Scientia Iranica E (2017) 24(6), 3394{3408

Sharif University of Technology

Scientia Iranica
Transactions E: Industrial Engineering
www.scientiairanica.com

Optimizing reliability and cost of system for aggregate

production planning in a supply chain
M. Ramyara , E. Mehdizadehb; and S.M. Hadji Molanaa
a. Department of Industrial Engineering, College of Engineering, Tehran Science and Research Branch, Islamic Azad University,
Tehran, Iran.
b. Faculty of Industrial and Mechanical Engineering, Qazvin Branch, Islamic Azad University, Qazvin, Iran.
Received 25 January 2016; received in revised form 9 September 2016; accepted 3 October 2016

KEYWORDS Abstract. In this paper, the researchers present a multi-objective model for multi-
Aggregate production product, multi-site aggregate production planning model in a supply chain. The goals are to
planning; minimize the total cost of the supply chain, including inventory costs, manufacturing costs,
Supply chain work force costs, hiring and ring costs, and also to maximize the minimum of suppliers'
management; reliability by considering probabilistic lead times to simultaneously improve the system
Reliability; performance. Since the problem is NP-Hard, a Pareto-based multi-objective harmony
Multi-objective search algorithm is proposed. To demonstrate the performance of the presented algorithm,
harmony; a Non-dominated Sorting Genetic Algorithm (NSGA-II) and a Non-dominated Ranking
NSGA-II & NRGA. Genetic Algorithm (NRGA) are applied. The results demonstrate the robustness of the
proposed algorithm to probe the Pareto solutions.

© 2017 Sharif University of Technology. All rights reserved.

1. Introduction between the rms is one of the most important issues

which can be conducted by Supply Chain Management
A supply chain is a set of di erent parties, including (SCM). In other words, SCM should be able to plan
suppliers, manufacturers, people, distribution centers, all the activities involved in producing a product from
transportation channels, etc. in which each rm has suppliers to costumers in a supply chain. Inappropriate
a speci c role in transforming raw materials to end management of a supply chain leads to the bankruptcy
products with regard to costumers' demand. In re- of the members and failure in global competitions.
cent years, improvement in globalization, competition, One of the medium-range planning in the SCM, which
integration, etc. has channelized rms to concentrate includes a time horizon from 3 to 18 months, is Aggre-
on the entire members involved in the supply chain, gate Production Planning (APP) in which aggregated
in addition to their individual production plans. Con- products are considered instead of individual products.
necting all members of a supply chain together and pro- APP, as a technical level planning, attempts to de-
viding production planning based on the relationships termine the optimal quantity of production, inventory
level, workforce, etc. in each period with regard to
some constraints to satisfy the uncertain demand of
*. Corresponding author. Tel/Fax: +98 28 33670051 all products [1]. The result of APP can be used as
E-mail addresses: [email protected] (M. Ramyar);
[email protected] (E. Mehdizadeh); [email protected] a base for other plans such as Capacity Requirements
(S.M. Hadji Molana) Planning (CRP), Master Production Schedule (MPS),
and Material Requirements Planning (MRP) [2]. In
doi: 10.24200/sci.2017.4398 a systematic view of APP, one can introduce capacity
 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408 3395

programming model to deal with APP considering two

con icting objectives simultaneously as well as the
uncertain nature of the supply chain. Mirzapour Al-e-
Hashem et al. [11] introduced a multi-objective model,
including minimizing the total cost, maximizing the
service level, and maximizing workers' productivity,
to deal with a multi-period, multi-product and multi-
site aggregate production planning problem under un-
certainty. Sadeghi et al. [12] considered minimizing
the total costs, carrying and backordering costs, and
rate of changes in workforce level simultaneously by a
multi-objective model for aggregate planning and used
Figure 1. Inputs and outputs of APP process. a goal programming approach to solve the proposed
model. Rodger [13] used real-world data to determine
constraints, demands, and rm's policies and strategies the structure of a Bayesian network. In this research,
as inputs of APP and determine the production levels, stochastic simulation based on Markov blankets was
inventory levels, workforce levels, subcontracting, etc. used to determine the distribution of backorders and
as outputs of the system depicted in Figure 1. customer wait time in supply chains. Fuzzy logic was
Based on a number of the objective function, con- employed to produce a funnel diagram, demonstrating
sidered in the models, the APP models can be classi ed that the Acquisition Advice Code, Acquisition Method
into two following categories: single-objective function Sux Code, Acquisition Method Code, and Controlled
problems and multi-objective function problems. A Inventory Item Code backorder performance metric of
common objective function in the APP models is to a trigger group dimension might change dramatically
minimize the total cost of the system. In addition, with variations in ALT, PLT, unit price, quantity
maximization of service level, minimization of changing ordered, and stock. Rodger et al. [14] reported on
of work force level, and minimization of the risk are the results of fuzzy induced linguistic ordered weighted
other objective functions which can be considered. averaging for group decision support evaluation of
Over the last decades, numerous single-objective backorder risk triggers to ensure that equipment is
APP models have been studied in the research studies. available and fully operational when needed. Risk
Various models have been developed to solve the APP factors were identi ed, the impact importance and
problem since Holt et al. [3] presented HMMS (Holt, probability metric performance ratings were deter-
Modigliani, Muth and Simon) rule. Bergstrom and mined via induced linguistic ordered weighted averag-
Smith [4] generalized the HMMS approach for multi- ing, and a risk mitigation strategy was used to identify
product system. Oh and Karimi [5] developed a multi- and predict Supply Chain Backorder Risk Triggers
product supply chain planning model with a drawback. (SCBORT). Finally, they presented supply chain risk
Guillen et al. [6] discussed a mixed-integer linear as a probability/impact matrix. Ramanathan [15]
programming model for tactical planning of multi- studied several supply chain management initiatives
product supply chains. such as vendor-managed inventory, continuous replen-
All the above mentioned research studies dis- ishment, and Collaborative Planning Forecasting and
cussed a single-objective APP model which tried to Replenishment (CPFR) that had been previously pro-
minimize the total cost, while the other objective func- posed in the literature to improve the performance of
tions can be considered in APP models. In other words, supply chains. In this research, collaborative perfor-
in the practical APP model, con icting objective mance measurement acts as a testing tool to identify
functions can be considered [7]. Wang and Liang [8] a helping environment to collaborate by pinpointing
proposed a multi-objective APP model including mini- areas requiring improvements before initializing col-
mizing the total costs, carrying and backordering costs, laboration. Gholamian et al. [16] proposed a fuzzy
and changing workforce level in a fuzzy environment. multi-objective mixed-integer nonlinear programming
Leung and Chan [9] proposed a multi-objective APP (FMOMINLP) to address a comprehensive multi-site,
model which attempted to maximize pro t, minimize multi-period and multi-product Aggregate Production
repairing costs, and maximize machine utilization re- Planning (APP) problem under uncertainty. They
garding di erent operational constraints. Baykasoglu developed a fuzzy programming method to solve the
and Gocken [10] used a direct solution method based presented model. Finally, the eciency of the model
on ranking methods of fuzzy numbers and Tabu search was investigated by a real industry SC case study.
to solve fuzzy multi-objective aggregate production Pasandideh et al. [17] proposed a bi-objective opti-
planning problem. Mirzapour Al-e-Hashem et al. [1] mization of a multi-product, multi-period three-echelon
proposed a multi-objective mixed integer nonlinear supply-chain-network problem. To bring the problem
3396 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408

closer to reality, the majority of the parameters in this algorithms, a widely used Pareto-based algorithm is
network were considered stochastic. The problem was an extended version of a Genetic Algorithm (GA)
rst formulated into the framework of a single-objective for multi-objective problems, called Non-dominated
stochastic mixed integer linear programming model. Sorting Genetic Algorithm (NSGA-II) proposed by
Then, it was reformulated into a bi-objective deter- Deb et al. [22]. This algorithm can be used in di erent
ministic mixed-integer nonlinear programming model. scopes of operational management. Harmony Search
To solve the complicated problem, a Non-dominated Algorithm (HSA), as a music-inspired algorithm, is
Sorting Genetic Algorithm (NSGA-II) was utilized and simple in concept and has just a few parameters. It
Non-dominated Ranking Genetic Algorithm (NRGA) is easy to be implemented and has been successfully
was used to validate the results obtained. Pasandideh applied to di erent problems including the mechanical
et al. [18] investigated a bi-objective optimization of a structure design [23], pipe network optimization [24],
multi-product multi-period three-echelon supply chain and inventory models [25].
network consisting of manufacturing plants, Distribu- In this paper, a multi-objective model considering
tion Centers (DCs) with uncertain services for each one a multi-period, multi-product, and multi-site aggregate
and customer nodes. The two objectives were the min- production planning problem is developed. The rst
imization of the total cost, while minimization of the objective function is minimizing the sum of the total
average number of products dispatched to customers cost in the supply chain, and the second one, considered
was followed by considering reliability indices assumed as a contribution, includes improving the performance
for the distribution centers. Rooeinfar et al. [19] of the system through maximizing the minimum sup-
proposed a new Supply Chain Network, and two novel pliers' reliability by considering probabilistic lead times
hybrid algorithms were developed to solve the model. not discussed in the literature of multi-objective APP
Each hybrid algorithm integrates the simulation tech- models. With regard to the complexity of the proposed
nique with two metaheuristic algorithms, including the APP model, we present a Multi-Objective Harmony
Genetic Algorithm (GA) and the Simulated Annealing Search Algorithm (MOHSA) and compare its results
Algorithm (SAA), namely HSIM-META. Computa- with those of the Non-dominated Sorting Genetic Algo-
tional results show that the proposed HSIM-META rithm (NSGA-II) and Non-dominated Ranking Genetic
has suitable accuracy and speed for use in real-world Algorithm (NRGA) as the best-developed Pareto-based
applications. Entezaminia et al. [20] proposed a multi- approaches in the literature.
objective, multi-period, multi-product and multi-site The rest of this paper is organized as follows:
Aggregate Production Planning (APP) model in a A mathematical formulation of the proposed multi-
green supply chain considering a Reverse Logistic (RL) objective APP problem is presented in Section 2. The
network. In this model, minimizing the total losses solution procedure and parameter tuning are proposed
and maximizing total environmental scores of products in Sections 3 and 4, respectively. Section 5 provides the
were the objective functions. They demonstrated the result analysis and comparisons. Finally, conclusion
trade-o between the con icting objective functions by and suggestion for future research are presented in
a set of Pareto-optimal solutions as generated by the Section 6.
LP-metrics method.
It is obvious that in a chain with various suppliers, 2. Problem formulation
producing the raw material for manufacturers can be
done in di erent ways. But, delivery time of the suppli- The proposed APP problem in a supply chain can
ers is not precise, and it is possible that delays occur in be described as follows: there are S suppliers, J
shipment of materials from suppliers to manufacturers. manufacturers, and C customers, as shown in Fig-
Hence, considering probabilistic lead time is more ure 2. Each supplier can produce several items with
realistic. This lead time can be a measure for de ning regard to its capacity constraint and transport them
the reliability of suppliers, and consideration of their to the manufacturers. Each manufacturer is able to
reliability can in uence manufacturers' production. In produce several products according to the needed raw
other words, the selection of suppliers that maximizes materials provided by suppliers and subcontracting is
the reliability of the whole system can be considered as allowed for the manufacturer. There is a restricted
another objective function in this paper. available time for manufacturing. For each supplier
Based on the complexity of the APP problems and and manufacturer, the production cost of each item can
the proposed problem associated with non-linear mixed be di erent. The warehouse capacity of manufacturers
integer programming models, using the exact or hard- for raw materials and end products is limited. The
computing methods will be time-consuming, especially produced end products are transported to costumers'
when the problem size increases [21]. This is proof of zones regarding their demands. Based on geographical
why we propose a meta-heuristic algorithm method to distribution of supply chain's members, including sup-
solve the proposed problem. Among multi-objective pliers, manufacturers, and costumers in an area, the
 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408 3397

Figure 2. Supply chain con guration.

transportation time between them can vary. Shortage I1mjt : Inventory holding cost for raw material
can occur for each product. The probabilistic lead m (1; 2; :::; M ) at factory j in period t
time for transporting raw materials from suppliers I2ijt : Inventory holding cost for nished
to manufacturers is considered, and the reliability of product i at factory j in period t
suppliers in delivering materials can be computed based
on this de nition. In other words, a supplier is reliable I3ict : Inventory holding cost for nished
if he can transport the needed raw materials to a product i in costumer's zone c in
manufacturer in a prede ned duration time. period t
This paper models the APP problem as a bi- T1sjt : Transportation cost for supplier s
objective programming with the following objective (1; 2; :::; S ) to factory j in period t
functions: T2ict : Transportation cost from factory j to
1. Minimizing the total cost of supply chain; demand point c in period t
2. Maximizing performance level by maximizing relia- Crsmt : Cost of raw material m provided by
bility in supplier selection process. supplier s in period t
im : Number of units of raw material m
2.1. Notations required for each unit of product i
The parameters in this model are as follows:
t : Fraction of the workforce variation
allowed in period t
Dict : Demand of product i (1; 2; :::; I ) in
demand point c (1; 2; :::; C ) in period t : Productivity of labors (0    1)
(1; 2; :::; T ) T Iqjt : Available regular time (q = 1),
Cqj : Production cost per hour in regular overtime (q = 2), and capacity of
time (q = 1), overtime (q = 2), and subcontracting (q = 3) in terms of time
subcontracting (q = 3) at factory j unit at factory j in period t
(1; 2; :::; J ) P1j : Raw material storage capacity at
Ljt : Manpower cost of labors at factory j factory j
in period t P2j : End product storage capacity at
aij : Production time of product i at factory factory j
j P3c : End product storage capacity in
Fjt : Firing cost of workers at factory j in customer's zone c
period t P4smt : Maximum number of raw material m
Hjt : Hiring cost of workers at factory j in supplier s could be provided in period
period t t
3398 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408

X X
LTsj : Probabilistic lead time for shipping + Fjt XFjt + Hjt XHjt
raw material from supplier s to factory j;t j;t
j X X
'LT : Maximum acceptable lead time of + I1mjt XMmjt + I2ijt XPijt
supplier for being reliable m;j;t i;j;t
LTjc : Lead time required for shipping end X X
products from factory j to demand + I3ict XIict + T1sjt XSsmjt
point c i;c;t s;m;j;t
ict : Shortage cost of product i in customer's X X
zone c in period t + T2ict Y Sijct + ict Bict ; (1)
i;j;c;t i;c;t
TC : Total Cost of supply chai
rsj = P (LTsj < 'LT ): Reliability of supplier s for MaxZ2 = Mins;j;t (rsj zsjt ) ; (2)
providing required raw materials of factory j
X X
XPijt = XPij (t 1) + Xijqt Y Sijct 8i; j; t;
Decision variables q c (3)
X
Xijqt Number of product i produced at XMmjt =XMmj (t 1) + XSsmj (t
factory j using method q in period t LTsj )
s
XLjt Number of workers at factory j in X
period t xijq t : im 8m; j; t; (4)
XFjt Number of workers at factory j red in q;i
period t
XHjt Number of workers at factory j hired XLjt = XLj (t 1) + XHjt XFjt 8j; t; (5)
in period t X
XMmjt Inventory level of raw material m at XIict =XIic(t 1) + Y Sijc(t LTsj ) Dict
j
factory j at the end of period t
XPijt Inventory level of end product i at Bic(t 1) 8i; c; t; (6)
factory j in period t

X
XIict Inventory level of end product i in XLjt  T I1jt + T I2jt  xijqt :aij 8j; t; (7)
customer's zone c in period t i;q2f1;2g
XSsmjt Number of units of raw material m X
shipped from supplier s to factory j xij 3t :aij  T I3jt 8j; t; (8)
Y Sijct Number of units of end product i i
provided by factory j for demand point X
c in period t XMmjt  P1j 8j; t; (9)
zsjt One if supplier s provides at least one m
of the raw materials for factory j in X
period t; 0 otherwise XPijt  P2j 8j; t; (10)
i
Bict Shortage of product i in demand point
c in period t X
XIict  P3c 8c; t; (11)
O One if the total cost of the supply m
chain is violated from a pre-speci ed

value ( ); 0 otherwise. (XFjt + XHjt )  (t 1) XLj (t 1) 8j; t; (12)
2.2. Multi-objective APP model X
The multi-objective APP model can be stated as XSsmjt  P4smt qquad8s; m; t; (13)
follows: j
X
MinZ1 = T C = aij Cqj Xijqt XSsmjt  M:zsjt 8m; s; j; t; (14)
i;j;q;t
X X zsjt 2 f0; 1g ; (15)
+ Crsmt XSsmjt + Ljt XLjt
s;m;j;t j;t XFjt ; XHjt ; XLjt  0 and integer 8j; t; (16)
 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408 3399

Xijqt ;XSjsmt ; XMmjt ; XPijt ; Y Sijct ; Bict ; XFjt ; try to nd the optimal front during the evolution
process, which is achieved in the last iteration of
XHjt ; XLjt  0 8i; j; c; s; m; t: (17) algorithm. This front is expected to have the most
convergence and the highest diversity [22].
In this model, Eq. (1) denotes the rst objective
function aiming to minimize the total cost of the supply 3.1. The MOHSA
chain including production cost of manufacturers and In HSA, the objective function is interpreted as har-
suppliers, workforce hiring and ring costs, inventory mony, and aesthetic estimation of the player helps
costs, transportation costs, and shortage costs. The him/her to nd good state of the harmony. Indeed, in
second objective function (Eq. (2)) attempts to improve this algorithm, the qualitative improvisation process is
performance level through maximizing the minimum turned into quantitative optimization process. There-
of suppliers' reliability. In this function, it should be fore, in this paper, multi-objective version of HSA
considered that the minimization operator is de ned for is developed and presented to the literature of APP.
fs; j jzsjt = 1g. By this consideration, the system tries The details of MOHSA are described in the following
to con rm a balance in supplier selection problems. subsections.
Eqs. (3) and (4) are balance constraints for inventory
of end product and raw material, respectively. Eq. (5) 3.1.1. Solution representation
ensures the workforce level balance. An inventory To code the solutions, we presented a bi-part solution
balance equation of demand point c is considered in representation structure. A 1  S random vector spec-
Eq. (6). Eq. (7) guarantees that for each manu- i es the priority of suppliers for transporting materials
facturer, the sum of regular time and overtime with into the manufacturer, and a 1  J random vector
regard to productivity of workforce limits the available speci es the priority of manufacturers for producing
production time. Eq. (8) is a subcontracting constraint. the products. Figure 3 schematically represents an
Eqs. (9) to (11) limit the raw material, end product in- example of this structure. In this structure, each gene
ventory levels of manufacturers, and costumer zones' to of vectors is a random number between zero and one.
their related inventory storage capacities, respectively. Besides, customers' demands will never exceed the ca-
Eq. (12) limits the change in workforce level by the pacity limitations throughout the supply chain. In this
proportion of workers in the previous period. Eq. (13) gure, S = 4; so, the random numbers are generated as
restricts the amount of shipments from supplier s by 1  4 vector; their positions are kept, and then sorted
the supplier capacity. Eq. (14) ensures that if supplier in an ascending order. Based on our capacity, two of
s provides at least one of the raw materials for factory the rst genes are selected. The positions of these
j in period t, its related binary variable (zsjt ) must be numbers are selected as suppliers (suppliers numbers
one. Eqs. (15)-(17) denote variable types. 3 and 1 are selected based on the corresponding
capacity). Moreover, the continuous decision vari-
ables, including Xijqt ; XSjsmt ; XMmjt ; XPijt ; Y Sijct ;
3. The solution procedures Bict ; XFjt ; XHjt ; and XLjt , are encoded based on
Pareto-based approaches have been widely used in upper bounds and are randomly generated between
research studies, especially in order to solve compli- zero and its upper bound.
cated models. Based on NP-hardness of aggregate To prevent violation of constraints, a penalty
production planning, we present a multi-objective function approach method is applied to penalize
Pareto-based meta-heuristic algorithm called MOHSA them [26]. Penalty values are considered for all of the
and compare it with two based-developed ones in the two-objective functions through an additive function.
literature, NSGA-II and NRGA, which are described 3.1.2. Improvising process
in this section. In an improvisation of the process of a musician, when
A multi-objective problem consists of some con- a musician improvises with an instrument, he or she
icting objectives f (~x) = [f1 (~x); :::; fm (~x)] with regard faces three possible options of:
to some constraints gi (~x)  0, i = 1; 2; :::; c, ~x 2 X
where ~x denotes n-dimensional vectors and X is the (I) Playing from his/her memory (with probability
feasible region. For a minimization model, we say that PHMCR );
solution ~a dominates solution ~b(~a; ~b 2 X ) if:
1. fi (~a)  fi (~b); 8i = 1; 2; :::; m and;
2. 9 i 2 f1; 2; :::; mg : fi (~a) < fi (~b).
There are some solutions that cannot dominate each
other. The set of these solutions is called Pareto
solutions set or Pareto front. Pareto-based algorithms Figure 3. An instance of DCs encoding.
3400 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408

(II) Adjusting the pitches slightly (with probability 3.1.3. Multi-objective operators of the MOHS
Ppa ); While the objective function value was used to rank
(III) Composing randomly (with probability Prand ). the solutions in a single-objective algorithm, the dom-
ination concept was utilized for ranking Pareto-based
These options are formalized into three quantita- multi-objective algorithms. In the NSGA-II algorithm,
tive operators in HSA called harmony memory, pitch as proposed by Deb et al. [22], the Fast Non-Dominated
adjusting, and randomization [27]. Therefore, the Sorting (FNDS) operator was employed for inserting
improvising process of the HSA is the combination of the dominance concept by searching the rst goal called
these three operators. Accordingly, the main steps of convergence. Smaller values of FNDS indicate better
the HSA are explained in the next subsections. To ranks. To search the second goal named diversity,
apply the improvising process in di erent iterations, another operator named Crowing Distance (CD) was
a random solution is selected rst, and then one/two considered in NSGA-II to estimate the density of
operator(s) of the HSA (based on their probabilities) similar rank solutions placed surrounding a particular
is/are used to improvise the selected solution. After solution. Larger values of CD show better solutions
improvising a new solution, the HSM is updated by lying in a less crowded area. Then, a binary tourna-
replacing the worse solution with the new solution. ment selection is performed according to the above two
Interested readers should refer to Geem et al. [24], operators, in which if solutions are from di erent ranks,
Geem [28], and Rahmati et al. [29] for additional the one with a smaller rank is selected. Otherwise, the
information. Figure 4 presents a schematic view of the one with the more value of CD is selected.
relationship between di erent HSA probabilities. The process is initiated by initializing the initial
Besides, to minimize the impact of using di erent population of harmony vectors, Pt . Then, the new
operators on the performance comparison process of operators, including HM, migration, and mutation, are
the algorithms, operators are designed identically. To implemented on Pt to create a new population, Qt .
do so, the pitch adjusting operator of MOHSA is The combination of Pt and Qt creates Rt for keeping
designed similar to the mutation operator of the GA as elitism in the algorithm. In this step, vectors of Rt are
a swap strategy [27,29]. We also represent the Pseudo- sorted in several fronts based on FNDS and CD [22].
code of HSA as well in Figure 5. Using the proposed selection method, population of the
next iteration, Pt+1 , is chosen to have a predetermined
size. Figure 6 demonstrates the evolution process of
the proposed MOHS, schematically.
To demonstrate performance of the proposed
MOHSA, two well-developed Pareto-based multi-
objective evolutionary algorithms, called NSGA-II and
NRGA, are applied, described in the following subsec-
Figure 4. The relationship among di erent probabilities tion.
of HAS.
3.2. The NSGA-II and NRGA
Non-dominated Sorting Genetic Algorithm (NSGA) is
one of the most popular evolutionary algorithms, used
to solve multi-objective problems. This non-dominated
based meta-heuristic algorithm, which is an extension
of the Genetic Algorithm (GA), was proposed by
Srinivas and Deb [30]. The main criticisms of the
NSGA, such as lack of elitism, the need for specifying
the sharing parameter, and high computational com-
plexity of non-dominated sorting, lead to introducing
an extension of NSGA, called NSGA-II, proposed by
Deb et al. [22].
In this paper, we apply an NSGA-II approach to
solve the proposed APP model and compare the results
with those of the presented MOHSA approach. As
another Pareto-based evolutionary algorithm, NRGA
is another multi-objective evolutionary algorithm intro-
duced by Al Jaddan et al. [31] to solve multi-objective
optimization problems. The crossover and mutation
Figure 5. HSA Pseudo-code [29]. operators of NSGA-II and NRGA are uniform crossover
 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408 3401

Figure 6. MOHSA evolution process.

Figure 7. NSGA-II and NRGA owcharts.

and swap operators, respectively [32]. Further, bi- depends on the parameters and operators, Experiments
nary tournament selection and roulette wheel selection designs are used to set the parameters. Experiments
strategies methods are applied in NSGA-II and NRGA, designs is an organized method to determine the
respectively. The owcharts of NSGA-II and NRGA relationship between input and output processes. In
are plotted in Figure 7. this section, parameters and operators of MOHSA are
determined. Due to the large number of parameters
4. Parameter tuning in HSA, nding a suitable composition of parameters
that improves the performance of algorithms is favor-
Since the quality of meta-heuristic algorithms largely able. Since there are many parameters, using complete
3402 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408

factorial method is insucient. To solve this problem, are performed on a two GHz laptop with eight GB
Taguchi method has been used. In this section, to set RAM to estimate the response functions.
the parameters, several large-scale problems have been To evaluate and compare the performances of the
selected and are presented in Figures 8-10 by using
major e ect chart. In the cases where the number
of agents and their levels is high, Taguchi method is
more ecient than complete factorial method. For
orthogonal array, L27 equaling 27 is much less than the
number required for complete factorial method [33,29].
In order to tune the parameters, the Mean Ideal
Distance (MID) is selected as the main response in
Taguchi analysis.
According to the main e ects plots of Figures 8-
10, the optimal value of the algorithm's parameters for
problem no. 10 is reported in Table 1.

5. Result analyses and comparisons

This section provides the application of the proposed
methodology and the performance comparisons of the
three meta-heuristic algorithms. The developed al- Figure 9. Setting the NRGA parameters by Taguchi.
gorithms are coded in MATLAB software (Version
7.10.0.499, R2010a) environment, and the experiments

Figure 8. Setting the MOHSA parameters by Taguchi. Figure 10. Setting the NSGA-II parameters by Taguch.

Table 1. Optimal parameters for the algorithms.

Algorithms Parameters Leveling Factor optimal
Population size (Pop) f20,30,50g 50
Inner Loop (Loop1) f20,30,40g 40
MOHSA Outer Loop (Loop2) f30,60,90g 60
Pitch adjusting operator (Paj) f0.1,0.3,0.5g 0.5
Harmony memory operator (Phcr) f0.7,0.8,0.9g 0.7
Population size (Pop) f25,50,100g 25
NRGA Maximum of Iteration (MaxIT) f100,200,300g 300
Probability of crossover (Pc) f0.7,0.8,0.9g 0.9
Probability of mutation (Pm) f0.1,0.25,0.4g 0.25
Population size (Pop) f25,50,100g 50
NSGA-II Maximum of Iteration (MaxIT) f100,200,300g 300
Probability of crossover (Pc) f0.7,0.8,0.9g 0.8
Probability of mutation (Pm) f0.1,0.25,0.4g 0.25
 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408 3403

Table 2. Generated test problem.

Problem no. I C T J
1 2 2 2 1
2 2 5 2 1
3 3 5 3 2
4 3 10 3 2
5 5 10 6 3
6 5 10 6 3
7 10 20 12 4
8 10 20 12 4
9 10 40 12 5
10 20 40 18 5
11 20 50 24 10
12 35 50 24 10
Figure 11. Box-plot comparisons of the algorithms in
Table 3. Input Parameters' values of the numerical terms of NOS metric.
examples.
Parameters and their values
Dict  U (100; 500) T2ict  U (10; 30)
Cqj  U (10; 40) Crsmt  U (80; 120)
Ljt  U (1000; 3000) im  U (3; 9)
aij  U (10; 20) t  U (0; 1)
Fjt  U (2000; 4000) T I1;jt  U (7; 9)
Hjt  U (6000; 10000) T I2;jt  U (0; 4)
I1mjt  U (100; 200) P1j  U (1000; 2000)
I2ijt  U (150; 250) P2j  U (500; 1000)
I3ict  U (150; 250) P3c  U (800; 1400)
n  U (100; 400)
ict P4smt  U (1000; 5000)
LTsj ; LTjc  Uniform(0; 4) 'LT = 1
T1sjt  U (10; 30)

solution methodologies under di erent environments, Figure 12. Box-plot comparisons of the algorithms in
the experiments are implemented on 12 problems, terms of MID metric.
which are reported in Table 2. These problems are
categorized based on the number of costumers (I ), the The results of comparisons in terms of all multi-
number of facilities (J ), and the maximum number of objective metrics for all algorithms are reported in
on-duty servers (P ). Moreover, the following informa- Table 6. Moreover, the algorithms are compared based
tion is also given in Table 3. on the properties of their obtained solutions. For
Then, these instance problems are solved by three these cases, all metrics are also plotted and graphically
algorithms. Furthermore, to eliminate uncertainties compared in Figures 11-15.
of the solutions obtained, each problem is used three We note that while bigger values are desired
times under di erent random environments. Then, the in terms of the diversity and NOS metrics, smaller
averages of these three runs are treated as the ultimate values are better for spacing, MID, and CPU time.
responses. The objective function values of problems Thus, according to the analysis of variance outputs
nos. 8 and 11 are reported in Tables 4 and 5. in Tables 7-11 and p-values, MOHSA shows better
In order to evaluate the performances of the three performance in terms of CPU Time. However, in terms
multi-objective meta-heuristic algorithms, four metrics of MID, Diversity, and Spacing metrics, the algorithms
are used as Zitzler and Thiele [34]: almost work in a similar way. This conclusion is
(I) Number of Pareto Solution (NOS); con rmed at a 95% con dence level.
(II) Mean Ideal Distance (MID);
6. Conclusion and future research
(III) Spacing;
A multi-objective model for multi-product, multi-site
(IV) Diversity; aggregate production planning model in a supply chain
(V) Computational time. was developed in this paper. The goals were to
3404 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408

Table 4. Pareto solutions of problem no. 8.

Pareto solution NSGA-II NRGA Proposed MOHSA
number
Objective 1 Objective 2 Objective 1 Objective 2 Objective 1 Objective 2
1 6.53E+07 0.5086 5.09E+07 0.3450 1.39E+07 0.3501
2 8.01E+07 0.6715 4.59E+07 0.2567 5.27E+07 0.6990
3 3.50E+07 0.1481 4.39E+07 0.2244 9.19E+07 0.9344
4 7.61E+07 0.6324 9.54E+07 0.8702 8.02E+07 0.8890
5 9.84E+07 0.8344 7.47E+07 0.6397 6.65E+07 0.8082
6 5.60E+07 0.4225 8.79E+07 0.7854 6.01E+07 0.7516
7 9.35E+07 0.8043 6.88E+07 0.5773 | |
8 4.92E+07 0.3671 7.78E+07 0.6733 | |
9 8.88E+07 0.7673 5.61E+07 0.4467 | |
10 4.00E+07 0.2109 9.16E+07 0.8267 | |
11 8.43E+07 0.7230 3.91E+07 0.1284 | |
12 7.23E+07 0.5756 6.34E+07 0.5210 | |
13 1.13E+08 0.8911 8.44E+07 0.7461 { |
14 5.89E+07 0.4494 1.12E+08 0.8988 | |
15 4.45E+07 0.3033 5.84E+07 0.4702 | |
16 | | 7.17E+07 0.6077 | |

Table 5. Pareto solutions of problem no. 11.

Pareto solution NSGA-II NRGA Proposed MOHSA
number
Objective 1 Objective 2 Objective 1 Objective 2 Objective 1 Objective 2
1 5.96E+10 0.5760 9.05E+10 0.8402 3.67E+10 0.3831
2 9.05E+10 0.8280 7.66E+10 0.7583 6.90E+10 0.5640
3 3.33E+10 0.3466 3.06E+10 0.4313 1.17E+11 0.8558
4 1.01E+11 0.9133 7.05E+10 0.7204 1.30E+11 0.9302
5 9.83E+10 0.8903 1.11E+11 0.9140 6.21E+10 0.5189
6 2.23E+10 0.1907 3.01E+10 0.4023 1.35E+11 0.9473
7 3.62E+10 0.3727 5.97E+10 0.6502 2.59E+10 0.2774
8 7.66E+10 0.7161 8.33E+10 0.7982 8.52E+10 0.6664
9 5.49E+10 0.5357 6.48E+10 0.6844 9.47E+10 0.7243
10 7.05E+10 0.6660 5.05E+10 0.5868 | |
11 1.21E+11 0.9394 1.15E+11 0.9177 | |
12 4.27E+10 0.4309 3.93E+10 0.5031 | |
13 1.82E+10 0.1398 2.17E+10 0.3496 | |
14 5.05E+10 0.4982 1.33E+10 0.2540 | |
15 3.06E+10 0.3223 3.62E+10 0.4779 | |
16 8.32E+10 0.7700 4.64E+10 0.5574 | |
17 | | 9.84E+10 0.8845 | |
 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408 3405

Table 6. Computational results of multi-objective metrics comparisons for three algorithms.

Problem Proposed MOHSA NSGA-II NRGA
no.
NOS MID Spacing Diversity Time NOS MID Spacing Diversity Time NOS MID Spacing Diversity Time
1 9 6E+06 7690 82465023 11.52 23 7E+06 7633 836454875 21.66 21 6E+07 7655 648729584 21.99
2 9 6E+06 8766 876348278 15.44 21 9E+06 8655 847638748 38.91 19 1E+07 6544 772354726 22.73
3 11 8E+07 7434 82736292 16.81 16 7E+07 5688 834768378 45.88 18 9E+07 9853 827368742 29.91
4 13 8E+07 4584 73492284 21.53 17 8E+07 8721 483784734 76.10 15 1E+08 9790 67485282 49.21
5 8 1E+08 18712 873684277 22.73 21 6E+07 87654 84384783 91.62 22 1E+08 87341 3642347 82.61
6 5 2E+08 87645 784268392 26.72 21 2E+08 77611 684348384 110.81 21 6E+08 98652 46234827 121.81
7 7 2E+08 89966 374920923 41.94 15 3E+08 54388 483784783 119.85 25 9E+08 67390 6348232 129.82
8 6 4E+08 123560 374928902 49.66 15 5E+08 876691 83847384 159.75 16 1E+09 897443 47835404 148.73
9 5 5E+08 6588799 374928980 61.87 17 8E+08 8544975 84384783 178.83 18 8E+08 8756532 65354504 171.91
10 9 8E+08 78767221 238476899 77.99 16 9E+08 87655501 87589979 233.70 17 8E+08 87996209 67374363 201.74
11 4 1E+09 984989893 562437223 121.10 14 1E+09 698212180 23720645 298.61 15 2E+09 678524489 87373583 278.53
12 3 3E+09 6798567589 647289022 189.81 14 4E+09 8974461871 837843878 412.87 14 1E+09 732434588 54353783 362.74

Figure 15. Box-plot comparisons of the algorithms in

Figure 13. Box-plotcomparisons of the algorithms in terms of Time metric.
terms of Spacing metric.

Table 7. Statistical outputs of NOS metric.

Algorithm DF SS MS F P -value
Algorithms 2 894.6 447.3 45.14 0.000
Error 33 326.83 9.90
Total 35 1220.89

times to simultaneously improve performance of the

system. Since the problem is NP-Hard, a Pareto-based
multi-objective harmony search algorithm is proposed.
To demonstrate the performance of the presented algo-
rithm, NSGA-II and NRGA are applied. The results
Figure 14. Box-plot comparisons of the algorithms in show that MOHSA has better performance in terms
terms of Diversity metric. of CPU Time. However, in terms of MID, Diversity,
and Spacing metrics, our proposed MOHSA has the
minimize the total cost of supply chain, including same capability as the best-developed algorithms in
inventory costs, manufacturing costs, work force costs, the literature. As for future research, one can develop
hiring, and ring costs, and to maximize the minimum the model by formulating the problem in competitive
of suppliers' reliability by considering probabilistic lead supply chains.
3406 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408

Table 8. Statistical outputs of MID metric.

Algorithm DF SS MS F P -value
Algorithms 2 1.43865E+17 7.19323E+16 0.09 0.910
Error 33 2.49970E+19 7.57484E+17
Total 35 2.51408E+19

Table 9. Statistical outputs of Spacing metric.

Algorithm DF SS MS F P -value
Algorithms 2 3.11980E+18 1.55990E+18 0.44 0.645
Error 33 1.15933E+20 3.51313E+18
Total 35 1.19053E+20

Table 10. Statistical outputs of Diversity metric.

Algorithm DF SS MS F P -value
Algorithms 2 3.94541E+17 1.97270E+17 1.86 0.172
Error 33 3.50060E+18 1.06079E+17
Total 35 3.89514E+18

Table 11. Statistical outputs of CPU Time metric.

and nancial risk as a key element for short term
Algorithm DF SS MS F P -value scheduling optimization", Ind. Eng. Chem. Res., 44,
Algorithms 2 62182 31091 3.35 0.047 pp. 557-575 (2005).
Error 33 306463 9287 7. Wang, R.C. and Liang T.F. \Application of fuzzy
multi-objective linear programming to aggregate pro-
Total 35 368644 duction planning", Computers & Industrial Engineer-
ing, 46(1), pp. 17-41 (2004).
References 8. Wang, R.C. and Liang, T.T. \Aggregate production
1. MirzapourAl-e-Hashem, S.M.J., Malekly, H. and planning with multiple fuzzy goals", International
Aryanezhad, M.B. \A multi-objective robust optimiza- Journal of Advance Manufacturing Technology, 25, pp.
tion model for multi-product multi-site aggregate pro- 589-597 (2005).
duction planning in a supply chain under uncertainty", 9. Leung, S.C.H. and Chan, S.S.W. \A goal programming
In. J. Production Economics, 134, pp. 28-42 (2011). model for aggregate production planning with resource
2. Ozdamar, L., Bozyel, M.A. and Birbil S. \LA hierar- utilization constraint", Computers & Industrial Engi-
chical decision support system for production planning neering, 56, pp. 1053-1064 (2009).
(with case study)", European Journal of Operational 10. Baykasoglu, A. and Gocken, T. \Multi-objective ag-
Research, 104, pp. 403-422 (1998). gregate production planning with fuzzy parameters",
3. Holt, C., Modigliani, F. and Simon, H. \A linear deci- Advances in Engineering Software, 41, pp. 1124-1131
sion rule for production and employment scheduling", (2010).
Management Science, 2(1), pp. 1-30 (1955). 11. MirzapourAl-e-hashem, S.M.J., Aryanezhad, M.B. and
4. Bergstrom, G. and Smith, B. \Multi-item production Sadjadi S.J. \An ecient algorithm to solve a multi-
planning-an extension of the HMMS rules", Manage- objective robust aggregate production planning in
ment Science, 16(10), pp. 614-629 (1970). an uncertain environment", International Journal of
5. Oh, H.C. and Karimi, I.A. \Global multiproduct pro- Advance Manufacturing Technology, 58, pp. 765-782
duction distribution planning with duty drawbacks", (2012).
AIChE. J., 50, pp. 963-989 (2004). 12. Sadeghi, M., Hajiagha, S.H.R. and Hashemi S.S. \A
6. Guillen, G., Bagajewicz, M., Sequeira, S.E., Espuna, fuzzy grey goal programming approach for aggregate
A. and Puigjaner, L. \Management of pricing policies production planning", International Journal of Ad-
 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408 3407

vance Manufacturing Technology, 64, pp. 1715-1727 design", International Journal of Modeling and Simu-
(2013). lation, 22, pp. 125-133 (2002).
13. Rodger, J.A. \Application of a fuzzy feasibility 25. Taleizadeh, A.A., Niaki, S.T.A. and Barzinpour, F.
Bayesian probabilistic estimation of supply chain back- \Multiple-buyer multiple-vendor multi-product multi-
orderaging, un lled backorders, and customer wait constraint supply chain problem with stochastic de-
time using stochastic simulation with Markov blan- mand and variable lead-time: A harmony search
kets", Expert Systems with Applications, 41, pp. 7005- algorithm", Applied Mathematics and Computation,
7022 (2014). 217, pp. 9234-9253 (2011).
14. Rodger, J.A., Pankaj, P. and Gonzalez, S.P. \Deci- 26. Yeniay, O. and Ankare, B. \Penalty function methods
sion making using a fuzzy induced linguistic ordered for constrained optimization with genetic algorithms",
weighted averaging approach for evaluating risk in a Mathematical and Computational Application, 10, pp.
supply chain", International Journal Advanced Manu- 45-56 (2005).
facturing Technology, 70, pp. 711-723 (2014). 27. Geem, Z.W., Kim, J.-H. and Loganathan, G.V. \A new
heuristic optimization algorithm: harmony search",
15. Ramanathan, U. \Performance of supply chain collab- Simulation, 76(2), pp. 60-68 (2001).
oration - A simulation study", Expert Systems with
Applications, 41, pp. 210-220 (2014). 28. Geem, Z.W. \Harmony search algorithm for solving
Sudoku", In Knowledge-Based Intelligent Information
16. Gholamian, N., Mahdavi, I., Tavakkoli-Moghaddam, and Engineering Systems, B. Apolloni, R.J. Howlett
R. and Nezam Mahdavi-Amiri, N. \Comprehensive and L. Jain, Eds., KES 2007, Part I. LNCS (LNAI),
fuzzy multi-objective multi-product multi-site aggre- 4692, pp. 371-378, Springer, Heidelberg (2007).
gate production planning decisions in a supply chain
under uncertainty", Applied Soft Computing, 37, pp. 29. Rahmati, S.H.A., Hajipour, V. and Niaki, S.T.A. \A
585-607 (2015). soft-computing Pareto-based meta-heuristic algorithm
for a multi-objective multi-server facility location prob-
17. Pasandideh, S.H.R., Akhavan Niaki, S.T. and Asadi, lem", Applied Soft Computing, 13(4), pp. 1728-1740
K. \Bi-objective optimization of a multi-product (2013).
multi-period three-echelon supply chain problem under 30. Srinivas, N. and Deb, K. \Multi-objective function
uncertain environments NSGA-II and NRG", Informa- optimization using non-dominated sorting genetic al-
tion Sciences, 292, pp. 57-74 (2015). gorithms", E Vol. Comput, 2(3), pp. 221-248 (1995).
18. Pasandideh, S.H.R., Akhavan Niaki, S.T. and Asadi, 31. Al Jaddan, O., Rajamani, L. and Rao, C.R. \No domi-
K. \Optimizing a bi-objective multi-period three ech- nated ranked genetic algorithm for solving constrained
elon supply chain network with warehouse reliability", multi-objective optimization problems", Journal of
Expert Systems with Applications, 42, pp. 2615-2623 Theoretical and Applied Information Technology, 5,
(2015). pp. 640-651 (2009).
19. Rooeinfar, R., Azimi, P. and Pourvaziri, H. \Multi- 32. Haupt, R.L. and Haupt, S.E., Practical Genetic Algo-
echelon supply chain network modelling and optimiza- rithms, 2nd Ed., John Wiley & Sons (2004).
tion via simulation and metaheuristic algorithms",
Scientia Iranica E, 23(1), pp. 330-347 (2016). 33. Chambari, A., Rahmati, S.H.R., Naja , A.A. and
Karimi, A. \A bi-objective model to optimize relia-
20. Entezaminia, A., Heydari, M. and Rahmani, D. \A bility and cost of system with a choice of redundancy
multi-objective model for multi-product multi-site ag- strategies", Computers & Industrial Engineering, 63,
gregate production planning in a green supply chain: pp. 109-119 (2012).
Considering collection and recycling centers", Journal 34. Zitzler, E. and Thiele, L. \Multi-objective optimiza-
of Manufacturing Systems, 40(1), pp. 63-75 (2016). tion using evolutionary algorithms: A comparative
21. Partovi, F. and Seifbarghy, M. \Service centers lo- case study", In A.E. Eiben, T. Back, M. Schoenauer
cation problem considering service diversity within and H.P. Schwefel, Eds., Fifth International Confer-
queuing framework", Scientia Iranica E, 22(3), pp. ence on Parallel Problem Solving from Nature (PPSN-
1103-1116 (2015). V), PP. 292-301, Berlin, Germany (1998).
22. Deb, K., Pratap, A., Agarwal, S. and Meyarivan, T.
\A fast and elitist multiobjective genetic algorithm: Biographies
NSGA-II", IEEE Transactions on Evolutionary Com-
putation, 6, pp. 182-197 (2002). Mohammad Ramyar is currently a PhD student at
College of Industrial Engineering, Iran University of
23. Lee, K.S. and Geem, Z.W. \A new structural optimiza- Science and research in Tehran. He obtained his MSc
tion method based on the harmony search algorithm", degree in Industrial Engineering from the Advanced
Computers and Structures, 82, pp. 781-798 (2004). Institute of Education & Research in Mangement &
24. Geem, Z.W., Kim, J.H. and Loganathan, G.V. \Har- Planning in Iran (2005) and his BSc degree in Applied
mony search optimization: application to pipe network Mathematics from the University Shiraz in Iran (2000).
3408 M. Ramyar et al./Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 3394{3408

His research interests include risk analysis and reliabil- production scheduling, fuzzy sets, and meta-heuristic
ity in supply chain management, aggregate production algorithms. He has several papers in journals and
planning and fuzzy sets. conference proceedings. Also, he is a Managing Editor
of International Journal of Optimization in Industrial
Esmaeil Mehdizadeh is currently an Associate Pro- Engineering.
fessor at the Department of Industrial engineering,
Islamic Azad University, Qazvin Branch, Iran. He Seyyed Mohammad Hadji Molana graduated with
received his PhD degree from Islamic Azad University, BSC from Sharif University of Technology and with
Science and Research Branch, Tehran in 2009, MSc MSC and PhD from AmirKabir University of Technol-
degree from Islamic Azad University, South Tehran ogy in Iran. He is the Head of Industrial Engineering
Branch in 1999, and BSc degree from Islamic Azad Department at Science and Research Branch of Islamic
University, Qazvin Branch in 1996 all in Industrial Azad University. His research interests are in supply
Engineering. His research interests are in the areas chain management, advanced inventory control and
of operation research, such as production planning, production planning.