rÞÞ1=v k :ðn1k : A þ n2k : ak Þ; k ¼ 1; 2; . . . ; K
!
X
K
1=v k
pffiffiffi pffiffiffiffiffi X
J j¼1
zk :ð1  ak :w:ð1 þ rÞÞ :ðnk : A þ nk : ak Þ 6 min
1 2
bj ; G ð5eÞ
k¼1 j¼1

ð5cÞ X
K
qi :yij 6 xjk :T < qiþ1 :yij þ M:ð1  yij Þ;
k¼1

i ¼ 1; 2; . . . ; I; j ¼ 1; 2; . . . ; J; q1 ¼ 0; qIþ1 ¼ M ð5fÞ

X
I
yij ¼ 1; j ¼ 1; 2; . . . ; J ð5gÞ
i¼1

A P Amin ; 0 6 h 6 1; 0 < T 6 1; w P f :ð1 þ dM Þ ð5hÞ

r P dD ; ak P akmin ; xjk P 0; yij ¼ 0 and 1 ð5iÞ

i ¼ 1; 2; . . . ; I; j ¼ 1; 2; . . . ; J; k ¼ 1; 2; . . . ; K
Expressions (5a) and (5b) are the upper-level and the lower-level
Fig. 2. Transportation cost levels. objective functions which reflect the profits of the manufacturer
 O. Amirtaheri et al. / Computers & Industrial Engineering 110 (2017) 527–537 531

and the distributor, respectively. The manufacturer’s profit is
obtained by subtracting setup cost, inventory holding cost, global
advertising expenditure and participation in local advertising
expenditure from his revenues. Similarly, the distributor’s profit is
equal to his revenues minus the sum of ordering cost, inventory
holding cost, local advertising expenditure, total transportation cost
from the manufacturer to the depots, and total transportation cost
from the depots to the markets. Both objective functions seek to
optimize profit.
Constraint (5c) guarantees that total generated demand by pric-
ing and advertising policies is completely met by the distributor
and the manufacturer. For this purpose, it shall exceed neither total
capacity of the distributor’s depots nor the production capacity of
the manufacturer. Constraints (5d) imply that the capacities of the
depots are not exceeded. Constraints (5e) ensure that the demand
of each market is entirely met by the depots. Constraints (5f) and
(5g) establish links between the sets of variables yij , xjk and T to
avoid incoherent solutions. Constraints (5h) and (5i) are for decla-
ration of variables.
2.2. Distributor-Stackelberg (DS) model
By exchanging the objective functions (profit functions) of the
members in MS model, we arrive at DS model. In DS model, the dis-
tributor and the manufacturer are considered as the leader and the
follower, respectively. Similar to the previous section, DS model
can be formulated as a bi-level program:
X
K pffiffiffi
max pD ¼ r:w: zk :ð1  ak :w:ð1 þ rÞÞ1=v k :ðn1k : A þ n2k : ak Þ
j
r;ak ;xjk ;yi k¼1
on GA and PSO to tackle MS and DS problems. These methodologies
solve the Stackelberg models directly by simulating the sequential
decision making procedure of BLPP. The proposed algorithms are
composed of two levels, upper-level algorithm (ULA) which is
designed to solve the leader’s problem based on the follower’s
response and lower-level algorithm (LLA) which is embedded in
ULA to solve the lower-level problem and obtain the optimal
response of the follower for each given set of values of leader’s
decision variables.
The ULA starts with generating initial solution for the leader’s
problem (x). The solution is then received by the LLA and reaction
of the follower (y ) is obtained and returned to the ULA. The initial
solution for the leader’s problem (x; y ) are evaluated based on the
follower’s response and the first iteration is completed. This proce-
dure is repeated in each iteration until solution (x ; y ) is obtained
for the upper-level problem. It is important to clarify that the solu-
tion returned by LLA is not always optimal for follower. Therefore,
the solution obtained by ULA for bi-level models could be near fea-
sible. Fig. 3 depicts the general steps of the proposed hierarchical
algorithms.
We construct four hierarchical algorithms, named as GG, PG, GP
and PP, by combining GA and PSO as the ULA and the LLA (see
Table 2).
In Section 3.1, we adopt two GAs, one for the manufacturer
problem (as ULA in MS model and as LLA in DS model) and another
for the distributor problem (as LLA in MS model and as ULA in DS
model). Similarly, two PSO algorithms are developed for the man-
ufacturer and the distributor problems in Section 3.2.
pffiffiffiffiffi
3.1. Genetic algorithms

X
J X
J X
K X
K
 1=T: uDj  w:T=2 xjk :qDj  ð1  hÞ: ak GA is a stochastic global search technique that solves problems
j¼1 j¼1 k¼1 k¼1 by imitating processes observed during natural evolution. Based on
the survival and reproduction of the fittest, GA gradually exploits
X
J X
I X
J X
K
 1=T: si :yij  cjk :xjk ð6aÞ new and better solutions without any presumptions, such as con-
j¼1 i¼1 j¼1 k¼1

St:ðw; A; T; hÞ 2 arg max pM

X
K pffiffiffi pffiffiffiffiffi
¼ ðw  f Þ: zk :ð1  ak :w:ð1 þ rÞÞ1=v k :ðn1k : A þ n2k : ak Þ
k¼1

X
K pffiffiffi pffiffiffiffiffi
 1=T:uM  T=2:qM :f : zk :ð1  ak :w:ð1 þ rÞÞ1=v k :ðn1k : A þ n2k : ak Þ
k¼1
X
K
 A  h: ak ð6bÞ
k¼1

As Constraints (6c)–(6i) are similar to constraints (5c)–(5i) in MS

model, they have been eliminated from the above formulation.
Due to nonlinear functions involved in the upper and lower level
problems, both MS and DS models are non-linear bi-level program-
ming problem. To tackle these problems, several solution proce-
dures are proposed in the following section.
In general, BLPP can be constructed in two ways. In the opti-
mistic formulation, the leader assumes that the follower will
choose the solution from solution set of the lower-level problem
Fig. 3. General steps of the proposed hierarchical algorithms.
that suits the leader best. In the pessimistic formulation, however,
the leader protects himself from the follower choosing the worst
possible solution. In this paper, we use optimistic approach for Table 2
MS and DS models. Proposed hierarchical algorithms.

ULA LLA
3. Solution procedures
GA PSO
GA GG GP
In consideration of the characteristics of sequential decision
PSO PG PP
making in BLPP, we propose four hierarchical algorithms based
532 O. Amirtaheri et al. / Computers & Industrial Engineering 110 (2017) 527–537

tinuity and convexity. The procedure of GA is a simulation of bio-
logical evolution behavior. It finds the approximate optimal solu-
tion through the processes of encoding, initialization, fitness
evaluation, selection, crossover, mutation and reproduction. The
processes of the GAs developed for the manufacturer and the dis-
tributor problems are separately presented as follows.
3.1.1. Encoding
Any solution of the manufacturer and the distributor problems
can be coded as a chromosome (see Fig. 4). The distributor problem
involves four sets of decision variables out of which xjk and yij can
be computed based on r and ak .
3.1.2. Initialization
This process randomly generates nPop initial feasible solutions.
We assume that population size is proportional to problem size
which is defined by the number of depots (J) and the number of
markets (K). Hence, we set nPop ¼ 10:ðJ þ KÞ. In order to avoid gen-
erating infeasible solutions, two main constraints of both MS and
DS models are considered based on expressions (2), (5c) and (6c):
w:ð1 þ rÞ 6 mink¼1;2;...;K f1=ak g
(Distributor in MS model and Manufacturer in DS model). Simi-
larly, the fitness function in ULA is the profit function of the leader
(Manufacturer in MS model and Distributor in DS model).
3.1.4. Selection
In both of the GAs, developed for the manufacturer and for the
distributor problems, we apply two selection operators. A roulette
wheel (Goldberg, 1989) operator with selection pressure is pro-
posed to choose parent chromosomes for crossover process, while
random selection operator is used for mutation process.
Selection pressure refers to the degree of utility provided by a
good solution. This parameter affects the convergence of GA. If it
is underestimated, it will take a long time for the algorithm to con-
verge. However, if it is set at a too high value, premature conver-
gence will happen and the algorithm will be trapped in local
optima (Sivanandam & Deepa, 2007). Denoting the selection pres-
sure parameter by b (b P 0), the probability of selection for solu-
tion i in roulette wheel operator can be determined as:
X
nPop
b b
PrðiÞ ¼ FitðiÞ = FitðiÞ ; i ¼ 1; 2; . . . ; nPop ð9Þ
ð7Þ i¼1

!
X
K
1=v k
pffiffiffi pffiffiffiffiffi X
J
where nPop is population size and FitðiÞ is fitness value of the i-th
zk :ð1  ak :w:ð1 þ rÞÞ ðn1k : A þ nk : ak Þ 6 min
2
bj ; G solution. From Eq. (9), it is obvious that:
k¼1 j¼1
8
ð8Þ > b¼0
< 1=nPop; if
By analyzing the relationship between the variables with respect to PrðiÞ ¼ 1; if b ¼ 1 and solution i is the best
>
:
the constraints (7) and (8), we propose allowable limits for the deci- 0; if b ¼ 1 and solution i is not the best
sion variables of the manufacturer and the distributor when they ð10Þ
act as leader and as follower (see Table 3). These limits ensure that
the constraints (7) and (8) are fully satisfied when generating ran-
dom solutions in both GAs, developed for the manufacture and
3.1.5. Crossover
the distributor problems.
In each iteration, N c ¼ P c :nPop solutions are selected by roulette
wheel operator, where Pc is crossover rate. We apply arithmetic
3.1.3. Fitness evaluation
crossover operator (Köksoy & Yalcinoz, 2008) to produce off-
In order to evaluate the fitness of the generated solutions, the
springs. Considering X ¼ ðx1 ; x2 ; . . . ; xn Þ and Y ¼ ðy1 ; y2 ; . . . ; yn Þ as
upper-level GA (ULA) heads into the lower-level GA (LLA) and
parent chromosomes, the offsprings are generated as follows:
obtains the optimal values of the follower’s decision variables.
The fitness function in LLA is the profit function of the follower xnew ¼ k:xi þ ð1  kÞ:yi ; i ¼ 1; 2; . . . ; n ð11Þ
i

ynew
i ¼ ð1  kÞ:xi þ k:yi ; i ¼ 1; 2; . . . ; n ð12Þ

where k is randomly selected from [c; 1 þ c] and c is a positive

constant which allows for generating higher or lower values than
that of the parent chromosomes, so as to increase diversification.
Fig. 4. Solution representation. We assume c ¼ 0:1 based on reports in the literature and pilot
testing.

Table 3
Allowable limits for the decision variables.

Decision variables Manufacturer-Stackelberg (MS) model Distributor-Stackelberg (DS) model

Manufacturer
Position: Leader Position: Follower
mink¼1;2;...;K f1=ak g mink¼1;2;...;K f1=ak g
f 6w6 1þdD f :ð1 þ dM Þ 6 w 6 1þr
PJ !2 0 nP o P 12
minf b ;Gg J K 1=v k pffiffiffiffi
j¼1 j min b ;G  z :ð1 k :w:ð1þrÞÞ
a :n2k : ak
06A6 PK 1=v k
; pmax ¼ mink¼1;2;...;K f1=ak g 06A6@
j¼1 j
PK
k¼1 k
A
z :ð1
k¼1 k
ak :pmax Þ :n1k z :ð1 k :w:ð1þrÞÞ1=v k
a :n1k
k¼1 k

0<T61 0<T61
06h61 06h61
Distributor
Position: Follower Position: Leader
mink¼1;2;...;K f1=ak g mink¼1;2;...;K f1=ak g
rL 6 r 6 w 1 06r6 f :ð1þdM Þ
1
06 ak 6 a
U 0 6 ak 6 a
U

⁄ Refer to Appendix A/⁄⁄ Refer to Appendix B/⁄⁄⁄ Refer to Appendix C.

 O. Amirtaheri et al. / Computers & Industrial Engineering 110 (2017) 527–537 533

3.1.6. Mutation Similar to initialization process, we take A ¼ 0 and w ¼ f :ð1 þ dM Þ

Within each iteration, N m ¼ Pm :nPop solutions are selected by for DS model.
random selection operator, where P m is mutation rate. The muta- Elitism approach is applied to form a new generation of solu-
tion operator randomly chooses nm ¼ n:um out of n components tions in both ULA and LLA, i.e. old and new solutions are merged
(genes) in the selected parent chromosome. As a GA parameter, and nPop better solutions are selected as new generation.
um denotes the mutation effectiveness rate (0 6 um 6 1). Furthermore, both ULA and LLA iterate until a maximum num-
The value of i-th component (xi ) is changed through the muta- ber of iterations (MaxIt ) is reached. We set MaxIt ¼ 7:ðK þ 3JÞ. The
tion process and a new value (xnew i ) is assigned randomly from a pseudo codes in Fig. 5 outline the steps of GAs, developed as ULA
Normal distribution function, xnew i  Nðxi ; r2 Þ. Standard deviation and LLA.
(r), called mutation step length, is proportional to the width of
search space: r ¼ l:ðxi  xi Þ where l is a positive constant ranging 3.2. Particle swarm optimization algorithms
from 0 to 1, while xi and  xi are lower and upper limits of the i-th
component, respectively. The longer the mutation step length PSO was inspired by the social behavior of animals such as fish
(r), the higher diversification and lower intensification will be schooling and bird flocking. The algorithm shares many similarities
realized, and vice versa. We found that l ¼ 0:1 in both problems with other evolutionary techniques such as GA. The system is ini-
can lead to best results. tialized with a population of random solutions and search for glo-
Table 4 lists the corresponding parameter setup for the cross- bal optima by updating the solutions from generation to
over and mutation operators in the proposed GAs. generation.
Unlike GA, the PSO algorithm has neither evolutionary operator,
3.1.7. Reproduction such as crossover and mutation, nor selection operator. All the
The new solutions generated in crossover and mutation pro- solutions, called particles, move through the search space by com-
cesses should be rectified before reproduction. Therefore, all deci- bining some aspect of the history of their own current and best
sion variables of the manufacturer (w; A; T; h) and the variable r of locations with those of other particles in the swarm.
the distributor are truncated to their allowed limits. In order to Original framework of PSO algorithm was designated by
correct the ak , one of the markets is randomly chosen and the ak Eberhart and Kennedy (1995). The i-th particle of the swarm is rep-
corresponding to that market is multiplied by the fixed value of resented by a vector, X i ¼ ðxi1 ; xi2 ; . . . ; xiD Þ, which is considered as a
0.99. This procedure continues until the constraint (8) is satisfied. potential solution to a problem in D-dimensional space. Each par-
ticle keeps a memory of its previous best position, Pbest i , and a
Table 4 velocity along each dimension, represented as
Parameter setup for the GAs. V i ¼ ðv i1 ; v i2 ; . . . ; v iD Þ. Position of the particle with best fitness
Parameter Value value in each iteration is designated as Gbest . In PSO algorithm, solu-
Manufacturer problem Distributor problem tions are updated by the following rules:
Pc 0.7 0.6 best
b 1.5 1.5
V new
i ¼ W:V old best
i þ c 1 :r 1 :ðP i  X old
i Þ þ c 2 :r 2 :ðG  X old
i Þ ð13Þ
Pm 0.4 0.3
um 0.5 0.3 X new
i ¼ X old new
i þ Vi ð14Þ

Fig. 5. Pseudo codes for the GAs.

534 O. Amirtaheri et al. / Computers & Industrial Engineering 110 (2017) 527–537
Fig. 6. Pseudo codes for the PSO algorithms.
where W is the inertia weight, c1 and c2 are cognitive and social
learning factors, while r1 and r 2 denote two random numbers uni-
formly distributed in the interval ½0; 1.
In the proposed PSO algorithms, we use constriction coefficient
(v) which has been introduced by Clerc and Kennedy (2002) to
guarantee the algorithm convergence and to avoid particles careen
toward infinity:
2
v¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; / ¼ /1 þ /2
/  2 þ /2  4/
where /1 and /2 are positive numbers, often set to 2.05. By using
the constriction coefficient, the PSO parameters can be defined as:
W ¼ v; c1 ¼ /1 :v; c2 ¼ /2 :v
Inertia weight (W) is an important parameter which significantly
affects the convergence and exploration-exploitation trade-off in
PSO algorithm. We use a damping factor (W damp ) to decrease W
from iteration to iteration (W ¼ W:W damp ; 0 < W damp < 1), so as to
have more exploration at the beginning and higher exploitation at
the end of PSO algorithm execution. Here, we set W damp ¼ 0:99,
which has shown good results in most of the PSO implementations.
Apart from updating rules, the processes of encoding, initializa-
tion, fitness evaluation and reproduction in both developed PSO
algorithms for the manufacturer and for the distributor problems
are similar to those of the GAs, described in previous section. The
outline of the PSO algorithms, proposed as ULA and LLA, are pre-
sented in Fig. 6.
4. Computational experiments
In this section, computational experiments are conducted in
order to compare the performance of the proposed solution proce-
dures. To the best of our knowledge, it is the first time that a decen-
tralized production-distribution supply chain with distinctive
features such as nonlinear demand function and cooperative
advertising is studied and bi-level programming models are pro-
posed to deal with those hierarchical decision making problems.
So, there are not related studies on the same problem and therefore
no data for comparison exist. Hence, we present the results of tests
conducted on real-world data derived from automotive spare parts
ð15Þ supply chain in Iran.
In order to evaluate the ability of the proposed hierarchical
algorithms to find optimal (near optimal) solutions for MS and
DS models, we generate different test problems, arranged in four
classes (Test Problem 1 to Test Problem 4). For each class, five
ð16Þ problem samples are defined based on the number of depots
(J ¼ 2, 3, 4, 5, 6) and the number of markets (K ¼ 3, 5, 7, 10, 15).
Furthermore, the parameters of the manufacturer problem and
the distributor problem in each class are randomly selected from
uniform distributions, defined by the experts in Iranian automotive
spare parts supply chain. Table 5 lists the domains of the parame-
ters corresponding to the manufacturer, the distributor and the
markets. Other parameters such as G, bj , si and qi are directly
dependent on the total induced demand. Therefore, they are sepa-
rately generated for each problem sample.
In order to validate the results of the proposed hierarchical
algorithms (GG, GP, PG, and PP), we adopt an exhaustive grid
search within the domains of the decision variables to find near
optimal solutions for small-size problem instances (J ¼ 2, 3 and
K ¼ 3, 5).
Exhaustive grid search method creates a grid by partitioning the
ranges of decision variables into jointed intervals and evaluates the
fitness function for all possible combinations. The intervals corre-
sponding to the best combination are more divided into subinter-
vals and the process is repeated until the length of the last
 O. Amirtaheri et al. / Computers & Industrial Engineering 110 (2017) 527–537 535

Table 5
Domains of problem parameters.

Manufacturer Distributor Markets

Parameter Domain Parameter Domain Parameter Domain

f [50, 150] cjk [0, 5] zk [20, 100]

uM [2000, 5000] uDj [50, 100] ak [0.002, 0.005]
qM [0.01, 0.03] qDj [0.01, 0.03] vk [0.5, 1.5]
dM 5% dD 5% n1k [0.003, 0.005]
n2k [0.2, 0.5]

subinterval is less than a predefined tolerance. Grid search is com-
putationally expensive and does not guarantee to find global
optima; however, it provides a very good lower bound for the ben-
efit functions. This method has been employed by other research-
ers to validate the results of meta-heuristic algorithms (Sadigh
et al., 2012).
For optimization of the follower’s problem, we use nonlinear
solver of MATLAB optimization tool box in each iteration of the
grid search.
Experiments are performed on Problem Test 1 to Problem Test 4
instances for both MS and DS models. Each experiment is repeated
5 times and the average result is reported. The performance of the
algorithms is measured in terms of the upper-level (leader) objec-
tive function value and the computing time involved. The results
are summarized in Tables 6 and 7. All the algorithms are imple-
mented in MATLAB 8.1.0 (R2013a) and tested on an AMD
2.00 GHz processor with 2.5 GB of main memory.
Experimental results in Table 6 indicate that the PG (PSO-GA)
exceeds the other proposed hierarchical algorithms (GG, GP, and
PP) as well as the grid search algorithm in finding optimal (near
optimal) solution for MS model. It reveals that PSO gives better
performance than GA for the manufacturer problem. It also implies
that PSO outperforms GA for the distributor problem. Considering
the conflict between the profit functions of the manufacturer and
the distributor, we find that poor performance of LLA can lead to
higher profit for the leader. As the manufacturer’s profit in PG
(PSO-GA) is higher than his profit in PP (PSO-PSO), we reach the
conclusion that PSO is more efficient than GA for the distributor
problem. The results for DS model, shown in Table 7, confirm the
above finding. Similar to MS model, PG (PSO-GA) shows better per-
formance for DS model in comparison with other algorithms.
As far as the computational time is concerned, PP (PSO-PSO) is
the most efficient algorithm in both Stackelberg models. However,
PG (PSO-GA) requires a quite reasonable amount of time such that
in the largest instances of MS and DS models, it takes an average of
193 s and 202 s, respectively, which are lower than the mean time
needed by GG (GA-GA) and GP (GA-PSO).
Moreover, Table 8 presents a comparison between the profits of
the manufacturer, the distributor and the system, obtained in MS
and DS models. The following conclusions can be made based on
the results:
(a) The system profits (total profit gained by the members) in
MS and DS models are almost equal.
(b) Each member (manufacturer or distributor) obtains more
profit when playing the leader role in the Stackelberg game
models.
The above findings are compatible with the real world prob-
lems. In a competitive environment, leaders who have priority in
decision-making usually obtain more share of profit from total
profit exist in the system. As shown in Table 8, total profit gained
by the members is almost equal in both Stackelberg Scenarios.
Decision about assigning the leader role to manufacturer or dis-
tributor is directly dependent on the aspect of facing with the
problem. As bi-level programming optimizes the objective function
of the leader (not the follower), the member whose profit is to be
maximized, will take the leader role.

Table 6
Test results for MS model.

Test problem J K Manufacturer’s profit GAP (%) Time (s)

Grid Search GG GP PG PP PG-GG PG-GP PG-PP GG GP PG PP
1 2 3 30,569 30,575 30,575 30,575 30,575 0 0 0 38 22 22 15
3 5 33,207 34,261 34,236 34,326 34,326 0.189 0.263 0 74 58 59 37
4 7 – 39,263 39,219 39,380 39,333 0.299 0.411 0.12 91 71 68 56
5 10 – 42,873 42,827 43,174 43,044 0.702 0.81 0.301 186 163 158 140
6 15 – 51,720 51,561 52,402 51,903 1.319 1.632 0.962 217 207 192 173
2 2 3 51,086 51,185 51,184 51,185 51,185 0 0.001 0 42 27 26 18
3 5 52,118 53,962 53,906 54,069 54,030 0.198 0.303 0.073 69 50 47 35
4 7 – 56,304 56,207 56,479 56,405 0.311 0.484 0.132 92 70 71 59
5 10 – 62,004 61,840 62,389 62,197 0.621 0.888 0.308 181 148 151 149
6 15 – 66,332 65,998 67,014 66,353 1.028 1.539 0.996 215 201 189 182
3 2 3 54,537 54,602 54,602 54,602 54,602 0 0 0 33 25 25 18
3 5 56,206 57,179 57,159 57,179 57,179 0 0.035 0 68 50 53 36
4 7 – 60,382 60,271 60,486 60,451 0.173 0.357 0.058 111 72 70 55
5 10 – 64,312 64,057 64,429 64,367 0.182 0.581 0.096 193 161 158 151
6 15 – 69,483 69,320 70,295 69,735 1.168 1.407 0.803 229 216 208 180
4 2 3 39,701 39,708 39,708 39,708 39,708 0 0 0 41 25 26 13
3 5 43,681 43,746 43,711 43,824 43,824 0.178 0.259 0 75 54 52 32
4 7 – 48,471 48,376 48,572 48,526 0.208 0.405 0.095 89 73 72 56
5 10 – 53,028 52,958 53,346 53,236 0.599 0.733 0.207 161 142 143 133
6 15 – 58,368 58,188 59,127 58,593 1.301 1.613 0.911 224 213 182 176
536 O. Amirtaheri et al. / Computers & Industrial Engineering 110 (2017) 527–537

Table 7
Test results for DS model.

Test problem J K Distributor’s profit GAP (%) Time (s)

Grid Search GG GP PG PP PG-GG PG-GP PG-PP GG GP PG PP
1 2 3 6378 6391 6391 6391 6391 0 0 0 37 30 29 25
3 5 6976 7026 7019 7026 7026 0 0.103 0 69 53 53 46
4 7 – 8352 8343 8367 8364 0.174 0.284 0.038 95 76 78 68
5 10 – 8603 8590 8631 8616 0.329 0.472 0.178 195 174 181 169
6 15 – 10,482 10,470 10,568 10,523 0.82 0.939 0.427 249 201 198 188
2 2 3 9322 9342 9342 9342 9342 0 0 0 40 27 27 22
3 5 11,052 11,382 11,355 11,386 11,386 0.038 0.271 0 71 54 56 45
4 7 – 12,089 12,067 12,127 12,105 0.318 0.495 0.179 101 78 81 69
5 10 – 13,121 13,106 13,199 13,129 0.597 0.707 0.533 189 162 160 153
6 15 – 14,402 14,364 14,563 14,442 1.115 1.388 0.836 264 224 211 181
3 2 3 10,962 11,230 11,229 11,230 11,230 0.002 0.008 0 42 27 26 23
3 5 11,003 11,836 11,809 11,878 11,877 0.359 0.587 0.005 70 54 55 45
4 7 – 12,611 12,582 12,691 12,644 0.637 0.863 0.371 97 82 80 71
5 10 – 13,362 13,332 13,471 13,391 0.819 1.045 0.599 184 160 163 156
6 15 – 14,702 14,665 14,884 14,738 1.237 1.493 0.991 268 219 201 182
4 2 3 7384 7403 7403 7403 7403 0 0 0 46 28 29 26
3 5 8417 8524 8515 8531 8531 0.083 0.186 0 76 55 58 53
4 7 – 9510 9499 9534 9521 0.256 0.371 0.139 89 74 77 70
5 10 – 10,332 10,327 10,391 10,353 0.573 0.619 0.371 205 195 184 173
6 15 – 11,732 11,717 11,848 11,756 0.986 1.116 0.785 271 206 197 185

Table 8
Benefit comparison in MS and DS models.

Test problem J K Manufacturer’s profit Distributor’s profit System’s profit

MS DS MS DS MS DS
1 2 3 30,575 22,350 12,411 20,415 42,986 42,765
3 5 34,326 23,781 15,862 26,341 50,188 50,122
4 7 39,380 27,461 20,215 31,639 59,595 59,100
5 10 43,174 28,306 20,228 35,379 63,402 63,685
6 15 52,402 37,122 23,218 38,408 75,620 75,530
2 2 3 51,185 40,996 11,906 22,365 63,091 63,361
3 5 54,069 45,060 17,900 27,419 71,969 72,479
4 7 56,479 44,059 20,254 33,106 76,733 77,165
5 10 62,389 49,000 23,973 37,010 86,362 86,010
6 15 67,014 57,460 30,640 40,294 97,654 97,754
3 2 3 54,602 46,137 14,142 22,473 68,744 68,610
3 5 57,179 49,176 19,467 27,043 76,646 76,219
4 7 60,486 52,945 26,864 34,225 87,350 87,170
5 10 64,429 57,261 31,175 38,152 95,604 95,413
6 15 70,295 53,691 24,126 40,886 94,421 94,577
4 2 3 39,708 34,905 15,757 20,894 55,465 55,799
3 5 43,824 31,596 13,649 25,939 57,473 57,535
4 7 48,572 39,101 22,077 31,648 70,649 70,749
5 10 53,346 48,716 31,931 36,717 85,277 85,433
6 15 59,127 49,369 30,504 40,117 89,631 89,486

5. Conclusions efficiency and effectiveness of the algorithms. The results demon-

This paper studies a decentralized manufacturer-distributor
supply chain within two Stackelberg models. The MS model
wherein the manufacturer has the leading power of the chain,
and the DS model which allows the distributor to act as leader.
We formulate the models using bi-level programming approach
to find optimal policies for the manufacturer (including wholesale
price, global advertising expenditure, production interval, and par-
ticipate rate in local advertising expenditures) and for the distrib-
utor (including markup percentage, local advertising expenditures,
and the assignment of markets demands to the depots).
Four solution procedures with hierarchical structures are devel-
oped for each Stackelberg model by combining GA and PSO. In
addition, an exhaustive grid search is adopted for validation pur-
poses. Computational experiments are carried out to evaluate the
strate that PSO exceeds GA in finding high quality solutions for
both the manufacturer and the distributor problems.
Comparing the profits obtained by the members, we find out
that the total profits in MS and DS models are almost equal and
the leader always gains more profit than the follower. These results
are clearly compatible with real world problems. However, they
can be compared with cooperative scenario in which the members
are willing to cooperate with each other. Based on the related
works in the literature, it can be predicted that the total profit of
the system will increase in cooperation approach but the impor-
tant ambiguity will be about the members’ profits in cooperation
scenario which will demonstrate if they are willing to cooperate
or not. In order to persuade the members to form cooperation, both
of them should gain higher profit than what they obtain in compe-
tition (Stackelberg) scenario. Therefore, developing incentive
 O. Amirtaheri et al. / Computers & Industrial Engineering 110 (2017) 527–537 537

mechanisms in order to divide surplus profit between members in Ben-Ayed, O., & Blair, C. E. (1990). Computational difficulties of bilevel linear
programming. Operations Research, 38(3), 556–560.
such a way that they become willing to cooperate could be among
Calvete, H. I., Galé, C., & Iranzo, J. A. (2014). Planning of a decentralized distribution
interesting future research directions. network using bilevel optimization. Omega, 49, 30–41.
Calvete, H. I., Galé, C., & Oliveros, M.-J. (2011). Bilevel model for production–
distribution planning solved by using ant colony optimization. Computers &
Appendix A. Calculate the lower limit for r Operations Research, 38(1), 320–327.
Clerc, M., & Kennedy, J. (2002). The particle swarm-explosion, stability, and
In MS model, ULA initializes w and A (as well as T and h) in each convergence in a multidimensional complex space. IEEE Transactions on
Evolutionary Computation, 6(1), 58–73.
iteration and presents their values to LLA for optimizing the fol- Dempe, S. (2002). Foundations of bilevel programming. Nonconvex optimization
lower’s reaction. Therefore, the values of w and A are known when and its applications. Dordrecht: Kluwer.
calculating the lower limit for r. Eberhart, R. C. & Kennedy, J. (1995). A new optimizer using particle swarm theory.
In Proceedings of the sixth international symposium on micro machine and human
We define the function f ðrÞ with respect to Constraint (8) as science. New York, NY.
follows: Gao, Y. et al. (2011). Particle swarm optimization for bi-level pricing problems in
( ) supply chains. Journal of Global Optimization, 51(2), 245–254.
XJ X
K pffiffiffi Goldberg, D. E. (1989). Genetic algorithms in search optimization and machine learning
f ðrÞ ¼ min bj ; G  zk :ð1  ak :w:ð1 þ rÞÞ1=v k :n1k : A (Vol. 412). Addison-wesley Reading Menlo Park.
j¼1 k¼1 Hansen, P., Jaumard, B., & Savard, G. (1992). New branch-and-bound rules for linear
bilevel programming. SIAM Journal on Scientific and Statistical Computing, 13(5),
It can be proved that the first derivative of f ðrÞ is always positive 1194–1217.
except for the upper limit of r defined in Table 3. Hence, the f ðrÞ Khilwani, N. et al. (2008). Fast clonal algorithm. Engineering Applications of Artificial
Intelligence, 21(1), 106–128.
is a strictly increasing function with respect to r: Koh, A. (2007). Solving transportation bi-level programs with differential evolution.
In IEEE congress on evolutionary computation, 2007 (CEC 2007). IEEE.
X
k
ak :w pffiffiffi
:zk :n1k : A:ð1  ak :w:ð1 þ rÞÞ1=v k P 0
0 Koh, A. (2013). A metaheuristic framework for bi-level programming problems with
f ðrÞ ¼
k¼1
vk multi-disciplinary applications. In Metaheuristics for bi-level optimization
(pp. 153–187). Springer.
Köksoy, O., & Yalcinoz, T. (2008). Robust design using Pareto type optimization: A
Applying Newton-Raphson algorithm, the root of the function (~r) is genetic algorithm with arithmetic crossover. Computers & Industrial Engineering,
calculated and the lower limit is determined as below: 55(1), 208–218.
Ma, W., Wang, M., & Zhu, X. (2014). Improved particle swarm optimization based
r L ¼ max f~r ; dD g approach for bilevel programming problem-an application on supply chain
model. International Journal of Machine Learning and Cybernetics, 5(2), 281–292.
Marinakis, Y., & Marinaki, M. (2008). A bilevel genetic algorithm for a real life
Appendix B. Calculate upper limit for ak in MS model location routing problem. International Journal of Logistics: Research and
Applications, 11(1), 49–65.
Mokhlesian, M., & Zegordi, S. H. (2014). Application of multidivisional bi-level
The values of w and A are predefined for LLA in MS model. In programming to coordinate pricing and inventory decisions in a multiproduct
addition, the value for r is generated within its allowed range. In competitive supply chain. The International Journal of Advanced Manufacturing
Technology, 71(9–12), 1975–1989.
order to maintain the Constraint (8), we propose below algorithm Rajesh, J. et al. (2003). A tabu search based approach for solving a class of bilevel
for determining the aU and generating ak (k ¼ 1, 2, . . ., K): programming problems in chemical engineering. Journal of Heuristics, 9(4),
307–319.
Roghanian, E., Sadjadi, S. J., & Aryanezhad, M.-B. (2007). A probabilistic bi-level
Step 1. Calculate f ðrÞ using the equation presented in Appendix linear multi-objective programming problem to supply chain planning. Applied
A. Mathematics and Computation, 188(1), 786–800.
Step 2. Select a market randomly and set its index as k. Ryu, J.-H., Dua, V., & Pistikopoulos, E. N. (2004). A bilevel programming framework
for enterprise-wide process networks under uncertainty. Computers & Chemical
Step 3. Calculate aU as below: Engineering, 28(6), 1121–1129.
Sadigh, A. N., Mozafari, M., & Karimi, B. (2012). Manufacturer–retailer supply chain
!2 coordination: A bi-level programming approach. Advances in Engineering
f ðrÞ
aU ¼ Software, 45(1), 144–152.
zk :ð1  ak :w:ð1 þ rÞÞ1=v k :n2k Sahin, K. H., & Ciric, A. R. (1998). A dual temperature simulated annealing approach
for solving bilevel programming problems. Computers & Chemical Engineering,
Step 4. Generate a random value for ak within the range of 23(1), 11–25.
SeyedEsfahani, M. M., Biazaran, M., & Gharakhani, M. (2011). A game theoretic
[0, aU ]. approach to coordinate pricing and vertical co-op advertising in manufacturer–
Step 5. Update f ðrÞ using below equation and repeat Steps 2–4 retailer supply chains. European Journal of Operational Research, 211(2),
until all ak are assigned. 263–273.
Sivanandam, S., & Deepa, S. (2007). Introduction to genetic algorithms. Springer
pffiffiffiffiffi Science & Business Media.
f ðrÞ ¼ f ðrÞ  zk :ð1  ak :w:ð1 þ rÞÞ1=v k :n2k : ak Swami, S., & Shah, J. (2011). Channel coordination in green supply chain
management: The case of package size and shelf-space allocation. Technology
Operation Management, 2(1), 50–59.
Von Stackelberg, H. (1952). The theory of the market economy. Oxford University
Appendix C. Calculate upper limit for ak in DS model Press.
Xie, J., & Neyret, A. (2009). Co-op advertising and pricing models in manufacturer–
In the ULA of DS model, w and A are not known. Therefore, we retailer supply chains. Computers & Industrial Engineering, 56(4), 1375–1385.
Xie, J., & Wei, J. C. (2009). Coordinating advertising and pricing in a manufacturer–
set these variables to their lower limits (w ¼ f :ð1 þ dM Þ and A ¼ 0) retailer channel. European Journal of Operational Research, 197(2), 785–791.
in order to guarantee the satisfaction of the Constraint (8) and Yue, J. et al. (2006). Coordination of cooperative advertising in a two-level supply
follow the algorithm proposed in Appendix B. chain when manufacturer offers discount. European Journal of Operational
Research, 168(1), 65–85.
Zhang, G., Lu, J., & Gao, Y. (2015). Bi-level pricing and replenishment in supply
References chains. In Multi-level decision making (pp. 325–336). Springer.
Zhang, J., Xie, J., & Chen, B. (2013). Cooperative advertising with bilateral
Aust, G., & Buscher, U. (2014). Cooperative advertising models in supply chain participation. Decision Sciences, 44(1), 193–203.
management: A review. European Journal of Operational Research, 234(1), 1–14.