You are on page 1of 12

Computers & Industrial Engineering 88 (2015) 366377

Contents lists available at ScienceDirect

Computers & Industrial Engineering


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

A possibilistic environment based particle swarm optimization


for aggregate production planning q
Ripon K. Chakrabortty a,, Md.A. Akhtar Hasin b, Ruhul A. Sarker a, Daryl L. Essam a
a
b

School of Engineering and Information Technology, University of New South Wales, Canberra, Australia
Department of Industrial & Production Engineering, Bangladesh University of Engineering & Technology, Bangladesh

a r t i c l e

i n f o

Article history:
Received 22 October 2014
Received in revised form 10 June 2015
Accepted 28 July 2015
Available online 4 August 2015
Keywords:
Particle swarm optimization
Possibilistic environment
Aggregate production planning

a b s t r a c t
Development of an Aggregate Production Plan (APP), the top most level in a hierarchical production
planning system, is a difcult task, especially when input and other production planning parameters
are uncertain because of their inherent impreciseness. This therefore makes generation of a master
production schedule highly complex. Regarding this point, in this paper, we present a scheme of a
multi-period and multi-product APP which is formulated as an integer linear programming model. The
proposed approach uses a triangular possibility distribution for handling all the imprecise operating
costs, demands, and also for the capacity data. The proposed approach uses the strategy of simultaneously minimizing the most possible value of the imprecise total costs, maximizing the possibility of
obtaining lower total costs and minimizing the risk of obtaining higher total costs. A modied variant
of a possibilistic environment based particle swarm optimization (PE-PSO) approach is used to solve
the APP model. A numerical model for demonstrating the feasibility of the proposed model is also carried
out. In the computational study, the considered case study data were experimented with and analyzed to
evaluate the performance of the PE-PSO over both a standard genetic algorithm (GA) and a fuzzy based
genetic algorithm (FBGA). The experimental results demonstrate that the PE-PSO variant provides better
qualities in the aspects of its accuracy when compared to the other two algorithms.
2015 Elsevier Ltd. All rights reserved.

1. Introduction
Aggregate Production Planning (APP) is a leading technique
which aggregates all the production information and which determines the best way to meet forecasted demand by utilizing the
available physical resources. Since it is usually impossible to consider every ne detail associated with the production process
while maintaining a long planning horizon, it is mandatory to
aggregate the information being processed. For that, APP is the
one among some other approaches such as Master Production
Scheduling (Elloumi & Fortemps, 2010), Capacity Requirement
Planning (CRP) and Material Requirement Planning (MRP). APP is
a medium term capacity planning technique which determines
the optimum level of production, workforce, inventory, subcontracting and backlog over a specic time horizon that ranges from
2 to 12, or even 18 months to satisfy uctuating demand requirements with limited capacity and resources (Al-e, Aryanezhad, &

This manuscript was processed by Area Editor Mitsuo Gen.

Corresponding author.

E-mail addresses: ripon.chakrabortty@student.adfa.edu.au,


yahoo.com (R.K. Chakrabortty).
http://dx.doi.org/10.1016/j.cie.2015.07.021
0360-8352/ 2015 Elsevier Ltd. All rights reserved.

ripon_ipebuet@

Sadjadi, 2012; Wang & Yeh, 2014). An efcient and successful


APP can even help to maintain supply chain effectiveness and
responsiveness. A Multi-criteria APP encompasses several costs,
for example: inventory costs, product purchasing costs, manufacturing costs, manpower maintain costs, extra subcontracting or
backordering costs. Since the ultimate output of aggregate
planning is to prepare a master production plan, any discrepancies
that occur in the planning stage can make large impacts on future
master scheduling. So proper planning is mandatory.
Numerous APP models and solutions with various degrees of
sophistication have been introduced since early 1950. There have
been a large number of research techniques that range from simple
graphical methods to more sophisticated search approaches. For
example the switching heuristic and other dynamic methods that
are considered in the review paper of Nam and Logendran
(1992). Also in the real manufacturing world, the practice of allowing mass customization is increasing, and the relevant operational
data is becoming more imprecise and fuzzy in nature. As a consequence, the difculty of handling uncertainties while designing a
successful APP plan is becoming more ambiguous and complex.
Al-e et al. (2012) have recently categorized earlier research
which considered uncertainty into four primary approaches:

R.K. Chakrabortty et al. / Computers & Industrial Engineering 88 (2015) 366377

(1) stochastic programming approach (Kazemi Zanjani, Ait-Kadi, &


Nourelfath, 2013; Mirzapour Al-e-hashem, Baboli, & Sazvar, 2013),
(2) fuzzy programming approach (Iris & Cevikcan, 2014; Liang,
Cheng, Chen, & Shen, 2011), (3) stochastic dynamic programming
approach (Kogut & Kulatilaka, 1994) and (4) robust optimization
approach (Al-e et al., 2012). The stochastic approach has been
applied to study the imprecise nature of decision models since
the late 1950s. In it, the input data are given probability function
distributions. However, with these stochastic approaches it is
impossible to translate the imprecise data from membership to
non-membership (Hsieh & Wu, 2000). On the other hand, when
historical data are not available to estimate a probability distribution, we have to invite some domain experts to evaluate their
degree of belief that each event will occur, which are vulnerable
to such experts giving larger variance than the real frequency.
Therefore, probability theory based approaches are inappropriate
to deal with real world APP problems (Ning, Liu, & Yan, 2013).
Among all the existing APP literatures, the mostly used
approach is the fuzzy programming approach. Gen, Tsujimura,
and Ida (1992) presented a method for solving multi-objective
APP problems with fuzzy parameters. They proposed an efcient
method that transforms a fuzzy multi-objective linear programming (MOLP) problem into a crisp MOLP model and an interactive
solution procedure that suggested the best compromise aggregate
production plan for fuzzy multi-period multi-objective APP
problems. Wang and Fang (2001) presented a novel fuzzy linear
programming method for solving the APP problem while considering fuzzy market demands and the ability to subcontract. Wang
and Liang (2004) developed a fuzzy multi-objective linear programming model for solving the multi-product APP problem in a
fuzzy environment. The handled problem had three objectives,
minimizing total production costs, carrying and backordering
costs, and minimizing the rate of change in labor levels.
Piecewise linear membership functions were specied for fuzzy
goals, while the other parameters of the problem were assumed
to be certain. The fuzzy multi-objective linear programming problem is converted into its crisp equivalent by using Zimmermanns
approach. Later on, Wang and Liang (2005) developed a novel
interactive possibilistic linear programming (PLP) approach for
solving multi-product APP decision problems with imprecise
forecast demand, related operating costs, and capacity. The
strategies of minimizing the most possible value of the imprecise
total costs, while maximizing the possibility of obtaining lower
total costs, and minimizing the risk of obtaining higher total costs,
were simultaneously used for developing the model.
For solving the APP problems, there exists many other methods,
such as the fuzzy grey goal programming approach (Sadeghi,
Hajiagha, & Hashemi, 2013), the multi-objective tabu search
method (Baykasoglu & Gocken, 2010; Baykasoglu & Gken,
2006), the hybrid of genetic algorithm and tabu search
(Ramezanian, Rahmani, & Barzinpour, 2012), the genetic algorithm
(Al-e et al., 2012), the linear relaxation approach (Zhang, Zhang,
Xiao, & Kaku, 2012), meta-heuristic approaches (Baykasoglu &
Gocken, 2010; Raa, Dullaert, & Aghezzaf, 2013), and possibilistic
linear programming approaches (Hsieh & Wu, 2000; Liang, 2007;
Sakall, Bayko, & Birgren, 2010; Vasant, Barsoum, &
Bhattacharya, 2008). Possibilistic linear programming models possess many advantages over other stochastic linear programming
models with respect to computational efciency and exible doctrines (Liang, 2007; Vasant et al., 2008). Distinguishing among
probability, possibility and fuzzy sets has always been a difcult
task. Consequently, a quantitative comparison of those measures
in terms of both efciency and expressiveness is not given except
in the earlier research work from Drakopoulos (1995). In his

367

research article, he claimed that possibilistic measures are the


most efcient of all to compute and have clear computational
advantages over innite domains.
Since large-scale problems cannot be tackled with mathematical programming solvers, meta-heuristics are increasingly used in
the literature to solve such problems (Raa et al., 2013). Among
the different meta-heuristic approaches, genetic algorithms (GA)
and particle swarm optimization (PSO) are the most common.
There are many successful applications in various domain problems, among them PSO has shown that it is a considerably promising, efcient and robust technique for practical applications.
Although there are many applications that used PSO for various
elds, it is however rare to nd any paper in the APP eld due to
its initial formulations as it was mainly developed for the
unconstrained and continuous cases. Recently, Wang and Yeh
(2014) proposed a modied PSO (MSPO) algorithm for the APP
problem which can alleviate the deciencies of traditional PSOs.
In their work, they experimented and analyzed their MPSO algorithm with GA and showed that the MPSO gained more accuracy,
reliability and convergence speed than the traditional PSO and
GA. Hybridizing PSO with other algorithms, such as by combining
fuzzy PSO and fuzzy GA to get better results, was focused on by
Chakrabortty, Sarker, and Essam (2014) which enlightened future
researchers to the possibilities of hybrid PSO algorithms, although
it was not for the APP problems. But these presented methods generally concentrate on the solution algorithm but not on a general
model. On the other hand, the consideration of all of the parameters in an APP model makes it more difcult. So researchers have
not presented a comprehensive and general model to formulate
real production environments. The majority of models in the APP
are relevant to single product and single stage systems and they
are not compatible to real production systems (Ramezanian
et al., 2012).
Therefore, a new model that considers the impreciseness of
operational coefcients is proposed in this paper. The proposed
multi-period and multi-product APP model was organized and formulated as an integer linear programming model which was later
solved by our modied PSO variant under possibilistic environments. All the forecasted demands, related imprecise operating
costs and capacity data were covered and tackled with a triangular
possibility distribution function. A novel interactive approach for
handing an imprecise objective function and technical constraints
has been developed and analyzed. The proposed approach
attempts to minimize total costs in terms of inventory levels, labor
levels, overtime, subcontracting and backordering levels, and labor,
machine and warehouse capacity. Due to imperfections in the traditional PSO algorithm for solving constrained problems, we have
used an important variant of PSO, linear reduction of inertia
weight, for solving this APP model. This unique way of representing PSO under a possibilistic environment, while optimizing and
analyzing the model with that PE-PSO, are the contributions of this
paper.
The structure of the paper is as follows: in Section 2, we cover
some important information on the basic PSO algorithm and its different variants. The basic APP problem structure and its associate
assumptions are discussed in Section 3. The terminologies, basic
APP model, its integer linear programming formulations, the relevant operational constraints, the ways of handling imprecise data
and the overall possibilistic APP model are described in
Section 4. In Section 5, solution approaches are discussed. The
experimental studies, along with the comparative results of tests
that measure the performance of different algorithms are
described in Section 6. In the nal section, we provide overall
conclusions.

368

R.K. Chakrabortty et al. / Computers & Industrial Engineering 88 (2015) 366377

2. Particle swarm optimization


The particle swarm optimization algorithm (PSO) is an evolutionary meta-heuristic technique that was introduced by
Eberhart and Kennedy (1995). It was stimulated by swarming
behavior, for example that of ocking birds or sh schools, in
which each individual movement is inuenced by its own experience and the overall swarm experience (Fahmy, Hassan, &
Bassioni, 2014). PSO resembles this behavior through a population
of swarm particles moving in space. This population of swarm particles is often known as a potential solution for PSO, while the
space is termed as its search space. In the iterative procedure,
the velocity of each particle is calculated based on the experience
of the particle (local optimum) as well as the overall swarm
experience (global optimum). In the original PSO algorithm, each
swarm particle was represented with two characteristics, position
and speed. The position of particle i was represented as
X i xi1 ; xi2 ; . . . ; xiD which is a potential solution to a problem in
D-dimensional space. Each particle keeps a memory of its previous
best solution Pbest, and a velocity along each dimension, represented as V i v i1 ; v i2 ; . . . ; v iD . At each iteration, the position of
the particle with the best tness value in the search space,
designed as g, and the P vector of the current particle are combined
to adjust the velocity along each dimension, and then that velocity
is used to compute a new position for the particle (Kuo & Yang,
2011). After realizing the potential of birds ocking and therefore
the simulation models ability to perform optimization, the early
version of PSO was dened by the following equation:

v new
v old
id
id c 1  rand1  P bestid  X id c 2  rand2  Gbestid  X id
1
new
X new
X old
id
id v id

where c1 and c2 (0 6 c1 6 2, and 0 6 c2 6 2) are weighting


factors that are called the cognitive and social parameters respectively (also called acceleration coefcients) and rand1 and rand2 are
random variables uniformly distributed within [0, 1]. During iterative processes, the trajectory of each particle i is adjusted by
dynamically altering the velocity of its own according to its own
best ying experience (P bestid ) and the best ying experience of
the swarm (Gbestid ) in D-dimensional search space. The rst part
of (1) is the previous velocity, which provides the necessary
momentum for particles to roam around the search space. The second part, known as the self-cognitive component, represents the
personal thinking of each particle. The cognitive component
encourages particles to move toward their own best positions
found so far. The third part, regarded as the social inuence component, expresses the collaborative effect of particles in discovering the global optimal solution (Wang & Yeh, 2014). The social
component always pulls particles to the global best position found
so far. It should be noted that the PSO algorithm has no knowledge
of the underlying objective function, and thus has no way of knowing if any of the candidate solutions are near to or far away from a
local or global maximum. The PSO algorithm simply uses the
objective function to evaluate its candidate solutions, and operates
upon the resultant tness values (Eberhart & Shi, 2001).
Apart from the canonical PSO algorithm described earlier, many
variations of the PSO algorithm exist. The early PSO performed satisfactorily for simple optimization problems (Eberhart and Shi
(2001)). However, for complex optimization problems with large
search spaces, its crucial inadequacy was revealed and therefore
further renements to the earlier PSO algorithms were required.
There exist numerous such variants, for example: Constant inertia
weight, linear reduction of inertia weight, constriction factor,
dynamic inertia and maximum velocity reduction, tracking of time

dependent minima and, discrete optimization. Amongst these the


most used variant is the insertion of inertia weight, w
(0 6 w 6 1.2), which is a proportional agent that is related with
the speed during the last time (Eberhart & Shi, 1998; Shi &
Eberhart, 1998). In that case, a linear reduction of inertia weight
has potential. This is because the larger inertia weight has more
exploitation ability at the beginning, and the later smaller inertia
weights have more exploitation ability (Kuo & Yang, 2011; Shi &
Eberhart, 1998). The updating rule after the inclusion of inertia
weight is as follows:

v new
wold  v old
id
id c 1  rand1  P bestid  X id c 2  rand2
 Gbestid  X id

wnew a  wold

new
X new
X old
id
id v id

Here, a is the decreasing factor or the determinant of the linear


decreasing of the inertia weight. The value of the inertial coefcient
w is typically between 0.8 and 1.2, which can either dampen the
particles inertia or accelerate the particle in its original direction
(Shi & Eberhart, 1998). Generally, lower values of the inertial
coefcient speed up the convergence of the swarm to an optima,
and higher values of the inertial coefcient encourage exploration
of the entire search space.
3. Problem description
The multi-period and multi-product APP problem can be
described as follows: Consider a company that manufactures N
types of products to satisfy some market demand over a planning
horizon T. Due to incomplete or unobtainable information in a
medium time horizon, the environmental coefcients and the
related operation parameters are typically uncertain in real world
APP decision problems. Therefore, the forecast demand, related
operating costs, and labor and machine capacity are imprecise over
the planning horizon (Wang & Liang, 2004). Most of the existing
APP models are either deterministic or only demand is considered
as a source of uncertainty. But there may be some inaccurate or
unreliable information regarding supplies, processing, transportation, shortage or human related costs (Al-e et al., 2012) which alter
the traditional APP decision problems to be more ambiguous.
Hence, assigning only a set of crisp values for the operational coefcients and related parameters is inappropriate for dealing with
such ambiguous APP decision problems. The primary endeavor of
this considered APP problem is to determine the optimum aggregate plan for meeting forecast demand by adjusting regular and
overtime production rates, inventory levels, labor levels, subcontracting and backordering rates, and other controllable variables.
Escalating factors for different operational variables are also considered here. Based on the above characteristics of the considered
APP problem, the mathematical model herein is developed on the
following assumptions.
i. The values of all parameters are certain over the next T
planning horizon
ii. The escalating factors in each of the costs categories are certain over the next T planning horizon.
iii. Actual labor levels, machine capacity and warehouse space
in each period cannot exceed their respective maximum
levels.
iv. The forecasted demand over a particular period can be either
satised or backordered, but the backorder must be fullled
in the next period.

369

R.K. Chakrabortty et al. / Computers & Industrial Engineering 88 (2015) 366377

4. The mathematical models


Primarily, we considered the Integer Linear Programming (ILP)
mode from Wang and Liang (2004), which considered production
costs, skills of workers, product life cycle, employment law, and
other forces, to minimize total production costs. The rst step in
constructing the APP model of the problem is to identify demand
requirements, related costs like backorder and subcontracting
costs, labor hours, inventory level, and the set of decision variables
whose values are to be determined as part of the planning (Wang &
Yeh, 2014). Some of the involved notations that are used to formulate the ILP model for APP are dened as follows:

T. The inventory level and warehouse spaces are also considered as


secondary elements to construct a comprehensive APP model. The
related costs coefcients in the objective function are frequently
impalpable in nature, because some information is incomplete or
unobtainable in a medium time horizon. Moreover, the surmised
demands and the work-hours required for each job is always
vulnerable for change. Considering all these aspects, the objective
function of the considered model is as follows:
(i) Regular time production cost
N X
T
X
rpnt RP nt 1 EF rp t 

n1 t1

4.1. Nomenclature

rpnt
BLnt
opnt
scnt
Dnt
icnt
bcnt
WLt
WHt
ILnt
EF rp
EF op
EF sc
EF ic
EF bc
EF hl
HLnt
ht
M tmax
lct
OPnt
RP nt
U nt
SV nt
WSnt
WStmax
Ltmax

Regular time production cost per unit for nth product


in period t (Tk./unit)
Backorder level for the nth product in period t (unit)
Overtime production cost per unit for the nth product
in period t (Tk./unit)
Subcontracting cost per unit of the nth product in
period t (Tk./unit)
Forecasted demand for the nth product in period t
(units)
Inventory carrying cost per unit of the nth product in
period t (Tk./unit)
Backorder cost per unit of the nth product in period t
(Tk./unit)
Workers laid off in period t (man-hour)
Worker hired in period t (man-hour)
Inventory level in period t for the nth product (units)
Escalating factor for regular time production cost (%)
Escalating factor for overtime production cost (%)
Escalating factor for subcontract cost (%)
Escalating factor for inventory carrying cost (%)
Escalating factor for backorder cost (%)
Escalating factor for hire and layoff cost (%)
Hours of labor per unit of the nth product in period t
(man-hour/unit)
Cost to hire one worker in period t (Tk./man-hour)
Maximum machine capacity available in period t
(machine-hour)
Layoff cost for one worker in period t (Tk./man-hour)
Overtime production for the nth product in period t
(units)
Regular time production for the nth product in period
t (units)
Hours of machine usage per unit of the nth product in
period t (machine-hour/unit)
Subcontracting volume for the nth product in period t
(units)
Warehouse spaces per unit of the nth product in
period t (ft2/unit)
Maximum warehouse space available in period t (ft2)
Maximum labor level available in period t
(man-hour)

4.2. Objective function


In real world APP decision problems, total costs are usually considered as the sole objective to minimize. The total costs are the
sum of the production costs, including overtime production costs
and the costs of changes in labor levels over the planning horizon

(ii) Overtime labor cost


N X
T
X
opnt OP nt 1 EF op t 

n1 t1

(iii) Subcontracting cost


N X
T
X
scnt SV nt 1 EF sc t 

n1 t1

(iv) Cost of holding inventory


N X
T
X
icnt ILnt 1 EF ic t 

n1 t1

(v) Cost of backordering


N X
T
X
bcnt BLnt 1 EF bc t 

10

n1 t1

(vi) Cost of hiring and lay off workers


T
X

ht WHt lct WLt

11

t1

The total cost within a planning horizon T can be estimated as


the summation of the previous costs from (6) to (11). Escalating
factors were also included for each of the cost categories. Thus
the objective function can be represented as follows:

Min Z

N X
T
X

rpnt RP nt 1 EF rp t opnt OPnt 1 EF op t

n1 t1

scnt SV nt 1 EF sc t icnt ILnt 1 EF ic t


bcnt BLnt 1 EF bc t 

T
X
ht WHt lct WLt

12

t1

4.3. Formulation of constraints


4.3.1. Constraints on carrying inventory

ILnt1  BLnt1 RP nt OP nt SV nt  ILnt BLnt


 nt for 8n; 8t
D

13

As mentioned earlier, the forecasted demands are vulnerable to


^ nt denotes that imprecise
change or are imprecise in nature. Here, D
forecast demand of the nth product in period t. The basic assumption of this APP model is that the demand over a particular planning period should be met by allowing backordering for any
incomplete amounts. But the backorder must be fullled in the
subsequent period. Therefore, the sum of the resultant inventory

370

R.K. Chakrabortty et al. / Computers & Industrial Engineering 88 (2015) 366377

level after the nth period (ILnt1  ILnt ), the resultant backordered
amounts (BLnt  BLnt1 ), subcontracted amounts (SV nt ), regular
and overtime production quantity should equal the market
demand as shown in constraint Eq. (13).
4.3.2. Constraints on labor levels
N
X
HLnt1 RP nt1 OP nt1 WHt  WLt
n1

N
X
HLnt RP nt OP nt for 8t

14

n1
N
X
HLnt RP nt OP nt 6 Ltmax

for 8t

15

n1

For any particular time period t, the labor levels required for regular
and overtime production should be equal to the labor level that
already exists in period (t  1) with the addition of new hired
employees, while subtracting the number of lay off employees as
shown in constraint Eq. (14). Again the actual labor levels cannot
exceed the maximum available labor levels in each period, as
shown in Eq. (15). Owing to uncertain labor market demand and
supply, the maximum available labor levels are also imprecise in
nature (Wang & Liang, 2005).
4.3.3. Constraints on machine capacity & warehouse space
N
X
^ nt OP nt RP nt 6 M
^ tmax
U

for 8t

16

n1
N
X
WSnt ILnt 6 WStmax

for 8t

17

n1

^ nt ),
Again, for the nth product the machine usage per unit (U
^
along with the maximum available machine capacity (M tmax ) in
period t, are also imprecise as the usage rate may vary with product specications, process modications, machine depreciation
etc. Therefore, using the deterministic values for them will give
suboptimal results. Constraint Eqs. (16) and (17) represent the limits of the actual machine and warehouse capacity in each period
respectively.

of the triangular possibility distribution for all the imprecise


coefcients, as similar to some earlier literature (Chakrabortty &
Hasin, 2013a; Wang & Liang, 2004, 2005). Fig. 1 presents the
triangular possibility distribution of the imprecise number
o
f
rp nt rppnt ; rpm
nt ; rpnt . In practice, the triangular possibility distribution is constructed by three prominent data, such as the most
pessimistic value, the most possible value and the most optimistic
value. The basic construction is somehow a secondary representation of the fuzzy triangular membership function, but the exception is that without using the [0, 1] range membership values,
the parameters can take any three values that follow triangular
characteristics. In this paper, all the imprecise parameters were
tackled with this triangular possibility distribution. For example,
a decision maker can model f
rp nt based on the three prominent values, as follows. The most pessimistic value (rppnt ) that has a very
low likelihood of belonging to the set of available values (possibility degree = 0 if normalized), the most possible value (rpm
nt ) that
denitely belongs to the set of available values (possibility
degree = 1 if normalized) and the most optimistic value (rpont ) that
has a very low likelihood of belonging to the set of available values
(possibility degree = 0 if normalized) (Wang & Liang, 2005). The
f nt ; h
e t ; lc
e t; U
e nt ; bc
e nt ; D
e nt ; e
f
L tmax ;
p nt ; s~cnt ; ic
other imprecise values o
e tmax were also modeled in this manner.
M

4.4.1. Strategy for solving the imprecise objective function


As the data associated with the objective function is imprecise
in nature, it has been modeled by a triangular possibility distribution. As shown in Fig. 2, geometrically, this imprecise objective is
fully dened by three prominent points (zp , 0) (zm , 1) and (zo , 0).
For minimizing the imprecise objective value (i.e., total costs) the
three projecting points should be pushed towards left. For the sake
of stability, the vertical coordinates which should be xed at either
1 or 0, the horizontal coordinates should be the only concern for
minimizing the objective value. Therefore, solving the imprecise
objective requires minimizing zp ; zm and zo simultaneously. For all
these three important cases, the three new crisp objective functions should be those in Eqs. (18)(20):

Min Z 1 Z m

N X
T
X

t
t
m
rpm
nt RP nt 1 EF rp opnt OP nt 1 EF op

n1 t1
m

t
t
t
scm
nt SV nt 1 EF sc icnt ILnt 1 EF ic bcnt BLnt 1 EF bc

T
X
m
m
ht WHt lct WLt 1 EF hl t

18

t1

Non-negativity Constraints on decision variables are:

Min Z 2 Z p
RP nt ; OP nt ; SV nt ; ILnt ; BLnt ; WHt ; WLt P 0 for 8n; 8t

N X
T
X
rppnt RP nt 1 EF rp t oppnt OPnt 1 EF op t
n1 t1
p

scpnt SV nt 1 EF sc t icnt ILnt 1 EF ic t bcnt BLnt 1 EF bc t


4.4. Model development
Probabilities, possibilities, and fuzzy sets are all measures to
formalize and quantify uncertainty. In spite of some ongoing
debate regarding the appropriateness of each measure in formalizing uncertainty, Kosko (1990) and Klir (1989) put some arguments
in favor of possibilities and fuzzy sets (Slimen, Ayachi, & Amor,
2013). Possibility theory which was introduced by Negoita,
Zadeh, and Zimmermann (1978) and was further developed by
Dubois, Foulloy, Mauris, and Prade (2004), lies at the crossroads
between fuzzy sets and probability. Among the different types of
possibility distribution, the triangular possibility distribution is
an optimal transform of the uniform probability distribution and
it is the upper envelope of all the possibility distributions that
are transformed from symmetric probability densities with the
same support (Dubois et al., 2004). This study adopts the pattern

T
X
p
p
ht WHt lct WLt 1 EF hl t
t1

Fig. 1. The triangular possibility distribution of f


rp nt .

19

371

R.K. Chakrabortty et al. / Computers & Industrial Engineering 88 (2015) 366377


N
X
U 0nt;b RPnt OP nt 6 M0tmax;b

8t

25

n1

Using the fuzzy decision-making for all the imprecise data, and
following the fuzzy programming method from Wang and Liang
(2005); (Zimmermann, 1978), the complete equivalent
single-goal LP model for solving the APP decision problem can be
formulated as follows:

e
Min Z
Subject to;

 zp ; zm ; zo .
Fig. 2. The Strategy to minimize the total costs, Z

ILnt1  BLnt1 RP nt OPnt SV nt  ILnt BLnt

N X
T
X
Min Z 3 Z
rpont RP nt 1 EF rp t opont OP nt 1 EF op t

o
w1 Dpnt;b w2 Dm
nt;b w3 Dnt;b

n1 t1
o

scont SV nt 1 EF sc t icnt ILnt 1 EF ic t bcnt BLnt 1 EF bc t


T
X
o
o

ht WHt lct WLt 1 EF hl t

20

t1

w2 Dm
nt;b

w3 Dont;b

8n; 8t

N
X
HLnt RPnt OPnt 0 8t;

N
X
o
int RP nt OP nt 6 w1 Lptmax;b w2 Lm
tmax;b w3 Ltmax;b

8t

22

n1

To solve these imprecise right-hand sides, this paper employs


the concept of the most likely values that was proposed by Lai
and Hwang (1992) and Wang and Liang (2005), which assume
w2 = 4/6 and w1 = w3 = 1/6. The reason for using the most likely
values here is that the most possible values are usually the most
important ones, and thus should be assigned more weight.
Moreover, to solve problems with an imprecise machine usage rate
~ tmax , the approach proposed
(Unt ) as well as available resources M
here converts these imprecise inequality constraints into crisp
ones using the fuzzy ranking concept (Lai & Hwang, 1992;
Tanaka, Guo, & Zimmermann, 2000). Consequently, the auxiliary
inequality, constraints (16), can be represented as follows:
N
X
U pnt;b RP nt OPnt 6 M ptmax;b

8t

23

8t

24

n1
N
X
m
Um
nt;b RP nt OP nt 6 M tmax;b

N
X
o
int RP nt OPnt 6 w1 Lptmax;b w2 Lm
tmax;b w3 Ltmax;b 8t
n1
N
X
U pnt;b RP nt OP nt 6 M ptmax;b 8t
n1
N
X
m
Um
nt;b RP nt OP nt 6 M tmax;b 8t
n1
N
X
U 0nt;b RP nt OP nt 6 M 0tmax;b 8t
n1
N
X
WSnt ILnt 6 WStmax

8t;

n1

06L61

21

where w1 + w2 + w3 = 1, w1, w2, w3 denote the weights of the most


pessimistic, most possible and most optimistic values of the imprecise demand, respectively. The weights w1, w2 and w3 can be determined subjectively by the experience and knowledge of
practitioners (Wang & Liang, 2005). Similarly, for the imprecise
maximum labor level, if the minimum acceptable possibility b is
given, then the auxiliary crisp inequality, constraints (15), can be
represented as follows:

n1

RP nt ; OPnt ; SV nt ; ILnt ; BLnt ; WHt ; WLt P 0 for 8n; 8t

ILnt1  BLnt1 RP nt OP nt SV nt  ILnt BLnt

n1

n1

4.4.2. Handling imprecise technological coefcients and available


resources
As mentioned earlier, the conjectured demand data, Dnt , is
imprecise in nature and therefore it is assumed to have triangular
possibility distributions with the most and least possible values. A
possible interval for that imprecise data is not too difcult for the
practitioners with suitable experience and knowledge. In contrast,
obtaining a crisp representation number for that imprecise
demand is very hard. This paper simulates the weighted average
method proposed by Lai and Hwang (1992) to convert Dnt into
crisp number. If the minimum acceptable possibility, b, is given,
then the auxiliary crisp equality, constraint (13), can be represented as follows:

w1 Dpnt;b

8n; 8t

N
X
HLnt1 RPnt1 OPnt1 WHt  WLt

5. Solution approach
This study proposes to use Possibilistic Environment based
Particle Swarm Optimization (PE-PSO) order to solve the considered aggregate production planning problem. As mentioned in
the problem statement section, all the imprecise data are tackled
by the fuzzy triangular membership function. Escalating factors
of 1% for each case (EF rp ; EF op ; EF sc ; EF ic ; EF bc ; EF hl ) was also considered for further allowance of the approximate data.
Fig. 3 shows the original PSO operations for the conventional
case. With all of the initial swarm cases, if the stopping criteria
(that is the maximum number of loops or computational time)
are satised, the problem stops. The details of this PSO approach
are discussed in Section 2. A modied PSO variant, linear reduction
of inertia weight was considered in this problem for faster convergence and better results, where the inertial weight (w) was equal
to 0.9 for the rst run. The value of decrement (a) for the inertia
weight was 0.005. The step-wise procedures for our modied
PE-PSO algorithm are depicted in the following owchart shown
in Fig. 4. At rst, all the imprecise parameters were incorporated
with their designed and model specic escalating factors. After
that, all of those parameters were modeled by using a triangular
possibility distribution. Meanwhile, similar to most of the earlier
literature, the cognitive and social coefcients for PSO algorithm
were considered as 1 for equation (1), and the random numbers
were uniformly distributed. The later portion of this modied
PSO is quite similar to that of the conventional PSO algorithm.
The algorithm was run for 1000 iterations. We solved this
PE-PSO algorithm by using the commercial C programming

372

R.K. Chakrabortty et al. / Computers & Industrial Engineering 88 (2015) 366377

Fig. 3. PSO operations.

compiler, Code Blocks v10.05, which was executed on an Intel core


i7 processor with 16.00 GB RAM and a 3.40 GHz CPU. For better
understanding, a summary of the solution procedure is given
below:
Step 1: The problem was formulated as an integer linear programming one by considering the costs of production, overtime
labor, subcontracting, holding inventory, backordering and
workers hiring and lay off. The basic target was to minimize
their resultant costs (i.e., total production costs).
Step 2: As a result of modern volatile markets and enhanced
customization, the forecasted demands, actual labor levels,
maximum available machine capacity and machine usage rate,
become more and more vulnerable to change or become more
imprecise in nature. Therefore, to cope with real time scenarios,
imprecise characteristics of those variables were imposed in the
model constraints.
Step 3: For handling all sorts of possible uncertainties, the
model parameters, such as per unit regular time production
cost, overtime cost, subcontracting cost, inventory carrying

Fig. 4. Operations of PE-PSO.

costs, backordering cost, workers hiring or lay off costs etc,


were also considered to be imprecise, and so adopted the pattern of the triangular possibility distribution (details are discussed in Section 4.4).
Step 4: Due to the impreciseness of the associated data, the
objective function itself also becomes imprecise, and so has also
been modeled by a triangular possibility distribution. According
to Fig. 2 (in Section 4.4.1), the least value among (zp ; zm ; zo )
should be the optimum result for this APP problem.
Step 5: For handling imprecise conjectured demand data (Dnt )
and available resources such as maximum labor level, machine
usage rate, and number of available machines, some weighted
parameters were employed. A weighted average method was
simulated to modify all those imprecise constraints into their
crisp equality or auxiliary inequalities (details are shown in
Section 4.4.2).
Step 6: Finally, as shown in Fig. 4, all those possibilistic environment based parameters were then solved with our modied
PSO algorithm (PE-PSO), with the model dened parameters.

6. Experimental results and analysis


6.1. Case description
Comt Composite Knit Limited (CCKL) was used as a case study
to demonstrate the practicality of the proposed methodology.
Comt Composite Knit Limited is the sister concern of Youth
Group, which is one of the pioneer companies of the Ready Made
Garments (RMG) sector in Bangladesh. This company readily produced knitwear items and among them some were fancy & some
were expensive. The jacket items, as well as cardigan items, are
the most expensive and most time & cost incurring manufacturing
items. So it needs a lot of precise observations & perfect manufacturing practices to catch up to the market & satisfy buyers within
specied lead times. Since they are the most expensive items, the
cases of one particular style of hooded jacket (Product 1) and
another special type of ladies cardigan (Product 2) were
considered.
The APP decision problem presented here for CCKLs Knit garments manufacturing plant, focuses on developing an interactive
PE-PSO algorithm approach for minimizing total costs. The demand
requirements for the hooded jacket and cardigan items from consumers are highly seasonal. Generally, from April to November of
each scal year, the demand of these types of winter clothes are
high and therefore requires a lot of overtime production and even
subcontracting for more amounts than the actual production
capacity level. Although there is no limit on overtime production,
subcontracting, stock outs and backlogs, the company has to obey
the labor rules which state that no workers could work more than
the limit of legal labor hours at each period. The forecasted
~ 1t : demand of product
demand for each product at each period (D
~ 2t : demand of product 2 at period t) along with
1 at period t, and D
their relevant operational data (maximum man hour available,
machine capacity and warehouse limitations) are given in
Table 1. Since all the demand and operating data are imprecise in
nature, it should be tacked with triangular possibility distribution
according to the solution methodology. As a consequence, all the
collected data from CCKL had three values that followed the characteristics of that triangular possibility distribution. All the relevant operating cost data are shown in Table 2, which also follows
the same characteristics. In this study, we consider May and June
as our planning periods, in order to illustrate the acceptability of
our proposed algorithm in the peak demand period for the CCKL
case. Other relevant data are as follows.

373

R.K. Chakrabortty et al. / Computers & Industrial Engineering 88 (2015) 366377


Table 1
Forecasted demand and other operating data for the CCKL Case.
Items

1st Period

~ 1t (units)
D
~ 2t (units)
D

2nd Period

Items

1st Period

2nd Period

(1200, 1400, 1500)

(2750, 3000, 3200)

(180, 225, 250)

(180, 225, 250)

(1400, 1600, 1800)

(700, 800, 1000)


Ltmax (Man-hours)

Ltmax (Machine-hours)
 tmax (ft2)
WS

Table 2
Related Operating cost for the CCKL Case.
Product

f
rp nt
(tk./unit)

f
o
p nt
(tk./unit)

e nt
sc
(tk./unit)

e nt
ic
(tk./unit)

f nt
bc
(tk./unit)

1
2

(20, 22, 25)


(18, 20, 22)

(38, 40, 44)


(36, 40, 42)

(25, 27, 30)


(28, 30, 32)

(3, 3.5, 4)
(3.5, 4, 4.5)

(40, 42, 44)


(45, 47, 50)

Table 3
The problems variables and upper bound values.
Item (Product 1)

RP1t (Units)
OP1t (Units)
SV1t (Units)
IL1t (Units)
BL1t (Units)
WHt (man-hours)

Period

Items (Product 2)

569
569
182
430
2
43

1033
1019
331
522
580
169

RP2t (Units)
OP2t (Units)
SV2t (Units)
IL2t (Units)
BL2t (Units)
WLt (man-hours)

Period
1

621
682
94
150
146
0

447
446
96
211
96
160

I. Initial inventory in period 1 is 500 units of product 1 and 200


units of product 2. The end inventory in period 2 is 400 units
of product 1 and 300 units of product 2.
II. Initial labor level is 225 man-hours. The costs associated
with hiring and layoffs are Tk. (20, 22, 25) and Tk. (6, 8,
10) per worker per hour, respectively.
III. Hours of labor per unit for any periods are xed to 0.033
man-hours for product 1 and 0.05 man hours for product
2. Hours of machine usage per unit for each of the two planning periods are (0.09, 0.1, 0.11) machine-hours for product
1 and (0.07, 0.08, 0.09) machine-hours for product 2.
Warehouse spaces required per unit are 1 square feet for
product 1 and 1.5 square feet for product 2.
IV. The expected escalating factor in each of the costs categories
is 1%.
After putting all the relevant data into our proposed PE-PSO
algorithm using designated compiler, the results for different decision variables are summarized in Table 3. For all the PSO variants
the tness value gives approximately similar values. After 137 iterations the PSO algorithm gives the minimum result. It receives its
minimum tness or objective value Tk. 234,606.63 after 137 iterations and the remaining iterations give the same value. For this
minimum objective value the problems variables and upper bound
values for the multi period and multi product APP plan are shown
in Table 3.
6.2. Evaluation of the PE-PSO with Other Algorithms
In order to evaluate the performance of the PE-PSO algorithm
for the APP problem, at rst we made use of that case study data.
For better comparison, we considered three different algorithms;
Fuzzy based Genetic Algorithm (FBGA), Genetic Algorithm (GA)
and our proposed PE-PSO algorithm. The rst two algorithms were
coded with Matlab R2010b and were executed on a PC with an
Intel core i7 processor, 16.00 GB RAM and a 3.40 GHz CPU.

(360, 400, 430)

(450, 500, 540)

1000

1000

To carry out the FBGA approach, we considered three different


scenarios with different GA parameters. All the imprecise data
were rst tacked with triangular possibility distributions and then
three different crisp objective functions were derived. Later on,
those crisp objective functions were simulated as fuzzy triangular
membership functions to derive the decision makers satisfaction
level, which should be within [0, 1]. In this experiment, we set
the parameters of the population size, the crossover rate, the mutation rate and the generations according to Table 4, and the maximum satisfaction level (scenario 3), and the total cost, along with
other decision variables, are summarized in Table 5. As of
Table 5, the objective function value for the third scenario represents the least total cost, and therefore it will be considered for validating our proposed PE-PSO model. Meanwhile for the GA
approach, we employed ve distinct scenarios according to the
parameter settings as shown in Table 6. Similar to the FBGA
approach and PE-PSO approach, here the imprecise data were rst
tackled with triangular possibility distributions. After which a simple GA was used to solve all those scenarios and the best results
among all scenarios (fourth scenario) are summarized in Table 7.
More analysis and different solution methods for both the FBGA
and GA approaches can be found in the earlier literature from
Chakrabortty and Hasin (2013b) and Chakrabortty and Hasin
(2013a) respectively.
From Table 7 it is quite clear, that to choose the best algorithm
for solving APP problems is not an easy task. Different approaches
combine different sets of parameters and functions, also different
approaches have different features. In the case of FBGA, the results
are quite satisfactory but it can have a higher cost value. From
Table 7, it has been observed that the FBGA approach gives the
imprecise objective function value of Tk. (217,238; 235,086;
256,457) which means that the total cost for the APP plan would
be within tk. 217, 238 to tk. 256, 457. On the contrary, for the
GA approach, the objective function gives quite satisfactory results
with higher cost. Comparing to these two algorithms, our proposed
PE-PSO algorithm gives the total cost as tk. 234, 606.63. It is important to note, that if the practitioners were not lucky enough, then
the total cost of the proposed APP plan could be tk. 256, 457, which
is much higher than the results obtained by GA and PE-PSO. Again
comparing to GA, our proposed PE-PSO gives the least objective
value. Moreover, if the objective function bears only linear terms,
then the possibilitistic linear programming approach might give
better results. But for nonlinear objective functions such as this
case, PE-PSO is the most suitable. The proposed approach also provides information on alternative strategies for overtime, subcontracting, inventory, backorders, and hiring and layoffs workers, in
response to variations in forecast demand. Additionally, the proposed model considers the actual limitations in labor, machine,
and warehouse capacity. Notably, the goal values of the FBGA
approach should be imprecise, because the unit cost and related
coefcients are always imprecise in nature. On the other hand,
the analytical results demonstrate that replacing the imprecise
data with the most possible value or the average of the possibility
distribution is unreasonable (Wang & Liang, 2005). Therefore, relying on the FBGA approach is not always a wiser decision for the
decision maker. Finally, the proposed PE-PSO approach is the most

374

R.K. Chakrabortty et al. / Computers & Industrial Engineering 88 (2015) 366377

Table 4
Different parameter setting for FBGA under three scenarios.
Scenario

Population size

Creation Function

Crossover (Fraction)

Mutation

Migration (Fraction)

Generations needed

1
2
3

360
360
360

Constraint dependent
Feasible Population
Constraint dependent

Intermediate (0.8)
Scattered (0.8)
Arithmetic (0.5)

Constraint dependent
Constraint dependent
Adapt Feasible

Forward (0.2)
Forward (0.2)
Both (0.5)

107
116
136

Table 5
Imprecise Objective Function values for different scenarios under FBGA approach.
Imprecise Objectives

Scenario 1

Scenario 2

Scenario 3

Z1
Z2
Z3
Total Cost (Tk.)

2,17,272.30
2,35,121.88
2,56,493.00
7,08,886.20

2,17,273
2,35,123.37
2,56,497.16
7,08,893.55

2,17,238.30
2,35,086.98
2,56,457.00
7,08,782.32

The bold result for scenario 3 is representing the minimum total cost among all
three scenarios.
Table 6
Different parameter setting for GA under ve scenarios.
Scenario

Population
size

Crossover

Crossover
rate

Mutation
rate

Generations

1
2
3
4
5

50
50
360
360
360

Two point
Heuristic
Arithmetic
Scattered
Single point

0.8
0.5
0.5
0.8
0.8

0.2
0.2
0.5
0.2
0.5

103
260
134
111
152

practical for solving APP decision problems because it can generate


better decisions than the other models.
In order to compare the convergence rate for the proposed
PE-PSO algorithm with our case study data, we performed experiments with different combinations of PSO variants to determine
the appropriate set of parameters, and to judge their effect on
the solutions. Inertia weight plays a key role in the process of providing balance between the exploration and exploitation processes
and determines the contribution rate of a particles previous velocity, to its velocity at the current time step. So it is important to nd

an appropriate set of inertia weights, along with social and cognitive coefcients.
As shown in Table 8, we considered 8 combinations with inertia
weight varied from 0.9 to 0.55 with a decrement of 0.05, social (C1)
and with the cognitive coefcients (C2) from 1 to 4.5 with an increment of 0.5. To see how the quality of solutions varies with both
the changing inertia weight, and the social and cognitive coefcients, we executed a series of experiments. Comparing with other
parameter sets, for parameter set 1, our algorithm provided the
best solution after 580 iterations; this showed that the convergence rate and result of PE-PSO with set 1 is better than the others.
As a consequence, parameter set 1 was used to compare the
PE-PSO algorithm with FBGA and GA for our concerned case study
data, as shown in Table 7.
Furthermore, to evaluate the performance of this proposed
PE-PSO algorithm for the concerned APP problem, we made use
of an articial dataset of 8 instances with large equality constraints, among which 4 were small-sized and 4 were large-sized.
For comparing these algorithmss performance, these instances
were also coded and solved by the LINGO optimization solver to
discover their global optimum which hence the lower bound
(LB). In order to present the diversity of performance between
the global optimum (LB) and the objective values of the results
of these algorithms, a quality measurement called percent deviation, denoted as %Dev, is dened as the following equation:

%Dev

BOVPEPSO;FBGA;GA  LBLINGO
 100%
LBLINGO

26

where BOVPEPSO;FBGA;GA is the best objective value derived from


PE-PSO, FBGA and GA, and LBLINGO is the lower bound originated
from LINGO.

Table 7
Comparative results for different approaches under CCKL case.
Product

Item

FBGA results (third scenario)

GA results (fth scenario)

PE-PSO results

1st Period

2nd Period

1st Period

2nd Period

1st Period

2nd Period

1026
1010
333
524
582

571
571
186
432
5

1034
1021
334
524
583

569
569
182
430
2

1033
1019
331
522
580

664
670
85
159
140
104.4
0
2,36,893.43

446
445
95
213
95
170.67
162.7

621
682
94
150
146
43
0
2,34,606.63

447
446
96
211
96
169
160

Product 1

RP1t (Units)
OP1t (Units)
SV1t (Units)
IL1t (Units)
BL1t (Units)

544
544
186
446
5

Product 2

RP2t (Units)
OP2t (Units)
SV2t (Units)
IL2t (Units)
BL2t (Units)

664
454
670
454
85
95
160
212
140
94
102.65
173.40
0
163.49
(2, 17,238; 2, 35,086; 2, 56,457)

WHt (man-hours)
WLt (man-hours)
Objective values (Tk.)

Table 8
Combination of different PSO variants under CCKL case.

Inertia weight
C1
C2

Com-1

Com-2

Com-3

Com-4

Com-5

Com-6

Com-7

Com-8

0.9
1
1

0.85
1.5
1.5

0.80
2
2

0.75
2.5
2.5

0.70
3
3

0.65
3.5
3.5

0.60
4
4

0.55
4.5
4.5

375

R.K. Chakrabortty et al. / Computers & Industrial Engineering 88 (2015) 366377


Table 13
Comparative performances of PE-PSO, FBGA and GA.

Num. of products

Num. of periods

LINGO optimum

1
2
3
4
5
6
7
8

2
4
6
8
15
18
20
25

2
2
4
4
6
6
12
12

234,508.63
425,275.87
635,275.38
879,300.68
1,098,658.32
1,325,598.52
1,510,740.58
1,725,040.68

Table 10
Statistical results of PE-PSO experiments.
Test
No.

PE-PSO Best

1
234,606.63
2
425,662.87
3
635,306.38
4
879,897.68
5
1,098,994.32
6
1,325,860.52
7
1,511,031.58
8
1,725,380.68
Average

Mean

Std.
deviation

%Dev

Com. Time
(sec)

235,085.14
429,867.29
642,218.82
888,609.60
1,110,852.40
1,329,803.45
1,527,053.30
1,787,031.55

1158.85
5867.97
8241.49
1075.52
10,071.99
15,211.17
20,141.75
22,696.55
10,558.161

0.04179
0.091
0.00488
0.067895
0.030583
0.019765
0.019262
0.01971
0.03686

0.79750
0.8752
1.0024
1.0142
1.7550
1.8705
1.8957
2.0142
1.403088

Table 11
Statistical results of FBGA experiments.
Test
No.

FBGA Best

1
234,606.63
2
425,680.82
3
635,312.38
4
879,890.68
5
1,099,452.32
6
1,326,128.52
7
1,511,345.58
8
1,725,369.68
Average

Mean

Std.
deviation

%Dev

Com. time
(sec)

235,447.86
432,356.18
643,416.84
894,372.10
1,117,306.50
1,330,918.86
1,536,165.10
1,792,089.17

1375.01
6601.45
9014.12
9048.61
12,436.58
16,445.46
21,977.87
24,277.28
12,647.0475

0.04179
0.095221
0.005824
0.067099
0.07227
0.039982
0.040047
0.019072
0.047663

0.87
0.97
1.05
1.43
1.89
2.04
2.35
2.69
1.66125

Table 12
Statistical results of GA experiments.
Test
No.

GA Best

Mean

Std.
deviation

%Dev

Com. time
(sec)

1
2
3
4
5
6
7
8
Average

234,686.58
425,613.25
635,290.58
879,910.68
1,099,719.32
1,326,746.52
1,511,301.58
1,726,080.68

237,156.78
436,270.30
650,762.66
899,428.92
1,126,404.33
1,345,713.09
1,544,125.89
1,810,692.04

1607.92
7439.95
9906.27
17,130.52
14,987.84
17,887.90
24,055.91
26,185.60
14,900.24

0.075882
0.079332
0.002393
0.069373
0.096572
0.086602
0.037134
0.060288
0.06

1.03
1.25
1.58
1.68
2.02
2.35
2.59
3.01
1.94

In this experiment, for the PE-PSO, FBGA and GA algorithms,


each of those 8 instances was run for 30 trials with 1000 iterations
as its stopping criterion. The statistical result of each instance was
only collected from those that found feasible results among 30 trials of execution. As for the FBGA and GA algorithms, the parameter
settings remaining the same as mentioned in Tables 4 and 6
respectively. Again, for the PE-PSO algorithm, the linearly decreasing time dependent inertia weight was set as 0.9, with the social
and cognitive coefcients each set as 1. The experimental statistical results obtained by LINGO, PE-PSO, FBGA and GA, are shown in

PE-PSO
FBGA
GA

Percentage Deviation

Test No.

Std. Dev. (%)

% Dev.

Com. time (sec)

100
120
141

100
129
172

1
1.18 times
1.38 times

0.12
0.1
0.08
0.06

PE-PSO %Dev.

0.04

FBGA %Dev.
GA %Dev.

0.02
0

Test Sets
Fig. 5. Variation of percentage of deviations for different algorithms under different
test sets.

Computing time (sec)

Table 9
Parameters and statistical results of LINGO.

3.5
3
2.5
2
1.5
1
0.5
0

PE-PSO Com. time


FBGA Com. Time
GA Com. Time

Test Sets
Fig. 6. Variation of computing times (sec) for different algorithms under different
test sets.

Tables 912. Each instance of the APP problem is given a test number. Table 9 includes key parameters of each test instance of the
APP problems, namely the number of products and periods. Of
these, the rst four instances are small-sized and the last four
are large-sized. It also displays the lower bound from LINGO.
Tables 1012 show the statistical results of PE-PSO, FBGA and
GA in the test, including the best, mean, standard deviation, percent deviation of their object values, and their executing time for
100 iterations. These statistical results show that the best objective
values of each instance were differentiate from the corresponding
optimum. All of them also have some disparities among their
mean, standard deviation, percent deviation and execution time.
Note that, for FBGA approach, we took the medium values of each
triangular fuzzy objective function value. The comparative performances for all algorithms are also summarized in Table 13. It is
worth mentioning that in the average of standard deviation, percent deviation, successful rate, and execution time, PE-PSO provides better performance than FBGA and GA. Considering 100% as
the PE-PSO results, the percentage deviation gain increased by
29% and 72% for FBGA and GA respectively. Meanwhile, the searching discrepancy of the particles standard deviation was 20% and
41% smaller than FBGA and GA respectively. The execution time
also increased by 1.18 and 1.38 times for FBGA and GA
respectively.
For better illustration, all the comparative analyses on different
performance parameters for PE-PSO, FBGA and GA are depicted in
Figs. 5 and 6. Here in Fig. 5, it has well addressed that despite of
some anomalies, on most cases PE-PSO algorithm shows least percentage of deviation from the LINGO optimum results than the

376

R.K. Chakrabortty et al. / Computers & Industrial Engineering 88 (2015) 366377

Table 14
Comparison of the APP models.
Factor

PLP (Wang and Liang, 2005)

MPSO (Wang and Yeh, 2014)

FGGP (Sadeghi et al., 2013)

The proposed PE-PSO

Objective function
Objective value
Constraint
Product item
Escalating factors
Degree of satisfaction
Planning horizon
Market demand
Machine capacity
Cost of regular time & overtime production
Subcontracting
Backordering
Hiring cost
Layoff cost
Labor level

Single, linear
Imprecise
Imprecise
Product family
Not Included
Presented
Multi-period
Imprecise
Limited
Imprecise
Considered
Considered
Imprecise
Imprecise
Deterministic

Single
Deterministic
Deterministic
Single
Not included
Not Presented
Multi-period
Deterministic
Limited
Deterministic
Considered
Considered
Deterministic
Deterministic
Deterministic

Multiple
Fuzzy
Fuzzy
Product family
Not included
Not Presented
Multi-period
Fuzzy
Limited
Fuzzy
Considered
Considered
Fuzzy
Fuzzy
Fuzzy

Single
Imprecise
Imprecise
Product family
Included
Not presented
Multi-period
Imprecise
Limited
Imprecise
Considered
Considered
Imprecise
Imprecise
Imprecise

other two mentioned algorithms which also enforced the consistency of this algorithm over other heuristic algorithms. In addition
of that computing time or executing time for PE-PSO algorithm is
also in tandem with its suitability over other two algorithms as
depicted in Fig. 6. For each and every problem instances, PE-PSO
algorithm took lesser amount of time to solve those instances compare with other algorithms which also predominantly proves its
competency over those algorithms specically on these sorts of
problem structures. Therefore, we conclude that the PE-PSO variant possesses better qualities in accuracy, reliability, and convergence speed for optimizing the APP problem, than both of FBGA
and GA.
Meanwhile, Table 14 also compares the proposed PE-PSO
approach presented here to some other APP models, like the
Possibilistic Linear Programming (PLP) (Wang & Liang, 2005),
Modied Particle Swarm Optimization (MPSO) (Wang & Yeh,
2014), and Fuzzy Grey Goal Programming (FGGP) (Sadeghi et al.,
2013) models. There are several features which distinguish the
proposed PE-PSO approach from other APP models. The proposed
PE-PSO approach can solve most real-world APP problems that
involve imprecise parameters through an interactive decision making process. The proposed model constitutes a systematic framework that facilitates the decision-making process. This proposed
PE-PSO approach includes escalating factors for all the imprecise
parameters. This feature provides more allowances to the uncertain data, which are vulnerable to change, and thus allows the decision maker to have more meaningful results. Rather than relying
on deterministic data for every operating parameter, this proposed
approach focuses on the real world uncertainties and considers
them as being imprecise in nature. Moreover, the proposed
PE-PSO approach covers more wide ranging decision information
than the others which predominantly focus on multi-period and
multi-product (product family) problems in an APP decision making process. The proposed approach is also distinct from the traditional fuzzy based and stochastic based approaches. Basically, the
FGGP is based on the subjective preferred concept for establishing
membership functions with fuzzy data, while the PE-PSO is based
on the objective degree of event occurrence that is required to
obtain possibilistic distributions with imprecise data. On the contrary, the main problem in applying the stochastic programming
models is the lack of computational efciency and the inexible
probabilistic doctrines in which the real imprecise meaning of
the decision making might be impossible to model (Wang &
Liang, 2005). Alternatively, the possibilitic approaches not only
provide more computational efciency, but also supports possibilistic decision making in an uncertain environment (Lai &
Hwang, 1992; Tang, Wang, & Fung, 2000).

7. Conclusions
In this paper, a PE-PSO model for APP problems was developed. PSO, which was inspired by the behavior of birds and sh,
has certain advantages, including simple structure, easy implementation, immediate accessibility, short search time, and
robustness. However, there has been limited study of PSO to
address the uncertainties found in the aggregate production
planning problem. This research work therefore presented a
novel PE-PSO method for solving a complex aggregate planning
problem with a single objective, which included the minimization of inventory costs, production costs and manpower costs.
In this work, the linear reduction of inertia weight, an important
variant of PSO, was used for modifying the particles velocity and
movement. This method provides a novel approach to consider
the natural uncertainty of the parameters in an APP problem.
Triangular possibility distributions were employed to tackle the
model parameters uncertainties. For justifying the feasibility of
the proposed approach, an industrial case was demonstrated.
This approach can be applied in ambiguous and indeterminate
circumstances of real world production planning and scheduling
problems with ill dened data. For better justication of the
proposed PE-PSO approach, a comparative analysis on the performances of FBGA, GA and PE-PSO was carried on. The experimental results showed that the PE-PSO variant had better accuracy in
comparison to the FBGA and GA approaches for this sort of APP
problems.
In spite of the increasing applicability of the traditional PSO
algorithms, it still has some of the limitations from when it was
originally introduced for unconstrained and continuous problems.
That is why the application of PSO in the APP literature is so
small. Moreover, to the best of our knowledge, there is no earlier
research on PSO for APP, which covers operating parameters
uncertainties. Therefore, in response to these shortcomings, we
used PE-PSO along with a PSO-variant. Future work could be to
nd other ways of representing how the conventional PSO algorithm is tuned for APP problems. Alternative and complementary
algorithms include dynamic evolutionary programming, differential evolution strategies, bee colony and ant colony optimizations
and methods based on integer linear programming. These can
also be applied for the APP problem. Other variants of PSO could
be another important area to focus on. More possible topics for
further study include modication of the particle position representation, particle movement, and particle velocity. Issues related
to Pareto optimality, such as a solution maintenance strategy and
performance measurement, are also topics worthy of future
study.

R.K. Chakrabortty et al. / Computers & Industrial Engineering 88 (2015) 366377

References
Al-e, S. M. J. M., Aryanezhad, M. B., & Sadjadi, S. J. (2012). An efcient algorithm to
solve a multi-objective robust aggregate production planning in an uncertain
environment. The International Journal of Advanced Manufacturing Technology,
58(58), 765782.
Baykasoglu, A., & Gken, T. (2006). A tabu search approach to fuzzy goal programs
and an application to aggregate production planning. Engineering Optimization,
38(2), 155177.
Baykasoglu, A., & Gocken, T. (2010). Multi-objective aggregate production planning
with fuzzy parameters. Advances in Engineering Software, 41(9), 11241131.
Chakrabortty, R. K., Sarker, R., & Essam, D. (2014). Event based approaches for solving
multi-mode resource constraints project scheduling problem. Paper presented at
the 13th international conference on computer information systems and
industrial management applications, Ho Chi Minh City, Vietnam.
Chakrabortty, R., & Hasin, M. (2013a). Solving an aggregate production planning
problem by using multi-objective genetic algorithm (MOGA) approach.
International Journal of Industrial Engineering Computations, 4(1), 112.
Chakrabortty, R. K., & Hasin, M. A. A. (2013b). Solving an aggregate production
planning problem by fuzzy based genetic algorithm (FBGA) approach.
International Journal of Fuzzy Logic Systems (IJFLS), 3(1), 116.
Drakopoulos, J. A. (1995). Probabilities, possibilities, and fuzzy sets. Fuzzy Sets and
Systems, 75(1), 115.
Dubois, D., Foulloy, L., Mauris, G., & Prade, H. (2004). Probability-possibility
transformations, triangular fuzzy sets, and probabilistic inequalities. Reliable
Computing, 10(4), 273297.
Eberhart, R. C., & Kennedy, J. (1995). A new optimizer using particle swarm theory.
Paper presented at the proceedings of the sixth international symposium on
micro machine and human science.
Eberhart, R. C., & Shi, Y. (1998). Comparison between genetic algorithms and particle
swarm optimization. Paper presented at the Evolutionary Programming VII.
Eberhart, R. C., & Shi, Y. (2001). Particle swarm optimization: developments,
applications and resources. Paper presented at the evolutionary computation,
2001. Proceedings of the 2001 Congress on.
Elloumi, S., & Fortemps, P. (2010). A hybrid rank-based evolutionary algorithm
applied to multi-mode resource-constrained project scheduling problem.
European Journal of Operational Research, 205(1), 3141.
Fahmy, A., Hassan, T. M., & Bassioni, H. (2014). Improving RCPSP solutions quality
with Stacking Justication-Application with particle swarm optimization.
Expert Systems with Applications, 41(13), 58705881.
Gen, M., Tsujimura, Y., & Ida, K. (1992). Method for solving multiobjective aggregate
production planning problem with fuzzy parameters. Computers & Industrial
Engineering, 23(1), 117120.
Hsieh, S., & Wu, M.-S. (2000). Demand and cost forecast error sensitivity analyses in
aggregate production planning by possibilistic linear programming models.
Journal of Intelligent Manufacturing, 11(4), 355364.
Iris, C., & Cevikcan, E. (2014). A fuzzy linear programming approach for aggregate
production planning. Supply chain management under fuzziness. Springer
(pp. 355374). .
Kazemi Zanjani, M., Ait-Kadi, D., & Nourelfath, M. (2013). A stochastic programming
approach for sawmill production planning. International Journal of Mathematics
in Operational Research, 5(1), 118.
Klir, G. J. (1989). Is there more to uncertainty than some probability theorists might
have us believe?. International Journal of General System, 15(4), 347378.
Kogut, B., & Kulatilaka, N. (1994). Operating exibility, global manufacturing, and
the option value of a multinational network. Management Science, 40(1),
123139.
Kosko, B. (1990). Fuzziness vs. probability. International Journal of General System,
17(23), 211240.
Kuo, R., & Yang, C. (2011). Simulation optimization using particle swarm
optimization algorithm with application to assembly line design. Applied Soft
Computing, 11(1), 605613.

377

Lai, Y.-J., & Hwang, C.-L. (1992). A new approach to some possibilistic linear
programming problems. Fuzzy Sets and Systems, 49(2), 121133.
Liang, T.-F. (2007). Application of interactive possibilistic linear programming to
aggregate production planning with multiple imprecise objectives. Production
Planning and Control, 18(7), 548560.
Liang, T.-F., Cheng, H.-W., Chen, P.-Y., & Shen, K.-H. (2011). Application of fuzzy
sets to aggregate production planning with multiproducts and multitime
periods. IEEE Transactions on Fuzzy Systems, 19(3), 465477.
Mirzapour Al-e-hashem, S., Baboli, A., & Sazvar, Z. (2013). A stochastic aggregate
production planning model in a green supply chain: Considering exible lead
times, nonlinear purchase and shortage cost functions. European Journal of
Operational Research, 230(1), 2641.
Nam, S.-J., & Logendran, R. (1992). Aggregate production planningA survey of
models and methodologies. European Journal of Operational Research, 61(3),
255272.
Negoita, C., Zadeh, L., & Zimmermann, H. (1978). Fuzzy sets as a basis for a theory of
possibility. Fuzzy Sets and Systems, 1, 328.
Ning, Y., Liu, J., & Yan, L. (2013). Uncertain aggregate production planning. Soft
Computing, 17(4), 617624.
Raa, B., Dullaert, W., & Aghezzaf, E.-H. (2013). A matheuristic for aggregate
productiondistribution planning with mould sharing. International Journal of
Production Economics, 145(1), 2937.
Ramezanian, R., Rahmani, D., & Barzinpour, F. (2012). An aggregate production
planning model for two phase production systems: Solving with genetic
algorithm and tabu search. Expert Systems with Applications, 39(1), 12561263.
Sadeghi, M., Hajiagha, S. H. R., & Hashemi, S. S. (2013). A fuzzy grey goal
programming approach for aggregate production planning. The International
Journal of Advanced Manufacturing Technology, 64(912), 17151727.
Sakall, . S., Bayko, . F., & Birgren, B. (2010). A possibilistic aggregate production
planning model for brass casting industry. Production Planning & Control, 21(3),
319338.
Shi, Y., & Eberhart, R. (1998). A modied particle swarm optimizer. Paper presented at
the Evolutionary Computation Proceedings, 1998. IEEE World Congress on
Computational Intelligence. The 1998 IEEE International Conference on.
Slimen, Y. B., Ayachi, R., & Amor, N. B. (2013). Probability-possibility transformation.
Fuzzy Logic and Applications. Springer (pp. 122130). .
Tanaka, H., Guo, P., & Zimmermann, H.-J. (2000). Possibility distributions of fuzzy
decision variables obtained from possibilistic linear programming problems.
Fuzzy Sets and Systems, 113(2), 323332.
Tang, J., Wang, D., & Fung, R. Y. (2000). Fuzzy formulation for multi-product
aggregate production planning. Production Planning & Control, 11(7),
670676.
Vasant, P. M., Barsoum, N. N., & Bhattacharya, A. (2008). Possibilistic optimization in
planning decision of construction industry. International Journal of Production
Economics, 111(2), 664675.
Wang, R.-C., & Fang, H.-H. (2001). Aggregate production planning with multiple
objectives in a fuzzy environment. European Journal of Operational Research,
133(3), 521536.
Wang, R.-C., & Liang, T.-F. (2004). Application of fuzzy multi-objective linear
programming to aggregate production planning. Computers & Industrial
Engineering, 46(1), 1741.
Wang, R.-C., & Liang, T.-F. (2005). Applying possibilistic linear programming to
aggregate production planning. International Journal of Production Economics,
98(3), 328341.
Wang, S.-C., & Yeh, M.-F. (2014). A modied particle swarm optimization for
aggregate production planning. Expert Systems with Applications, 41(6),
30693077.
Zhang, R., Zhang, L., Xiao, Y., & Kaku, I. (2012). The activity-based aggregate
production planning with capacity expansion in manufacturing systems.
Computers & Industrial Engineering, 62(2), 491503.
Zimmermann, H.-J. (1978). Fuzzy programming and linear programming with
several objective functions. Fuzzy Sets and Systems, 1(1), 4555.

You might also like