2014 IEEE Congress on Evolutionary Computation (CEC)

July 6-11, 2014, Beijing, China

A Chaotic Particle Swarm Optimization Algorithm

for the Jobshop Scheduling Problem*
YAN Ping JIAO Minghai
School of Economics and Management Computing Center
Shenyang Aerospace University Northeastern University
Shenyang, China Shenyang, China
[email protected] [email protected]

Abstract—An improved Chaotic Particle Swarm Optimization continuous and discrete problems. Recently, PSO algorithms
(CPSO) algorithm for a jobshop scheduling problem, with were successfully applied to a wide range of applications such
minimization of makespan as the criterion, is proposed in this as aggregate production planning by Wang and Yeh [6], data
research. A real-valued encoding scheme based on a matrix envelopment analysis by Meng [7], trajectory planning
representation is developed, which converts the continuous problem of underactuated spacecrafts by Zhuang and Huang
position value of particles in PSO to the processing order of job [8], and medical technology by Yang et al. [9]. However, the
operation. A compound chaotic search strategy that integrates converging processes of the standard version of PSO algorithm
both Tent and Logistic chaotic search process is employed to the tend to be too slow for practical-scale JSSPs. In this paper,
global best particle to enhance the local searching ability of PSO.
some effort has been devoted to the modification and
In addition, a gaussian disturbance technology is embedded in
improvement of PSO algorithm that is applied to the JSSP.
the CPSO algorithm to improve the diversity of the particles in
the swarm. The performance of CPSO is compared with the The reminder of the paper is organized as follows: In the
standard PSO algorithm on a benchmark instance of jobshop following section, details of the JSP are described. Then we
scheduling problems. The results show that the proposed CPSO give a brief introduction to PSO algorithm and its inspiration.
algorithm has a superior performance to the PSO algorithm. The structure of the proposed PSO algorithm is presented in
Section 3. A series of comparative experiments are conducted
Keywords—scheduling; particle swarm optimization; chaotic
in Section 4 to evaluate the performance of the proposed
search
algorithm. Finally, conclusions are given in Section 5.
I. INTRODUCTION
II. PROBLEM DESCRIPTION
Among all types of scheduling problems, job shop
scheduling problem (JSSP) has important applications in Scheduling for job shops is an important topic in production
different industries. This problem is known to be NP-hard in management. Generally for the job-shop scheduling problem
the strong sense [1]. Thus, it is unlikely to obtain the optimal there are a set of n jobs are waiting to be processed on a set of
schedule through polynomial time-bounded algorithms. Over m machines where the processing of each job consists of m
the years there has been a great deal of research to develop operations performed on these machines in a specified
efficient approaches for the problem. A new hybrid swarm sequence. In this paper we consider the deterministic and
intelligence algorithm to solve the job-shop scheduling static job-shop scheduling problem. Some basic assumptions
problem is proposed by Tsung-Lieh Lin et al. [2]. Roshanaei et are in the following:
al. [3] develop new solution methodologies including 1) All the jobs are available at time zero.
mathematical modelling and a meta-heuristic for the flexible 2) Each machine can process only one job at a time.
job shop scheduling problem. Mencia et al. [4] apply a hybrid
3) Each job can be processed by only one machine at a time.
search algorithm that interleaves best-first and depth-first
search to the job shop scheduling problem. Yuan and Xu [5] 4) Once started operations cannot be interrupted.
integrate hybrid harmony search (HHS) and large 5) All the machines are continuously available throughout the
neighborhood search (LNS) for the flexible job shop production stage.
scheduling problem with makespan criterion. Among these 6) The transportation time to deliver relevant jobs between
solution methodologies of JSSP, intelligent optimization different machines is neglected.
algorithms are relatively easy to implement and they could 7) The setup time for the machines to switch between different
conveniently be adapted for different kinds of scheduling jobs is neglected.
problems. This has made the research on them increasingly The manufacturing procedure to be performed on any one
popular in the recent years. PSO is one of the latest machine is called an operation of the job. The operation of job i
metaheuristic methods in the literature. Based on the metaphor has to be performed on machine j with deterministic processing
of social interaction and communication such as bird flocking time tij. The goal for optimization is to minimize the makespan
and fish schooling, PSO was introduced to optimize both (Cmax). In other words, the scheduling objective considered in
this paper is to determine the processing sequence of the
This work was supported in part by Humanistic social science fund of
China's Ministry of education (NO.12YJC630312),Science and Technology
Projects of Zhejiang Province under Grant #2011C03004.

978-1-4799-1488-3/14/$31.00 ©2014 IEEE 218

operations on each machine such that the completion time of Step 4. Select the operation r* with the largest priority value
the last operation could be minimized, i.e., from Ψ and schedule it. If more than one operation
exists according to the priority values, tie is broken by a
Minimize F = Max (Ci ) random choice.
i∈{1,", n}

where Ci is the completion time of job i. Step 5. Delete the operation r* from Φ, and put its immediate
III. CPSO ALGORITHM FOR JOB SHOP SCHEDULING successor into Φ.
Step 6. Update eij in Φ and return to step 2 until all operation
A. Standard PSO are scheduled.
Particle swarm optimization (PSO), originally designed by An initial population is generated randomly where xij is
Kennedy and Eberhart [10] in 1995, is inspired by observing drawn from [ xijlow , xijup ]. xijlow and xijup are the lower and upper
the bird flocking or fish school. In the PSO algorithm, birds bounds of xij, respectively. As for the range of particle velocity,
are called particles, each representing a potential solution. All
we clip the range of particle velocities vij within [ − ( xijup − xijlow ) ,
particles have their position, velocity, and fitness values. To
find the optimal solution, each particle adjusts its flying xijup − xijlow ].
according to its own flying experience and its companions’
flying experience. For n-dimensional search space, each C. A Compound Chaotic Search Technology for the Global
particle in the swarm population has the following attributes: a Best Particle
current position represented as X ik = [ xik1 , xik2 ,", xink ] ; a current
Chaos is a deterministic, random-like process found in non-
velocity represented as Vi k = [vik1 , vik2 ,", vink ] ; a current personal linear, dynamical system, which is non-period, non-
best position represented as Pi k = [ pik1 , pik2 ,", pink ] . Let converging and bounded. The two main characteristics of the
chaos method concentrate on randomicity and ergodicity that
G k = [ g1k , g 2k ,", g nk ] denote the current global best position.
will strengthen the performance of the traditional PSO
During the evolution process of the swarm, the new position
algorithm. In order to emphasize exploitation ability, a
and velocity of each particle i are determined by the following
compound chaotic search for the global best particle is
equations:
embedded in the proposed PSO algorithm. The chaotic search
vijk +1 = w k vijk + c1r1 ( pijk − xijk ) + c2 r2 ( g kj − xijk ) (1) process integrates Tent and Logistic chaotic maps which are
xijk +1 = xijk + vijk +1 (2) defined as equations (3) and (4) respectively.
xn +1 = 1 − 2 | xn |, − 1 < xn ≤ 1 (3)
B. Coding Scheme and Initial Population Generation
xn+1 = λxn (1 − xn ), 0 < xn < 1, 0 < λ ≤ 4 (4)
In solving the jobshop scheduling by PSO, first task is to
A compound chaotic map can be obtained by incorporating
represent a solution of a problem as a particle. We utilize a
the above two maps as follows:
priority-based representation where a particle is encoded as a
xn+1 = 1 − 2λxn (1 − xn ), 0 < xn ≤ 1 (5)
matrix. Suppose n jobs are to be scheduled on m machines. An
n×m matrix is given by a particle as follows: Some good results have been shown in some applications
when λ is set to 2 [12]. Let xijlow and xijup be the lower and
⎡ x11 x12 " x1m ⎤
⎢ ⎥ upper bounds of xij in a particle matrix respectively. The local
x x22 " x2 m ⎥ ,
X = ⎢ 21 search algorithm is described in the following:
⎢ # # # # ⎥
⎢ ⎥ Step 1. Set k = 1.
⎢⎣ xn1 xn 2 " xnm ⎥⎦
Step 2. Map matrix variables xijk to chaotic variables cxijk in
where xij denotes the priority value of job i which is processed
on machine j. Then, we use Giffler and Thompson's heuristic terms of (6):
[11] to create a feasible schedule based on the priorities values xijk − xijlow (6)
cxijk =
of job operations for each particle. Below is a brief outline of xijup − xijlow
the G&T algorithm for obtaining schedules. Let Φ contain the
Step 3. Generate chaotic variables cxijk +1 based on the
schedulable operation of each job and eij denote the earliest
time at which the operation (i, j) can start. compound chaotic map (5) where λ is set to 2, i.e.:
Step 1. Let eij =0, for all operation (i, j) in Φ. cxijk +1 = 1 − 4cxijk (1 − cxijk ), 0 < cxijk ≤ 1 (7)
Step 2. Compute f * = min {e + t } where tij is the Step 4. Translate the chaotic variables cxijk +1 into matrix
( i , j )∈Φ ij ij
processing time of operation (i, j). Let m* denote the variables xijk +1 :
machine on which the minimum is achieved.
Step 3. Find out the conflict set Ψ of all operations (i, m*) on xijk +1 = xijlow + cxijk +1 ( xijup − xijlow ) (8)
machine m* such that eim* < f * .

219
Step 5. Decode the matrix representation of the particle into a IV. COMPUTATIONAL EXPERIMENTS
feasible processing sequence of the operations on all
machines. A. Experiment Design
Step 6. Compute the fitness fk+1of the particle. To test the performance of the proposed CPSO algorithm
Step 7. Update the best position of the particle if a better for solving JSSP, some comparative experiments are carried
fitness (fk+1< fk) is obtained by the chaotic search. out. A benchmark instance of JSSP which is prefixed by LA01
Step 8. Set k = k +1. Go to step 2 until the maximum number of and described by Lawrence [13] is employed in the
iteration for chaotic search is achieved. experiments. There are 10 jobs to be processed on 5 machines
in LA01 and the optimal objective function value of the test
D. Gaussian Disturbance Strategy problem is 666 as reported in the literature for the makespan
One of the major drawbacks of the PSO is its premature criterion. Table I and II list the processing machine sequence
convergence, especially while handling problems with more and processing time of each job on all the machines in details.
local optima. Particles tend to be stagnant when their In the experiment, the parameters for the proposed CPSO
velocities are near to zero. In this paper, a gaussian algorithm are assigned as follow. The number of the
disturbance strategy is introduced in the proposed PSO to population size of the particles is set to 30. Default values for
solve the global optimization problem. When matrix variables the parameters c1 and c2 have been used: c1 = c2 = 2. The
xij of a particle are unchanged in μ consecutive iterations, the inertia weight w is set to decrease linearly from 0.9 to 0.4
disturbance strategy will be applied to the particle in the during a run of the PSO. The time decreasing inertia weight
swarm. Suppose G k = [ g ijk ]n×m be the current global best allows the PSO to explore a large area at the start of the run,
and to refine the search later by using a smaller inertia weight.
position in the kth iteration of CPSO. Formula (9) gives the To clip the range of positions and velocities for the encoding
gaussian disturbance process. scheme of particles, we set xijup (i=1, 2, …, n; j=1, 2, …, m) to
⎧ g ijk × [1 + C × N (0,1)], 0 ≤ rnd ≤ 0.98
⎪ be 4 while xijlow is set to −4. The maximum number of iteration
k ⎪ (9) for compound chaotic search is set to 10. The gaussian
xij = ⎨ k k
r
⎪ ij , rij is a random real number disturbance parameter C is set to 0.9. The termination criterion
⎪ and rijk ∈ [ xijlow , xijup ], otherwise of the proposed CPSO is the maximum of 1000 iterations in
⎩ population evolution.
where C is a predefined constant parameter. N(0,1)
represents the standard normal distribution where mean is set TABLE I. PROCESSING MACHINE SEQUENCES OF JOBS
to 0 and variance is set to 1. For each matrix variable xij to be
disturbed, a random value rnd is generated uniformly from [0, M1 M2 M3 M4 M5
1]. Whether a gaussian disturbance or a random immigrant J1 2 1 5 4 3
technique will be applied in terms of the value of rnd. J2 1 4 5 3 2
J3 4 5 2 3 1
E. The Framework of Proposed CPSO J4 2 1 5 3 4
J5 1 4 3 2 5
Based on the above design, the procedure of the proposed
J6 2 3 5 1 4
CPSO algorithm for solving JSSP is summarized as follows.
J7 4 5 2 3 1
Step 1. Generate randomly an initial population within the
J8 3 1 2 4 5
range of position and velocity values.
J9 4 2 5 1 3
Step 2. Translate matrix representations of particles into
J10 5 4 3 2 1
feasible scheduling schemes.
Step 3. Evaluate all particle individuals in the population
TABLE II. PROCESSING TIME OF JOBS
based on the fitness.
Step 4. Update the personal best particle Pbest as well as M1 M2 M3 M4 M5
global best particle Gbest in the swam. J1 21 53 95 55 34
Step 5. Apply the gaussian disturbance strategy to each J2 21 52 16 26 71
particle in the current population. Update the personal J3 39 98 42 31 12
best position and the current global best position J4 77 55 79 66 77
whenever a better fitness is obtained by the gaussian J5 83 34 64 19 37
disturbance. J6 54 43 79 92 62
Step 6. Perform the chaotic local search algorithm on the J7 69 77 87 87 93
current global best particle Gbest. J8 38 60 41 24 83
Step 7. Update the velocities and positions of particles J9 17 49 25 44 98
according to equations (1) and (2). J10 77 79 43 75 96
Step 8. Go back to Step 2, unless the termination condition is
met.

220
 All of the compared algorithms are coded in Visual C++ Then, we compare the experimental results of proposed
and the experiments are executed on a Pentium PIV 3.0 GHz CPSO with standard PSO. Table III summarizes the results of
PC with 512 MB memory. the experiments. From Table III it can be seen that the best and
average solutions found by PSO are 671 and 672.6
B. Computational Results respectively while the results of CPSO is much better than the
Firstly, some experiments are designed in order to find an original design. CPSO obtains the optimal solution 666 and its
optimal value for parameter μ in the gaussian disturbance average solution is 668 which are much closer to the optimal
strategy. The experiments of the JSSP optimization are done solution. The convergence trends of PSO and CPSO are shown
three times using the CPSO algorithm by adjusting the values in Fig. 2. The proposed CPSO algorithm can reach the best
of μ from 1 to 100. Three values including 1, 10 and 100 are solution in 60 iterations. Whereas the particles of PSO do not
tested for μ. The results of the experiments are presented in converge to global optima after 1000 iterations. The
Figure 1. The results show that the performance of CPSO performance of CPSO can be attributed to the compound
varies when μ set to different values. When μ is set to 100, chaotic search and gaussian disturbance improvements which
CPSO is to be stagnant in local minima and does not converge enhance search behavior of PSO and allow it to avoid local
to the global minima in the end. CPSO with μ=1 converges optima.
much faster than CPSO with μ=10 although both of them
converge to the optimal makespan value. It seems that a TABLE III. COMPUTATIONAL RESULTS OF STANDARD PSO AND CPSO
smaller μ will achieve more improvement in solution than a
PSO CPSO
larger one. Thus, we select μ=1 for CPSO.
Best Worst Average Best Worst Average
671 724 672.6 666 740 668

V. CONCLUSION
In this research, the idea of applying chaotic search
principle in particle swarm optimization to the jobshop
scheduling problems has been explored. By applying a special
encoding mechanism, the continuous position values of
particles in PSO are converted to the processing order of job
operation. A compound chaotic local search strategy is
designed to perform exploration for promising solutions within
the entire region. The approach also incorporates a gaussian
disturbance technique that is used to enhance diversity to the
population. Simulation results and comparisons demonstrate
the effectiveness of the proposed CPSO in terms of searching
quality and robustness. The future work is to investigate the
applications of the proposed method to some other kinds of
combinatorial optimization problems.
Fig. 1. The results of CPSO with different μ.
REFERENCES
[1] J.K. Lenstra, A.H.G.R. Kan, and P. Brucker, “Complexity of machine
scheduling problems,” Anal. Discrete Math. , vol. 7, no. 1, pp. 343–362,
1977.
[2] T. Lin, S. Horng, T. Kao, Y. Chen, R. Run, R. Chen, J. Lai, and I. Kuo,
“An efficient job-shop scheduling algorithm based on particle swarm
optimization,” Expert Syst Appl. , vol. 37, no. 3, pp. 2629-2636, 2010.
[3] V. Roshanaei, A. Azaba, and H. ElMaraghy, “Mathematical modelling
and a meta-heuristic for flexible job shop scheduling,” Int J Prod Res.,
vol. 51, no. 20, pp. 6247-6274, 2013.
[4] C. Mencia, M. R. Sierra, and R. Varela, “An efficient hybrid search
algorithm for job shop scheduling with operators,” Int J Prod Res., vol.
51, no. 17, pp. 5221-5237, 2013.
[5] Y. Yuan and H. Xu, “An integrated search heuristic for large-scale
flexible job shop scheduling problems,” Comput Oper Res., vol. 40, , no.
12, pp. 2864–2877, December 2013.
[6] S. Wang and M. Yeh, “A modified particle swarm optimization for
aggregate production planning,” Expert Syst Appl. , vol. 41, pp. 3069–
3077, 2014.
[7] M. Meng, “A hybrid particle swarm optimization algorithm for
satisficing data envelopment analysis under fuzzy chance constraints,”
Fig. 2. The representative convergence curves for PSO and CPSO. Expert Syst Appl. , vol. 41, pp. 2074–2082, 2014.

221
[8] Y. Zhuang and H. Huang, “Time-optimal trajectory planning for under [11] J. Giffler and G. L. Thompson, “Algorithms for solving production
actuated spacecraft using a hybrid particle swarm optimization scheduling problems,” Oper. Res., vol. 8, pp. 487–503, July-August
algorithm,” Acta Astronaut. , vol. 94, pp. 690–698, 2014. 1960.
[9] C. Yang, H. Chang, Y. Cheng, and L. Chuang, “Novel generating [12] X. Yi and K. Yang, “Improvement and performance research of hybrid
protective single nucleotide polymorphism barcode for breast cancer algorithm based on compound chaotic-artificial fish swarm,” Computer
using particle swarm optimization,” Cancer Epidemiology, vol. 33, pp. Engineering & Science, vol. 35, no. 8, pp. 89-95, 2013.
147–154, 2009. [13] S. Lawrence, “Resource constrained project scheduling: an experimental
[10] J. Kennedy and R. C. Eberhart, “Particle swarm optimization,” in Conf. investigation of heuristic scheduling techniques (supplement),” in
Rec. 1995 IEEE Int. Conf. on neural networks , New York: Piscataway, Technical Report, Graduate School of Industrial Administration,
pp. 1942–1948. Carnegie Mellon University, Pittsburgh, 1984.

222