Proceedings of the 7th

World Congress on Intelligent Control and Automation

June 25 - 27, 2008, Chongqing, China

Adaptive Genetic Algorithms for The Job-Shop

Scheduling Problems *
Gui Yang, Yujun Lu and Ren-wang Li Jin Han
Advanced Mechanical Technology Inst Sichuan tobacco industry co., ltd.
Zhejiang Sci-Tech University
Hangzhou, Zhejiang, 310018, China Chengdu, Sichuan, 610066, China
[email protected]

Abstract - In order to solve the feeble adaptability and the the theory of evolution. In this paper, an adaptive genetic
imbalance between random search and local search in the Job- algorithms (AGA) was proposed to conquer the limitation
Shop Scheduling Problem, a new adaptive genetic algorithm after analyzing the search property of GA in section III. The
(AGA) was presented in this paper. The superiority of this implementing of this method is also sketched. In section IV,
algorithm was the adaptation achieved by adjusting the crossover
the model of the AGA and arithmetic operators are presented.
rate and mutation rate. At the same time, the search property has
been balanced by restricting crossover and mutation. To insure Then, the algorithm is tested by Muth-Thompson problems
the best chromosome pass to the next generation, we immediately and getting the computational result. Finally, we obtained the
reserved the best chromosome. Operation-based representation conclusion.
was adopted. Therefore, work piece position-based Crossover
II. JSSP AND GA
and search region-based Mutation was applied in this paper. The
developed algorithm had been tested by benchmark problems. A. JSSP Definition
Computational results show this Adaptive Genetic Algorithm In this paper, the JSSP is defined as “n/m/G/*” [17]. In
(AGA) has an effective search behavior. This can get away from this model each of n jobs need process at most m different
local optimal and avoid premature convergence. Also the
convergence speed increases.
tasks. Each task is finished by one of m machines. The
execution of task nth job J i by machine M j is called
Index Terms – Job-Shop Scheduling Problem. Genetic operation Oij . The processing times of all operations Oij are
Algorithm. Adaptation.
known and denoted as pij . The parameter G indicates that the
I. INTRODUCTION jobs are constrained by a process called precedence constraint.
The job-shop scheduling problem (JSSP) is a typical non- It keeps to the order of tasks in the process. The symbol “*”
deterministic polynomial problem, which has many means a performance measure of optimization. In this paper,
applications in manufacturing engineering. The object of this we adopt the minimum make span, so “*= Cmax ”.
task is to find a sequence of operations of n jobs on each of m B. GA for JSSP
machines with given objective, such as minimizing the GA is generated by using evolution theory’sϘsurvival of
average tardiness or minimizing the make-span. This is to get the fittest ϙ .this is a random search method. It operates
high efficiency, flexibility and reliability. The academic structured object without the restriction of differential
significance of JSSP is rested on the property of N-P hard. At coefficient and continuity. Because the introductions of the
the same time, the fine scheduling strategy could decrease probability search, it could capture and guide the search space
work-span and storage, improve the production efficiency and automatically, adjust the search direction adaptively. These
benefit. Therefore a lot of research in the filed has been done. qualities have been applied widely in combinatorial
Especially in recent years, with the increasing emergence of optimization, machine learning, signal processing etc. It is one
intelligent algorithms, it is turning to a hotspot. Because of the of key technologies relating to intelligent computation.
characteristic of Genetic AlgorithmsΔit allows JSSP to work. As far back as 1985, GA was used to research JSSP by
[1]. Lawrence Davis [1]. Due to the complexity, it is difficult to
Although many approaches about AG were developed to code the problem in bit. Thus, many non-bit string
solve JSSP[9] Δ especially combining the hybrid representation have been proposed, such as operation-based,
algorithms[2,4,6] and heuristic algorithms[5], the parameters process-based, preferential process-based, process rule-based,
of these methods are almost constant. However, these and machine-based representation etc [9].
methods lack adaptability. In the general GA, the crossover Because of the simplicity of coding technology and
rate and mutation rate are designed by different testing in genetic operation, GA has become a universal optimize
accordance with different problems. It is hard to fine a proper algorithm. But GA is not sufficiently capable of handling the
value randomly. The invariability of the parameter is against large scale combinatorial optimization with large search space.

*
This work is supported by the National Natural Science Foundation, China (Grant NO. 50675208), Zhejiang Natural Science Foundation, China
(Grant NO. Y107575; Y605187) and Scientific Research Item of Zhejiang Province Education, China (Grant No. 20070333).

978-1-4244-2114-5/08/$25.00 © 2008 IEEE. 4501

thorized licensed use limited to: SVKM¿s NMIMS Mukesh Patel School of Technology Management & Engineering. Downloaded on July 29,2022 at 19:43:46 UTC from IEEE Xplore. Restrictions app
 At the same time, it is not good at neighbourhood search. k1 f m
Therefore, some hybrid algorithms have arisen. Beatrice et al. pc = pc max − (1)
f max
[2] combined the GA and Local search thereby got the trait of
local search and global search. To avoid plunging the local § ktf ·
pm = pm max ¨1 − 2 c ¸ (2)
optimization, José Ramón et al [3] integrated GA and TS ¨ Gf ¸
(taboo search). In a multiple objection condition, Sergio et al. © max ¹

[4] united the Earliest Due Date and GA, got the better p
Where the c max is the maximum value of crossover rate,
solution than Branch and Bound. X.-D. Zhang et al. [5] got the p
feasibly solution by heuristics scheduling rule at first, then the m max is the maximum value of mutation rate. The fitness
optimize it by GA. In this way, it is not only to increase the f max is the estimate value. The fitness f c belongs to the best
solving efficiency, but also satisfy the feasibility of solution. chromosome of crossover operation in a generation. The
Genetic arithmetic operators complete the random search
f
and local search in search space by select, crossover and fitness m belongs to the best chromosome of mutation
mutation, so some researchers proceed with the algorithm by operation in a generation. The G is the generation size. The t is
improving the arithmetic operators. Such as adding immune the present generation. The constant k1 , k 2 assure the validity
algorithm in to mutation [6], D. Alisantoso et al. [6] applied and change rate of pc and pm . We can value it in according
this method in the scheduling of a flexible PCB flow shop.
to fact problem.
The Reference [8] not only mends the coding but also order-
based crossover arithmetic operators, thus obtain better The parameter pm has the trend of decrease when the
offspring. generation increases. In order to balance the random search
and local search, we take the strategy which is restricting the
III. ADAPTATION OF GA crossover and mutation rate each other. The main strongpoint
A. The Adoption of Parameter of strategy is that when the crossover operator can not get
The capability of GA is related to the search behavior in preferable solution, the mutation rate will increase
search space. The random search process exploration and automatically to give prominence to the local search, and vice
could get away from local optimal solution, the local search versa. In theory, the global and local search restricting each
handle in exploitation and could climb to optimal solution. In other could balance the search behaviors.
generally, the information of exploitation is accumulated by IV. REALIZATION OF AGA FOR JSSP
select and the new area is explored by other arithmetic
operators in standard GA [9]. But the balance between local A. Representation
search and global search is associated with population size, To represent the chromosome, we take the operation-
generation size, mutation rate and crossover rate etc. Cheng et based representation that uses an un-partitioned permutation
al. [10] pointed that crossover and mutation could be designed with m-repetitions of job numbers. The operation-based
to balance the search behaviors. It is one for random search, representation would not produce infeasible solution. It is a
another for local search. pattern of repeating job number as many times as its member
From the last section, we can see that almost all the of operation. Each gene represents one operation and it is
literature adopt constant as parameter. In theory, GA is assigned to machines in the represented order.
dynamic, adaptive and generated from Darwinism. So, the For example, a problem involving three jobs and three
invariable parameter is contravened the theory. Changing the machines is listed in table. It includes the process order and
tactical parameter in the process of algorithm could accord to enduring time.
the theory of evolutionism largely and possess more We can represent a chromosome as “2 1 1 3 2 1 3 3”. The
intelligence. “2” represents the job two. The appearance time of “2”
B. Realization of Adaptability represents the operator number of job two. Then, we can
The adaptability of GA could be realized by modifying decode the chromosome according to decode rule, produce
the parameter by :( 1) introduce a rule, (2) feedback the search scheduling scheme [13].
state, (3) adaptation mechanism. The parameter amended by TABLE I
3 JOBS AND 3 MACHINES PROBLEM
fixity rule is deterministic adaptation. The feedback parameter machine(work time)
has the ability of adaptation. Adaptation mechanism has the jobs
operator 1 operator 2 operator 3
property of self-adaptation, and the parameter evolves by the
different generation [11]. j1 1(3) 2(2) 3(3)

In the iterative search procedure, the chromosome is j2 3(2) 1(3) 2(4)

mainly decided by crossover and mutation. The evaluating of
the chromosome depends on its fitness. So, the crossover rate j3 2(2) 3(2) 1(3)
and mutation rate should change with the fitness [15]. Decoding process:
p p Step 1: At current time, set the processed job number as 0. The
In this paper, the crossover rate c and mutation rate m
defined as: first operator noted as k (i ) = 1, (i = 1, 2,", n) .

4502

thorized licensed use limited to: SVKM¿s NMIMS Mukesh Patel School of Technology Management & Engineering. Downloaded on July 29,2022 at 19:43:46 UTC from IEEE Xplore. Restrictions app
 Step 2: Take out the gene xk of the chromosome in order, parents p1 and p2 according to crossover rate. Then, the
obtain the earliest start time of this operator k ( xk ) : c c
children 1 and 2 are generated by follows:
(1) Get the operator machine m of k ( xk ) according to Step 1: Choose randomly a job j from job set n.
table 1. Step 2: Get the position set K of job j.
(2) Get the next operator of xk , and set k (i ) = k (i ) + 1 Step 3: Setting the gene of position K in c1 as j.
(3) Process the xk in machine m as permitted earliest Step 4: Insert genes besides j from c2 to c1 in normal
time t: direction flow.
Ϥ1 : Because of the precedence constraint, the t is Step 5: Likewise, get the c2 .
decided by last operator finish time t x . If the current operator D. Mutation
is the first operator of job xk , set t x = 0 . Otherwise, Mutation operator has the equivalent important role as
crossover; especially when the crossover could not produce
t x evaluate the due time of last operator. good offspring, mutation always can improve the search
Ϥ
2 : The operation time of xk is pij . If there is not performance. We introduce neighbourhood search [14]. This
methods can extends search range, finally choose the best
processing queue in machine m , set t = tx , and this operator
solution from given candidate solution.
is the first operator in m . Otherwise, judge the t according to E. AGA for JSSP
following: The AGA approach introduces the thought of self-
­° max(a, t x ), max(a, t x ) + pij ≤ b (3) adaptation. The main step of AGA is similar to standard GA.
t =® The difference between AGA and GA is the parameter value.
°̄ max(t m , t x ), max(a, t x ) + pij > b (4)
In standard GA, the parameters are constant all the time. The
The idle time of machine m is [ a, b] . Formula (3) means this approach of adaptation is implemented by alterable crossover
operator could insert into the idle time of machine m. Formula rate pc and mutation rate pm .
(4) means this operator inserts into the last of the processing The change of the generation is considered to the
queue. (4) The processed job number adds 1, and register the adaptation. The parameters have the trend of decrease with the
start time and end time. generation increase. Besides this, we need take search balance
Step 3: repeat step 2 until finish the last gene. into consideration. In order to compensate the balance of
B. Fitness Functions and Select search behaviour; the algorithm must adjust the different
Fitness is the evaluation criterion of chromosome. The search behaviour: local search and global search. Generally
high fitness means the chromosome has strong adaptability speaking, the crossover operator is engaged in random search
and has large survival chance. Also, the fitness is the evidence of global, the mutation operator is engaged in local search. So
if final selection. In this paper, the objective function is the we can get the balance by amending crossover and mutation
minimum make span, the fitness function defined as: operators.
f = C − fi (5) When chose an operator strategy, it is difficult to change
C
Where is a constant which is bigger than the make- the method self. So, we handle it by change the crossover
(Function (1)) and mutation rate (Function (2)). . In addition,
span fi .
the pseudo structure and the detailed steps are described as
Superior animals are mostly used as seed animals to bring follows:
forth the young at domestic animal breeding farms. GA Procedure: AGA
introduces this method to operator. We adopt the Elitist Begin
strategy in which the best chromosome is automatically Initialization;
transferred to the next generation for reserving the excellent Evaluation;
offspring. While (not termination condition) do
C. Crossover Selection;
The crossover operator should maintain and evolve a Crossover;
good order relationship of chromosomes. It guides the search Mutation;
behaviour effectively. The operator processes the Evaluation;
chromosomes according to crossover rate. Calculate crossover rate and mutation rate;
From now on, researchers have presented some preferable End
crossover operator aiming at JSSP. Partially mapped End
crossover, Order crossover, Order-based crossover and Cycle Detailed steps: AGA
crossover etc have the application. Step 0: Set the parameters.
We adopt work piece position crossover which represent Input: population size N, the maximum crossover rate
good property in convergence speed and getting away from pc max , the maximum mutation rate pm max , the maximum
local optimal solution [13]. This method first produces two
estimate value f max , the constant k1 , k 2 and number of
generations G.

4503

thorized licensed use limited to: SVKM¿s NMIMS Mukesh Patel School of Technology Management & Engineering. Downloaded on July 29,2022 at 19:43:46 UTC from IEEE Xplore. Restrictions app
 Output: the optimal chromosome and the fitness value.
Step 1: Initialize the population. Execute the algorithm to
generate an initial population and calculate the fitness
value of chromosome.
Step 2: Perform a selection operation using the Elitist
approach.
Step 3: Execute a crossover operation. The probability of
crossover is determined by crossover rate function.
Step 4: Choose a better solution of two offspring to next
generation.
Step 5: Execute a mutation operation. The probability of
mutation is determined by crossover rate function.
Step 6: Calculate the crossover rate and mutation rate
according to function given in section III.
Step 7: Determine whether the stopping criterion has been
met.
If yes, then go to Step 8. Otherwise, return to Step 2 and
continue through Step 6.
Step 8: Display the optimal chromosome and the fitness value. Fig. 1 The computational results of best schedules over 10 runs
The converge speed of AGA enhanced obviously in Fig.
V. COMPUTATIONAL RESULTS
1. That is because the adaptation algorithm improved the
In this section, we take Muth-Thompson 10×10 and MT search capability. The restriction of crossover and mutation
20×5 problems to testing the performance of AGA and make each other balanced the global search and local search. So, the
comparison with the same example as those considered by superiority of AGA is: (1) speeds up the convergence; (2)
Tung-kuan Liu et al. [18], Wang et al. [13] and Zhang et al. restrain the premature.
[16].
VI. CONCLUSION
The main parameters were set as follows:
Population size: 100 individuals. This paper analyses the GA which adopts constant as its
Elitist select rate: 5%. parameter. The conclusion is that the shorting of adaptability
The maximum value of crossover rate: 0.9. is the general limitation of present GA. So we propose a
The maximum value of mutation: 0.1. modified GA SGA which introduces adaptability based on
k1 , k 2 : 0.2, 1.0. evolution. SGA associated evolution with generation and
The computational results comparison is showed as table fitness. That is when the generation is increasing the
2. The convergence curve comparison is draw as Fig. 1. probability of mutation is in decline; the crossover and
We obtained the optimal solution of chromosome by mutation compensate each other. When the crossover
running the AGA program 10 times. From the table 2, we can operating effect is feeble Δ the mutation operator will
see that AGA could converge to preferable solution. The strengthen automatically. When the algorithm plunges into
average value is 950. In figure (1), the curve AGA is the local solution, the crossover operator will intensify
convergence trend of GA in which the crossover rate and automatically so as to get away. The theory and algorithm of
mutation rate introduce the adaptation algorithm. The curve AGA is verified by benchmark problem, provided by Muth
GA is the convergence trend of GA in which the crossover and Thompson. The computational results proved that AGA
rate and mutation rate is constant. has the superiority of convergence speed and search behaviors.
TABLE II It has better and more robust result than other GA-based
COMPUTATIONAL COMPARISON algorithms.
Test problem
Search approach
MT10×10 MT20×5 REFERENCES
Conventional GA 965 1215
[1] Lawrence Davis. Job Shop Scheduling with Genetic Algorithms.
Giffler-Thompson GA 930 1184
Proceedings of the 1st International Conference on Genetic Algorithms,
Shifting-Bottleneck GA 938 1178 1985, Pages: 96 - 140.
Job-Pair based GA 937 1193 [2] Beatrice M. Ombuki and Mario Ventresca. Local Search Genetic
Machine-Constraint propagation GA 930 1165 Algorithms for the Job Shop Scheduling Problem. Applied Intelligence
Priority-Rule based GA 960 1294 no.21, pp: 99–109, 2004.
AGA 930 1178 [3] José Ramón Zubizarreta Aizpuru, Javier Arrieta. GA/TS: A Hybrid
Approach for Job Shop Scheduling in a Production System. Progress in
Artificial Intelligence, Volume 990/1995.
[4] Sergio Cavalieri and Paolo Gaiardelli. Hybrid genetic algorithms for a
multiple-objective scheduling problem. Journal of Intelligent
Manufacturing .no.9, pp: 361 – 367, 1998.

4504

thorized licensed use limited to: SVKM¿s NMIMS Mukesh Patel School of Technology Management & Engineering. Downloaded on July 29,2022 at 19:43:46 UTC from IEEE Xplore. Restrictions app
 [5] X.-D. Zhang, H.-S. Yan. Integrated optimization of production planning
and scheduling for a kind of job-shop [J]. Int J Adv Manuf Technol no.26,
pp: 876–886, 2005.
[6] XiuLi Wu, ShuDong Sun, GangGang Niu, YinNi Zhai. The Performance
Analysis of a Multi-Objective Immune Genetic Algorithm for Flexible
Job Shop Scheduling .Knowledge Enterprise: Intelligent Strategies in
Product Design, Manufacturing, and Management, Volume 207/2006.
[7] D. Alisantoso, L. P. Khoo, P. Y. Jiang. An immune algorithm approach to
the scheduling of a flexible PCB flow shop. Int J Adv Manuf Technol ,
no.22, pp: 819–827, 2003.
[8] Christian Bierwirth. A generalized permutation approach to job shop
scheduling with genetic algorithms [J]. OR Spektrum (1995) 17:87-92.
[9] Mitsuo Gen, Runwei Cheng.Genetic Algorithms and Engineering
Optimization. Beijing: Tsinghua University Press, 2004.
[10] Cheng, R. , and M, Gen, Evolution program for resource constrained
project scheduling problem, Proceedings of the IEEE International
Conference on Evolutionary Computation, IEEE Press, Piscataway,
NJ,1996.
[11][11] Schwefel H. Evolution and Optimum Seeking, Wiley, New York,
1995.
[12] [12] Christian Bierwirth A generalized permutation approach to job shop
scheduling with genetic algorithms. OR Spectrum, vol.17, no.2-3, pp: 87-
92, 1995.
[13]Baowen Wang, Chunyang Li, Wenyuan Liu and Dong Yang. Genetic
Algorithm for Job-shop Scheduling based on work piece position
crossover. Computer Engineering and Applications, 2007, 43(13):96-97.
[14] Cheng R. A Study on Genetic Algorithms-based Optimal Scheduling
Techniques Κ [ Ph D Dissertation ]. Tokyo: Tokyo Institute of
Technology,1997
[15]Srinivas M, Patnaik L M. Adaptive probabilities of crossover and
mutationing genetic algorithm.IEEE Trans System Man and Cybernetics,
vol.24, no.4, pp656-667, 1994.
[16]Yongchao Zahng, Yunqing Rao, Xianjun Liu and Peigen Li. An Improved
Genetic Algorithm for the Job Shop Scheduling Problem. Chinese
Mechanical Engineering, vol.23, no.12, pp: 299-2153, 2004.
[17]French S. Sequencing and scheduling - an introduction to the mathematics
of the job-shop. John Wiley, 1982: New York
[18] Tung-Kuan Liu, Jinn-Tsong Tsai and Jyh-Horng Chou. Improved genetic
algorithm for the job-shop scheduling problem. Int J Adv Manuf Technol,
no.27, pp: 1021–1029, 2006.

4505

thorized licensed use limited to: SVKM¿s NMIMS Mukesh Patel School of Technology Management & Engineering. Downloaded on July 29,2022 at 19:43:46 UTC from IEEE Xplore. Restrictions app