Expert Systems with Applications 41 (2014) 30693077

Contents lists available at ScienceDirect

Expert Systems with Applications

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

A modied particle swarm optimization for aggregate production

planning
Shih-Chang Wang a,, Ming-Feng Yeh b,1
a
Department of Business Administration, Lunghwa University of Science and Technology, Taoyuan, Taiwan
b
Department of Electrical Engineering, Lunghwa University of Science and Technology, Taoyuan, Taiwan

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

Keywords:
Particle swarm optimization (PSO)
Aggregate production planning (APP)
Integer linear programming model
Particle swarm optimization (PSO) originated from bird ocking models. It has become a popular research
eld with many successful applications. In this paper, we present a scheme of an aggregate production
planning (APP) from a manufacturer of gardening equipment. It is formulated as an integer linear pro-
gramming model and optimized by PSO. During the course of optimizing the problem, we discovered that
PSO had limited ability and unsatisfactory performance, especially a large constrained integral APP prob-
lem with plenty of equality constraints. In order to enhance its performance and alleviate the deciencies
to the problem solving, a modied PSO (MPSO) is proposed, which introduces the idea of sub-particles, a
particular coding principle, and a modied operation procedure of particles to the update rules to regu-
late the search processes for a particle swarm. In the computational study, some instances of the APP
problems are experimented and analyzed to evaluate the performance of the MPSO with standard PSO
(SPSO) and genetic algorithm (GA). The experimental results demonstrate that the MPSO variant provides
particular qualities in the aspects of accuracy, reliability, and convergence speed than SPSO and GA.
2013 Elsevier Ltd. All rights reserved.

1. Introduction
Aggregate production planning (APP) is an important technique
in Operations Management. Other essential approaches, such as
master production scheduling (MPS), capacity requirements plan-
ning (CRP) and material requirements planning (MRP), are closely
associated with it. APP is medium-term capacity planning which
determines ideal levels of workforce, production, inventory, sub-
contracting, 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 resource (Al-E-Hashem,
Aryanezhad, & Sadjadi, 2012; Graves, 2002; Stevenson, 2009). As
the name suggests, APP solves problems involving aggregate deci-
sions. It determines aggregate capacity level in factories for a given
amount of periods, while without determining the quantity of each
individual stock-keeping unit will be produced. The level of details
makes APP a useful tool for thinking about decisions with an inter-
mediate time frame that is too early to determine production levels
Corresponding author. Address: Department of Business Administration, Lung-
hwa University of Science and Technology, 300, Sec. 1, Wanshou Rd., Guishan
District, Taoyuan County 33306, Taiwan. Tel.: +886 2 82093211x6513; fax: +886 2
82093211x6510.
E-mail addresses: scwang@mail.lhu.edu.tw (S.-C. Wang), mfyeh@mail.lhu.
edu.tw (M.-F. Yeh).
1
Department of Electrical Engineering, Lunghwa University of Science and
Technology, 300, Sec. 1, Wanshou Rd., Guishan District, Taoyuan County 33306,
Taiwan. Tel.: +886 2 82093211x5501; fax: +886 2 82094650.
by stock-keeping unit and too late to arrange for additional capac-
ity. The goal of APP is normally to meet forecasted uctuating de-
mand requirements during a specic period in cost-effective
manner. Typical costs include the costs of production, inventory,
subcontracting, backlog, payroll, hiring, layoff, regular-time, and
overtime (Silva, Figueira, Lisboa, & Barman, 2006).
1.1. Literature on APP domain
Many APP models and solutions with various degree of sophis-
tication have been introduced since early 1950. The pioneers of the
eld, Holt, Modigliani, and Simon (1955) and Holt, Modigliani, and
Muth (1956), initially revealed the importance and obstacles of this
domain, and focused on the resolution of the aggregate planning
problem. They formalized and quantied an aggregate problem
by using a quadratic approximation to the criterion function
involving costs of inventory, overtime, and employment. They also
calculated a generalized optimal solution of the problem in the
form of a linear decision rule, commonly known as the LDR model.
The proposed approach was applied to a paint factory to generate a
production plan by using a quadratic approximation to the actual
operational costs of the factory. Hanssmann and Hess (1960)
developed a model based on the linear programming approach
using a linear cost structure of decision variables. It focused on
the resolution that minimizing the total cost of regular payroll
and overtime, hiring and layoffs, inventory and shortages incurred

0957-4174/$ - see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.eswa.2013.10.038
3070 S.-C. Wang, M.-F. Yeh / Expert Systems with Applications 41 (2014) 30693077
during a given planning horizon. Lanzenauer and Haehling (1970)
extended the model of Hanssmann and Hess (1960) to a multi-
product, multi-stage production system, in which optimal disag-
gregate decisions can be made under capacity constraints. Rakes,
Franz, and James Wynne (1984) presented a chance-constrained
goal programming approach to the problem. It is a special case of
stochastic programming to production scheduling which incorpo-
rates probabilistic product demand. A sophisticated overview of
earlier research is given in Nam and Logendran (1992). They com-
piled the research literature of APP that is consisted of 140 journal
articles and 14 books from 17 journals, presenting a classication
scheme and summarizing various existing techniques into a frame-
work. The techniques of those researches range from simple graph-
ical methods to more sophisticated search, switching-heuristic and
other dynamic methods, can be broadly categorized into two
types: those that guarantee an exact optimal solution and those
that do not.
In recent decades, depending on more assumptions made and
advanced modeling approaches invented, the APP problem has be-
come quite complex and large scale. There is a trend in the re-
search community to solve the large complex problems by using
modern heuristic optimization techniques. This is mainly due to
the time-consuming and unsuitability of classical techniques in
many circumstances.
Paiva and Morabito (2009) proposed an optimization model to
support decisions in the APP of sugar and ethanol milling compa-
nies. The model is a mixed integer programming formulation based
on the industrial process selection and the production lot-sizing
model. Also, in their APP real case study, the application of the
model results in 12,306 variables, where 5796 are binary and
6902 constraints. Sillekens, Koberstein, and Suhl (2011) presented
a mixed integer linear programming model for an APP problem of
ow shop production lines in automotive industry. In contract to
traditional approaches, the model considered discrete capacity
adaptions which originated from technical characteristics of
assembly lines, work regulations and shift planning. A solution
framework containing different primal heuristics and preprocess-
ing techniques is embedded into a decision support system. Zhang,
Zhang, Xiao, and Kaku (2012) built a mixed integer linear program-
ming model which characterize an APP problem with capacity
expansion in a manufacturing system including multiple activity
centers. They used a heuristic method based on capacity shifting
with linear relaxation to solve the problem. Ramezanian, Rahmani,
and Barzinpour (2012) considered multi-period, multi-product and
multi-machine systems with setup decisions, developed a mixed
integer linear programming model for general two-phase APP sys-
tems. Due to the NP-hard class of the APP model, they imple-
mented a genetic algorithm and tabu search for solving the model.
In addition to the integer linear programming models of APP
problems, more complicated models have also been proposed.
Mirzapour Al-E-Hashem, Malekly, and Aryanezhad (2011)
addressed a multi-site, multi-period, multi-product APP problem
under uncertainty in supply chain, proposed a robust multi-
objective mixed integer nonlinear programming model, consider-
ing two conicting objectives simultaneously to deal with the
problem. In their research, cost parameters and demand uctua-
tions are subject to uncertainty, then the problem can transform
into a multi-objective linear one, and to be solved as a single-
objective mixed integer programming model applying the
LP-metrics method. Adil Baykasoglu and Gocken (2010) presented
a fuzzy multi-objective APP model and proposed a direct solution
method based on ranking methods of fuzzy numbers and tabu
search to solve the model. Sakalli, Baykoc, and Birgoren (2010)
discussed an APP model with possibilities for a blending problem
in a brass factory. Their possibilistic linear programming model is
solved by fuzzy ranking concept relaxed by using Either or
constraints. The approach successfully solved the multi-blend
problem for brass casting and determines the optimum raw mate-
rial purchasing policies.
1.2. Literature implemented PSO on various elds
Over the past decade, a number of computational swarm-based
systems have been developed. Some of them become very popular
optimization techniques in many domain researches soon after-
wards. One is particle swarm optimization (Kennedy & Eberhart,
1995a), abbreviated as PSO. PSO originated from bird ocking
models and has become an exciting new research eld still in its
infancy compared to other paradigms in articial intelligence.
Baltas, Tsafarakis, Saridakis, and Matsatsinis (2013) introduced
a PSO variant to a service design and diversication problem. They
designed and implemented genetic algorithm and PSO to stated-
preference data derived from conjoint consumer preferences for
service attributes in a retail setting. Their method has valuable
implications for managers aiming to improve how they design
their services. Tsafarakis, Saridakis, Baltas, and Matsatsinis (2013)
presented a new hybrid PSO approach to design an optimal indus-
trial product line. The hybrid PSO searches for an optimal product
line in a large design space which consists of discrete and contin-
uous design variables. The approach illustrated through an applica-
tion to a simulated dataset of industrial cranes. It also yielded
important implications for strategic customer relationship and
production management. Ramazanian and Modares (2011) intro-
duced a multi-objective goal programming model for a multi-prod-
uct multi-step multi-period APP problem in the cement industry.
The model was reformulated as a single objective nonlinear pro-
gramming model. It was solved by using the expanded objective
function method and a proposed PSO variant whose inertia weight
was set as a function. The simulation comparing with GA in the -
nal showed that PSO gains satisfactory results than GA.
With many successful applications in various domain problems,
PSO has shown that it is a considerably promising, efcient and ro-
bust technique for practical applications. For examples, PSO had
been successfully applied to scheduling problems (Chen, 2011;
Liao, Tseng, & Luarn, 2007), game theory problems (Lung & Dumit-
rescu, 2009; Pavlidis, Parsopoulos, & Vrhatis, 2005), optimization
on continuously changing environments (Parsopoulos & Vrahatis,
2001), and detection of periodic orbits (Skokos, Parsopoulos, Patsis,
& Vrahatis, 2005).
Although there are many applications implemented PSO on var-
ious elds, however, we seldom found it applying to the APP eld.
The reason may be attributed to that PSO was originally introduced
for unconstrained and continuous optimization problems. Its oper-
ations imply the existence of unrestricted and continuous explora-
tions in search space, which may have limited ability in dealing
with constrained integral APP problems. Afterwards, during the
process of optimizing APP problems, we did nd that PSO gains
limited ability and inefciency in dealing with the problems, espe-
cially a large constrained integral APP problem with plenty of
equality constraints. Therefore, in response to ease these short-
comings, we developed an effective modied mechanism for PSO,
which introduced the concept of sub-particles, a particular coding
principle, and a modied operation procedure of particles to the
update rules to regulate the search processes for a particle swarm.
And we found that the MPSO variant gains satisfactory perfor-
mance in the aspects of accuracy, reliability, and convergence
speed than SPSO and GA. Also, there are advantages of the MPSO
to the APP problem solving than other approaches: (i) only a few
parameters need to be adjusted; (ii) be able to speed up the con-
vergence to the optimal solution; (iii) can be applied to optimize
most of APP problems.
 S.-C. Wang, M.-F. Yeh / Expert Systems with Applications 41 (2014) 30693077
Besides, different companies or industries have various charac-
teristics of aggregate decisions in aggregate production planning.
Increasingly sophisticated parameters and assumptions into APP
models only make them hindering and impracticable for practical
applications. Therefore, a comprehensive APP model that is easy
to expand and adjust, and to be optimized by the modied mech-
anism of PSO we developed, is what we want to discover in this
study.
In this paper, we rst propose a general APP model from a real-
world problem, which is organized and formulated as an integer
linear programming model. During the course of implemented
PSO to the constrained and discrete APP problem, we found that
PSO has some imperfections, especially a large constrained integral
APP problem with plenty of equality constraints. Then a modied
scheme of PSO is proposed. The discussions of all the processes
about proposing an APP model, optimizing the model by PSO, its
difculties and ndings, the proposed MPSO, the introduced sub-
particles, and its examinations, etc., are the contribution of the
paper.
The rest of the paper is organized as follows. In Section 2, we
introduce some variants of PSO from early standard to contempo-
rary hybrids. Section 3 presents a general APP problem from a real-
world manufacturer of gardening equipment. It is formulated as an
integer linear programming model. In Section 4, the optimization
of the APP model by PSO is revealed. The features of the proposed
modied PSO for optimization of the APP model are claried in
Section 5. The experiments and comparisons for the modied
PSO with standard PSO and GA are displayed in Section 6, some
conclusions and discussions are also addressed in the nal
Section 7.
2. Particle swarm optimization
The precursors of particle swarm optimization were simulators
of social behavior for visualizing bird ocks. After realizing the po-
tential of their simulation models to perform optimization, the
early version of PSO was published by Eberhart and Kennedy in
1995, dened by the following equations (Eberhart & Kennedy,
1995; Eberhart, Simpson, & Dobbins, 1996; Kennedy & Eberhart,
1995b):
mij t 1 mij t c1 R1 pij t  xij t c2 R2 pgj t  xij t 1
xij t 1 xij t mij t 1
where i = 1, 2, . . . , N and j = 1, 2, . . . , n. N be the number of particles
and n the dimension of search space A  Rn; t denotes the iteration
counter; R1 and R2 are random variables uniformly distributed
within [0, 1]; c1, c2 are weighting factors called the cognitive and so-
cial parameter respectively (also called acceleration coefcients). A
swarm of particles is dened as a set S = {x1, x2, . . . , xN} represented
as potential solutions of an objective function to be optimized.
The position of ith particle be xi = (xi1, xi2, . . . , xin)T 2 A and each par-
ticle is assumed to move within the search space A iteratively. This
is possible by adjusting their position using a proper position shift
vi = (vi1, vi2, . . . , vin)T, called velocity, which is also adapted iteratively
to render particles capable of potentially visiting any region of A.
pij(t) represents the position with the best tness for the ith particle
in the jth dimension of search space, and pgj(t) the best position dis-
covered by the whole particles.
During iterative processes, the trajectory of each particle is ad-
justed by dynamically altering the velocity of its own according to
its own best ying experience (pBest) and the best ying experi-
ence of the swarm (gBest) in search space. In (1), pi(t) represents
the position with the best tness found so far for ith particle, and
pg(t) the best position discovered by whole particles. The rst part
3071
of (1) is the previous velocity, which provides the necessary
momentum for particles to roam around the search space. The sec-
ond 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 com-
ponent, expresses the collaborative effect of particles in discover-
ing the global optimal solution. The social component always
pulls particles to the global best position found so far.
The early PSO performed satisfactorily for simple optimization
problems (Eberhart & Shi, 2001; Kennedy & Eberhart, 2001). How-
ever, its crucial deciencies were revealed as soon as that it was
applied on complex optimization problems with large search
spaces. Therefore, renements developed to ease the deciencies
of the early PSO model are discussed and veried by several
researchers, and later, a new parameter, w, called inertia weight,
was introduced into (1), resulting in a new PSO variant (Eberhart
& Shi, 1998; Shi & Eberhart, 1998):
mij t 1 wmij t c1 R1 pij t  xij t c2 R2 pgj t  xij t 3
The rest of the equation and parameters remains the same as
the early PSO model. A very common choice is the initialization
of w to a value slightly greater than 1.0 to promote explorations
in early optimization stages, and a linear decrease towards a posi-
tive lower bound to eliminate oscillatory behaviors in later stages.
In general, a linearly decreasing scheme for w can be mathemati-
cally described as a linearly decreasing time-dependent inertia
weight as follows:
t
wt wmax  wmax;  wmin 4
T max
where t stands for the iteration counter and Tmax be the total num-
ber of iterations; wmax and wmin are the desirable upper and lower
bounds of w, usually set to 0.9 and 0.4, respectively (Shi & Eberhart,
1998; Shi & Eberhart, 1999). Also, researches presented that the PSO
variant with a nonlinearly varying inertia weight was applicable
and successfully applied to the optimization of practical problems
(Venter & Sobieszczanski-Sobieski, 2004). The amazing perfor-
mance improvement of the PSO variant gained by using inertia
weight with velocity clamping, rendered it the most popular PSO
approach for a few years. Then it is so-called standard PSO (SPSO).
However, although particles were able to avoid swarm explo-
2 sion and converge around the best position, they were still easily
getting trapped in local minima, especially in complex, multimodal
problems. This deciency can be attributed to the global informa-
tion exchange scheme which allows each particle to know the
overall best position instantly at each iteration. All particles as-
sume new positions in regions related to the same overall best po-
sition, reducing the exploration capabilities of the swarm. In order
to overcome the problem, the concept of neighborhoods was intro-
duced. The main idea of it was the reduction of the global informa-
tion exchange scheme to local one, where information is diffused
only in small parts of the swarm, called neighbors, at each itera-
tion. The Eq. (3) then be modied as:
mij t 1 wmij t c1 R1 pij t  xij t c2 R2 pgi j t  xij t 5
where gi denotes the index of the best particle in the neighborhood
of xi. The only difference between Eqs. (3) and (5) is the index of the
best position in the second parenthesis. According to Eq. (5), a
particle will move towards its own best position and the best
position of its neighborhood, instead of the overall best position
in Eq. (3). The introduction of neighborhoods enhanced the perfor-
mance of PSO signicantly, offering a boost to research towards the
development of more competitive and sophisticated variants that
3072 S.-C. Wang, M.-F. Yeh / Expert Systems with Applications 41 (2014) 30693077

incorporated all presented concepts so far. To distinguish the Sets and indexes
approaches, the PSO variant that employed the overall best position nped number of periods in planning horizon
is called the global PSO variant, denoted as gBest, while the one with nwkr number of worker types
strictly smaller neighborhoods is called the local PSO variant, t index of periods in planning horizon,
denoted as lBest. t = 1, 2, . . . , nped
The good performance of the presented PSO variants attracted j index of worker types, j = 1, 2, . . . , nwkr
the interest of the scientic community. The blossoming research Parameters
prompted researchers to investigate the theoretical properties of cprd unit material cost
PSO, offering better understanding of its operations and dynamics, cinv inventory holding cost per unit of
as well as the mathematical traits for proper parameter congura- product at each period
tion. Clerc and Kennedy (2002) considered different generalized cblg backlog cost per unit of product at
PSO models and performed a dynamical system analysis of their each period
convergence. The analysis offered a solid theoretical background csct subcontracting cost per unit of product
to the algorithm. They established an investigated model as the de- chir hiring and training cost per worker of
j
fault contemporary PSO variant by transforming Eq. (1) as: type j
  clof layoff cost per worker of type j
mij t 1 v mij t c1 R1 pij t  xij t c2 R2 pgj t  xij t 6 j
crt
j
regular time labor cost per hour of
worker type j at each period
where v is called constriction coefcient, while the rest of the equa-
tion and parameters remains the same as the early PSO model. The cot
j
overtime labor cost per hour of worker
type j at each period
variant is distinguished in literature due to its theoretical properties
hj
prd labor hours of a worker type j required
that imply the following explicit selection of its parameters (Clerc &
for producing a unit of product at each
Kennedy, 2002):
period
rt labor hours of a worker type j provided
2 hj
v p 7 for regular time at each period
j2  u  u2  4uj
hot legal labor hours of a worker limited
for overtime at each period
where u = c1 + c2 and u > 4. Based on this equation, the settings,
dt forecasted demand requirements at
v = 0.729 and c1 = c2 = 2.05, are currently considered as the default
period t
parameter set of the constriction coefcients of the variant.
Decision variables
Wjt workforce sizes of worker type j at
3. Integer linear programming model for APP period t
W hir workforce sizes of worker type j hired
jt
In this section, a scenario that formally states a general APP at the beginning of period t
problem from the extension of Red Tomato Tools, a manufacturer W lof workforce sizes of worker type j laid
jt
of gardening equipment with highly seasonal demand require- off at the beginning of period t
ments in Mexico (Chopra & Meindl, 2007), is presented and formu- Hot
jt
number of overtime hours of worker
lated as an integer linear programming model. Red Tomato Tools is type j worked at period t,
a manufacturer of gardening equipment, whose operations consist Pprd units of product produced at period t
t
of the assembly of purchased parts into a multipurpose gardening Pinv units of inventory of product at the
t
tool. The demand requirements for Red Tomato Tools gardening end of period t
tools are highly seasonal. Spring is the peak season as people plant Pbt lg units of product backlogged at the end
their gardens. Red Tomato Tools has decided to use aggregate pro- of period t
duction planning to overcome the obstacle of seasonal demand and Psct
t
units of product subcontracted at
maximize prots. The options Red Tomato Tools has for handling period t
the seasonality are as follows:

(1) Its production strategies to satisfy forecasted uctuating

demand requirements are adding workers or subcontracting 3.1. Objective function
some works out during the peak season, building up inven-
tories during the off season, and developing up backlog The demand requirements at period t, dt , are given by forecast
orders which will be delivered late to customers, as the and orders in hand. The objective function aims to minimize the to-
majority of practical use. tal cost incurred within a planning horizon. These costs include
(2) Its capacity of production operations is primarily deter- regular time labor cost, overtime labor cost, hiring cost, layoff cost,
mined by the total working labor hours. holding inventory cost, backlog cost, production cost, and subcon-
(3) As the reason of labor rules, workers worked in overtime tracting cost, etc., as follows:
cannot more than the limit of legal labor hours at each
period. (1) Regular time labor cost
XX rt
crtj  hj  W jt 8
The rst step in constructing the APP model of the problem is to t j
identify demand requirements, related costs and labor hours, and
the set of decision variables whose values are to be determined (2) Overtime labor cost
XX
as part of the planning. Some involved notations used to formulate cot ot
j  H jt 9
the problem are dened as follows: t j
 S.-C. Wang, M.-F. Yeh / Expert Systems with Applications 41 (2014) 30693077 3073

(3) Cost of hiring and layoff crete optimization problems. Thus, an adjustment mechanism is
XX hir lof lof
needed for PSO to alleviate the deciencies for the APP model.
chir
j  W jt c j  W jt 10 To rst issue, in order to solve the APP model by PSO restricted
t j
within plenty of equality and inequality constraints, we introduce
(4) Cost of holding inventory and backlog the penalty function method, which expands the objective function
X  of the APP model by importing an appropriate penalty function
v
cinv  Pin
t cb lg  Pbt lg 11 converted from all equality and inequality constraints in the mod-
t el. The original model is then reconstructed as a new model which
(5) Cost of production and subcontracting has expanded objective function with no constraints thereafter,
stated as follows:
X 
cprd  Pprd csct  P sct : 12 The APP model with many equality and inequality constraints
t t
t can be formally expressed as the formulation:

min f x 18
The total cost within a planning horizon can be formulated as s:t: g i x 6 0; i 1; . . . ; p 19
the sum of previous costs from (8) to (12). Thus, the objective func-
hj x 0; j 1; . . . ; q 20
tion can be expressed as follows:
XX rt
XX ot then it can be reconstructed as a new model, whose expanded
min crtj  hj  W jt cot
j  H jt
t t
objective function with no constraints is F(x), by importing a pen-
j j
XX hir lof lof
 alty function Px as:
chir
j  W jt c j  W jt
t j
min Fx f x Px 21
X v
 X 
cinv  Pin
t cb lg  Pbt lg cprd  P prd
t csct  Psct
t 13 where the penalty function P(x) is organized as:
t t
c
Px ckg xk2 khxk2 22
2
3.2. Constraints where g x g
T

1 x; . . . ; g p x ; g i x maxg i x; 0; hx
T
h1 x; . . . ; hq x , and c be the penalty parameter.
The values of decision variables in the objective function are Therefore, the APP model presented in Section 3 can be recon-
integer and cannot be arbitrarily set. They are subject to a variety structed as a new expanded objective function with no constraints,
of constraints. The next step in setting up the aggregate production expressed as follows:
planning model is to dene constraints linking decision variables XX XX XX 
rt ot hir lof lof
clearly, as follows: min crtj  hj  W jt cot
j  Hjt chir
j  W jt c j  W jt
t j t j t j
X  X 
(1) Workforce, hiring, and layoff constraints cinv  P inv b lg
 Pbt lg cprd  Pprd sct
 Psct
t c t c t
t t
W jt W jt1 W hir lof 8 "
jt  W jt 14 !#2
<X X h
rt ot
prd j  W jt Hjt
(2) Capacity constraints c max Pt  ;0
: t j hj
prd

X hrtj  W jt Hot )
jt XXh  i2
Pprd
t 6 prd
15 ot ot
max Hjt  h  W jt ; 0
j hj
t j
(
(3) Inventory balance constraints c XX 2
 W jt  W jt1  W hir lof
jt W jt
v b lg inv 2
Pin prd
t1 P t Psct
t dt P t1 P t  Pbt lg 16 t j
)
X inv b lg inv
2
(4) Overtime limit constraints Pt1 Pprd sct blg
t P t  dt  P t1  P t P t 23
t
ot
Hot
jt 6 h  W jt : 17
To second issue, an alteration method is adopted which transforms
an integral problem into its corresponding continuous problem. The
In the model, one can easily add parameters and constraints alteration is that each particle component is rounded to nearest
which limit the amount purchased from subcontractors at each integral value in PSO update rules. Hence, PSO also retained its
period, or the maximum number of employees to be hired or laid dynamics while produced solutions are integer.
off. Any other limited backlogs or inventories can also be accom- More specically, let xi = (xi1, . . . , xin)T be ith particle of the
modated in the model for practical need. swarm. Then its rounded counterpart zi = (zi1, . . . , zin)T can be de-
ned, where zij bxij 0:5c, and zij is used to replace xij. Therefore,
4. Optimization of APP model by PSO the components of a particle retain integer-valued, although the
particles velocity can be real-valued.
PSO was originally introduced for unconstrained and continu-
ous optimization problems. Its operations imply the existence of 5. The modied PSO
unrestricted and continuous explorations in search space. But to
the problem of optimizing the constrained integral APP model, The approach of the penalty function method with the alter-
there are two issues to limit PSOs performance raise: (i) restricted ation method can alleviate the deciencies of PSO to solve the con-
within plenty of equality and inequality constraints and (ii) explor- strained integral APP model. But during the course of optimizing
ing in integral subspace of n-dimensional Euclidean space. That is the reconstructed model, we found that the approach still gains
to say, PSO has limited ability in dealing with constrained and dis- unsatisfactory performance, especially a large constrained integral
3074 S.-C. Wang, M.-F. Yeh / Expert Systems with Applications 41 (2014) 30693077

APP model with plenty of equality constraints. This defect might be
attributed to that the particles have great randomness exploring in
large multi-dimensional integral subspace. They are inefcient to
exploit the locations where meet all of the equality constraints
transformed. That inuences its performance to solve the APP
problem.
In order to improve the problem solving by PSO aforemen-
tioned, a modied scheme of PSO is proposed, stated as follows:
In SPSO, the position of ith particle xi = (xi1, xi2, . . . , xin)T is indi-
cated as a vector which presents a potential solution of the APP
model in n-dimensional search space, where n = nped  ndv, nped
be the number of periods in a planning horizon, and ndv the num-
ber of decision variables in the model. Each particle moves within
the whole n-dimensional search space, adjusting its trajectory by
dynamically adopting the velocity of its own, mi = (mi1, mi2, . . . , min)T,
START
Initialize the swarm
Evaluate the fitness for each particle
YES stopping
criterion ?
NO
END Update pBest and gBest
iteratively, according to the best searching experience of its own
and the swarm. The update rules are as Eqs. (2) and (3), and the
operations of the PSO is depicted in Fig. 1.
In order to alleviate the deciency of solving the constrained
integral linear programming model aforementioned, we dene a
modied scheme of PSO. The MPSO introduces the concept of
sub-particles, a particular coding principle, and a modied opera-
tion procedure of particles to the update rules to regulate the
search processes for a particle swarm. The main idea of the MPSO
is the components of ith particle partitioning into nped sub-particles
 T
sequentially as xi xTi1 ; xTi2 ; . . . ; xTip ; . . . ; xTinped . Each sub-particle,
xip , is formed to a n dv -dimensional vector as
 T
xip xip1 ; xip2 ; . . . ; xipd ; . . . ; xipndv represents a potential solution
of the ndv decision variables at pth period of time in planning
horizon. Each sequentially adjacent sub-particle shows successive
relationship of decision variables at a period of time in planning
horizon. The updates of the position and velocity are by sub-
particles sequentially to meet all the constraints in the APP model.
Until the last sub-particle within a particle is updated, the particle
then renews. The update rules of the MPSO are shown in Eqs. (24)
and (25), and the operations of the MPSO is also illustrated in Fig. 2.
mipd t 1 wmipd tc1 R1d pipd txipd tc2 R2d pgd txipd t 24
xipd t 1 xipd t mipd t 1 25
6. Computational study
6.1. Case description

The test scenario of the APP problem is illustrated through the

Update velocity vi and position xi extension of the Red Tomato Tools, a small manufacturer of gar-
dening equipment with manufacturing facilities in Mexico afore-
Fig. 1. Operations of PSO. mentioned in Section 3. Red Tomatos products are sold through
retailers in the US. Its operations consist of the assembly of pur-
chased parts into a multi-purpose gardening tool. Because of
unlimited equipment and space required for assembly operations,
START
its capacity is determined mainly by the size of workforce.
The demand requirements for Red Tomatos gardening tools
Initialize the swarm from consumers are highly seasonal. Spring is the peak season as
people plant their gardens. The options it has for handling the sea-
sonality are adding workers or subcontracting some works out
Evaluate the fitness for each particle during the peak season, building up inventories during the off sea-
son, or building up backlog orders which will be delivered late to
customers. Its capacity of production operations is determined pri-
YES stopping marily by the total working labor hours. It also has no limits on
criterion ? subcontracting, inventories, stockouts, and backlogs. Because of la-
NO bor rules, no workers could work more than the limit of legal labor
hours at each period.
END Update pBest and gBest

6.2. Evaluation of the MPSO with SPSO and GA

p=1
In order to evaluate the performance of the MPSO algorithm for
the APP problem, we made use of an articial dataset that 8 in-
Update velocity vip and position xip stances of the problems, 4 are small-sized and 4 large-sized, with
large equality constraints are generated and experimented. These
instances with algorithms of MPSO, SPSO and GA are coded by
p = p +1
MATLAB R2012a and executed on a PC with Intel Pentium Dual
1.6 GHz CPU, 2 GB RAM, and Windows 7 operating system. For
NO comparing these algorithms performance, these instances are also
p > nped
coded by LINGO optimization solver of integer linear programming
YES to discover their global optimum as lower bound (LB), and to com-
pare with corresponding algorithm results. In order to present the
Fig. 2. Operations of MPSO. diversity of performance between the global optimum and objec-
 S.-C. Wang, M.-F. Yeh / Expert Systems with Applications 41 (2014) 30693077 3075

tive values with the results of these algorithms, a quality measure- Table 3
ment called percent deviation, denoted as %Dev, is dened as the Statistical results of SPSO experiment.

following equation: Test SPSO

SN best Mean Std. Dev. %Dev Suc. rate Exe. time
OV MPSO;  LBLINGO (%) (s)
SPSO or GA
%Dev  100% 26
LBLINGO 1 206,780 207526.86 1675.01 0.45 78 2.15E03
2 423,339 429356.18 7001.45 1.67 98 6.91E02
3 650,312 658416.84 9217.12 1.25 92 4.76E02
where OVMPSO, SPSO or GA be the objective value derived from MPSO,
4 899,766 907372.10 9068.62 1.91 76 1.27E01
SPSO or GA, and LBLINGO the lower bound stemmed from LINGO. 5 1,101,830 1,17306.50 13333.58 2.54 64 4.23E01
In the experiments of both PSO variants, each instance is run for 6 1,308,726 1326918.86 17345.46 1.40 50 3.06E01
50 trials. The acceptance level of objective value of a trial is set to 7 1,532,174 1561165.10 21997.87 2.18 48 8.51E01
5% above its corresponding lower bound. If the objective value of a 8 1,787,020 1825095.26 25276.28 2.22 76 5.91E01
Avg 13114.42 1.703 72.8 3.02E01
trial is in the midst of acceptance level, then the trial is called a
successful trial. The statistical result of each instance is collected
from the successful trials within 50 trials of execution. The linearly
Table 4
decreasing time-dependent inertia weight is set from 0.9 to 0.4, Statistical results of GA experiments.
and acceleration coefcients 2.05. The number of particles and iter-
Test GA
ations are 30 and 500 respectively. As for GA experiments, we set
SN best Mean Std. Dev. % Suc. rate Exe. time
the parameters of the population size, the crossover rate, the Dev (%) (s)
mutation rate and the generations as 100, 0.9, 0.01 and 1500
1 206,987 208873.06 1907.92 0.62 74 3.00E03
respectively. These experimentally statistical results obtained by 2 422,830 430781.41 7839.95 2.29 96 1.22E01
LINGO, MPSO, SPSO and GA are shown in Tables 14. Each instance 3 650,950 664564.64 10109.27 1.46 88 5.40E02
of the APP problem is given a test serial number. Table 1 includes 4 886,921 906666.42 17150.53 2.57 70 2.37E01
key parameters of each test instance of APP problems, containing 5 1,083,097 1119950.23 15884.84 3.18 58 5.32E01
6 1,311,880 1340597.68 18787.90 1.51 44 3.30E01
the number of periods and worker types. It also displays the global
7 1,528,130 1560014.09 24075.97 2.79 42 9.67E01
optimum as lower bound experimented by LINGO in the tests. 8 1,788,829 1838640.51 27184.60 2.53 68 6.76E01
Tables 2 and 4 show the statistical results of MPSO, SPSO and GA Avg 15367.62 2.118 67.5 3.65E01
in the test, including the best, the mean, the standard deviation,
and the percent deviation of their object values. The successful rate
(%) and execution time (s) are also included in these tables. Table 5
These statistical results show that MPSO, SPSO and GA provide Comparative performances of MSPO, SPSO and GA.
the best objective values of each instance that are differentiated to Std. Dev. % Dev Suc. rate (%) Exe. time (s)
the corresponding optimum. All of them also have disparities
MPSO 100% 100% 100% 100%
among mean, standard deviation, percent deviation and execution SPSO 118% 130% 90% 125%
time to some degree. The average percentages of successful rate for GA 139% 162% 84% 151%
MPSO and SPSO are decient but acceptable, as we early men-
tioned that the reason might be attributed to the great random
particles needing more efforts to exploit in a large multi-

Table 1
Parameters and experimental results of LINGO.

Test SN Num. of periods Num. of worker type LINGO

Optimum
1 3 1 206,583
2 6 2 422,275
3 9 3 650,275
4 12 4 890,300
5 15 5 1,089,628
6 18 7 1,308,597
7 21 9 1,527,740
8 24 11 1,785,340

Table 2
Statistical results of MPSO experiments.

Test MPSO
SN Best Mean Std. Dev. %Dev Suc. Exe.
Rate(%) Time(sec) Fig. 3. Variation of objective values over iterations.
1 206,780 207,164.14 1,458.85 0.28 90 1.32E03
2 422,662 426,867.29 6,267.97 1.08 98 1.64E02 dimensional integral subspace, where exists plenty of equality
3 650,306 657,218.82 8,444.79 1.06 96 4.17E02 constraints. Also, the execution time is effective in these ap-
4 892,897 901,609.60 1,095.53 1.27 82 1.86E02
proaches. The comparative performances are summarized in Ta-
5 1,091,372 1,110,852.40 10,968.99 1.94 74 3.18E01
6 1,308,994 1,325,803.45 16,111.17 1.31 62 2.85E01 ble 5. The convergence characteristics over iterations for these
7 1,527,860 1,552,053.30 20,161.75 1.59 56 7.44E01 approaches is depicted in Fig. 3, that each curve represents the var-
8 1,787,031 1,820,037.64 23,696.55 1.94 86 5.12E01 iation of objective values over iterations for corresponding
Avg 11,025.70 1.309 80.5 2.42E01
approach.
3076 S.-C. Wang, M.-F. Yeh / Expert Systems with Applications 41 (2014) 30693077
It is worth mentioning that in the average of standard deviation,
percent deviation, successful rate, and execution time, MPSO pro-
vides better performance than SPSO, and greater than GA. Appar-
ently, as a result of sub-particles introduced into the MPSO
variant, the searching discrepancies of particles in standard devia-
tion and percent deviation gain decreases of 18% and 30% respec-
tively than SPSO. The execution time also gets a decrease of 25%
while the successful rate an increase of 10% than SPSO. Also, these
performances of MPSO are greater than GA.
Therefore, we conclude that the MPSO variant possesses partic-
ular qualities in accuracy, reliability, and convergence speed for
optimizing the APP problem than SPSO and GA.
7. Conclusions and discussions
7.1. Conclusions
In this paper, we rst explore the important issues of PSO from
early precursors to contemporary standard variants. Some hybrids
and variants are also presented for comparison. In next context, we
propose a general APP model which states a real-world APP prob-
lem from the extension of a manufacturer of gardening equipment
in Mexico. The illustration is organized and formulated as an inte-
ger linear programming model which can be easily expanded with
adding parameters, decision variables, and constraints as needed
for practical use in industries. To the optimization of the APP mod-
el, we rarely found PSO applying to the eld of APP problems in re-
searches. We also discovered that PSO has some imperfections in
optimizing discrete constrained APP problems, especially a large
constrained integral problem with plenty of equality constraints.
Therefore, in succeeding context, a modied scheme of PSO is pro-
posed which introduces the concept of sub-particles to the update
rules of PSO to alleviate the deciency of solving the constrained
integral model. In the nal, 8 instances of APP problems with large
equality constraints are implemented and experimented by MAT-
LAB and LINGO, to evaluate the performance of MPSO, SPSO and
GA. The experimental results show that the MPSO variant gains
particular qualities in accuracy, reliability, and convergence speed
than SPSO and GA. The characteristics of this study comparing with
recent researches are depicted in Table 6.
7.2. Discussions
Although PSO has become a promising technique in solving
many practical applications and there are many successful applica-
tions implemented PSO on various elds. However, we seldom
found it applying to the eld of APP problems. The reason may
be attributed to that PSO was originally introduced for uncon-
strained and continuous problems. Its operations imply the exis-
tence of unrestricted and continuous explorations in search
space. It has limited ability in dealing with constrained integral
APP problems, especially a large constrained integral problem with
plenty of equality constraints. Therefore, in response to the
Table 6
Characteristics of this study with current researches.
Research Model Objective
This paper Integer Single
Ramezanian et al. (2012) mixed integer Single
Zhang et al. (2012) Mixed integer Single
Al-E-Hashem et al. (2012) Mixed integer Multiple
Sillekens et al. (2011) Mixed integer Single
Paiva and Morabito (2009) Mixed integer Single
Baykasoglu (2001) Goal Multiple
Koroec, Bole, and Papa (2013) Goal Multiple
Ramazanian and Modares (2011) Goal Multiple
shortcomings, we developed an effective modied mechanism for
PSO to alleviate the deciencies. And the experimental studies re-
vealed that the MPSO gains satisfactory performance in the aspects
of accuracy, reliability, and convergence speed than SPSO and GA.
The discussions of all the processes about the proposed APP model,
optimizing the APP model by PSO, the difculties and ndings, the
proposed MPSO, the introduced sub-particles, and its examina-
tions, etc., are the focus and the contribution of this study.
Some research implications and managerial implications have
to be mentioned when implementing these works in the future:
(1) On the technical level, the variations that this study differs
from other PSO applications are the idea of the sub-particles,
the particular coding principle and the modied operation
procedure of particles, as shown in Section 5 and Figs. 1
and 2. In order to alleviate the deciency of PSO to the APP
problem solving, the MPSO introduced the concept of sub-
particles, a particular coding principle, and a modied oper-
ation procedure of particles to the update rules to regulate
the search processes for a particle swarm.
(2) The MPSO belongs to the population-based search tech-
nique, just like GA. But the differences of both are that the
MPSO has no manipulations of crossover, mutation, and
selection. It is based on the particular coding principle, the
operation procedure, and the speed of its sub-particles to
determine the search. Unlike others, these sub-particles
have memories for their search. Also, there are three advan-
tages of the MPSO to the APP problem solving than others:
(i) only a few parameters needed to be adjusted; (ii) be able
to speed up the convergence to the optimal solution; (iii) can
be applied to optimize most of APP problems.
(3) Different companies or industries have various characteris-
tics of aggregate decisions in aggregate production planning.
Increasingly sophisticated parameters and assumptions into
an APP problem only make it hindering and impracticable
for practical applications. Therefore, this study proposed a
comprehensive APP model that is easy to adjust and to be
optimized by algorithms.
(4) Aggregate production plans are mainly based on forecasts of
future demand requirements. It is important that these fore-
casts are always wrong to some degree in practice. The
aggregate production planning needs to have some exibil-
ity built into it if it is to be more useful. We recommend that
a manager had better to perform sensitivity analysis on the
input parameters after implementing the work. It will
enable the manager to choose the best solution for the range
of possibilities that could occur in practical applications.
(5) Despite this study inferred the MPSO variant gaining
particular qualities than SPSO and GA. In order to ensure
the performance of the problem solving of it, we suggest
further researchers applying the MPSO to solve more
large complex APP problems and comparing to more
algorithms.
Linearity Goal Methodology Decision
Linear Min. cost PSO Determinist
Linear Min. cost Genetic & Tabu Determinist
Linear Min. cost Heuristic Determinist
Non-linear Min. cost & Max. satisfaction LP-metrics Uncertainty
Linear Min. cost DSS with Heuristics Uncertainty
Linear Min. cost MLP Determinist
Linear Min. cost Tabu Determinist
Linear non-linear Multiple Heuristic Determinist
Non-linear Multiple PSO Determinist
 S.-C. Wang, M.-F. Yeh / Expert Systems with Applications 41 (2014) 30693077 3077

References Lung, R., & Dumitrescu, D. (2009). Evolutionary multimodal optimization for Nash
equilibria detection. In N. Krasnogor, M. Melin-Batista, J. Prez, J. M. Moreno-
Vega, & D. Pelta (Eds.), Nature inspired cooperative strategies for optimization
Al-E-Hashem, S. M. J. M., Aryanezhad, M. B., & Sadjadi, S. J. (2012). An efcient
(NICSO 2008). Berlin Heidelberg: Springer, pp. 227237.
algorithm to solve a multi-objective robust aggregate production planning in an
Mirzapour Al-E-Hashem, S. M. J., Malekly, H., & Aryanezhad, M. B. (2011). A multi-
uncertain environment. International Journal of Advanced Manufacturing
objective robust optimization model for multi-product multi-site aggregate
Technology, 58, 765782.
production planning in a supply chain under uncertainty. International Journal of
Baltas, G., Tsafarakis, S., Saridakis, C., & Matsatsinis, N. (2013). Biologically inspired
Production Economics, 134, 2842.
approaches to strategic service design: Optimal service diversication through
Nam, S.-j., & Logendran, R. (1992). Aggregate production planning A survey of
evolutionary and swarm intelligence models. Journal of Service Research, 16,
models and methodologies. European Journal of Operational Research, 61,
186201.
255272.
Baykasoglu, A. (2001). MOAPPS 1.0: Aggregate production planning using the
Paiva, R. P. O., & Morabito, R. (2009). An optimization model for the aggregate
multiple-objective tabu search. International Journal of Production Research, 39,
production planning of a Brazilian sugar and ethanol milling company. Annals of
36853702.
Operations Research, 169, 117130.
Baykasoglu, A., & Gocken, T. (2010). Multi-objective aggregate production planning
Parsopoulos, K. E., & Vrahatis, M. N. (2001). Particle swarm optimizer in noisy and
with fuzzy parameters. Advances in Engineering Software, 41, 11241131.
continuously changing environments. In Articial intelligence and soft computing.
Chen, R.-M. (2011). Particle swarm optimization with justication and designed
Pavlidis, N. G., Parsopoulos, K. E., & Vrhatis, M. N. (2005). Computing Nash equilibria
mechanisms for resource-constrained project scheduling problem. Expert
through computational intelligence methods. Journal of Computation and
Systems with Applications, 38, 71027111.
Applied Mathematics, 175, 113136.
Chopra, S., & Meindl, P. (2007). Aggregate planning in a supply chain. In Supply chain
Rakes, T. R., Franz, L. S., & James Wynne, A. (1984). Aggregate production planning
management: Strategy, planning, and operation. Upper Saddle River, NJ: Pearson
using chance-constrained goal programming. International Journal of Production
Education.
Research, 22, 673684.
Clerc, M., & Kennedy, J. (2002). The particle swarm explosion, stability, and
Ramazanian, M., & Modares, M. (2011). Application of particle swarm optimization
convergence in a multidimensional complex space. IEEE Transactions on
algorithm to aggregate production planning. Asian Journal of Business
Evolutionary Computation, 6, 5873.
Management Studies, 2, 4454.
Eberhart, & Shi, Y. (2001). Particle swarm optimization: Developments, applications
Ramezanian, R., Rahmani, D., & Barzinpour, F. (2012). An aggregate production
and resources. In Proceedings of the 2001 congress on evolutionary computation
planning model for two phase production systems: Solving with genetic
2001 (Vol. 1, pp. 8186 vol. 81).
algorithm and tabu search. Expert Systems with Applications, 39, 12561263.
Eberhart, R., & Kennedy, J. (1995). A new optimizer using particle swarm theory. In
Sakalli, U. S., Baykoc, O. F., & Birgoren, B. (2010). A possibilistic aggregate production
Proceedings of the sixth international symposium on micro machine and human
planning model for brass casting industry. Production Planning and Control, 21,
science (pp. 3943). Nagoya, Japan: IEEE.
319338.
Eberhart, R. C., & Shi, Y. (1998). Comparison between genetic algorithms and
Shi, Y., & Eberhart, R. C. (1999). Empirical study of particle swarm optimization. In
particle swarm optimization. In Proceedings of the 7th international conference on
Proceedings of the 1999 congress on evolutionary computation (Vol. 3, pp. 1950,
evolutionary programming VII (pp. 611616). Springer-Verlag.
Vol. 1953).
Eberhart, R. C., Simpson, P. K., & Dobbins, R. W. (1996). Computational intelligence PC
Shi, Y., & Eberhart, R. (1998). A modied particle swarm optimizer. In IEEE
tools. Academic Press Professional, Inc.
international conference on evolutionary computation (pp. 6973). IEEE.
Graves, S. C. (2002). Manufacturing planning and control. In P. M. Pardalos & M. G. C.
Sillekens, T., Koberstein, A., & Suhl, L. (2011). Aggregate production planning in the
Resende (Eds.), Handbook of applied optimization. New York 10016 USA: Oxford
automotive industry with special consideration of workforce exibility.
University Press.
International Journal of Production Research, 49, 50555078.
Hanssmann, F., & Hess, S. W. (1960). A linear programming approach to production
Silva, C. G. d., Figueira, J., Lisboa, J., & Barman, S. (2006). An interactive decision
and employment scheduling. Management Technology, 1, 4651.
support system for an aggregate production planning model based on multiple
Holt, C. C., Modigliani, F., & Muth, J. F. (1956). Derivation of a linear decision rule for
criteria mixed integer linear programming. Omega-International Journal of
production and employment. Management Science, 2, 159177.
Management Science, 34, 167177.
Holt, C. C., Modigliani, F., & Simon, H. A. (1955). A linear decision rule for production
Skokos, C., Parsopoulos, K. E., Patsis, P. A., & Vrahatis, M. N. (2005). Particle swarm
and employment scheduling. Management Science, 2, 130.
optimization: An efcient method for tracing periodic orbits in three-
Kennedy, J., & Eberhart, R. (1995b). Particle swarm optimization. In Proceedings IEEE
dimensional galactic potentials. Monthly Notices of the Royal Astronomical
international conference on neural networks 1995 (Vol. 4, pp. 19421948, vol.
Society, 359, 251260.
1944).
Stevenson, W. J. (2009). Operations management (10th ed.). McGraw-Hill/Irwin.
Kennedy, J., & Eberhart, R. (1995a). Particle swarm optimization. In Proceedings of
Tsafarakis, S., Saridakis, C., Baltas, G., & Matsatsinis, N. (2013). Hybrid particle
IEEE international conference on neural networks (pp. 19421948). Perth, WA,
swarm optimization with mutation for optimizing industrial product lines: An
Australia: IEEE. Vol. 4.
application to a mixed solution space considering both discrete and continuous
Kennedy, J., & Eberhart, R. C. (2001). Swarm intelligenc. Morgan Kaufmann
design variables. Industrial Marketing Management, 42, 496506.
Publishers Inc..
Venter, G., & Sobieszczanski-Sobieski, J. (2004). Multidisciplinary optimization of a
Koroec, P., Bole, U., & Papa, G. (2013). A multi-objective approach to the application
transport aircraft wing using particle swarm optimization. Structural and
of real-world production scheduling. Expert Systems with Applications, 40,
Multidisciplinary Optimization, 26, 121131.
58395853.
Zhang, R. Q., Zhang, L. K., Xiao, Y. Y., & Kaku, I. (2012). The activity-based aggregate
Lanzenauer, V., & Haehling, C. (1970). Production and employment scheduling in
production planning with capacity expansion in manufacturing systems.
multistage production systems. Naval Research Logistics Quarterly, 17, 193198.
Computers and Industrial Engineering, 62, 491503.
Liao, C.-J., Tseng, C.-T., & Luarn, P. (2007). A discrete version of particle swarm
optimization for owshop scheduling problems. Computers and Operations
Research, 34, 30993111.