Sie sind auf Seite 1von 21

http://www.paper.edu.

cn
519

A novel multi-objective particle swarm optimization


algorithm for no-wait flow shop scheduling problems
Q-K Pan1*, L Wang2, and B Qian2
1
College of Computer Science, Liaocheng University, Liaocheng, Peoples Republic of China
2
Department of Automation, Tsinghua University, Beijing, Peoples Republic of China
The manuscript was received on 25 August 2007 and was accepted after revision for publication on 11 December 2007.
DOI: 10.1243/09544054JEM989

Abstract: The current paper presents a novel multi-objective particle swarm optimization
(MOPSO) algorithm for solving no-wait flow shop scheduling problems with makespan and
maximum tardiness criteria. First, in the algorithm, particles are represented as job permutations and updated directly in the discrete domain. Second, the concept of Pareto dominance
is applied to compare different solutions of multi-objective optimization, and a set is employed
to hold and to update the obtained non-dominated solutions, where a randomly selected nondominated solution is assigned as the global best particle to maintain the diversity of the
searching direction and to speed up the convergence process to the Pareto front. Third, a
new multi-objective heuristic, named the PWQ (PanWangQian) heuristic, is proposed to produce a population of initial particles with relatively good performances. Fourth, a simple but
effective multi-objective local search algorithm is developed to embed in the MOPSO algorithm
for stressing the balance between global exploration and local exploitation. In addition, two
speed-up methods are devised to evaluate a job permutation and its insert neighbourhood
respectively. Computational simulation results based on the well-known benchmarks and statistical performance comparisons are provided. It is shown that the proposed algorithm is
superior to a recently published multi-objective hybrid differential evolution (MHDE) algorithm in terms of searching quality, diversity level, robustness, and efficiency.
Keywords: particle swarm optimization, no-wait flow shop, multi-objective, insert neighbourhood, local search, speed-up

1 INTRODUCTION
The flow shop scheduling problem is one of the most
famous machine scheduling problems with extensive
engineering backgrounds, representing nearly a
quarter of manufacturing systems, assembly lines,
and information service facilities in use nowadays
[1, 2]. It is well known that in a large number of real
problems, production scheduling managers have to
face multiple, often conflicting, decision criteria.
Thus, it is very important to develop effective, efficient, and advanced scheduling technologies and
approaches for the multi-objective flow shop scheduling problem.

*Corresponding author: School of Computer Science, Liaocheng


University, Liaocheng 252059, Peoples Republic of China.
email: qkpan@lcu.edu.cn

During the last two decades, there has been an


increasing interest in solving the multi-objective flow
shop scheduling problems. Daniels and Chambers
[3] considered the trade-off between makespan and
maximum tardiness. Rajendran [4] presented a
branch-and-bound algorithm and two heuristics for
the two-stage flow shop scheduling problems to minimize total flow time with a constraint on makespan.
Ravindran et al. [5] proposed three heuristics for the
flow shop scheduling problems with makespan and
total flow time criteria. Toktas et al. [6] studied the
two-machine flow shop scheduling problems with
makespan and maximum earliness criteria. With the
development of computer technology, more and
more researchers extended meta-heuristics to the
multi-objective flow shop scheduling problems.
TKindt et al. [7] presented an ant colony optimization
approach for solving the two-machine flow shop scheduling problems with the objective of minimizing


520

http://www.paper.edu.cn
Q-K Pan, L Wang, and B Qian

total completion time and makespan. Loukil et al. [8]


designed a multi-objective simulated annealing algorithm for the one machine, parallel machines, and
permutation flow shop scheduling problems, where
the weighted function from a diversified set of weight
vectors was employed to compute acceptance probability. Rahimi-Vahed and Mirghorbani [9] devised a
multi-objective particle swarm optimization algorithm for the bi-criteria flow shop scheduling problems to minimize weighted mean completion time
and weighted mean tardiness. Based on the features
of a biological immune system and bacterial optimization, Tavakkoli-Moghaddam et al. [10] provided a
multi-objective immune algorithm for the flow shop
scheduling problems. Ponnambalam et al. [11] presented an evolutionary algorithm for the flow shop
scheduling problems, where the weighted sum of
makespan, mean flow time, and machine idle time
was used. Murata et al. [1214] developed a class of
multi-objective genetic local search (MOGLS) algorithms for the multi-objective flow shop problems
where the weight vectors were generated randomly
or systematically and used to evaluate the quality of
generated solutions as well as to guide the procedure
of local search. In reference [15], Arroyo and Armentano presented another MOGLS algorithm by applying the concept of Pareto dominance to classify the
individuals in a population, and a mechanism to
assign suitable fitness values to promote the dispersion of the population. In reference [16], Li and
Wang proposed a hybrid quantum-inspired genetic
algorithm for the multi-objective flow shop scheduling problems, where both a randomly weighted linear-sum function and a non-dominated sorting
technique including classification of Pareto front and
fitness assignment were applied to evaluate individuals. More recently, Qian et al. [17] presented a
multi-objective hybrid differential evolution (MHDE)
algorithm for the flow shop scheduling problems
with multiple objectives by fusing differential evolution and a simple but efficient local search, and the
authors empirically showed that their algorithm outperformed the MOGLS [1214] algorithm on a set of
typical problems. A comprehensive state-of-the-art
survey on the multi-objective scheduling problems
can be found in reference [18].
The no-wait flow shop scheduling problem is a typical scheduling problem with the additional restriction
that the processing of each job has to be continuous,
i.e. once a job is started on the first machine, it must be
processed through all machines without any interruption. Such a no-wait constraint is usually a result of the
specific requirements of the production process or
the absence of adequate storage capacity between
the operations of a job. There are wide applications
of no-wait flow shop scheduling problems, especially
in chemical processing [19], food processing [20],

concrete ware production [21], and pharmaceutical


processing [22]. According to the research work [23],
the no-wait condition secures that any no-wait flow
shop schedule must be a permutation schedule, and
it is well known that a no-wait flow shop scheduling
with more than two machines is strongly nondeterministic polynomial-bounded (NP)-hard [23].
Therefore, in the past two decades, most research
work on no-wait flow shop scheduling problems
focused on developing heuristics for the reason that
they generally only need acceptable time and memory
requirements to obtain a sub-optimal or optimal solution. For the single-objective no-wait flow shop scheduling problems, solution techniques can be broadly
classified into two groups: constructive heuristics and
meta-heuristics. For example, many constructive
heuristics can be found in references [2428]. As for
meta-heuristics, they have grown quickly with the
development of computer technology, e.g. the simulated annealing algorithm [29, 30], the genetic algorithm [29], the hybrid genetic algorithm and simulated
annealing algorithm [31], the variable neighbourhood
search algorithm [31], the descending search algorithm
[32], the tabu search algorithm [30, 32], the iterated
greedy algorithm [33], the hybrid particle swarm optimization algorithm [34], and the discrete particle
swarm optimization algorithm [35, 36]. For the multiobjective no-wait flow shop scheduling problems,
Tavakkoli et al. [37] developed a multi-objective
immune algorithm to minimize both weighted mean
completion time and weighted mean tardiness. The
paradox in the no-wait flow shop scheduling problems
was also investigated. Spieksma and Woeginger [38]
showed that increasing the speed of some machines
might worsen the optimal makespan. It was observed
by Kalczynski and Kamburowski [39] that under the
no-wait condition, makespan was not an increasing
function of job processing times. In addition, a comprehensive survey on the no-wait flow shop scheduling
problems can be found in reference [40].
The particle swarm optimization (PSO) algorithm,
which was first developed for continuous non-linear
functions [41], is one of the latest evolutionary optimization methods. In the PSO algorithm, each individual is called a particle with position and velocity.
During the search procedure, each particle continuously updates its position and velocity by its own
experience, the experience of its neighbours, and
the experience of the whole swarm. That is, individual improvement, population cooperation, and
population competition direct the population
towards the best position in the search space. Owing
to the simplicity, easy implementation, and quick
convergence, the PSO algorithm has gained much
attention and a wide range of successful applications
[4250]. However, the applications of the PSO algorithm to scheduling problems are still considerably

http://www.paper.edu.cn
A novel multi-objective particle swarm optimization algorithm

limited [35, 36]. To the best of the present authors


knowledge, there has been almost no published work
dealing with multi-objective no-wait flow shop
scheduling problems using the PSO algorithm. Therefore, in the current paper, a novel multi-objective
PSO (MOPSO) algorithm is proposed for the no-wait
flow shop scheduling problems to minimize both
makespan and maximum tardiness. First, particles
are represented as discrete job permutations and are
updated by using job-permutation-based position
updating operations. Second, the concept of Pareto
dominance is applied to compare different solutions
and a set is used to hold the non-dominated solutions
found so far, where a randomly selected nondominated solution is used as the global best particle.
Third, a new multi-objective constructive heuristic is
proposed, named the PWQ (PanWangQian) heuristic, to produce an initial population with relatively
good performances and incorporate a multi-objective
local search algorithm is incorporated in the MOPSO
algorithm to stress the balance between global
exploration and local exploitation. Fourth, two
speed-up methods are devised for the evaluations of
a job permutation and its insert neighbourhood to
reduce the computational time requirements. Last,
the effectiveness, efficiency, and robustness of the
proposed algorithm are demonstrated by simulated
results and comparisons.
The rest of the paper is organized as follows. In section 2, the no-wait flow shop scheduling problem
with makespan and maximum tardiness criteria are
stated and formulated. In section 3, a speed-up
method for the insert neighbourhood structure is
proposed. The traditional PSO algorithm is introduced and the MOPSO algorithm is presented in
detail in section 4. The computational results and
comparisons are provided in section 5. Conclusions
are presented in section 6.

2 NO-WAIT FLOW SHOP SCHEDULING


PROBLEM

If a solution is not dominated by any other solutions


in V, then it is called a Pareto-optimal solution or
non-dominated solution. The set of all the Paretooptimal solutions is called the Pareto-optimal set, or
the efficient set. The corresponding images of the
Pareto-optimal set in the objective space are called
the Pareto-optimal front.
2.2 No-wait flow shop scheduling problem
The no-wait flow shop scheduling problem can be
described as follows: Each of n jobs from set J {1,
2, . . ., n} will be sequenced through m machines (k
1, 2, . . ., m). Job j 2 J has a sequence of m operations
(oj1,oj2,. . .,ojm) and a given due date d(j). Operation
ojk corresponds to the processing of job j on machine
k during an uninterrupted processing time p(j,k). At
any time, each machine can process at most one
job and each job can be processed on at most one
machine. To follow the no-wait restriction, the completion time of the operation ojk must be equal to
the earliest start time of the operation oj,k 1 for k
1, 2, . . ., m1). In other words, there must be no waiting time between the processing of any consecutive
operations of each of n jobs. The problem is, then,
to find a schedule such that the processing order of
the jobs is the same on each machine and the given
criteria are minimized.
2.3 Objective functions
In the current paper, the first objective to be minimized is the maximum completion time or makespan. Suppose that a job permutation p {p1, p2,
. . . pn} represents a schedule of the jobs to be processed. Let e(pj1, pj) be the minimum delay on the
first machine between the start of jobs pj1 and pj
restricted by the no-wait constraint when job pj is
directly processed after job pj1, and the completion
time C(pj,m) of job pj on machine m can be computed by the following formula [33].
Cp1 ; m

Given a q-objective optimization problem f (f1,f2,. . .,


fq) defined on a finite set V
1

where x 2 V
Suppose x1, x2 2 V, then x1 is said to be dominated by
x2, if and only if
fi x2 6 fi x1 ;

8i 1; 2; . . . ; n

9i 1; 2; . . . ; n

and
fi x2 <fi x1 ;

m
X

pp1 ; k

k1

2.1 Pareto optimal solution

Minimize f x f1 x; f2 x; . . . ; fq x

521

Cpj ;m

j
X

epi1 ;pi

i2

m
X

ppj ;k;

j 2;3;...;n

k1

5
and e(pj1,pj) can be computed as follows [33]
epj1 ; pj Fj1;j pj ; m

m
X

ppj ; k for j 2; . .. ; n

k1

6
where Fj1,j(pj,m) denotes the makespan of jobs pj1
and pj in the two-machine permutation flow shop


522

http://www.paper.edu.cn
Q-K Pan, L Wang, and B Qian

scheduling problem. So, the makespan of the schedule p {p1,p2,. . .pn} is Cmax(p) C(pn,m).
The second objective to be minimized is the maximum tardiness, which is related to the job due dates.
The maximum tardiness of the schedule p {p1, p2,
. . . pn} is given as follows
n

Tmax p maxfmax0; Cpj ; mdpj g


j1

Owing to the multi-objective property, the no-wait


flow shop scheduling problem with makespan and
maximum tardiness criteria is required to find a set of
non-dominated solutions. According to reference
[51], usually three main aspects should be considered
to evaluate the obtained non-dominated solutions;
that is, the number of the obtained non-dominated
solutions, the distance of the obtained non-dominated
front to Pareto front, and the spread and distribution
of the obtained non-dominated solutions.
2.4 Short cut to evaluate a job permutation
P
Since m
k1 ppj ; k and e(p j1,p j) can be computed
in advance and used for the evaluation of a job permutation, the time complexity of equation (6) is O
(n) and that of equation (7) is O(n2). Therefore, the
total time complexity to evaluate a job permutation
is O(n2) by using equations (6) and (7) directly. However, based on the similarity of these two equations, a
short cut is presented to evaluate a job permutation
in time O(n), which is given as follows.
Step 1. Let d1 0, and calculate di di1 e(pi1, pi),
i 2, 3, . . ., n.
Step 2. Calculate C(pi, m) di TP(pi), i 1, 2, . . ., n,
P
where TP(pi) denotes m
j1 pp i; j.
Step 3. Calculate l(pi) C(pi,m)d(pi), i 1,2,. . .,n.
Step 4. Let Cmax(p) (pn,m).
Step 5. Let Tmax p maxnj1 fmax0; lpj g.

Fig. 1 Calculate makespan after inserting job 3


Table 1

Representation of position information and the


job permutation

Dimension j of particle Xi

Job permutation

Table 2

An example of the PTL crossover operator

PTL crossover

3 SPEED-UP METHOD BASED ON INSERT


NEIGHBOURHOOD
Insert neighbourhood of a job permutation p is
widely used for flow shop scheduling problems in
the literature [2], which is defined by considering all
possible insert moves. The number of neighbours in
an insert neighbourhood is (n1)2, so the time complexity to evaluate the whole insert neighbourhood
is O(n3) by using the short cut proposed in section
2.4. Based on the similarity of a permutation and its
neighbours, however, a speed-up method is presented to reduce the time complexity to O(n2), whose
procedure is given as follows.

P1
P2
O1
O2

5
3
3
1

1
5
5
4

4
4
2
3

PTL crossover
2
2
1
5

3
1
4
2

P1
P2
O1
O2

5
5
5
1

1
1
2
4

4
4
3
5

2
2
1
2

3
3
4
3

Step 1. Let k 1.
00
00
00
00
Step 2. Let p fp1 ; p2 ; . . . ; pn1 g be a partial permutation generated by removing job pk from a permutation p. Execute the following steps.
Step 2.1. Let d1 0, and calculate di di1
ep00i1 ; p00i ,
i 2, 3, . . ., n1;
00
00
i 1,
Step 2.2. Calculate Cpi ; m di TPpi ,
2, . . ., n1;
00
00
00
i 1,
Step 2.3. Calculate lpi Cpi ; mdpi ,
2, . . ., n1;

http://www.paper.edu.cn
A novel multi-objective particle swarm optimization algorithm

Table 3

523

Initial particles and their personal best positions

Xi0

Cmax Xi0

Tmax Xi0

Pi0

Cmax Pi0

3526714890

11 706

3526714890

11 706

883

14

2567891430

10 455

2567891430

10 455

994

37

2135698704

9582

2135698704

9582

3113

46

3215698074

9765

3215698074

9765

3993

50

3215698074

9765

883
...
994
...
3113
...
3993
...
3993

3215698074

9765

3993

Table 4

00

Step 2.4. Let ln1 lpn1 , and calculate li1


00
maxfli ; lpi g, i n2, n3,. . .1.
Step 3. Let p0 be a permutation generated by inserting job pk into position k0 of p00 . Repeat the following steps until all possible positions k0 ,
0
0
= fk; k1g, of p00 are considk 2 f1; 2; :::; ng^ k 2
ered.
Step 3.1. As seen in Fig. 1, the maximum completion time Cmax(p0 ) is calculated as follows
8
00
k0 0
Cmax p00 epk ;p0
>
>
00
>
00
>
< Cmax p epk0 1 ;pk
00
00
00
Cmax p0
epk ;pk0 epk0 1 ;pk0 0<k0 <n
>
00
>
00
>
>
k
: Cmax p epn1 ;p
00
k0 n
TPpk TPpn1
Step 3.2. The maximum tardiness can be calculated as follows
8
00
maxf0; l1 epk ;p0 ; EPpk
>
>
>
>
>
k0 0
dpk g
>
<
00
0
;
d
ep
;p

maxf0;
l
1 k 1
k
k 0 1
Tmax p0
>
EPpk dpk ; lk0
>
>
> C p0 C p00 g
>
0<k 0 <n
max
max
>
:
maxf0; l1 ;Cmax p0 dpk g
k0 n
Step 4. Let k k 1. If k > n, stop the procedure;
otherwise go back to Step 2.
There are n iterations for Step 2 and Step 3, and
both Step 2 and Step 3 can be executed in time O(n).
Consequently the time complexity of the proposed
speed-up method is O(n2).

4 MULTI-OBJECTIVE PARTICLE SWARM


OPTIMIZATION (MOPSO) ALGORITHM
4.1 Brief introduction to the traditional PSO
algorithm
The traditional PSO algorithm, introduced by
Eberhart and Kennedy [41], is a parallel direct search
method. Starting from an initial swarm of PS particles
with position and velocity, the swarm moves towards

Initial non-dominated solutions

Permutation

1
2
3
4

2
2
2
2

5
5
5
1

6
8
8
3

7
6
6
5

9
7
7
6

1
9
9
9

Tmax Pi0

4
1
1
8

8
4
3
7

3
3
0
0

0
0
4
4

Makespan

Maximum tardiness

11 213
10 393
9668
9582

190
932
2426
3113

the best position in the search space by individual


improvement, population cooperation, and populat
t
t
; xi2
; ::; xiz
g and
tion competition. Let Xit fxi1
t
t
t
; vi2
; ::; viz
g denote the position and velocity
Vit fvi1
respectively of the ith particle in z-dimensional search
space at iteration t. In addition, a current personal
best position is represented as Pit fpti1 ; pti2 ; :::; ptiz g,
and a current global best position is represented as
G t fg1t ; g2t ; :::; gzt g. During the search procedure, the
velocity and position of the jth dimension of the ith
particle are updated as follows:
t1
vijt w t1 vijt1 c1 randpt1
ij xij
c2 rand gjt1 xijt1

xijt xijt1 vijt

where wt1 is the inertia weight, and c1 and c2 are


acceleration coefficients, and rand() denotes a uniform random number between 0 and 1. Then Pit
and Gt will be correspondingly updated if Xit is better
than its personal best position and the global best
position respectively.
The procedure of the gbest model [42] of the PSO
algorithm is given as follows.
Step 1. Set w, c1, c2, and PS, and let t 0.
Step 2. Initialize PS particles.
Step 3. Evaluate each Xit, i 1,2,. . .,PS.
Step 4. Let Pit Xit, i 1,2,. . .,PS, and Gt be equal to
the best initial position.
Step 5. Let t t 1.
Step 6. Update all the particles according to equations (8) and (9).
Step 7. Evaluate all the particles.
Step 8. For i 1, 2, . . ., PS, if Xit is better than Pit, let
Pit Xit.

http://www.paper.edu.cn

524

Q-K Pan, L Wang, and B Qian

Table 5

Particles and their personal best positions after update operation

Xi1

Cmax Xi1

Tmax Xi1

Pi1

Cmax Pi1

6352714890

12 967

3526714890

11 706

883

14

6791432580

11 397

2567891430

10 455

994

37

2135697804

9570

2135697804

9570

2328

46

2586791304

9668

2586791304

9668

2426

50

3256807914

10 075

2149
...
6814
...
2328
...
2426
...
2833

3215698074

9765

3993

Step 9. For i 1, 2, . . ., PS, if Xit is better than Gt, let


Gt Xit.
Step 10. If a stopping criterion is met, output Gt;
otherwise, go back to Step 5.
4.2 The MOPSO algorithm
This section presents a novel MOPSO algorithm for
the no-wait flow shop scheduling problems with
makespan and maximum tardiness criteria after
explaining solution representation, a job-permutation-based position updating operation, personal
best position updating operation, update of a nondominated solution set and the global best position,
the PWQ heuristic, and initialization.
4.2.1 Solution representation
One of the key issues to design the PSO algorithm lies
in its solution representation. Since the job-permutation-based representation bears the necessary information related to the flow shop scheduling
problems and it has been widely used in the literature
[2], this representation is adopted in the current
paper. An example of the job-permutation-based
representation is given in Table 1.
It can be seen from Table 1 that this encoding
scheme cannot be directly applied in the traditional
PSO algorithm in that the traditional position updating equation works in the continuous space. Therefore, the next section presents a novel position
updating operation for this representation.
4.2.2 Job-permutation-based position updating
operation
Since the new position of a particle is determined
by the global best position, its personal best position, and its current position and velocity, a jobpermutation-based position updating operation can
be given as follows
n
h
i
o
Xit c2  F2 c1  F2 w  F1 Xit1 ; Pit1 ; G t1
10
where w 2 [0,1] is the inertia weight, and c1 2 [0,1] and
c2 2 [0,1] are acceleration coefficients.

Table 6

Tmax Pi1

The non-dominated solutions after update


operation

Permutation

Makespan

Maximum tardiness

1
2
3

2567914830
2586791430
2135697804

11 213
10 393
9570

190
932
2328

The neighbours of p {5 8 6 1 7 9 3 2 4 0}
with respect to job 5

Table 7
i

Permutation

1
2
3
4
5
6
7
8
9

8
8
8
8
8
8
8
8
8

5
6
6
6
6
6
6
6
6

6
5
1
1
1
1
1
1
1

1
1
5
7
7
7
7
7
7

7
7
7
5
9
9
9
9
9

9
9
9
9
5
3
3
3
3

3
3
3
3
3
5
2
2
2

2
2
2
2
2
2
5
4
4

4
4
4
4
4
4
4
5
0

0
0
0
0
0
0
0
0
5

Table 8

Makespan

Maximum tardiness

12 356
11 834
12 161
12 454
11 987
12 117
12 403
12 022
12 813

6859
6337
6664
6957
6490
6731
7603
8350
9693

The solutions in Q

Permutation

Makespan

Maximum tardiness

8651793240

11 834

6337

Table 9

The last non-dominated solutions

Permutation

1
2
3
4
5
6
7
8
9
10

2
2
2
2
2
4
2
4
4
4

5
5
1
1
5
2
1
2
2
2

6
7
3
5
7
5
5
5
5
1

7
6
5
3
6
6
3
8
6
5

4
4
6
0
4
7
0
6
7
3

1
1
7
7
1
8
9
7
9
6

8
3
9
9
8
9
8
9
1
7

9
9
8
8
9
1
6
1
3
9

3
8
0
6
3
3
7
3
8
8

0
0
4
4
0
0
4
0
0
0

Makespan

Maximum tardiness

10 662
10 238
9496
9159
10 484
9983
9421
9921
10 154
9812

190
572
2254
4308
449
918
3649
1319
892
2001

The above position updating operation consists of


three components. The first component represents
the current position and velocity of the particle, that is

F1 Xit1 if rand < w
t
t1
li w  F1 Xi
otherwise
Xit1
11

http://www.paper.edu.cn
A novel multi-objective particle swarm optimization algorithm

525

1200
PWQ heuristic
Randomization method
1000

Maxmimum Lateness

800

600

400

200

0
700

800

900

1000

1100

1200

1300

Makespan

Fig. 2 Comparison between the PWQ heuristic and randomization method

where lit is a temporary individual, and Fl indicates


an insert operator, and rand() is a uniform random
number between 0 and 1.
The second component is the cognition part of
the particle, representing the private thinking of the
particle itself, that is

F2 lti ;Pit1 if rand < c1
t
t
t1
di c1 F2 li ;Pi
lti
otherwise
12
where is a temporary individual, and F2 represents
the Pan-Tasgetiren-Liang PTL crossover operator
[35]. In the PTL crossover shown in Table 2, a block
of jobs is randomly determined from the first parent
by two cut points. This block is either moved to the
right or left corner of the permutation. Then the offspring permutation is filled out with the remaining
jobs from the second parent. This procedure will
always produce two distinctive offspring even from
the same two parents. In the present paper, one of
these two unique offspring is chosen randomly with
an equal probability.
The third component is the social part of the particle, representing the collaboration among particles,
that is

F2 dti ;Gt1 if rand<c2
t
t
t1
Xi c2 F2 di ;G
dti
otherwise
dti

13

4.2.3 Personal best position updating operation


In order to solve multi-objective optimization problems, the personal best position of a particle is
updated according to the concept of Pareto dominance, that is
 t
if Xit dominate Pit1
Xi
14
Pit
t1
Pi
otherwise
4.2.4 Update of non-dominated solution
set and global best position
To avoid the loss of any non-dominated solution, a
tentative set S is employed to hold and to update
the non-dominated solutions found so far. At each
generation, a solution in S is selected randomly as
the current global best particle to guide the search
procedure of the whole swarm. If a new obtained
position dominates its current personal best position
and is not dominated by any solutions in S, this position will be added to S and all solutions dominated by
the added one will be eliminated from S.
4.2.5 PWQ heuristic and initialization
NawazEnscoreHam (NEH) [52] and the earliest
due date (EDD) heuristics are regarded as the effective constructive methods for flow shop scheduling
problems with makespan and maximum lateness

http://www.paper.edu.cn

526

Q-K Pan, L Wang, and B Qian

Table 10
Instance
Name
Car01
Car02
Car03
Car04
Car05
Car06
Car07
Car08
Hel1
Hel2
Rec01
Rec03
Rec05
Rec07
Rec09
Rec11
Rec13
Rec15
Rec17
Rec19
Rec21
Rec23
Rec25
Rec27
Rec29
Rec31
Rec33
Rec35
Rec37
Rec39
Rec41
Mean

Comparison of D1R between the MOPSOLS and MHDE algorithms


MOPSOLS
MAX

n m

AVG

MIN

115
134
125
144
104
89
77
88
10010
2010
205
205
205
2010
2010
2010
2015
2015
2015
3010
3010
3010
3015
3015
3015
5010
5010
5010
7520
7520
7520

0.00
0.00
0.00
0.06
0.00
0.00
0.00
0.00
16.82
0.29
1.20
0.46
0.77
0.71
0.36
1.37
0.06
1.83
0.23
2.98
6.07
1.82
1.86
2.53
2.44
3.96
7.41
4.63
6.49
7.47
7.27
2.55

0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
11.86
0.00
0.05
0.05
0.48
0.00
0.03
0.68
0.00
0.00
0.08
2.12
2.65
1.30
0.92
1.62
0.58
3.08
5.70
3.13
4.66
5.32
5.28
1.60

0.00
0.00
0.06
0.39
0.00
0.00
0.00
0.00
20.55
0.65
1.83
1.12
1.24
1.47
1.13
3.31
0.24
5.91
0.70
3.78
8.78
2.37
2.71
3.58
3.48
5.15
9.55
6.35
7.94
10.34
9.68
3.62

criteria respectively. In the current paper, a multiobjective heuristic based on the NEH and EDD heuristics, named the PWQ heuristic, is presented for the
considered problems, whose procedure is given as
follows.
Step 1. Set i 0 and the number of the permutations
N to be produced.
i
i
; 1 N 1
.
Step 2. Set a weight vector (w1, w2) as N 1
Step 3. The following steps are used to generate a
seed sequence p0.
Step 3.1. Generate a job permutation p1 by ordering jobs in descending sums of their processing
times.
Step 3.2. Generate another job permutation p2 by
using the EDD heuristic.
Step 3.3. Compute the weighted sum of job position values fj w1 wj p1 w2 wj p2 for j
1,2,. . .,n, where wj(p1) and wj(p2) denote the
position values of job j in permutation p1 and
p2 respectively.
Step 3.4. Generate a permutation p0 fp01 ;
p02 ; :::; p0n g by ordering jobs in ascending
weighted sum of their position values fj.
Step 4. The first two jobs of p0 are taken and the two
possible sub-sequences of these two jobs are

MHDE
SD

AVG

MIN

MAX

SD

0.00
0.00
0.01
0.14
0.00
0.00
0.00
0.00
2.17
0.23
0.48
0.31
0.25
0.40
0.33
0.71
0.11
2.71
0.17
0.44
1.34
0.33
0.57
0.50
0.64
0.59
1.27
0.82
0.84
1.31
1.12
0.57

8.82
7.39
7.53
8.05
2.97
2.41
2.58
0.00
53.90
8.87
6.12
6.60
11.62
7.06
8.86
8.28
9.22
13.46
5.63
9.06
16.16
11.22
10.09
11.11
10.68
17.82
23.06
18.11
20.49
22.96
20.52
11.96

2.51
1.28
1.50
3.24
0.00
0.00
0.00
0.00
45.69
4.90
3.72
4.52
7.94
4.34
5.06
5.89
5.42
9.95
3.17
7.29
13.56
8.18
7.53
9.60
8.50
15.59
19.17
14.80
17.33
21.38
18.35
8.72

33.40
12.42
9.50
14.94
9.65
12.05
13.92
0.00
58.39
12.00
8.76
10.80
16.22
9.57
17.62
14.73
12.88
19.53
7.68
10.19
20.98
13.63
14.53
13.30
14.22
20.76
25.74
20.19
22.97
25.83
24.29
16.80

6.51
3.45
2.10
3.11
3.36
3.45
4.01
0.00
3.70
2.17
1.45
1.43
2.72
1.52
3.39
2.67
1.94
2.37
1.09
0.76
2.37
1.65
1.46
1.14
1.87
1.53
1.89
1.57
1.52
1.34
1.61
2.23

evaluated by the weighted sum of the two objective values f p w1 Cmax p w2 Tmax p,
and then the better sub-sequence is selected as
the current sequence.
Step 5. Repeat the following steps until all jobs are
sequenced. Take job pj0, j 3, 4, . . ., n, and find
the sub-schedule with the minimum weighted
sum of the two objective values by placing it in
all possible positions of the sub-sequence of the
jobs that have been already scheduled. The best
sub-sequence is selected for the next generation.
Step 6. If i < N, let i i 1 and go back to Step 2;
otherwise, stop the procedure.
It is obvious that Step 3 can be performed in
time O(nlog2n), and Steps 4 and 5 in time O(n2)
by using the speed-up method proposed in section
3 to evaluate the [n(n 1)/2]1 sub-sequences.
Therefore, the time complexity of the PWQ heuristic is O(Nn2). In the current paper, advantage is
taken of the PWQ heuristic to produce a population of initial particles.
4.2.6 Procedure of the MOPSO algorithm
Based on the above solution representation, the jobpermutation-based position updating operation, the

http://www.paper.edu.cn
A novel multi-objective particle swarm optimization algorithm

Table 11
Instance
Name
Car01
Car02
Car03
Car04
Car05
Car06
Car07
Car08
Hel1
Hel2
Rec01
Rec03
Rec05
Rec07
Rec09
Rec11
Rec13
Rec15
Rec17
Rec19
Rec21
Rec23
Rec25
Rec27
Rec29
Rec31
Rec33
Rec35
Rec37
Rec39
Rec41
Mean

527

Comparison of NNDS between the MOPSOLS and MHDE algorithms


MOPSOLS
MAX

nm

AVG

MIN

115
134
125
144
104
89
77
88
10010
2010
205
205
205
2010
2010
2010
2015
2015
2015
3010
3010
3010
3015
3015
3015
5010
5010
5010
7520
7520
7520

8.00
18.00
17.95
13.85
10.00
10.00
8.00
5.00
33.15
23.00
26.55
37.85
26.25
23.95
29.00
40.65
15.00
18.85
43.85
36.05
23.20
39.30
42.45
46.70
34.45
43.10
37.95
46.80
61.50
44.30
46.75
29.40

8.00
18.00
17.00
13.00
10.00
10.00
8.00
5.00
20.00
23.00
25.00
36.00
24.00
23.00
28.00
38.00
15.00
18.00
42.00
29.00
20.00
36.00
36.00
42.00
30.00
33.00
29.00
36.00
45.00
33.00
40.00
25.48

8.00
18.00
18.00
14.00
10.00
10.00
8.00
5.00
38.00
23.00
28.00
40.00
29.00
25.00
30.00
43.00
15.00
20.00
45.00
44.00
30.00
46.00
52.00
55.00
39.00
54.00
49.00
60.00
74.00
67.00
57.00
34.00

personal best position updating operation, the


update of the non-dominated solution set and the
global best position, the PWQ heuristic, and the initialization, the procedure of the MOPSO algorithm is
presented as follows.
Step 1. Initialize w, c1, c2, and PS. Set t 0 and S F.
Step 2. Produce PS particles using the PWQ heuristic.
Step 3. Evaluate each Xit, i 1, 2, . . ., PS.
Step 4. Let Pit Xit, i 1, 2, . . ., PS.
Step 5. Determine all the non-dominated solutions in
the initial population and put them in S.
Step 6. Randomly select a solution from S and assign
it to Gt.
Step 7. t t 1.
Step 8. Update all the particles according to equation
(10).
Step 9. Evaluate all the particles.
Step 10. For i 1, 2, . . ., PS, if Xit dominates Pit, let Pit
Xit and update S.
Step 11. If a stopping criterion is reached, output S;
otherwise, go back to Step 6.
It can be found that the MOPSO algorithm adopts
the job-permutation-based encoding scheme, and
employs a job-permutation-based position updating

MHDE
SD

AVG

MIN

MAX

SD

0.00
0.00
0.22
0.37
0.00
0.00
0.00
0.00
4.30
0.00
0.76
1.60
1.45
0.51
0.79
1.57
0.00
0.67
0.93
4.03
3.00
2.89
4.25
2.85
1.96
5.34
5.65
6.07
9.01
7.96
5.08
2.30

4.95
11.60
13.85
8.50
8.95
8.70
7.30
5.00
0.00
6.85
5.75
4.35
2.45
6.20
7.85
5.95
9.85
7.05
8.25
0.25
0.30
0.10
0.45
1.40
0.75
0.05
0.05
0.00
0.00
0.00
0.00
4.41

4.00
10.00
11.00
7.00
7.00
6.00
5.00
5.00
0.00
4.00
2.00
2.00
1.00
3.00
5.00
3.00
8.00
4.00
4.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
2.94

7.00
14.00
15.00
9.00
10.00
10.00
8.00
5.00
0.00
10.00
8.00
9.00
5.00
11.00
11.00
9.00
12.00
11.00
14.00
1.00
1.00
1.00
1.00
4.00
3.00
1.00
1.00
0.00
0.00
0.00
0.00
6.16

0.76
1.14
1.31
0.61
0.94
1.22
0.86
0.00
0.00
1.98
1.65
1.76
1.19
1.96
1.76
1.73
1.23
1.64
2.88
0.44
0.47
0.31
0.51
1.31
0.91
0.22
0.22
0.00
0.00
0.00
0.00
0.94

operation, and applies the concept of Pareto dominance to decide the personal best positions, and
uses a set to hold the obtained non-dominated solutions. So the MOPSO algorithm can be easily applied
to the multi-objective no-wait flow shop scheduling
problems.
4.3 Improvement of the MOPSO algorithm
To overcome the parameter dependence and the
problem of being trapped in non-global local nondominated solutions, an improvement of the MOPSO
algorithm is presented by embedding a multi-objective
local search algorithm based on the insert neighbourhood to perform exploitation. The multi-objective local
search algorithm is applied to the global best position
Gt at the end of each iteration t, whose procedure is
described as follows.
Step 1. Generate a job permutation p by executing
several insert moves to Gt.
Step 2. Let p* p.
Step 3. Repeat the following steps until all jobs are
considered.
Step 3.1. Randomly remove a job from p without
repetition, and then insert it into all the other

http://www.paper.edu.cn

528

Q-K Pan, L Wang, and B Qian

Table 12
Instance
Name
Car01
Car02
Car03
Car04
Car05
Car06
Car07
Car08
Hel1
Hel2
Rec01
Rec03
Rec05
Rec07
Rec09
Rec11
Rec13
Rec15
Rec17
Rec19
Rec21
Rec23
Rec25
Rec27
Rec29
Rec31
Rec33
Rec35
Rec37
Rec39
Rec41
Mean

Comparison of RNDS between the MOPSOLS and MHDE algorithms

nm

AVG

MIN

115
134
125
144
104
89
77
88
10010
2010
205
205
205
2010
2010
2010
2015
2015
2015
3010
3010
3010
3015
3015
3015
5010
5010
5010
7520
7520
7520

1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00

1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
0.97
0.96
1.00
1.00
1.00
1.00
1.00
1.00
0.97
0.96
0.97
0.97
0.98
1.00
0.93
0.95
1.00
1.00
1.00
1.00
0.99

MOPSOLS
MAX
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00

MHDE
SD

AVG

MIN

MAX

SD

0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.01
0.01
0.00
0.00
0.00
0.00
0.00
0.00
0.01
0.01
0.01
0.01
0.00
0.00
0.02
0.01
0.00
0.00
0.00
0.00
0.00

0.92
0.95
0.99
0.87
0.96
0.99
1.00
1.00
0.00
0.54
0.42
0.26
0.28
0.50
0.66
0.34
0.91
0.80
0.40
0.02
0.04
0.01
0.04
0.12
0.07
0.01
0.01
0.00
0.00
0.00
0.00
0.42

0.75
0.77
0.86
0.67
0.73
0.88
1.00
1.00
0.00
0.31
0.13
0.11
0.09
0.23
0.42
0.16
0.73
0.60
0.17
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.31

1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
0.00
0.83
0.55
0.56
0.63
0.79
0.88
0.56
1.00
1.00
0.64
0.13
0.17
0.10
0.11
0.44
0.38
0.11
0.10
0.00
0.00
0.00
0.00
0.55

0.10
0.06
0.04
0.09
0.08
0.04
0.00
0.00
0.00
0.16
0.12
0.12
0.14
0.15
0.13
0.12
0.09
0.10
0.15
0.04
0.06
0.03
0.05
0.13
0.10
0.02
0.02
0.00
0.00
0.00
0.00
0.07

possible positions in p respectively. Let Q


denote a non-dominated solution set obtained
by these insert moves.
Step 3.2. If a solution in Q is not dominated by any
solutions in the non-dominated solution set S,
update S.
Step 3.3. If a solution in Q dominates p, let p be
equal to this solution. If there are ties, break
them randomly.
Step 4. If p dominates p*, go back to Step 2; otherwise, stop the procedure.

scheduling problems, local optima and global optima


are located closely to each other [53], so, an insert
operator is suitable for performing a thorough search.
Last, the speed-up method proposed in section 3 can
be used to evaluate the insert neighbours. Therefore,
the local search tends to guide the search to the promising region in relatively short time.
Based on the above multi-objective local search
and the MOPSO-based search, an improved algorithm, called MOPSOLS algorithm, is developed as
follows.

This simple local search method is very efficient


because of the following reasons. First, in Step 1,
the incumbent solution is the perturbation of Gt,
but not Gt itself, so the local search can avoid cycling
and getting trapped into non-global local non-dominated solutions. Second, in Step 3.2, any newly
obtained non-dominated solution will be added to
the set S to avoid the loss of the found non-dominated solutions. Third, in Step 3.3, the incumbent
solution is updated only by its dominance solution,
thus, the local search can be fast to reach a local
non-dominated solution. Fourth, the distance of the
old solution and the new one caused by insert move
is only one, and in the landscape of flow shop

Step 1. Initialize w, c1, c2, and PS. Set t 0 and S F.


Step 2. Produce PS particles by the PWQ heuristic.
Step 3. Evaluate each Xit, i 1, 2, . . ., PS.
i 1, 2, . . ., PS.
Step 4. Let Pit Xit,
Step 5. Determine all the non-dominated solutions in
the initial population and put them in S.
Step 6. Select a solution from S randomly and assign
it to Gt.
Step 7. t t 1
Step 8. Update all the particles according to equation
(10).
Step 9. Evaluate all the particles.
Step 10. For i 1, 2, . . ., PS, if Xit dominates Pit, let
Pit Xit and update S.

http://www.paper.edu.cn
A novel multi-objective particle swarm optimization algorithm

529

100
MOPSO LS
MHDE

Emperical cumulative frequency

90
80
70
60
50
40
30
20
10
0

10
D1 R

12

14

16

18

20

(a) Empirical cumulative distribution function of D1R

Emperical cumulative frequency

100
MOPSOLS
MHDE

90
80
70
60
50
40
30
20
10
0

10

15

20

25
NNDS

30

35

40

45

50

(b) Empirical cumulative distribution function of N NDS


100
MOPSO LS
MHDE

Emperical cumulative frequency

90
80
70
60
50
40
30
20
10
0

0.5

R NDS

1.5

(c) Empirical cumulative distribution function of R NDS


Fig. 3 Plots of empirical cumulative distribution function for the MOPSOLS and MHDE algorithms

http://www.paper.edu.cn

530

Q-K Pan, L Wang, and B Qian

600
MOPSOLS
MHDE
500

Maxmimum Lateness

400

300

200

100

0
720

740

760

780

800

820

840

860

880

Makespan

Fig. 4 Non-dominated solutions of the MOPSOLS and MHDE algorithms

Step 11. Apply the multi-objective local search to Gt.


Step 12. If a stopping criterion is reached, output S;
otherwise, go back to Step 6.
It can be seen that the MOPSOLS algorithm not only
applies the PSO-based evolutionary searching mechanism effectively to perform exploration for promising
solutions within the whole solution space, but it also
employs the effective multi-objective local search to
perform exploitation for solution improvement in
sub-regions. Since both the balance of exploration
and exploitation are stressed, it is expected to achieve
good results for the multi-objective no-wait flow shop
scheduling problems. In section 5, the performance
of the proposed MOPSO/MOPSOLS algorithm is
investigated by computational simulations and comparisons.
4.4 Illustrative example
In order better to understand the details of the proposed MOPSO/MOPSOLS algorithm, a complete
example is given that matches with the steps of the
proposed algorithm in section 4.3. In this example,
the benchmark problem Car05 with ten jobs and
four machines is considered. All illustrative data are
achieved by a typical run of the proposed algorithm.

Step 1. Set the parameters the same as those in section 5.3. That is, w 0.2, c1 0.2, c2 0.8, PS 50,
and maximum running time T 0.4 s. Set t 0
and S F.
Steps 23. Produce 50 initial particles by applying the
PWQ heuristic and evaluate them by using the
method proposed in section 2.4. The results are
shown in Table 3.
Step 4. Let Pi0 Xi0, i 1, 2, . . ., 50 (shown in Table 3).
Step 5. Determine the initial non-dominated solutions and put them in S (shown in Table 4).
Step 6. G0 {2 5 8 6 7 9 1 4 3 0} is randomly selected
from S.
Step 7. t t 1.
Step 8. Update all the particles using equation (10).
The update process of particle X460 {3 2 1 5 6 9
8 0 7 4} is given as follows.
Step 8.1. A random number rand() 0.085 is generated. Since rand() < w, an insert move is performed on X460, and a temporary individual
l461 {3 6 2 1 5 9 8 0 7 4} is obtained.
Step 8.2. A random number rand() 0.541 is generated. Since rand()>c1, according to equation
(12), a temporary individual d461 is equal to
l461 {3 6 2 1 5 9 8 0 7 4}.
Step 8.3. A random number rand() 0.688 is
generated. Since rand() < c2, perform the PTL

http://www.paper.edu.cn
A novel multi-objective particle swarm optimization algorithm

15
MOPSOLS
MHDE

D1R

10

10

20

30

40

50

60

70

80

mn
100

90

Computational Time
(a) Overall mean value of D1R

40
35
30
MOPSO LS
MHDE

25

NNDS

crossover operation on G0 and d461, and a new


particle X461 {2 5 8 7 9 1 3 0 4} is obtained.
Step 9. Evaluate all the new particles by using the
method proposed in section 2.4. The results are
shown in Table 5.
Step 10. For i 1, 2, . . ., 50, update Pil and S. The
results are shown in Tables 5 and 6 respectively.
Step 11. Apply the multi-objective local search to G0.
Step 11.1. Perform several insert moves to G0 and
obtain a permutation p {5 8 6 1 7 9 3 2 4 0}
with Cmax(p) 11 933 and Tmax(p) 6436.
Step 11.2. Let p* p.
Step 11.3. Repeat the following steps until all the
jobs are considered.
Step 11.3.1: Job 5 is randomly selected from p
and inserted into the other nine positions
respectively. The obtained solutions are shown
in Table 7, and the non-dominated solutions
of those are put in set Q (shown in Table 8).
Step 11.3.2: Since all the solutions in Q are
dominated by some solutions in S, S does
not need to update.
Step 11.3.3: Since a solution {8 6 5 1 7 9 3 2 4 0} in
Q dominates p {5 8 6 1 7 9 3 2 4 0}, let p {8
6 5 1 7 9 3 2 4 0}.
Step 11.4. A solution p {2 8 6 7 9 5 3 1 0 4} with
Cmax(p) 9916 and Tmax(p) 4398 is obtained.
Since p dominates p*, go back to Step 11.2.
Step 12. Repeat Steps 6 to 11, until the running time is
not less than 0.4 s. The final non-dominated solutions are given in Table 9.

531

20
15
10
5
mn

0
10

20

30

40

50

60

70

80

90

100

Computational Time
(2) Overall mean value of NNDS

MOPSO LS
MHDE

RNDS

5 COMPUTATIONAL RESULTS

0.5

5.1 Comparison metrics


To evaluate comprehensively the non-dominated
solutions obtained by the MOPSO/MOPSOLS algorithms, three metrics are adopted in this paper. Let
S* denote the reference solution set, and Sj (j 1, 2,
. . .) the non-dominated solution set obtained by
algorithm j, and S [Sj the union of the nondominated solution sets obtained by all the algorithms, then the performance measures are explained
as follows.
5.1.1 Average distance D1R

10

20

30

40

50

60

70

80

90

100

mn

Computational Time
(3) Overall mean value of RNDS

Fig. 5 Comparison of the MOPSOLS and MHDE algorithms


with running time increasing

Let dy(Sj) denote the shortest normalized distance


from a reference solution y to a solution set Sj, which
is given as follows
dy Sj min
x2Sj
q

0
0
0
0
f1 xf1 y2 f2 xf2 y2 :::fq0 xfq0 y2
15

where fi is the ith objective normalized by using


the reference solution set S*. Suppose that fimin
and fimax are the minimum and maximum value
0
of the i th objective in S* respectively, then fi can
be calculated with the following formula

fi

fi fimin
100
max
fi fimin

16

http://www.paper.edu.cn

532

Q-K Pan, L Wang, and B Qian

Table 13
Instance
Name
Car01
Car02
Car03
Car04
Car05
Car06
Car07
Car08
Hel1
Hel2
Rec01
Rec03
Rec05
Rec07
Rec09
Rec11
Rec13
Rec15
Rec17
Rec19
Rec21
Rec23
Rec25
Rec27
Rec29
Rec31
Rec33
Rec35
Rec37
Rec39
Rec41
Mean

Comparison of D1R between the MOPSOLS and MOPSO algorithms


MOPSOLS
MAX

nm

AVG

MIN

115
134
125
144
104
89
77
88
10010
2010
205
205
205
2010
2010
2010
2015
2015
2015
3010
3010
3010
3015
3015
3015
5010
5010
5010
7520
7520
7520

0.00
0.00
0.00
0.06
0.00
0.00
0.00
0.00
16.82
0.29
1.20
0.46
0.77
0.71
0.36
1.37
0.06
1.83
0.23
2.98
6.07
1.82
1.86
2.53
2.44
3.96
7.41
4.63
6.49
7.47
7.27
2.55

0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
11.86
0.00
0.05
0.05
0.48
0.00
0.03
0.68
0.00
0.00
0.08
2.12
2.65
1.30
0.92
1.62
0.58
3.08
5.70
3.13
4.66
5.32
5.28
1.60

0.00
0.00
0.06
0.39
0.00
0.00
0.00
0.00
20.55
0.65
1.83
1.12
1.24
1.47
1.13
3.31
0.24
5.91
0.70
3.78
8.78
2.37
2.71
3.58
3.48
5.15
9.55
6.35
7.94
10.34
9.68
3.62

The average distance D1R(Sj) is the average value of


those shortest normalized distances from all reference solutions to Sj [54]; that is
1 X
fdy Sj g
17
D1R Sj
jS*j y2S*
According to reference [14], D1R(Sj) can be used to
evaluate the spread and distribution of Sj as well as
the proximity of Sj to S*. Obviously, the smaller
D1R(Sj) is, the better Sj is. If D1R(Sj) 0, all the reference solutions in S* are included in the solution set
Sj. In other words, the algorithm j can find all the
reference solutions.
5.1.2 Number of non-dominated solutions
This metric counts the number of solutions in Sj, which
are not dominated by any solution in S. This metric is
denoted as NNDS(Sj) with the following formula
NNDS Sj jSj fx 2 Sj j9y 2 S : xygj

18

where xy means that the solution x is dominated by


y. Clearly, the larger the metric is, the better Sj is.
5.1.3 Ration of non-dominated solutions
The ratio of solutions in Sj that are not dominated by
any solution in S is given as follows

MOPSO
SD

AVG

MIN

MAX

SD

0.00
0.00
0.01
0.14
0.00
0.00
0.00
0.00
2.17
0.23
0.48
0.31
0.25
0.40
0.33
0.71
0.11
2.71
0.17
0.44
1.34
0.33
0.57
0.50
0.64
0.59
1.27
0.82
0.84
1.31
1.12
0.57

1.41
10.56
6.00
18.71
17.60
7.29
15.88
13.65
71.54
11.49
11.65
11.84
13.22
18.13
12.45
7.58
10.57
13.68
5.59
18.72
13.90
17.69
9.41
13.88
15.86
26.25
43.50
35.60
34.05
33.42
33.65
18.54

0.00
10.00
4.68
17.91
14.44
7.29
3.82
6.79
70.38
9.00
8.21
10.29
10.74
9.32
8.83
6.53
9.00
11.69
4.90
16.71
12.75
15.52
8.03
12.59
13.97
25.12
40.39
32.36
33.52
31.48
33.24
16.11

4.57
11.02
7.17
19.94
18.06
7.29
19.14
15.81
71.82
13.68
14.21
14.43
18.04
24.77
14.84
9.76
12.94
15.21
6.28
19.84
15.33
19.29
11.30
15.03
16.51
26.75
45.12
38.09
34.37
33.91
33.78
20.27

1.18
0.29
0.64
0.46
0.91
0.00
4.70
1.86
0.37
1.31
1.62
1.05
1.93
4.62
1.88
0.86
1.01
1.25
0.41
0.73
0.64
0.90
0.82
0.65
0.66
0.44
1.41
1.42
0.31
0.65
0.15
1.13

RNDS Sj

NNDS Sj
jSj j

19

This metric can be used to evaluate the quality of


the solution set Sj. It is obvious that RNDS(Sj) 0
means every solution in Sj is dominated by some
solutions in S. On the contrary, RNDS(Sj) 1 means
that all the solutions in Sj are not dominated by any
solution in S. Therefore, the higher RNDS(Sj) is, the
better Sj is.

5.2 Testing problems


In this paper, 31 well-known flow shop benchmarks
with different sizes from [5557] are modified as the
no-wait flow shop problems with makespan and
maximum tardiness criteria. Since the due dates of
jobs are not given in the original problems, they are
specified as follows.
Step 1. Randomly generated a job permutation p
{p1,p2,. . .,pn}.
Step 2. Calculate C(pj,m) for j 1,2,. . .,n.
Step 3. Specify the due date of job j by
dpj Cpj ;mbrand21Cpj ;m/10c;
j1;2;...;n:

http://www.paper.edu.cn
A novel multi-objective particle swarm optimization algorithm

Table 14
Instance
Name
Car01
Car02
Car03
Car04
Car05
Car06
Car07
Car08
Hel1
Hel2
Rec01
Rec03
Rec05
Rec07
Rec09
Rec11
Rec13
Rec15
Rec17
Rec19
Rec21
Rec23
Rec25
Rec27
Rec29
Rec31
Rec33
Rec35
Rec37
Rec39
Rec41
Mean

533

Comparison of NNDS between the MOPSOLS and MOPSO algorithms


MOPSOLS

MOPSO

nm

AVG

MIN

MAX

SD

AVG

MIN

MAX

SD

115
134
125
144
104
89
77
88
10010
2010
205
205
205
2010
2010
2010
2015
2015
2015
3010
3010
3010
3015
3015
3015
5010
5010
5010
7520
7520
7520

8.00
18.00
17.95
13.85
10.00
10.00
8.00
5.00
32.90
22.60
26.55
37.60
26.25
23.95
28.55
40.60
14.95
18.85
43.55
36.10
23.15
39.00
42.05
45.85
34.10
40.90
38.05
46.80
60.70
43.80
45.75
29.14

8.00
18.00
17.00
13.00
10.00
10.00
8.00
5.00
20.00
21.00
25.00
35.00
24.00
23.00
27.00
38.00
14.00
18.00
42.00
29.00
20.00
35.00
36.00
41.00
30.00
30.00
29.00
36.00
45.00
32.00
39.00
25.10

8.00
18.00
18.00
14.00
10.00
10.00
8.00
5.00
38.00
23.00
28.00
40.00
29.00
25.00
30.00
43.00
15.00
20.00
45.00
44.00
30.00
45.00
51.00
54.00
39.00
51.00
49.00
60.00
71.00
67.00
57.00
33.71

0.00
0.00
0.22
0.37
0.00
0.00
0.00
0.00
4.46
0.75
0.76
1.76
1.41
0.51
0.89
1.60
0.22
0.67
0.89
4.06
3.10
2.75
4.36
2.85
2.15
5.04
5.74
6.07
8.99
7.96
5.28
2.35

7.05
5.20
5.05
2.95
2.50
5.00
3.60
3.05
0.30
1.10
0.25
0.45
1.10
0.15
1.05
0.85
2.30
2.25
0.30
0.00
0.35
0.30
0.55
1.25
0.35
1.20
0.00
0.00
1.60
1.45
1.10
1.70

6.00
4.00
3.00
2.00
2.00
5.00
3.00
3.00
0.00
0.00
0.00
0.00
1.00
0.00
0.00
0.00
0.00
1.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
1.00
0.00
0.00
0.00
0.00
0.00
1.00

8.00
6.00
7.00
5.00
4.00
5.00
6.00
4.00
2.00
3.00
2.00
2.00
2.00
3.00
2.00
2.00
5.00
3.00
1.00
0.00
1.00
1.00
1.00
2.00
1.00
2.00
0.00
0.00
4.00
3.00
3.00
2.90

0.51
0.77
0.94
0.89
0.69
0.00
0.82
0.22
0.57
0.97
0.55
0.69
0.31
0.67
0.51
0.75
1.69
0.72
0.47
0.00
0.49
0.47
0.51
0.55
0.49
0.41
0.00
0.00
0.94
1.00
0.64
0.59

where bc is the integer part of the corresponding


real number.
5.3 Experimental set-up
The proposed MOPSO/MOPSOLS algorithms are
coded in C, which adopt the parameters as
follows: w 0.2, c1 0.2, c2 0.8, PS 5n, and maximum running time T 10 mn milliseconds. For each
benchmark, every algorithm is independently run 20
replications on a Pentium P-IV 3.0 GHz PC with
512 MB memory, and in each replication the above
performance metrics are computed. The average
value (AVG), minimum value (MIN), maximum value
(MAX), and the standard deviation (SD) of these performance metrics in 20 replications are calculated as
the statistics for the performance measures.
5.4 Performance of the PWQ heuristic
To test the performance of the proposed PWQ heuristic, 500 solutions are produced for each testing problem by the PWQ heuristic to compare with 500
solutions generated randomly. Since almost the
same comparison plots are obtained for all the problems, only the comparison plot for problem Hel1 is
shown in Fig. 2. This figure shows that all the solutions

generated randomly are dominated by many solutions of the PWQ heuristic, and no solution of the
PWQ heuristic is dominated by any solution generated randomly. So it is concluded that the PWQ heuristic can provide a solution set with relatively good
performances. Therefore, advantage is taken of this
heuristic to produce an initial population.
5.5 Comparison of the MOPSOLS and MHDE
algorithms
A MHDE algorithm was presented by Qian et al. [17]
for the permutation flow shop scheduling problems
with makespan and maximum tardiness criteria,
and it was shown that the MHDE algorithm outperformed the famous IMMOGLS2 algorithm [14], which
was well-known for its abilities efficiently to find
Pareto solutions with better performances than both
the strength Pareto evolutionary algorithm (SPEA)
[58] and the revised non-dominated sorting genetic
algorithm (NSGAII) [59]. Thus, the present authors
algorithm is compared with the MHDE algorithm.
Since the MHDE algorithm was applied to the classic
permutation flow shop scheduling problems in reference [17], in this paper it is applied to the no-wait
flow shop scheduling problems by modifying the

http://www.paper.edu.cn

534

Q-K Pan, L Wang, and B Qian

Table 15
Instance
Name
Car01
Car02
Car03
Car04
Car05
Car06
Car07
Car08
Hel1
Hel2
Rec01
Rec03
Rec05
Rec07
Rec09
Rec11
Rec13
Rec15
Rec17
Rec19
Rec21
Rec23
Rec25
Rec27
Rec29
Rec31
Rec33
Rec35
Rec37
Rec39
Rec41
Mean

Comparison of RNDS between the MOPSOLS and MOPSO algorithms


MOPSOLS

MOPSO

nm

AVG

MIN

MAX

SD

AVG

MIN

MAX

SD

115
134
125
144
104
89
77
88
10010
2010
205
205
205
2010
2010
2010
2015
2015
2015
3010
3010
3010
3015
3015
3015
5010
5010
5010
7520
7520
7520

1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
0.99
0.98
1.00
0.99
1.00
1.00
0.98
1.00
1.00
1.00
0.99
1.00
1.00
0.99
0.99
0.98
0.99
0.95
1.00
1.00
0.99
0.99
0.98
0.99

1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
0.94
0.91
1.00
0.97
0.96
1.00
0.93
0.97
0.93
1.00
0.98
1.00
0.95
0.96
0.95
0.98
0.97
0.90
1.00
1.00
0.96
0.95
0.93
0.97

1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00

0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.02
0.03
0.00
0.01
0.01
0.00
0.03
0.01
0.01
0.00
0.01
0.00
0.01
0.01
0.02
0.01
0.01
0.03
0.00
0.00
0.01
0.02
0.02
0.01

1.00
0.85
0.70
0.55
0.80
1.00
1.00
1.00
0.01
0.14
0.03
0.07
0.16
0.02
0.18
0.09
0.34
0.42
0.03
0.00
0.04
0.03
0.05
0.10
0.04
0.07
0.00
0.00
0.07
0.07
0.05
0.29

1.00
0.67
0.43
0.33
0.67
1.00
1.00
1.00
0.00
0.00
0.00
0.00
0.11
0.00
0.00
0.00
0.00
0.17
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.05
0.00
0.00
0.00
0.00
0.00
0.21

1.00
1.00
0.88
1.00
1.00
1.00
1.00
1.00
0.06
0.33
0.20
0.40
0.40
0.43
0.40
0.25
0.83
0.60
0.11
0.00
0.14
0.13
0.13
0.18
0.17
0.13
0.00
0.00
0.17
0.17
0.13
0.43

0.00
0.12
0.12
0.18
0.17
0.00
0.00
0.00
0.02
0.12
0.06
0.12
0.07
0.10
0.10
0.08
0.25
0.15
0.05
0.00
0.06
0.05
0.05
0.04
0.06
0.03
0.00
0.00
0.04
0.05
0.03
0.07

objective evaluation. The MHDE algorithm is coded


in C and uses the parameters provided in reference [17]. For fair comparisons, the MHDE algorithm
is executed on the same personal computer (PC) and
uses the same stopping criteria as those of the MOPSOLS algorithm. Since it is very difficult to find a reference solution set S* for a testing problem, an
approximate reference solution set is obtained by
using the MHDE and MOPSOLS algorithms with a
long running time. More specifically, both algorithms
are carried out 30 replications for each testing
problem with computation time T 100 mn milliseconds, and only non-dominated solutions are chosen
as reference solutions from the 60 obtained nondominated solution sets.
The statistics of D1R, NNDS(), and RNDS() produced
by the MOPSOLS and MHDE algorithms are shown in
Tables 10, 11, and 12 respectively. It can be seen from
Table 10 that the MOPSOLS algorithm achieves much
better AVG, MIN, MAX, and SD than the MHDE algorithm in terms of the performance measure D1R, and
even MAX obtained by the MOPSOLS algorithm is
better than MIN by the MHDE algorithm for every
instance. In particular when the size of instance is
relatively small, such as car01 to car08, the MOPSOLS
algorithm can almost find all the reference solutions.

As the size of the instance grows, the difference of


AVG between the MOPSOLS and MHDE algorithm
increases. Therefore, it is concluded that the proposed MOPSOLS algorithm is more robust, effective,
and efficient than the MHDE algorithm to find a uniformly distributed and high-quality approximate Pareto-optimal set.
It can be found from Tables 11 and 12 that the
MOPSOLS algorithm yields much better NNDS() and
RNDS() than the MHDE algorithm for every instance.
As the size of the instance grows, the superiority of
the MOPSOLS algorithm over the MHDE algorithm
is more obvious. When the size of instance is larger
than 50 10, both NNDS() and RNDS() produced by
the MHDE algorithm are equal to 0. This means
that for these instances, every solution of the HMDE
algorithm is dominated by some solutions of the
MOPSOLS algorithm, and no solution of the MOPSOLS
algorithm is dominated by any solution of the MHDE
algorithm. Therefore it is concluded that the MOPSOLS
algorithm can produce much more non-dominated
solutions with high quality than the MHDE algorithm.
The aforementioned results depict the overall
performance of the MOPSOLS algorithm. However, it
can be observed from Tables 10, 11, and 12 that
some instances are very easy to solve, whereas others

http://www.paper.edu.cn
A novel multi-objective particle swarm optimization algorithm

5.6 Comparison of the MOPSOLS and MOPSO


algorithm
Next, the MOPSOLS algorithm is compared with the
MOPSO algorithm to show the effectiveness of the
fuse of the PSO-based search and local search. First,
both algorithms, 20 independent replications with
maximum running time T 10mn milliseconds are
carried out for each testing problem. The statistics

Emperical cumulative frequency

100
MOPSO LS
MOPSO

90
80
70
60
50
40
30
20
10
0

10

12

14

16

18

20

D1R

(a) Empirical cumulative distribution function of D1R


100
MOPSO LS
MOPSO

Emperical cumulative frequency

90
80
70
60
50
40
30
20
10
0

10

15

20

25

30

35

40

45

50

NNDS

(b) Empirical cumulative distribution function of NNDS


100
Emperical cumulative frequency

are very difficult. To justify further the performance


of the MOPSOLS algorithm in solving difficult problems, the hardest testing problem Hel1 is selected
and the MHDE and MOPSOLS algorithms 100 independent trials are run on the same PC mentioned
before. The cumulative empirical distribution functions of D1R, NNDS(), and RNDS() are given in Fig. 3.
It can be seen from the plots that the solution quality distribution of the MOPSOLS algorithm is to the
left of the MHDE algorithm in terms of D1R, and to
the right of the MHDE algorithm in terms of both
NNDS() and RNDS(). This implies that the MOPSOLS
algorithm has a much higher probability to reach a
good approximate Pareto-optimal set for difficult
problems than the MHDE algorithm.
To understand better the performance of the MOPSOLS algorithm, typical results of a replication for
both algorithms based on problem Hel1 are shown
in Fig. 4. It can be easily seen that the MOPSOLS algorithm performs much better than the HDE algorithm.
The conclusion is similar for other problems.
Next, the MOPSOLS algorithm is further compared
with the MHDE algorithm as the maximum running
time increases. More specifically, the maximum running time for both algorithms is set from 10mn milliseconds to 100mn milliseconds with a step of 5mn
milliseconds, and their other parameters remain
unchanged. Each testing problem is independently
run 20 replications, and the overall mean values of
D1R, NNDS(), and RNDS() are given in Fig. 5.
It can be observed from Fig. 5 that the MOPSOLS
algorithm performs much better than the MHDE
algorithm at the same maximum running time T.
Especially, the MOPSOLS algorithm with T 10mn
milliseconds even outperforms the MHDE algorithm
with T 100mn milliseconds. This means that the
MOPSOLS algorithm not only can produce much
more non-dominated solutions with high quality at
a given maximum computational time, but it also is
of a higher rate of convergence to Pareto-optimal
set than the MHDE algorithm.
Based on the above comparisons, it is concluded
that the proposed MOPSOLS algorithm is superior to
the MHDE algorithm in terms of searching quality,
diversity level, robustness, and efficiency for the nowait flow shop scheduling problems with makespan
and maximum tardiness criteria.

535

MOPSO LS
MOPSO

90
80
70
60
50
40
30
20
10
0

0.5

1.5

RNDS
(c) Empirical cumulative distribution function of RNDS

Fig. 6 Plots of empirical cumulative distribution function


for the MOPSOLS and MOPSO algorithms

of those performance measures are reported in


Tables 13, 14, and 15 respectively. From these tables,
it is shown that the statistical results generated by the
MOPSOLS algorithm are much better than those by
the MOPSO algorithm, which means that by embedding a simple local search algorithm, the MOPSO
algorithm becomes much more robust, efficient,
and effective in finding an approximate Paretooptimal set with high quality.

http://www.paper.edu.cn

536

Q-K Pan, L Wang, and B Qian

600
MOPSO LS
MOPSO
500

Maxmimum Lateness

400

300

200

100

0
720

740

760

780

800

820

840

860

880

Makespan

Fig. 7 Non-domiated solutions of the MOPSOLS and MOPSO algorithms

Second, both algorithms are run for testing problem Hel1 to compare their performances in solving
difficult problems with large size. The cumulative
empirical distribution functions of 100 independent
runs are given in Fig. 6. This figure illustrates that
the solution quality distribution of the MOPSOLS
algorithm is much superior to that of the MOPSO
algorithm, indicating once again that by fusing a simple local search algorithm, the performance of the
MOPSO algorithm is remarkably improved in solving
difficult problems with large size.
Typical results of a replication for the MOPSOLS
and MOPSO algorithms based on problem Hel1 are
shown in Fig. 7. It can be seen that the solutions
obtained by the MOPSOLS algorithm dominate all
the solutions by the MOPSO algorithm. The conclusion is similar for other problems.
Last, the MOPSOLS algorithm is further compared
with the MOPSO algorithm as the maximum running
time increases from 10mn milliseconds to 100mn
milliseconds with a step of 5mn milliseconds. The
statistics of 20 independent replications are shown
in Fig. 8. This figure shows that the MOPSOLS algorithm produces much better statistical results than
the MOPSO algorithm at T 10mn milliseconds.
Even if the maximum running time T of the MOPSO

algorithm is increased to 100mn milliseconds, it still


cannot generate comparable results. So it is concluded that with a local search the MOPSO algorithm
has a much higher convergence rate to Paretooptimal set.
In brief, the above comparisons demonstrate the
effectiveness of incorporating a local search into the
MOPSO algorithm. That is to say, the superiority of
the MOPSOLS algorithm in terms of searching quality,
diverse level, and robustness is attributable to the
combination of global search and local search or the
balance of exploration and exploitation.

6 CONCLUSIONS
To the best of the current authors knowledge, this
is the first report to propose a multi-objective particle
swarm optimization (MOPSO) algorithm for the nowait flow shop scheduling problems. Unlike the traditional PSO algorithm, the proposed MOPSO algorithm
directly worked in the discrete domain by employing a
job-permutation-based encoding scheme and a
job-permutation-based position updating operation.
In order to find the Pareto-optimal set, the MOPSO
algorithm used the concept of Pareto dominance to

http://www.paper.edu.cn
A novel multi-objective particle swarm optimization algorithm

25
MOPSO LS
MOPSO
20

D1R

15

10

10

20

30

40
50
60
70
Computational Time

80

mn
100

90

(a) Overall mean value of D1R

40
35
MOPSO LS
MOPSO

30
NNDS

25

537

process. Moreover, a new multi-objective heuristic,


the PWQ heuristic, was proposed to produce a population of initial particles with relatively good performances, and a multi-objective local search algorithm
was developed to embed in the MOPSO algorithm
for stressing the balance between global exploration
and local exploitation. In addition, two speed-up
methods for the evaluation of a job permutation and
its insert neighbourhood were devised to reduce the
running time requirements. Simulation results and
comparisons demonstrated the superiority of the proposed algorithm in terms of searching quality, diversity level, robustness, and effectiveness. The future
work is to extend the current authors bi-objective
insert-neighbourhood-based speed-up method to
the problems with three or more objectives, and they
aim to present more effective heuristics and to extend
their MOPSO for three-objective no-wait flow shop
scheduling problems. In addition, they will develop
the adaptive MOPSO algorithm and design effective
PSO-based multi-objective algorithms for other types
of combinatorial optimization problems.

20
15

ACKNOWLEDGEMENTS

10
5
0
10

20

30

40
50
60
70
80
Computational Time
(2) Overall mean value of NNDS

1.5

90

100

mn

MOPSOLS
MOPSO

REFERENCES

RNDS

0.5

0
10

20

30

40
50
60
70
Computational Time

80

90

100

This research is partially supported by National


Science Foundation of China under Grants
60774082, National 863 Hi-Tech R&D Plan under
Grant 2007AA04Z155, and the Project-sponsored by
SRF for ROCS, SEM, and Postdoctoral Science Foundation of China under Grants 20070410791.

mn

(3) Overall mean value of RNDS

Fig. 8 Comparison of the MOPSOLS and MOPSO with running time increasing

compare different solutions and dynamically updated


the obtained non-dominated solution set, where a
non-dominated solution was randomly chosen as
the global best particle to maintain the diversity of
searching direction and to speed up the convergence

1 Pinedo, M. Scheduling: theory, algorithms and systems,


2002, (Prentice-Hall, New Jersey).
2 Wang, L. and Zheng, D. Z. An effective hybrid heuristic
for flow shop scheduling. Int. J. Advd Mfg Technol., 2003,
21, 3844.
3 Daniels, R. L. and Chambers, R. J. Multiobjective flowshop scheduling. Nav. Res. Logist., 1990, 37, 981995.
4 Rajendran, C. Two-stage flowshop scheduling problem
with bicriteria. J. Opl. Res. Soc., 1992, 43(9), 871884.
5 Ravindran, D., Noorul, Haq, A., Selvkuar, S. J., and
Sivaraman, R. Flow shop scheduling with multiple
objective of minimizing makespan and total flow
time. Int. J. Advd Mfg Technol., 2005, 25, 10071012.
6 Toktas, B., Azizoglu, M., and Koksalan, S. K. Twomachine flow shop scheduling with two criteria: maximum earliness and makespan. Eur. J. Opl. Res., 2004,
157, 286295.
gt, D.
7 TKindt, V., Monmarche, N., Tercinet, F., and Lau
An ant colony optimization algorithm to solve a
2-machine bicriteria flowshop scheduling problem.
Eur. J. Opl. Res., 2002, 142, 250257.
8 Loukil, T., Teghem, J., and Tuyttens, D. Solving multiobjective production scheduling problems using metaheuristics.. Eur. J. Opl. Res., 2005, 161, 4261.


538

http://www.paper.edu.cn
Q-K Pan, L Wang, and B Qian

9 Rahimi-Vahed, R. A. and Mirghorbani, S. M. A multiobjective particle swarm for a flow shop scheduling problem. J. Comb. optimization, 2007, 13, 79102.
10 Tavakkoli-Moghaddam, R., Rahimi-Vahed, A., and
Mirzaei, A. H. Solving a bi-criteria flow shop problem
using immune algorithm. In Proceedings of the First
IEEE Symposium on Comput. Intelligence. Honolulu,
Hawaii, 2007, Vol. 1, pp. 4956.
11 Ponnambalam, S. G., Jagannathan, H., Kataria, M.,
and Gadicherla, A. A TSP-GA multi-objective algorithm
for flow-shop scheduling. Int. J. Advd Mfg Technol.,
2004, 23(11), 909915.
12 Murata, T., Ishibuchi, H., and Tanaka, H. Multiobjective genetic algorithm and its applications to flowshop
scheduling. Comput. Ind. Engng, 1996, 30(4), 957968.
13 Ishibuchi, H. and Murata, T. A multi-objective genetic
local search algorithm and its applications to flowshop
scheduling. IEEE Trans. Syst. Man, Cybernetics, C Appl.
Rev., 1998, 28(3), 392403.
14 Ishibuchi, H., Yoshida, I., and Murata, T. Balance
between genetic search and local search in memetic
algorithms for multiobjective permutation flowshop
scheduling. IEEE Trans. Evolutionary Comput., 2003, 7
(2), 204223.
15 Arroyo, J. E. C. and Armentano, V. A. Genetic local
search for multi-objective flowshop scheduling problems. Eur. J. Opl Res., 2005, 167, 717738.
16 Li, B. B. and Wang, L. A hybrid quantum-inspired
genetic algorithm for multiobjective flow shop scheduling. IEEE Trans System Man Cybernetics B, Cybernetics,
2007, 37, 576591.
17 Qian, B., Wang, L., Rong, H., Wang, W. L., Huang, D.
X., and Wang, X. A hybrid differential evolution for
permutation flow-shop scheduling. Int. J. Advd Mfg
Technol., 2007.
18 TKindt, V. and Billaut, J. C. Multicriteria scheduling:
theory, models and algorithms, 2002, (Springer-VerlagBerlin).
19 Rajendran, C. A no-wait flowshop scheduling heuristic
to minimize makespan. J. Opl Res. Soc., 1994, 45, 472
478.
20 Hall, N. G. and Sriskandarayah, C. A survey of machine
scheduling problems with blocking and no-wait in process. Opl Res., 1996, 44, 510525.
21 Grabowski, J. and Pempera, J. Sequencing of jobs in
some production system. Eur. J. Opl Res., 2003, 125,
535550.
22 Raaymakers, W. and Hoogeveen, J. Scheduling multipurpose batch process industries with no-wait restrictions by simulated annealing. Eur. J. Opl Res., 2000,
126, 13151.
23 Rock, H. The three-machine no-wait flowshop problem
is NP-complete. J. Assoc. Comput. Mach., 1984, 31, 336
345.
24 Bonney, M. C. and Gundry, S. W. Solutions to the constrained flowshop sequencing problem. Opl Res. Q.,
1976, 24, 869883.
25 King, J. R. and Spachis, A. S. Heuristics for flowshop
scheduling. Int. J. Prod. Res., 1980, 18, 343357.
26 Gangadharan, R. and Rajedran, C. Heuristic algorithms
for scheduling in no-wait flowshop. Int. J. Prod. Econ.,
1993, 32, 285290.

27 Rajendran, C. A no-wait flowshop scheduling heuristic to


minimize makespan. J. Opl Res. Soc., 1994, 45, 472478.
28 Rajendran, C. and Chaudhuri, D. Heuristic algorithms
for continuous flow-shop problem. Nav. Res. Logistics,
1990, 37, 695705.
29 Aldowaisan, T. and Allahverdi, A. A new heuristics for
no-wait flowshops to minimize makespan. Comput.
Opl Res., 2003, 30, 12191231.
30 Fink, A. and Vo, S. Solving the continuous flow-shop
scheduling problem by metaheuristics. Eur. J. Opl Res.,
2003, 151, 400414.
31 Schuster, C. J. and Framinan, J. M. Approximative procedure for no-wait job shop scheduling. Opl Res. Lett.,
2003, 31, 308318.
32 Grabowski, J. and Pempera, J. Some local search algorithms for no-wait flow-shop problem with makespan
criterion. Comput. Opl Res., 2005, 32, 21972212.
33 Pan, Q. K., Wang, L., and Zhao, B. H. An improved iterated greedy algorithm for the no-wait flow shop scheduling problem with makespan criterion. Int. J. Advd
Mfg Technol., 2007.
34 Liu, B., Wang, L., and Jin, Y. H. An effective hybrid
particle swarm optimization for no-wait flow shop
scheduling. Int. J. Advd Mfg Technol., 2007, 31, 1001
1011.
35 Pan, Q. K., Tasgetiren, M. F., and Liang, Y. C. A discrete
particle swarm optimization algorithm for the no-wait
flowshop scheduling problem. Comput. Ops Res., 2006.
36 Pan, Q. K., Wang, L., Tasgetirend, M. F., and Zhao,
B. H. A hybrid discrete particle swarm optimization
algorithm for the no-wait flow shop scheduling problem
with makespan criterion. Int. J. Advd Mfg Technol., 2007.
37 Tavakkoli-Moghaddam, R., Rahimi-Vahed, A., and
Mirzaei, A. H. A hybrid multi-objective immune algorithm for a flow shop scheduling problem with biobjectives: weighted mean completion time and
weighted mean tardiness. Inf. Sci., 2007, 177(22), 5072
5090.
38 Spieksma, F. C. R. and Woeginger, G. J. The no-wait
flow-shop paradox. Ops Res. Letts., 2005, 33(6), 603608.
39 Kalczynski, P. J. and Kamburowski, J. On no-wait and
no-idle flow shops with makespan criterion. Eur. J. Opl
Res., 2007, 178, 677685.
40 Hall, N. G. and Sriskandarayah, C. A survey of machine
scheduling problems with blocking and no-wait in process. Opl Res., 1996, 44, 510525.
41 Eberhart, R. C. and Kennedy, J. A new optimizer using
particle swarm theory. Proceedings of the sixth international Symposium on Micro Mach. Human Sci. Nagoya,
Japan, 1995, pp. 3943.
42 Kennedy, J., Eberhart, R. C., and Shi, Y. Swarm intelligence, 2001, (Morgan Kaufmann, San Mateo, CA, USA).
43 Liu, D. S., Tan, K. C., Goh, C. K., and Ho, W. K. A multiobjective memetic algorithm based on particle swarm
optimization. IEEE Trans Syst., Man and Cybernetics:
Part B (Cybernetics), 2007, 37(1), 4250.
44 Salman, A., Ahmad, I., and Al-Madani, S. Particle
swarm optimization for task assignment problem.
Microprocessors and Microsystems, 2003, 26, 363371.
45 Tasgetiren, M. F. and Liang, Y. C. A binary particle
swarm optimization algorithm for lot sizing problem.
J. Econ. Soc. Res., 2003, 5, 120.

http://www.paper.edu.cn
A novel multi-objective particle swarm optimization algorithm

46 He, Q. and Wang, L. A hybrid particle swarm optimization with a feasibility-based rule for constrained
optimization. Appl. Math. Computer, 2007, 186,
14071422.
47 Liu, B., Wang, L., and Jin, Y. H. An effective pso-based
memetic algorithm for flow shop scheduling. IEEE
Trans Syst. Man and Cybernetics, B, Cybernetics, 2007,
37, 1827.
48 Tasgetiren, M. F., Liang, Y. C., Sevkli, M., and Gencyilmaz,
G. Particle swarm optimization algorithm for makespan and total flowtime minimization in permutation
flowshop sequencing problem. Eur. J. Opl Res., 2007,
177, 19301947.
49 Onwubolu, G. C. and Clerc, M. Optimal operating path
for automated drilling operations by a new heuristic
approach using particle swarm optimization. Int. J.
Prod. Res., 2004, 42(3), 473491.
50 Lian, Z., Gu, X., and Jiao, B. A similar particle swarm
optimization algorithm for job-shop scheduling to
minimize makespan. Appl. Math. Comput., 2006, 183,
10081017.
51 Landa Silva, J. D., Burke, E. K., and Petrovic, S. An
introduction to multobjective metaheuristics for scheduling and timetabling. Lecture Notes Econ. Math Syst.,
2004, 535, 91129.

539

52 Nawaz, M., Enscore, E. E. J., and Ham, I. A heuristic


algorithm for the m-machine, n-job flow shop sequencing problem. OMEGA, 1983, 11, 9195.
53 Tommaso, S. and Thomas, S. A review of metrics on
permutations for search landscape analysis. Comput.
Opl Res., 2007, 34(10), 31433153.
54 Knowles, J. D. and Corne, D. W. On metrics for comparing nondominated sets. In Proceedings of 2002 Congress on Evolutionary Computation. Honolulu, HI
1217 May 2002, pp. 711716.
55 Carlier, J. Ordonnancements a contraintes disjonctives.
RAIRO Recherche Operationelle, 1978, 12, 333351.
56 Reeves, C. A genetic algorithm for flowshop sequencing.
Comput. Opl Res., 1995, 22, 513.
57 Heller, J. Some numerical experiments for an M J flow
shop and its decision-theoretical aspects. Opl Res., 1960,
8, 178184.
58 Zitzler, E. and Thiele, L. Multiobjective evolutionary
algorithms: a comparison case study and the strength pareto approach. IEEE Trans Evolutionary Comput., 1999, 3
(4), 257271.
59 Knowles, J. D. and Corne, D. W. Approximating the
nondominated front using the Pareto archive
evolution strategy. Evolutionary Comput., 2002, 8(2),
149217.

Das könnte Ihnen auch gefallen