Expert Systems with Applications 38 (2011) 3161–3171

Contents lists available at ScienceDirect

Expert Systems with Applications

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

FMS scheduling with knowledge based genetic algorithm approach

A. Prakash a, Felix T.S. Chan b,⇑, S.G. Deshmukh a
a
Indian Institute of Technology Delhi, Hauz Khas, New Delhi 110016, India
b
Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hung Hom, Hong Kong

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

Keywords: In this paper a complex scheduling problem in flexible manufacturing system (FMS) has been addressed
Scheduling with a novel approach called knowledge based genetic algorithm (KBGA). The literature review indicates
Flexible manufacturing system that meta-heuristics may be used for combinatorial decision-making problem in FMS and simple genetic
Genetic algorithm algorithm (SGA) is one of the meta-heuristics that has attracted many researchers. This novel approach
Knowledge management
combines KB (which uses the power of tacit and implicit expert knowledge) and inherent quality of SGA
KBGA
for searching the optima simultaneously. In this novel approach, the knowledge has been used on four
different stages of SGA: initialization, selection, crossover, and mutation. Two objective functions known
as throughput and mean flow time, have been taken to measure the performance of the FMS. The useful-
ness of the algorithm has been measured on the basis of number of generations used for achieving better
results than SGA. To show the efficacy of the proposed algorithm, a numerical example of scheduling data
set has been tested. The KBGA was also tested on 10 different moderate size of data set to show its robust-
ness for large sized problems involving flexibility (that offers multiple options) in FMS.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction ity has been addressed by many researchers, e.g. Upton (1994),
Wadhwa and Browne (1989), Wadhwa, Rao, and Chan (2005),
Owing to the globalization of the market, increasing demands of etc. Wadhwa and Rao (2000) have defined the flexibility as the
the customized products and rapidly changing needs of customers, ability to deal with change by judiciously providing and exploiting
the manufacturers are finding it difficult to survive under the controllable options dynamically. Due to this flexibility, some deci-
forces of increased competition and increased customers’ expecta- sion-making problems have occurred in the system. In order to run
tions. Therefore, to sustain in the global market, their focus is to the system efficiently, the decision points and their importance
develop a manufacturing system that can fulfil all the demanded should be defined and assessed very carefully (Wadhwa & Bhag-
requirements within due dates at a reasonable cost. Among all wat, 1998; Wadhwa & Browne, 1989; Caprihan & Wadhwa, 1997,
the existing manufacturing system, they require a manufacturing etc.). According to Stecke (1983) and Gen, Lin, and Zhang (2009),
system having the flexibility to make the customized product with there are four stages of decision problems for the FMS: designing,
medium volume. Therefore, they are motivated to consider flexible planning, scheduling, and control. The center of attention of
manufacturing system (FMS), which is a compromise between job authors is on the scheduling problem after considering that design-
shop manufacturing system and batch manufacturing system. ing and planning phase has been over. In the field of manufacturing
Flexible manufacturing system is a system, which is equipped with control problems of an FMS, scheduling is an extensive area for re-
the several computer-controlled machines, having the facility of search which is still alluring many a researchers.
automatic changing of tools and parts. The machines are intercon- Scheduling of operations is one of the most critical issues in the
nected by automatic guided vehicles (AGVs), pallets and several planning and managing of manufacturing processes. Scheduling
storage buffers. These components are connected and governed problem is an assignment problem, which can be defined as the
by computer using the local area network (LAN). The exquisiteness assigning of available resources (machines) to the activities (oper-
of this system is that it gleaned the ideas both from the flow shop ations) in such a manner that maximizes the profitability, flexibil-
and batch shop manufacturing system. The prominent literature ity, productivity, and performance of a production system. This
has several descriptions of FMS and its inherent feature of flexibil- problem will increase with the augmentation of rapid changing
of the product mix ratio, which is defined as the proportion of each
⇑ Corresponding author. job in each production lot. To find the best schedule can be very
E-mail addresses: [email protected] (A. Prakash), [email protected] easy or very difficult, depending on the shop environment, the pro-
(F.T.S. Chan), [email protected] (S.G. Deshmukh). cess constraints, and the performance indicator (Pinedo, 2002). The

0957-4174/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2010.09.002
3162 A. Prakash et al. / Expert Systems with Applications 38 (2011) 3161–3171

typical performance indicators for FMS are the mean flow time and 2. Problem description
throughput, make-span, tardiness etc. Due to flexibility, the
scheduling problem in such type of manufacturing system is a very In a given planning perspective, the fixed number of machines,
complex problem to solve which is known as NP-hard problem pre-known number of operations at each part type with processing
(Garey, Johnson, & Sethi, 1976; Ghedjati, 1993, Chaturvedi, 1993; time, number of part types for scrutinizing the FMS scheduling
Kumar, Tiwari, & Shankar 2003; Lee & DiCesare 1994). problem. This problem works as selecting the part type and allo-
To solve scheduling and assignment problems of FMS, several cates the operation of the selected part type to the resources (ma-
past studies have adopted mathematical programming techniques. chines) for achieving the desired performance measures of the
Liang and Dutta (1990) developed a mixed-integer programming system. Due to the inherent flexibility of FMS, a machine can per-
formulation to solve this problem along with the problems of pro- form more then one operation which produces moderately large
cess planning and machine loading. While Hutchinson, Leong, Syn- alternatives of part type routes for the allocation of operations on
der, and Ward (1991) provided an optimal solution procedure to the different machines. The detailed problem has been described
investigate the effect of routing flexibility on job shop FMS. In some as following.
real world problem instances, it has been seen that the traditional In the present paper, we work on the identical problem model
techniques can not reach up to the optimal solutions. Thus, it be- which has been taken by Su and Shiue (2003). In this problem,
came evident that to solve such real world FMS problems, the an FMS consist of three machine families (F1, F2, and F3), three
researchers are concerning about the use of artificial intelligence load/unload stations (L1, L2, and L3), three automatic guided vehi-
optimization methods known as meta-heuristics; these include ge- cles (A1, A2, A3), an input buffer, a centralized buffer, which is used
netic algorithms (GAs) (Etiler, Toklu, Atak, & Wilson, 2004; Chang, for avoidance of deadlock. A local area network is used for inter-
Chen, & Lin, 2005; Moon, Lee, & Bae, 2008; Pezzellaa, Morgantia, & connecting all the equipments. Eleven different parts type are
Ciaschettib, 2008; Ruiz, Maroto, & Alcaraz, 2006; Shiue, Guh, & being processed in the FMS. The processing time for each part at
Tseng, 2009), tabu search (Grabowski & Wodecki, 2004), ant colony each load/unload station and at each machine family is identical
optimization (Rajendran & Ziegler, 2004), particle swarm optimiza- as taken by Montazeri and Van Wassenhove (1990). The continu-
tion (Guo, Li, Mileham, & Owen, 2009). ously changing five different part mix ratios are also taken into
Genetic algorithms have proven to be effective for FMS schedul- the account for seeking the different conditions of bottlenecks,
ing problems. Many of the studies that have used GAs for scheduling which is given in the Table 1 and these are changing per 20000
problems have been summarized by Gen and Cheng (1997). Park, time units for showing the effects of rapidly changing market
Choi, and Kim (1998) applied GAs to a job shop system with alterna- environment.
tive process plans. Zhou, Feng, and Han (2001) developed a hybrid Various assumptions, which have been taken in this research,
GA for job shop scheduling problems. They applied priority rules are summarized as below:
such as the shortest processing time for the genetic search to devise
a hybrid GA. Cheng and Chang (2007) have proposed a combined ap- (1) There is no delay in the availability of raw material.
proach of Taguchi experimental design with GA to solve the schedul- (2) Each job order has only one part type, with individual due
ing problem in flexible flow shop. Whereas, Shiau, Cheng, and Huang date for each part.
(2008) has offered a hybrid GA for the same. Thus, it can be clearly (3) At a time only one job order will be executed on each
shown that GA has been extensively used for scheduling problem. machine.
From the prominent literature review related to GA applications (4) All the processing time is deterministic and pre-selected.
in FMS scheduling, it is comprehensible that GAs have a great po- (5) Each part type is processed by a predetermined sequence.
tential to solve the complex combinatorial problems. To improve (6) After finishing the each machining operation it must go back
the performance of GA, one of the major research thrusts is to to the load/unload station, otherwise go to central buffer.
use the knowledge of experts within the GA framework. The cur- (7) The movement of materials in which AGVs are not utilized,
rent work pursues research in which GA procedure is combined are taken as negligible.
with experts’ knowledge. In the present paper, a knowledge based
genetic algorithm (KBGA) has been developed to solve the schedul- The above described problem has two performance measures
ing problems for an FMS. The KBGA is a stochastic search technique known as throughput and mean flow time. The throughput (TP)
with the inherent ability of GA and strength of knowledge to en- is measured by the number of jobs, which have been completed
hance the performance of system and algorithm concurrently. To during the scheduling period. Whereas mean flow time (MF) is
demonstrate the effectiveness of the proposed algorithm, it has the measure of mean of the flow times of jobs in one scheduling
been tested on a benchmark problem and compares the results period. Flow time is measured by the difference of the completion
with conventional rules and SGA. The KBGA is also tested on 10 time and arrival time of the job.
moderate size problems to show the efficacy and computational
superiority over the existing classical genetic algorithm on the ba-
Table 1
sis of performance criteria. From the results, it proves that the pro-
Part mix ratio.
posed approach provides the better results within lesser number of
generations. Part type Part mix ratios (%)
The remainder of this paper is described as following: Section 2 Mix 1 Mix 2 Mix 3 Mix 4 Mix 5
delineates the complexity of the problems whereas the mathemat- 1 11.00 14.00 6.00 9.00 14.00
ical model has been described in Section 3. Section 4 describes the 2 11.00 14.00 6.00 9.00 14.00
background of simple genetic algorithm. The role of knowledge 3 11.00 15.00 6.00 9.00 14.00
management has been revealed in Section 5. The proposed algo- 4 12.00 10.00 15.00 8.00 15.00
5 6.00 12.00 15.00 13.00 7.00
rithm knowledge based genetic algorithm (KBGA) has been dis- 6 8.00 8.00 9.00 12.00 5.00
cussed in Section 6 and the detailed procedure of KBGA has been 7 8.00 5.00 8.00 3.00 5.00
portrayed in Section 7. Section 8 illustrates the computational 8 7.00 3.00 8.00 9.00 4.00
experience of the given problem by using the proposed algorithm. 9 7.00 3.00 7.00 8.00 4.00
10 2.50 1.00 4.00 1.00 6.00
In Section 9, the paper has been concluded with some issues and
11 16.50 15.00 16.00 19.00 12.00
future scope of the research.
 A. Prakash et al. / Expert Systems with Applications 38 (2011) 3161–3171 3163

To formulate such scheduling problem in FMS environment, a Inequity Eq. (3) states that the successive operation has to be
well defined mathematical model is required. The mathematical started after the completion of its precedent operation of the same
model of the above given problem has been portrayed in the next job, which represents the operation precedence constraints. Eq. (4)
section. states that one machine must be selected for each operation. Eq. (5)
shows the set of decision variable, which shows that at any time the
3. Mathematical model a single operations (k) of any part (i) will be assigned to any ma-
chine (j) and for others its value will be zero. Whereas Eq. (6) de-
The problem is to assign each operation to an available machine scribes the non zero property of completion time.
and sequence the operations assigned on each machine and the
performance measure which are taken into consideration are 4. Background of SGA
throughput and mean flow time. The notations, objective function,
constraints used to formulate the mathematical model are summa- Simple GA is an ‘intelligent’ probabilistic search algorithm that
rized as below: simulates the process of evolution by taking a population of solu-
tions and applying genetic operators in each reproduction. Each
3.1. Notations solution in the population is evaluated according to some fitness
measure. Fitter solutions in the population are used for reproduc-
I number of part types i = 1, 2, 3, . . . , n tion. New ‘off spring’ solutions are generated and unfit solutions
j number of machines I = 1, 2, 3, . . . , m in the population are replaced. The cycle of evaluation–selection–
k number of operations k = 1, 2, 3, . . . , K reproduction is continued until a satisfactory solution is found
(Goldberg, Korb, & Deb, 1989; Michaelwicz, 1992). Holland
(1975) first described a GA, which is commonly called the classical
3.2. Parameters genetic algorithm (CGA). The working of the CGA can best be
understood by the following steps, which are shown in Fig. 1.
n total number of part types
m total number of machines Step 1. Generate the initial population. Determine the size of the
Ki total number of operations at job i population and the maximum number of the generation.
oik the kth operation at job i Step 2. Calculate the fitness value of each member of the initial
pikj the processing time of kth operation of job i at jth population.
machine
AMik set of available machines for i
cik completion time of kth operation of ith part types
ARi arrival time of ith part types
Ci completion time of ith part types Start
Fi flow time of ith part types (Ci  ARi)
|SFi| number of finished part of ith part types
TP Throughput Randomly generated
MF Mean Flowtime initial population of
chromosomes
3.3. Decision variables
 Evaluation of each
ai 1 if part type i is selected chromosome
0 otherwise

1 if kth operation of part typeiis assigned on jth machine
bikj
0 otherwise Selection
In the present study, the objective functions can be defined as
follows:
The first objective function is to maximize the throughput Crossover
(number of completed parts), which can be formulated as shown Genetic Operators
in Eq. (1).
X
n
Mutation
max TP ¼ jSFi j: ð1Þ
i¼1

The second objective is to minimize the mean flow time, which can
be stated as in Eq. (2): No Is satisfying the
P termination criteria?
i 2 SFF i
min MF ¼ : ð2Þ
jSFj
Yes
These are subjected to some system constraints, which are as follows:
cik  pikj  bikj  ciðk1Þ  0; k ¼ 2; 3; . . . ; K i ; 8i; j; ð3Þ Near Optimal Results
X m
aikj ¼ 1; 8k; i; ð4Þ
j¼1

aikj 2 ð1; 0Þ; 8i; j; k; ð5Þ Stop

cik  0; 8k; i; ð6Þ

Fig. 1. Flowchart of simple genetic algorithm (SGA).
3164 A. Prakash et al. / Expert Systems with Applications 38 (2011) 3161–3171

Step 3. Calculate the selection probability of each member of the often working in parallel to complete assignments spanning tradi-
initial population using the ratio of fitness value of that initial tional boundaries and functional areas. Knowledge management
population to the summation of the fitness values of the indi- (KM) provides processes to capture a part of tactic knowledge
vidual solutions. through informal methods and pointers and fairly high percentage
Step 4. Select a pair of members (parents) that can be used for of explicit knowledge, reducing the loss of organizational knowl-
reproduction using selection probability. edge (Nonaka & Takeuchi, 1995).
Step 5. Apply the genetic operators such as crossover, mutation, ‘‘KM is the formalization of and access to experience, knowledge
and inversion to the parents. Replace the parents with the new and expertise that create new capabilities, enable superior perfor-
o€ spring to form a new population. Check the size of the new mance, encourage innovation and enhance customer value” (Beck-
population. If it is equal to the initial population size, then go man, 1997). According to Tiwana (2000), knowledge management
to step 6, otherwise go to step 4. is the ability to create and retain greater value from core business
Step 6. If the current generation is equal to the maximum num- competencies. Beckman (1997) realizes that knowledge manage-
ber of the generation then stop, else move to step 2. ment is the systematic, explicit, and deliberate building, renewal,
and application of knowledge to maximize an enterprise’s knowl-
After searching a large amount of the literature in the area of GA edge-related effectiveness and returns from its knowledge assets
application, it has been found that there are a plethora of articles knowledge management is the formalization of and access to expe-
addressing the scheduling problems of FMSs. This research intends rience, knowledge, and expertise that create new capabilities, enable
to demonstrate the advantage of knowledge management in GA superior performance, encourage innovation, and enhance customer
applications in the area of the scheduling problem of a random value. Whereas, Tiwana and Balasubramaniam (2001) feel that
FMS that is known for its computational complexity (even for mod- knowledge management addresses business problems particular to
erate size FMS). business-whether it’s creating and delivering innovative products
or services or managing and enhancing relationship with existing
5. Background of KM and new customers, partners, and suppliers, or administrating and
improving work practices and processes. Nietok (2003) examines
As Francis Bacon said, ‘‘Knowledge is power”. To learn new that knowledge has a connotation of ‘potential for action’ and is dif-
things, maintain valuable heritage, create core competences, and ferent from information in terms of its more immediate link with
initiate new situations, the power of knowledge is a very important performance. It is linked to the values and experience of the user,
resource for both individual and organizations now and in the fu- and therefore takes many forms. One may have knowledge of certain
ture. According to Nonaka (1994), Knowledge has been defined as facts. A KM strategy can help tear down traditional cross functional
‘‘justified true belief’’ that increases an organization’s capacity for boundaries. KM entails helping people share and put knowledge into
effective action. It has two dimensions: explicit and tacit knowl- action by creating access, context, infrastructure, and simulta-
edge. Davenport and Prusak (1998) define knowledge as a fluid neously reducing learning cycles (Davenport, Delong, & Beers
mix of framed experience, values, contextual information, and ex- 1998; Davenport & Prusak 1998; O’Dell & Grayson, 1997 ).
pert insight that provides a framework for evaluating and incorpo- In the present paper, the knowledge based tool is motivated by
rating new experiences and information. They suggest that it the ideas proposed by Wadhwa and Saxena (2006). The creation of
originates and is applied only in the mind of knower and holders today’s knowledge base requires blending of knowledge from di-
of tacit knowledge in organizations. It is embodied in documents, verse disciplinary and personal skills based on perspectives where
repositories, organizational routines, processes, practices and creative cooperation is critical for innovation (Handfield & Nichols,
norms. To respond to competitive challenges, otherwise-indepen- 1998). An integrated framework of KM has been shown in the
dent firms have become more closely coupled than in the past, Fig. 2. It shows the conversion of information to knowledge and

Information providers

Verification & Adapt

Filtration
Attract
Acquire
Knowledge Base
Engage the
Systematize people

Create Learn
Knowledge

Conversion to Knowledge Knowledge Utilization

Knowledge Users

Fig. 2. An integrated framework of knowledge management.

 A. Prakash et al. / Expert Systems with Applications 38 (2011) 3161–3171 3165

integration of knowledge base with knowledge utilization. To con- of the environment. By employing the knowledge of the environ-
vert the information to knowledge, the process follows the various ment like FMS and the complexities, i.e. flexibilities, we can get
activities as verification, acquiring the filtered information, classi- the better result within lesser time than SGA. As it works with
fication and creation of the knowledge from this information. All the knowledge base, it is identified as KBGA. The proposed algo-
the acquired knowledge is stored in the knowledge base. After rithm works not only for improving the performance measures of
accumulation, the knowledge has been distributed to the knowl- the system like traditional genetic algorithm but the performance
edge users by following the steps like adaptation, attraction, of the algorithm (Fig. 3). To enhance this idea, the knowledge based
engaging the people and teaches them how to use this knowledge. initialization, knowledge based crossover, knowledge based muta-
The knowledge synergy based thinking showed in Fig. 1 can signif- tion, and knowledge based selection have also been incorporating
icantly benefit the KM guided manufacturing endeavors. in the algorithm. The procedure of the algorithm has been de-
scribed in the next section.
6. Proposed knowledge based genetic algorithm (KBGA)
7. Procedure of KBGA for scheduling problem
Although GA is a global search technique, its practical useful-
ness depends on the initialization of the problem, crossover and As stated in the previous section, it is clear that the strong point
mutation techniques and selection scheme for the next generation. of KBGA over SGA is the knowledge based generation of the initial
Therefore, a number of techniques have been developed for han- population instead of random generation. It is followed by the
dling all the above constraints. To improve the performance of knowledge based selection (KBS), knowledge based crossover
GA, there is a lot of research especially in the FMS scheduling prob- (KBC), and knowledge based mutation (KBM) to provide the wider
lems. Very few researches have addressed the constraints of algo- search space within lesser time. The full procedure of KBGA has
rithm itself instead of constraints of the problem or the been shown in Figs. 4–7. All the steps of the proposed algorithm
environment. (KBGA) are as follows:
In the present paper, we have introduced a concept of improv-
ing the performance of GA by exercising the knowledge based sys- 7.1. Knowledge based initialization (KBI)
tem, which will develop a faster algorithm for better performance
of the system. It will employ on the basis of both tacit and explicit In the first step of the algorithm, an initial population set of the
knowledge. For a search stratagem, it is very essential that it solutions has been generated on the basis of the knowledge based
should also handle the inherent characteristics and complexities system. In this step, firstly the information related to the system

Information Providers
Start from Shop Floor

Generate Initial Population

based on the knowledge

Evaluate each
chromosome

Knowledge based
Selection
Knowledge Base
Knowledge based System
Crossover (Selection scheme,
Crossover Scheme,
Mutation Scheme)
Knowledge based
Mutation

Is satisfying the
No Termination
Criteria

Yes
Near Optimal Result

Stop

Fig. 3. Flowchart of KBGA.

3166 A. Prakash et al. / Expert Systems with Applications 38 (2011) 3161–3171

Parts Operation Tool

Machine System
Information Information Information
Information Information

Input (Information
from all the sources)

Is information No
relevant for Close the case
scheduling?

Yes

Plan the work according

to the performance
measure and
requirements

Make the schedule on Traditional

the basis of traditional Scheduling Rules
rules according to the like SPT, EDD,
performance measures LPT, CR etc.

Review the outcomes

Yes Is there any other

information
required?

No

Knowledge Base Make the initial

(Scheduling Rules) population of KBGA

Fig. 4. Initialization process of KBGA.

environment like (manufacturing system, types of flexibilities etc.), The knowledge base provides the initial population for the pro-
machines (type of machine, machine performance, setup time etc.), posed algorithm. Thus, here all the constraints related to the sys-
part (part types, no. of parts to be scheduled etc.), operation (num- tem or problem has been taken into the consideration. Hence, it
ber of operations, type of operation, operation time etc.) has been can be said that the seed of initial population will work better than
collected and filtered. randomly generated population. After getting the initial popula-
After that the performance measures, on which the system con- tion, the evaluation process is started and it is described in the next
centrates, and the requirements have been decided. After this deci- section.
sion making, the appropriate traditional scheduling rules has been
selected on the basis of performance measures. For instance, if the 7.2. Evaluation
minimum make-span has been taken into the consideration as per-
formance measure, the SPT rule works better as the knowledge In this step, each sequence has been calculated according to the
captured from the previous information. So the schedule according evaluation criteria, which is a problem specific function. In the real
to SPT will be selected first then other better schedules are also se- world state of affairs, several objectives work at the same time.
lected and make a knowledge base (Fig. 4). Thus, the proposed algorithm provides the facility to specify
 A. Prakash et al. / Expert Systems with Applications 38 (2011) 3161–3171 3167

Selection
Schemes e.g.
4 2 5 8 1 m 1 roulette wheel, 4 2 5 8 1 m 1
2 tournament etc. 2
1 2 6 9 3 m 4 2 5 8 1 m
3 5 2 1 7 m 3 3 6 9 8 7 m 3

8 4 6 1 3 m 4 Knowledge 5 9 8 4 2 m 4
Based
Selection
(KBS)

Selection
Criteria
7 9 5 3 1 m n 8 3 5 7 1 m n

Initial Population Selected Population

Fig. 5. Knowledge based selection.

Crossover
Schemes e.g.
1 5 2 3 6 8 7 4 8 5 7 4 6 3 1 2
PMX, CLX etc.

7 4 5 8 6 3 1 2 2 3 5 1 6 8 7 4
Knowledge
Based
Crossover
(KBC)

1 3 7 5 2 4 6 8 1 3 7 5 8 2 4 6

3 8 5 7 1 2 4 6 Crossover 3 2 5 7 1 4 6 8
Probability
ranges
Parents Chromosomes Offspring Chromosomes
Fig. 6. Knowledge based crossover.

Mutation
Schemes e.g.
INS, INV etc.

Knowledge
Based
3 8 5 7 1 2 4 6 Mutation 3 2 4 7 1 5 6 8
(KBM)

Chromosomes Mutation Mutated Chromosomes

Probability
ranges

Fig. 7. Knowledge based mutation.

3168 A. Prakash et al. / Expert Systems with Applications 38 (2011) 3161–3171

several objectives. The user can specify the relative weighted aver- knowledge about the outcomes with the different range of the
age for each objective. In the proposed algorithm (KBGA), the mutation probability. According to Michaelwicz (1992), the
knowledge based system is highly efficient for sustaining the solu- selection of the appropriate value of the mutation probability
tion feasibility. is an art not a science. Hence, it is cleared now that knowledge
(explicit or implicit) can help to determine the value of genetic
parameter.
7.3. Knowledge based selection (KBS)

After the evaluation of all the sequences, a subset of the initial 7.6. Termination criteria
population is selected on random basis. It works on the basis of
Darwinism ‘‘survival of the fittest” but not as a greedy algorithm. After mutation, the selected populations, equal to the size of
In KBGA, the selection is also effected by the knowledge base sys- initial population, have to be entered to the next generation out
tem to improve the performance of the algorithm. of the extended population of the chromosomes. The whole pro-
In this algorithm, the selection is based on Neo Darwinism, cess will be repeated until satisfy the termination criteria. The
which can sub-divide the procedure of selection of three catego- termination criteria can be characterized by the number of gen-
ries: (a) directional selection, (b) steady selection and (c) unruly erations or the predefined level of the output.
selection. The directional selection is based on the mean value
(increasing or decreasing both) of the population whereas steady
selection, which is based on normalizing, eliminates the chromo- 8. Computational analysis
somes with excessive values. It is called steady selection due to
provide the steady state search space. The chromosomes are elim- In the past literature, a lot of decision-making problems in
inated according to the moderate values in unruly selection. To FMS such as scheduling, planning etc. have been addressed. To
execute all three types of selection, there are several methods of solve such type of NP-hard problems, GA has been used as a
the selection e.g. tournament selection, roulette wheel selection, powerful tool but the main endeavour is to minimize the
logarithmic scales selection etc. number of generations to obtain the near optimal or sub-optimal
In the knowledge base system, all types of selection schemes results. Thus, it is highly essential to develop such type of algo-
with their characteristics and their performance in different sys- rithm which can decipher the large sized combinatorial problem
tems have been placed. According to this knowledge, the suitable within very few numbers of generations by utilizing the less CPU
selection scheme has been applied for the selection of a subset of time. In the SGA, there is no interaction or controlling by the hu-
the initial population for the next stage of the algorithm. man being during the processing of algorithm or in other words,
the human knowledge can not be used to improve the
7.4. Knowledge based crossover (KBC) performance of the system as well as algorithm. To consider
all of the above, a new algorithm has been proposed known as
Following the KBS, the surviving chromosomes are selected to knowledge based genetic algorithm (KBGA).
form the new off springs to explore the wider search space. The The proposed KBGA used the knowledge based initialization,
knowledge based crossover gives the inherent characteristics to selection, crossover, mutation instead of randomization. This pro-
the off springs from parents. Initially, a subset of survived chromo- vides a better initial solution seed rather than any random solu-
somes has been randomly selected according to the crossover tion. Thus it can converge at faster rate than simple GA. Some
probability. To perform the crossover, there are several crossover other knowledge based operators are also used to minimize the
schemes e.g. single point crossover, partial mapping crossover, computational hurdles to achieve the sub-optimal results within
uniform crossover, cycle crossover etc. with some specific few generations.
characteristics. To evaluate the quality of solutions, the proposed algorithm has
The characteristics of each crossover scheme and their perfor- been employed on two performance measures: throughput (TP)
mance for different types of system environments and problems and mean flow time (MF). Maximization of throughput and mini-
have been kept in the knowledge base and it will be updated as mization of mean flow time is the maximum use of the resources
increasing the knowledge. The tuning of the crossover probability within minimum time. The proposed KBGA maximizes the
is also a concern in the proposed algorithm. The performance of throughput (5265.83units) and minimizes the mean flow time
the system for the different environments at the various crossover with 1025.6 units. The comparative study with the other tradi-
probabilities has been captured and placed in the knowledge base. tional scheduling strategies has been shown in the Table 2 with
Another facility to check the unfeasible new offspring has been mean and standard deviation of both measures of performance in
provided to the user. If any parent reproduce an unfeasible solution a random type FMS. From the table, it is clearly shown that KBGA
that will be checked and discarded with the help of knowledge yields the better result then simple GA with less value of standard
base system. deviation.
To show the efficacy of KBGA, a comparative study of the
convergence of simple GA and KBGA has been taken into consid-
7.5. Knowledge based mutation eration. The comparison of both the algorithm for the first mea-
sure of performance (throughput) has been shown in Fig. 8,
Following the above step, the next genetic operator, named as while the graph for mean flow time has been shown in Fig. 9.
mutation, empowers the algorithm to explore the search space. From Fig. 8, it can be concluded that the quality of initial solu-
It modifies single chromosome by altering the genes or bits in- tions of KBGA is better than GA as shown the large differentia-
stead of recombining the two chromosomes. In the proposed tion between the values of throughput (5251.22 vs. 5117.42)
algorithm, a knowledge base has been created to store the after the 10 generations. After 20 generations, the values of
knowledge about the performance of various mutation operators throughput are 5259.14 (KBGA) and 5131.65 (SGA). The KBGA
e.g. inversion, insertion, displacement etc in the different system converged after 23 generations with the value of 5265.83
environments with a variety of objective function. It also has the whereas SGA converges after 112 generations with the value of
 A. Prakash et al. / Expert Systems with Applications 38 (2011) 3161–3171 3169

Table 2
Comparative results of the mean and standard deviation between KBGA and other scheduling strategy.

Scheduling strategy TP MF
Mean SD Mean SD
KBGA 5265.83 18.86 1025.61 117.42
SGA 5248.73 19.8771 1066.48 206.40
Decision tree (DT) 5246.76 19.81 1151.1685 270.78
First in first out (FIFO) 5226.90 56.96 1769.90 541.35
Shortest processing time (SPT) 5243.53 22.70 1086.01 221.68
Shortest imminent operation time (SIO) 4236.96 37.87 1237.88 286.81
Shortest remaining processing time (SRPT) 5245.50 22.50 1165.33 358.70
Critical ratio (CR) 5200.73 48.94 1669.06 304.98
Dynamic slack (DS) 5219.96 72.81 1163.14 346.71
Earliest due date (EDD) 5231.40 56.29 1536.28 585.81

The scheduling rules are given in Appendix A.

Comparison of Conversion Rate of KBGA and GA Convergence of KBGA with different Crossover
5300 Probabilities (Throughput)
5250 5270
5265
Throughput

5200
5260 Pc=0.5

Throughput
GA
5150 5255
KBGA Pc=0.55
5250
5100 Pc=0.6
5245
5050 5240 Pc=0.65

5000 5235
5230
0

0
40

60

80
10

23

5 10 15 20 23 25 28 30 35 37 40
10

11

13

No. of Generations No. of Generation

Fig. 8. Comparison of convergence rate of KBGA and SGA (performance Fig. 10. Convergence rate of KBGA with different crossover probabilities
criteria = throughput). (performance criteria = throughput).

Comparison of Conversion Rate of KBGA and GA Convergence of KBGA with different Crossover
1800 Probabilities (Mean Flow Time)
1600 1100
1400
Mean Flow Time

1080
Mean Flow Time

1200 Pc=0.5
1000 KBGA 1060
800 Pc=0.55
GA 1040
600 Pc=0.6
400 1020
Pc=0.65
200 1000
0
980
5 10 15 20 23 25 28 30 35 37 40
0

0
40

60

80
10

23

10

11

13

No. of Generations No. of Generation

Fig. 9. Comparison of convergence rate of KBGA and SGA (performance Fig. 11. Convergence rate of KBGA with different crossover probabilities
criteria = mean flow time). (performance criteria = mean flow time).

5248.73. With the other performance measure (mean flow time), According to De Jong (1975), some parameters in GA e.g. popu-
the KBGA achieved the value of 1063.01 in 10 generations in- lation size, crossover probability, mutation probability etc. have
stead of 1592.63, which is obtained by SGA. The KBGA converged the significant effect on the performance of the algorithm. The val-
in 23 generations with improved result of 1025.61 whereas SGA ues of these parameters vary according to the environment, oper-
converged in 112 generations with the value of 1066.48. Thus it ating conditions and problem constraints. Thus these parameters
can be concluded that the initial solution of the proposed KBGA cannot be set as a fixed value for each problem. Among all these
considerably reduces the computational time while its other parameters, the crossover probability has more significant effect
knowledge based operators diminish the other computational as it ensures the recombination of the building blocks of the algo-
hurdles to attain the near-optimal result in less generations. rithm. To keep it in mind, the proposed algorithm has also been
These comparative studies of KBGA and SGA show the effective- tested on the different values of the crossover probabilities for both
ness and efficacy of the proposed algorithm and also show the of the performance measures (Figs. 10 and 11). From the Fig. 10, it
power of knowledge. can be taken as a conclusion that at the value of crossover
3170 A. Prakash et al. / Expert Systems with Applications 38 (2011) 3161–3171

Table 3
Comparison of KBGA and SGA for 10 random problems.

S. Part Machines SGA KBGA

No. types
TP MF Number of TP MF Number of
generations generations
Mean Standard Mean Standard Mean Standard Mean Standard
deviation deviation deviation deviation
1 15 3 5283.46 18.24 1117.48 211.40 126 5308.27 19.36 1073.61 113.42 28
2 18 5 5367.13 18.86 1183.15 211.57 137 5409.34 19.11 1151.1685 115.64 30
3 20 5 5398.58 18.34 1241.88 211.32 150 5513.77 19.03 1182.32 115.12 31
4 25 7 5532.22 18.67 1417.45 211.63 154 5664.63 19.26 1252.88 116.50 35
5 28 10 5616.75 19.12 1536.28 226.15 168 5743.88 19.22 1285.27 118.38 36
6 32 12 5720.91 19.07 1669.06 226.26 179 5895.21 20.34 1396.25 121.85 38
7 35 15 5766.43 25.66 1702.37 229.04 181 5971.93 20.47 1423.42 118.33 42
8 40 17 5991.36 26.02 1815.69 231.83 208 6315.82 22.92 1578.16 115.07 55
9 50 20 6383.72 31.87 2237.53 238.77 227 6522.45 22.05 1769.90 118.68 58
10 60 20 6578.76 32.09 2873.82 247.21 253 6739.55 24.20 2125.08 123.36 64

probability as 0.6 (Pc = 0.6), it explores the search space in better Appendix A
way. On increasing the value, it again lessens the performance of
KBGA. Dispatching Description Mathematical
The identical results are obtained from Fig. 11. From the above rule formulation
results, it can be inferred that as the value of Pc increases, the per-
FIFO Select the job minieSJ{ARi}
formance improves up to a certain limit (Pc = 0.6) and then impairs
according to the
the quality of the solution as well as performance of the algorithm.
first in first out rule
From the obtained results, it is clear that the proposed heuristic is nP o
SPT Select the job with mini2SJ
capable to obtain the better results on measuring the all perfor- j P ij
the shortest
mance criteria due to its superiority.
processing time
To ascertain the generic efficiency of the KBGA, it was also
SIO Select the job with minieSJ{Pij}
tested on ten random problems of moderate size. For these prob-
the shortest
lems, near-optimal solutions have been achieved in less number
imminent
of generations, compared to the simple GA procedure. The results
operation time
have been shown in Table 3. nP o
SRPT Select the job with mini2SJ
The proposed heuristics, which is based on the genetic algo- j2SRi P ij
the shortest
rithm and Knowledge base system, has been coded in C++ language
remaining
and the programme was run on a 2.83 GHz Pentium IV PC sup-
processing time
ported by Windows XP operating system. ( )
CR Select the job with
mini2SJ PDi t
the minimum ratio P ij
j2SRi
9. Conclusion between time now
until due date and
The present paper shows the implication of a new approach its remaining
for scheduling problem in FMS environment. In this paper, a processing time
knowledge based genetic algorithm (KBGA) has been proposed n P o
DS Select the job with mini2SJ Di  t  j2SRi P ij
to improve the performance of the system. It combines a simple the minimum slack
GA (SGA) with (tacit and explicit) expert knowledge that makes time
the knowledge base. This is an attempt to improve the effective- EDD Select the job with mini2SJ fDi g
ness for FMS scheduling problem with the throughput and mean the earliest due
flow time as the key performance measures. The initialization date
scheme of the proposed approach enhances the potentials of
the conventional GA, which has been reported in the previous
section, by using both explicit and implicit knowledge. In the
proposed algorithm, the operators are also used with the knowl-
edge base to reduce the computational obstacles that minimizes Notation Description
the CPU time to reach the optimal/sub-optimal solution. To
show the efficacy of the proposed KBGA, a comparative study i job index
has also been made with SGA. This study shows the faster con- j operation index
vergence rate of KBGA with better quality solution. In the pres- k machine index (K = number of machines)
ent study, the emphasis is also placed on the deciding on the t time when decision is to be made
crossover probability, which is also playing a vital role in diver- Pij processing time of the jth operation of job i
sifying the search and computational time. ARi time when job i arrives at the system
As a future scope, this research can be stretched out to various Di due date of job I
problems of the flexible system environment that cover the balanc- SJ set of jobs within the system
ing or allocation of resources. This research can also be employed SRi set of remaining operation of job i
P
for the multi-criterion decision-making problems in FMS environ- SLi slack time of job i ðSLi ¼ Di  t  j2SRi P ij Þ
ment as well as flexible supply chain environment.
 A. Prakash et al. / Expert Systems with Applications 38 (2011) 3161–3171 3171

References Moon, I., Lee, S., & Bae, H. (2008). Genetic algorithms for job shop scheduling
problems with alternative routings. International Journal of Production Research,
46(10), 2695–2705.
Beckman, T. (1997). A methodology for knowledge management. In Proceedings of the
Nietok, M. (2003). From R&D management to knowledge management: An
IASTED international conference on artificial intelligence and soft computing,
overview of studies of innovation management. Technological Forecasting and
ASC’97 (pp. 29–32). Banff: IASTED ACT Press.
Social Change, 70, 135–161.
Caprihan, R., & Wadhwa, S. (1997). Impact of routing flexibility on the performance
Nonaka, I. (1994). A dynamic theory of organizational knowledge creation.
of an FMS – A simulation study. International Journal of Flexible Manufacturing
Organization Science, 5(1), 14–37.
Systems, 9, 273–298.
Nonaka, I., & Takeuchi, H. (1995). The knowledge-creating company: How Japanese
Chang, P.-C., Chen, S.-H., & Lin, K.-L. (2005). Two-phase sub population genetic
companies create the dynamics of innovation. New York, NY: Oxford University
algorithm for parallel machine-scheduling problem. Expert Systems with
Press.
Applications, 29(3), 705–712.
O’Dell, C., & Grayson, J. (1997). If we only knew what we know: Identification and
Chaturvedi, A. R. (1993). FMS scheduling and control: Learning to achieve multiple
transfer of internal best practices. Houston, TX: American Productivity and
goals. Expert System with Applications, 6, 267–286.
Quality Center.
Cheng, B.-W., & Chang, C.-L. (2007). A study on flow shop scheduling problem
Park, J., Choi, H., & Kim, Y. (1998). A genetic algorithm approach to job shop
combining Taguchi experimental design and genetic algorithm. Expert Systems
scheduling considering alternative process plans. Journal of Korean Institute of
with Applications, 32, 415–421.
Industrial Engineers, 24, 130–145.
Davenport, T. H., Delong, D. W., & Beers, M. C. (1998). Successful knowledge
Pezzellaa, F., Morgantia, G., & Ciaschettib, G. (2008). A genetic algorithm for the
management projects. Sloan Management Review, 39(2), 43–57.
flexible job-shop scheduling problem. Computers & Operations Research, 35,
Davenport, T., & Prusak, L. (1998). Working knowledge: How organizations manage
3202–3212.
what they know. Boston, MA: Harvard Business School Press.
Pinedo, M. (2002). Scheduling: Theory, algorithms and systems. Englewood Cliffs, NJ:
De Jong, K. A. (1975). An Analysis of the Behavior of a class of Genetic Adaptive Systems.
Prentice-Hall.
PhD thesis, University of Michigan, Ann Arbour. Department of Computer and
Rajendran, C., & Ziegler, H. (2004). Ant-colony algorithms for permutation flow shop
Communication Sciences.
scheduling to minimize makespan/total flowtime of jobs. European Journal of
Etiler, O., Toklu, B., Atak, M., & Wilson, J. (2004). A genetic algorithm for flow shop
Operational Research, 155, 426–438.
scheduling problems. Journal of the Operations Research Society, 55, 830–835.
Ruiz, R., Maroto, C., & Alcaraz, J. (2006). Two new robust genetic algorithms for the
Garey, M. R., Johnson, D. S., & Sethi, R. (1976). The complexity of flowshop and
flow shop scheduling problem. Omega – International Journal of Management
jobshop scheduling. Mathematics of Operations Research, 1, 117–129.
Science, 34, 461–476.
Gen, M., & Cheng, R. (1997). Genetic algorithms and engineering design. New York:
Shiau, D.-F., Cheng, S.-C., & Huang, Y.-M. (2008). Proportionate flexible flow shop
Wiley.
scheduling via a hybrid constructive genetic algorithm. Expert Systems with
Gen, M., Lin, L., & Zhang, H. (2009). Evolutionary techniques for optimization
Applications, 34, 1133–1143.
problems in integrated manufacturing system: State-of-the-art-survey.
Shiue, Y. R., Guh, R. S., & Tseng, T. Y. (2009). GA-based learning bias selection
Computers & Industrial Engineering, 56(3), 779–808.
mechanism for real-time scheduling systems. Expert System with Applications,
Ghedjati, F. (1993). Genetic algorithms for the job-shop scheduling problem with
36, 11451–11460.
unrelated parallel constraints: Heuristic mixing method machines and
Stecke, K. E. (1983). Formulation and solution of non-linear integer production
precedence. Computers & Industrial Engineering, 37, 39–42.
planning problem for flexible manufacturing system. Management Science,
Goldberg, D. E., Korb, B., & Deb, K. (1989). Messy genetic algorithm: Motivation
29(3), 273–288.
analysis and first results. Complex system. Complex System Publications. pp.
Su, C. T., & Shiue, Y.-R. (2003). Intelligent scheduling controller for shop floor control
493–530.
systems: A hybrid genetic algorithm/decision tree learning approach.
Grabowski, J., & Wodecki, M. (2004). A very fast tabu search algorithm for the
International Journal of Production Research, 41(2), 2619–2641.
permutation flowshop problem with makespan criterion. Computers and
Tiwana, A. (2000). The knowledge management toolkit. Prentice-Hall.
Operations Research, 31, 1891–1909.
Tiwana, A., & Balasubramaniam, R. (2001). A design knowledge management
Guo, Y. W., Li, W. D., Mileham, A. R., & Owen, G. W. (2009). Applications of particle
system to support collaborative information product evolution. Decision Support
swarm optimisation in integrated process planning and scheduling. Robotics
Systems, 31, 241–262.
and Computer-Integrated Manufacturing, 25, 280–288.
Upton, D. (1994). The management of manufacturing flexibility. California
Handfield, R. B., & Nichols, E. Z. (1998). Introduction to SCM. New York: Prentice-Hall.
Management Review, 36(2), 72–89.
Holland, H. H. (1975). Adaptation in natural and artificial systems. Detroit MI:
Wadhwa, S., & Bhagwat, R. (1998). Judicious increase in flexibility and decision
University of Michigan Press.
automation in semi-computerized flexible manufacturing (SCFM) systems.
Hutchinson, J., Leong, K., Synder, D., & Ward, P. (1991). Scheduling approaches for
Studies in Informatics and Control, 2(8).
random job shops flexible manufacturing systems. International Journal of
Wadhwa, S., & Browne, J. (1989). Modeling FMS with Petri nets. International Journal
Production Research, 29(5), 1053–1067.
of flexible Manufacturing Systems, 1, 255–280.
Kumar, R., Tiwari, M. K., & Shankar, R. (2003). Scheduling of flexible manufacturing
Wadhwa, S., & Rao, K. S. (2000). Flexibility: An emerging meta-competence for
system: An ant colony optimization approach. Journal of Engineering
managing high technology. International Journal of Technology Management,
Manufacturing – Part B, 217, 1443–1453.
19(7/8).
Lee, D. Y., & DiCesare, F. (1994). Scheduling flexible manufacturing systems using
Wadhwa, S., Rao, K. S., & Chan, F. T. S. (2005). Flexibility-enabled lead-time
Petri nets and heuristic search. IEEE Transaction on Robotics and Automation,
reduction in flexible system. International Journal of Production Research, 43(15),
10(2), 123–132.
3131–3163.
Liang, M., & Dutta, S. (1990). A mixed-integer-programming approach to the
Wadhwa, S., & Saxena, A. (2006). Supply chain flexibility and knowledge
machine loading and process planning problem in a process layout
management: A decision knowledge sharing focus. Studies in Informatics and
environment. International Journal of Production Research, 28, 1471–1484.
Control, 15(1).
Michaelwicz, Z. (1992). Genetic algorithm + data structure = evaluation programs.
Zhou, H., Feng, Y., & Han, L. (2001). The hybrid heuristic genetic algorithm for job
New York: Springer-Verlag.
shop scheduling. Computers & Industrial Engineering, 40, 191–200.
Montazeri, M., & Van Wassenhove, L. N. (1990). Analysis of scheduling rules for an
FMS. International Journal of Production Research, 28(4), 785–802.