Вы находитесь на странице: 1из 17

ComputersOpsRes. Vol.24, No. 4, pp.

335-351, 1997
© 1997 ElsevierScience Ltd
Pergamon PII: S0305-0548(96)00061-5 Printed in Great Britain.All fights reserved
0305-0548/97 $17.00+0.00

A GENETIC ALGORITHM APPROACH TO THE SIMULTANEOUS


SCHEDULING OF MACHINES AND AUTOMATED GUIDED
VEHICLES

Gtindtiz Ulusoyt~t, Funda Sivrikaya-~;erifo~lu§,Omit Bilg~


Department of Industrial Engineering, Bo~azi~i University, Bebek, 80815 Istanbul, Turkey

(Received September 1995; in revisedform July 1996)

Scope and Purpee~ In this article, a flexible manufacturing system (FMS) consisting of several machine
centers and one or more identical automated guided vehicles (AGVs) is considered. Both the scheduling of
operations on machine centers as well as the scheduling of AGVs are essential factors contributing to the
efficiency of the overall system. An increase in the performance of the FMS under consideration would be
expected as a result of making the scheduling of AGVs an integral part of the overall scheduling activity. Thus,
this article addresses the problem of concurrent scheduling of machines and a number of identical AGVs in an
FMS. At an increasing rate, genetic algorithms (GAs) are being used to solve difficult combinatorial problems.
The subproblems of machine scheduling and AGV (vehicle) scheduling which together make up the problem
addressed, are both known to be NP-hard. An additional difficulty is caused by the highly multimodai fitness
landscape. Therefore, this problem is a challenging one for GAs. A GA with a two allelic representation scheme
and a special uniform crossover operator is designed and tested extensively. For testing beyond the 82 test
problems already available in the literature, an easily computable lower bound is proposed which is rather
successful in that it represents the optimum in a relatively large percentage of the additional 180 test problems
considered. Computational results show that the GA developed is an effective solution method for this
problem.

Abstract--This article addresses the problem of simultaneous scheduling of machines and a number of identical
automated guided vehicles (AGVs) in a flexible manufacturing system (F'MS) so as to minimize the makespan.
For solving this problem, a genetic algorithm (GA) is proposed. Here, chromosomes represent both operation
sequencing and AGV assignment dimensions of the search space. A third dimension, time, is implicitly given by
the ordering of operations of the chromosomes. A special uniform crossover operator is developed which
produces one offspring from two parent chromosomes. It transfers any patterns of operation sequences and/or
AGV assignments that are present in both parents to the child. Two mutation operators are introduced; a bitwise
mutation for AGV assignments and a swap mutation for operations. Any precedence infeasibility resulting from
the operation swap mutation is removed by a repair function. The schedule associated with a given chromosome
is determined by a simple schedule builder. After a number of problems are solved to evaluate various search
strategies and to tune the parameters of the proposed GA. 180 test problems are solved. An easily computable
lower bound is introduced and compared with the results of GA. In 60% of the problems GA reaches the lower
bound indicating optimality. The average deviation from the lower bound over all problems is found to be 2.53%.
Additional comparison is made with the time window approach suggested for this same problem using 82 test
problems from the literature. In 59% of the problems GA outperforms the time window approach where the
reverse is true only in 6% of the problems. © 1997 Elsevier Science Ltd. All rights reserved

I. I N T R O D U C T I O N
T h e c u r r e n t t r e n d in m a n u f a c t u r i n g t e c h n o l o g y is t o w a r d s c r e a t i n g fully a u t o m a t e d a n d i n t e g r a t e d
m a n u f a c t u r i n g e n v i r o n m e n t s . A c a r e f u l l y d e s i g n e d a n d efficiently m a n a g e d m a t e r i a l h a n d l i n g s y s t e m is
o f c r u c i a l i m p o r t a n c e in a c h i e v i n g the r e q u i r e d integration.

"["To whom correspondence should be addressed (e-mail: ulusoyg@boun.edu.tr).


:[:GandUz Ulusoy is Professor of Industrial Engineering at Bo~azi¢i University, Istanbul. He received his BSME from Robert
College, Istanbol; MSME from University of Rochester and PhD from VPI&SU. His current research focuses on advanced
manufacturing systems and project and machine scheduling. He is currently an Associate Editor of the European Journal of
Operational Research. He has published research articles in Interfaces. Journal of the OperationalResearch Society, European
Journal of Operational Research, International Journal of Production Economics, International Journal of ~'roduction
Research, liE Transactions, International Journal of Production and Operations Management, Journal of Operations
Management.
§ Funda Sivrikaya-SerifoJlu is a doctoral student in the Department of Industrial Engineering at BoJaziqi University, lstanbul. She
holds a B.S. degree in Industrial Engineering from that university and an M.S. degree in Engineering Economic Systems from
Stanford University. She is currently conducting doctoral research in genetic algorithm applications in machine and project
scheduling.
(1 Omit Bilge is an Assistant Professor of Industrial Engineering at Bo[azi¢i University, Istanbui. Her research interests include
design and analysis of computer integrated manufacturing systems, flexible manufacturing systems, automated guided vehicle
systems, manufacturing information systems and production planning and control. Dr. Bilge has published in Operations
Research and InternationalJournal of Production Research.
24-4-

335
336 GUndtiz Ulusoy et al.

Automated Guided Vehicles Systems (AGVS) are among various advanced material handling
techniques that are finding increasing applications in today's computer integrated manufacturing (CIM)
settings. They are capable of transporting a variety of part types from point to point without human
intervention. They can be interfaced to various other production and storage equipment and controlled
through an intelligent computer control system. This flexibility and compatibility make AGVs a feasible
alternative to traditional material handling methods especially in flexible manufacturing environments.
This article addresses the problem of concurrent scheduling of machines and a number of identical
AGVs in a flexible manufacturing system (FMS). The importance of the material handling system for the
efficiency of the overall system has been emphasized by several researchers [1-6]. Attempts to improve
the AGVS' design process have been reported in literature [7-13]. An area of equal importance is that of
making scheduling of AGVs an integral part of the overall scheduling activity. Obviously, an increase in
the performance of the FMS would be expected as a result of the coordination of the machine and the
material handling system during the machine scheduling phase.
The simultaneous scheduling problem can be defined in the following general form:
Given a particular FMS environment and a set of jobs, determine the starting and completion times of
operations for each job and the trips between work stations together with the vehicle assignment
according to some system performance measure.
Two basic approaches to this problem are real-time scheduling and off-line scheduling. Both aspects
are studied by several researchers. Wu and Wysk [14], Ro and Kim [15], Sabuncuo~lu and
Hommertzheim [16], and Sawik [17] develop on-line dispatching and control rules for machines and
AGVs. The case of a special material handling transporter in a real time environment is treated by Han
and McGinnis [4]. Taghaboni and Tanchoco [18] develop an intelligent real-time controller for free-
ranging AGVs. Tanchoco and Co [19] introduce real-time control strategies for multiple-load AGVs.
Karabtik and Sabuncuo~lu [20] introduce a beam search based algorithm for the simultaneous
scheduling of machines and AGVs. A deterministic off-line scheduling model formulated as an integer
programming problem and a solution procedure based on concepts of project scheduling under resource
constraints are presented by Raman e t al. [ 1]. Their assumption that vehicles always return to the load/
unload station after transferring a load reduces the flexibility of the AGV and its influence on the
schedule. Blazewicz e t al. [21] consider an FMS with parallel identical machines arranged in a loop.
Pandit and Palekar [22] present a number of variants of a shifting bottleneck heuristic for minimizing
makespan with a single vehicle. Another off-line model for makespan minimization is presented by Bilge
and Ulusoy [23] who investigate the problem for multiple AGVs. They formulate the problem as a mixed
integer programming problem. In this formulation, the AGVs don't have to return to the load/unload
station after each delivery which increases the complexity of the problem. The overall problem is
decomposed into two subproblems, and an iterative solution procedure is developed.
In this article, a genetic algorithm (GA) application to the simultaneous scheduling problem defined
above will be presented. The objective is the minimization of makespan denoted by (;'max.The makespan
is defined as the time interval between the pick up of the first part from the load/unload station to the
finishing time of the last operation processed.
In the following section, the FMS environment will be described. Next, after a brief discussion of
genetic search approach to scheduling problems, the particular GA developed will be explained. Lastly,
the experiments and numerical results will be presented together with a proposal for a lower bound to this
problem.

2. THE FLEXIBLE MANUFACTURING ENVIRONMENT


The environment within which the FMS under consideration operates can be described as follows:
(1) The types and number of machines are known. Operations are nonpreemptive. There is sufficient
input/output buffer space at each machine.
(2) Processing, set-up, loading, unloading times are available and are deterministic.
(3) Number of AGVs is given and the AGVs are all identical in the sense that they have the same
speed and load carrying characteristics.
(4) Flow path layout is given and travel times on each segment of the path are known.
(5) A load/unload (L/U) station serves as a distribution center for parts not yet processed and as a
collection center for parts finished. All vehicles start from the L/U station initially and return to
there after accomplishing all their assignments. There is sufficient input/output buffer space at the
A genetic algorithm approach 337

(L/U) station.
(6) AGVs carry a single unit-load at a time. They move along predetermined shortest paths, with the
assumption of no delay because of congestion. Preemption of trips is not allowed. The trips are
called loaded or deadheading (empty) trips depending whether a part is carded or no part is carded
during that trip, respectively. The durations for the deadheading trips are sequence dependent and
are not known until the vehicle route is specified.
(7) It is assumed that all the design and set-up issues within the hierarchy of OR/MS problems in an
FMS as suggested by Stecke [24], have already been resolved. Machine loading, i.e., the allocation
of tools to machines and the assignment of operations to machines, is made. Pallets and other
necessary equipment are allocated to parts. The set of part types to be produced during the
planning period and the routing of each part type are available before making scheduling
decisions. In other words, routing flexibility is not considered. The routing for a part type can be
selected based on considerations of technological feasibility and processing efficiency, or by
formulating the set-up phase problems in a manner that can also handle the routing decisions
[25].
(8) Ready-times of all jobs are known. Initially, partially processed parts might be available at
machines waiting for further processing, and they can be treated as jobs having zero ready times
and their routing consists of the remaining operations.
(9) Such issues as traffic control, congestion, machine failure or downtime, scraps, rework, and
vehicle dispatches for battery changer are ignored here and left as issues to be considered during
real-time control.

3. GENETIC ALGORITHMS AND MACHINE SCHEDULING


Genetic algorithms (GAs) were developed by John Holland and his students as artificial adaptive
systems that simulate natural evolution [26]. As GAs are able to search very large spaces effectively and
efficiently, they are increasingly used to attack inherently intractable problems called NP-hard problems.
A large portion of machine scheduling problems belong to the class of NP-hard problems. Thus,
optimizing routines suggested for such problems explode rather quickly with increasing problem size. As
a consequence, more emphasis is placed on developing heuristic procedures which usually do not have
a claim for reaching either a local or a global optimum. Their aim is to obtain near optimal solutions
within a reasonable computation time. This results in the restriction of the search space in some or the
other way, leaving some parts totally untouched. Although GAs are heuristic procedures themselves, GAs
test a wealth of samplings from different regions of the search space for fitness simultaneously, and sort
out and exploit regions of interest very quickly. GAs are randomized but they are not random walks. The
fact that GAs use probabilistic transition rules and not deterministic ones also contributes to a more
effective search of the solution space.
Several researchers have successfully used GAs for machine scheduling problems. Mainly two
different approaches for chromosome representations are used: Indirect and direct representations.
In an indirect approach, the chromosome is an abstraction of the actual schedule. For example, only
the ordering of job orders is given by the chromosome while a schedule builder performs the rest of the
search. Syswerda [27] advocates this approach because of the advantage that the GA can be limited to
use one of the several variants of operators for ordered lists. Holsapple et al. [28] use an indirect
representation in their hybrid scheduler that generates static schedules in flexible manufacturing contexts.
The job sequences generated by the GA are fed into a filter beam search scheme that generates associated
schedules.
In a direct representation scheme, a chromosome represents all the information relevant to the problem.
In other words, the schedule itself becomes the chromosome. The advantage of this representation is that
the search is done entirely by the GA; but more complex genetic operators and usually also repair
functions for validating illegal offspring are needed. Empirical work has, in general, shown the
superiority of direct schemes over indirect ones. Uckun et al. [29] study both of these schemes in their
schedule optimizer prototype VSOP in the context of job shop scheduling. Their conclusion is that the
more problem specific information is included in the representation, the better is the performance. Kanet
and Sridharan [30] and Bruns [31] both use such a direct scheme; the former in parallel machine
scheduling, the latter in a general production scheduling environment.
338 GfindUzUlusoy etal.

There are also some representations which cannot be strictly classified as either direct or indirect. The
work of Fang et al. [32] involves a chromosome representing all the information relevant to the
scheduling problem but the explicit time settings. A simple schedule builder generates the schedule
according to the ordering of genes on the chromosome.
Lee et al. [33] use an adaptive representation scheme for solving two related problems, that of lot
sizing and sequencing, concurrently. The chromosome gives both lot sizes and sequences, and itself
undergoes an evolution.

4. A GENETIC ALGORITHM FOR THE SIMULTANEOUS SCHEDULING OF MACHINES AND AGVS

4.1. Representation
In this application, chromosomes are fixed-length strings representing both of the dimensions of the
search space: operation sequencing and AGV assignment. The third dimension, time, is implicitly given
by the ordering of operations on the chromosomes; the operations are assumed to be scheduled at their
earliest starting time resulting in a nondelay schedule. Each operation denotes the processing of a specific
part family on a specific machine center, and is expressed by an ASCII character starting from lower case
letters. The corresponding assignment of AGVs that will carry the workpiece from and to the machine
centers are given by integer numbers that identify AGVs. On the chromosome, lower case letters and
integer numbers alternate. The smallest meaningful piece of the chromosome, a gene, is made up of one
operation and its corresponding AGV assignment.
A small example is given in Fig. 1. It illustrates a scheduling situation with 3 workcenters, 3
workpieces and 2 AGVs. The first workpiece has 3 operations and the other 2 workpieces have 2
operations each. Note that as there are 7 operations the chromosome will consist of 7 genes. The
workcenters where each operation will be performed are as indicated. The travel times and process times
are given in Figs l(b) and (c). Consider the chromosome [d2, fl, al, e2, b2, gl, c2], which describes a
schedule. In the chromosome above, AGV-1 is to carry operations f, a, and g; and AGV-2 the operations
d, e, b, and c, in the given order on the chromosome. For example, workpiece 3 will be carded by AGV-1
from the L/U station to M3 for its first operation, f, to be performed. Similarly, workpiece 1 will be
carded by AGV-2 from M1 to M3 for operation b to be performed. For this chromosome, the schedule
results in the Gantt chart provided in Fig. l(d). On this Gantt chart, the numbers (l) and (2) in front of
each operation refer to the AGV that is assigned to carry the part involved.
Deadlock situations arise when AGVs have to wait for each other as the predecessors of the currently
assigned operations are both not considered yet. In this application, chromosomes representing such
deadlock situations are penalized with a very high makespan value.
A simple, computationally cheap schedule builder computes the makespan associated with any
chromosome by building the nondelay schedule represented by the ordering of operations on that
chromosome. The main reason for designing a schedule builder which generates nondelay schedules has
been the relative success of nondelay scheduling strategy over active scheduling strategy for this
particular problem as reported in [23]. Parts are processed at a workcenter following the FCFS rule.
Besides being a good performing rule, FCFS is also appealing in practice from the operational point of
view. When two parts arrive at the same workcenter at the same time, then SPT rule is used to break the
tie in the ordering of these parts for processing.

4.2. Operators
A shift from the traditional binary representation to a problem specific one is almost always coupled
with a shift from simple binary operators to more complex, problem specific ones. This is also the case
here.
4.2.1. Crossover operator. A special uniform crossover is developed that creates one offspring from
two parent chromosomes. Starting from the first operations on the parents, iteratively, one of the parents
is chosen randomly and its next unconsidered operation becomes the next operation on the child. If the
AGV selected for that operation is the same on both parents, then that selection is also made on the child;
if not, one of the AGV selections of the parents is chosen randomly. Figure 2 below gives a more detailed
account of the crossover operator. This operator possesses one much desirable characteristic of
recombination operators: It enables the offspring to inherit good partial solutions contained in the parent
A genetic algorithm approach 339

chromosomes. For the problem at hand, both the absolute and relative positioning of operations and of
operation-AGV pairs comprise important information to be propagated to the progeny. The suggested
operator facilitates the building block propagation in all those dimensions. In addition, the operator never
violates the precedence constraints. If the two parent chromosomes are feasible in terms of precedence
constraints, so will be the child. This claim can easily be proven as follows.
Proof. The first object on the child is the first object from either parent. Given that the parents are
precedence feasible, this first object is one without any predecessors. At any iteration of the algorithm,
the next object on the child becomes the first yet unconsidered object on the randomly chosen parent. This
object has all its predecessors already assigned to the child as otherwise the chosen parent would be
infeasible.

Workpieces 1 2 3

/> /--->
Operations 1 2 3 1 2 1 2

Workcenters
II
MI M3 M2
IIII
M1 M3 M3 M2

Representation a b c d e f g

(a) Job set

M1 M2 M3
0 4 8 10
18 0 4 6
20 14 0 8
M3 12 8 6 0
(b) Traveltime (0 matrix

PI P2
op p math op p math T w
a 12 MI d 14 MI
b 8 M3 e 18 M3 g 15
c 15 M2
(c) ProcessingTimes

MI
4 18 26 38

M2
38 53 58 73

M3
1o 20 24 42 44 52

(d) The Oanttch~ for the schedule

Fig. I. Exampleproblemfor the illustration of the chromosome[d2, fl, al, e2, b2, gl, c2]
340 Gilndliz Ulusoy et al.

START Choose one of the parentsrandomly.


The first operation on the first gene of the child becomes the first operation of the
c h ~ e n lmrenc
F'md this same operation on the other parent. If the AGV selections are the same on
both parents,make that sameselectionalso on the child; if not, chooseone of the AGV
selectionsof the parentsrandomlyand recordit on the child.
The fast gene of the child is completed.Proceedto the next geneon the child.
LOOP Choose one of the parents randomly.
Find the first operation on the chosen parent which is not assigned to the child yet.
That ol~ration becomes the next operation on the child.
If the AGV selections made for that operation are the same on both parents, then make
this same selection on the child too. If not, choose one of the AGV selections of the
parents randomlyand recordit on the child.
If all the genes of the child chromosome m-e completed,STOP
else proceedto the nextgeneon the child and repeatLOOP.
Fig. 2. Algorithmof the proposedcrossoveroperator.

4.2.2. Mutation operators. There are two mutation operators, one for each dimension of the search.
Operation swap mutation chooses two random positions on the chromosome and swaps the operations
associated with those positions with a user-specified probability. Note that the operation swap mutation
operator may cause infeasibilities in terms of the precedence relations. The simple repair algorithm
explained below complements the mutation operators in generating feasible mutations. In the bitwise
AGV mutation, for each gene in the population of chromosomes mutation occurs with a user-specified
probability. If mutation is to occur at a gene, one of the AGVs is chosen randomly and is assigned to that
gene to carry the part associated with the operation located in that gene. This may result in the same AGV
assignment for the particular gene, and aims to prevent the loss of any good assignments.

4.2.3. Random solution generator. A random solution generator is developed for providing solutions
to the initial population. The initial population can be generated entirely by this generator or can be
seeded with any number of seed solutions provided by the user. This generator constructs a chromosome
gene by gene. For an operation to be assigned to a gene, it needs to be eligible. An operation is said to
be eligible once all its predecessors are assigned. A set of eligible operations to be scheduled next is kept.
This set is initially formed by the first operations of each workpiece. At each iteration, one of the
operations in this set is chosen randomly and put onto the next position on the child. Then, one of the
AGVs is randomly chosen to complete the gene. If the chromosome is not yet complete, the eligible set
is updated and the process continues.

4.2.4. The repair function. A repair function is developed that validates chromosomes with any
precedence violations. Although some problem specific heuristics are incorporated, the repair function is
not designed to be too smart to prevent overly good repairs that lead to high performing children from
poorly performing parents. When repairing, care is taken not to create other infeasibilities. Repair is used
only to validate offspring generated by operation swap mutation.
Figure 3 below details the main part of the repair function. This part is repeatedly applied to the
chromosome until it is precedence feasible.

4.2.5. The flexible structure. The GA developed provides the user with a flexibility to configure the
GA runs. This is due to its flexible pool management as given in Fig. 4. Generation gaps are formed
following the structure provided by Grefenstette [34]. These gaps determine the degree to which
successive generations overlap. Here, a user-specified parameter, called generation gap and denoted by
G, determines the proportion of the population that will be carried over in tact into the next generation.
Explicit mating pool formation is also possible. Again, a user-specified rate, called mating pool formation

Find positions of the operations which violatethe ~ncc relations;


If the distance inbetween is smallerthan half the chromosome length
then
swap violatingoperations
else
choose one of the operations randomly;, lake it out and reinsert it right
before/afterthe other one dependingm the precedencerelations.
Fig. 3. Repair function.
A genetic algorithm approach 341

old oal I~ new ooi


N°G individuals
gap strategy

N*C individuals
mating pool strategy

recombination
mating pool
Fig. 4. Pool management.

rate and denoted by C, gives the proportion of the population that will be selected to form a mating pool
from where parent solutions are chosen to undergo recombination. There are three strategies for the
selection of individuals for both generation gap and mating pool formation: (i) fitness-based; (ii) elitist,
where only the best solution is preserved and rest is chosen fitness-based; (iii) elitist, where all the best
(N'G) or (N'C) solutions are preserved with N being the population size.

5. EXPERIMENTATION WITH THE GENETIC ALGORITHM


5.1. Generation of test problems
The genetic algorithm developed has been tested on 180 different problems generated. Each problem
is a combination of a layout and a job which consists of a set of parts to be processed. The layout and
the job are drawn from a set of layouts and from a set of jobs respectively. Three different layouts and
four different job sets are generated for this purpose.
The layouts are displayed in Figs 5, 6, and 7 where the travel distances in meters are indicated on each
arc. In all the layouts there are four machines. The travel times are calculated from the distances by
assuming an average speed of 40 m/rain for the AGVs. The loading and unloading times of the parts onto
and from the AGVs are taken to be 0.5 min each. In all problems, two AGVs are assumed to be
available.

6 6

6 6 6

-NEE 6 6 6
B
9

Fig. 5. Layout 1 (not to scale).

12 F --]12
8

16

Pq
Fig. 6. Layout2 (not to scale).
342 GtindlizUlusoy et al.

28
6

6
B
6
8 6~.q 6 8

Fig. 7. Layout3 (not to scale).

As each job consists of a set of parts, first two different part pools are generated. It is assumed here
that the operations needed to manufacture a part have to be performed in the given order, i.e. there is a
strict linear precedence relation between operations. Further, no alternative machines are assumed to be
available. Thus a part routing is defined by the sequence of machine numbers in the same order these
machines are visited by the part. In the following two different part pools generated, the part routings in
the second part pool display a unidirectional flow in machine numbers whereas the part routings in the
first part pool are of a mixed nature in this aspect.
The first part pool: {4-2-3}, {3-1-4}, {2-4-3}, {2-1-3}, {3-4-2-1}, {4-2-1}, {1-2-4}, {3-2-4}, {1-3-2},
{ 1-2-3-4}, {4-3-1 }.
The second part pool: {3-4}, {1-2-3-4}, {1-2-4}, {1-2}, {1-3-4}, [1-3}, 11-2-3}, {2-4}, {1-4},
{2-3-4}.
Four job sets are generated; the first two using the first part pool, and the last two using the second part
pool. The second and fourth job sets consist of jobs with a larger number of parts and thus a larger total
number of operations compared to the jobs in the first and third job sets. Each job set is applied to all
three layouts.
The first job set. This job set contains 30 jobs. The jobs consist of 6--10 parts randomly selected with
replacement from the first part pool.
The second job set. There are 5 jobs in this set. The jobs consist of 15-20 parts randomly selected with
replacement from the first part pool.
The third job set. This job set contains 20 jobs. The jobs consist of 6-10 parts randomly selected with
replacement from the second part pool.
The fourth job set. There are 5 jobs in this set. The jobs consist of 15-20 parts randomly selected with
replacement from the second part pool.
The processing times for the operations are randomly generated as integers from the interval [3,15]
minutes. When generating the problems an important characteristic taken into account is the relative
magnitude of the travel times and the processing times. This characteristic is represented as the ratio (t//5)
where t is the average of all the nonzero entries in the travel time matrix and/~ is the average of all
processing times. The problems are generated with a (t//~) ratio in the range [0.05, 0.20] in order to reflect
the practice into the test problems.

5.2. Tuning the parameters of the genetic algorithm


To evaluate various search strategies and to tune the parameters of the proposed GA, experiments are
conducted on 20 of the generated problems. In these experiments, I0 parameters are investigated
regarding their effects on the performance of the GA. These parameters are population size, number of
generations, crossover, bitwise mutation and operation swap mutation rates, generation gap and mating
pool formation rates and strategies, and the form of the fitness function. For population size, various
levels above 50 are tried. Number of generations is varied between 30 and 400; crossover rate between
0.50 and 1.00; bitwise mutation rate between 0.001 and 0.03; and operation swap mutation rate between
0.00 and 0.50. Several levels between 0 and 1 are tried for the generation gap formation rate G and
mating pool formation rate C. Also, all three strategies are investigated for both generation gap formation
and the mating pool formation processes.
A genetic algorithm approach 343

These experimentations have shown the viability of a well established and widely used search strategy
for off-line optimization of difficult problems as suggested by Goldberg [35]: Given a fixed number of
evaluations, running smaller sized GAs many times on the same problem is better than running a larger
GA for a few times. Each random restart yields a different sampling of the search space; and over many
replications, the probability of locating the global optima increases. Also, it has been observed that, even
when large populations are evolved over longer generations, the GA converges very early on. Once the
GA finds a good region, the rest of the generations is mainly used to homogenize the population around
that region. This may in part be the result of the crossover operator developed in this application; but
several researchers report similar observations.
The crossover operator developed for this application is more 'conservative' in nature when compared
to other order-based crossover operators. That is, it is more geared towards construction/exploitation than
towards destruction/exploration of partial solutions. This effect is counterbalanced by high mutation
rates.
One important characteristic of the fitness landscape is its multimodality. To complicate things further,
there are a large number of representatives for each (local) optimum. High mutation rates also help to
escape those local optima.
As the evolution proceeds the variation in the population decreases and it becomes more and more
difficult to distinguish between individuals of the population. Goldberg suggests the use of a fitness
scaling scheme. Here, as a substitute to a scaling scheme, two different fitness functions are tried; namely,
fix)= 1/z(x) andJ(x)= l/z2(x), wheref(x) denotes the fitness of an individual x and z(x) is its payoff or
objective function value. The latter form has been more effective in sorting out better individuals,
therefore it has been utilized in the GA runs.
As a result of these preliminary experimentations, values for GA parameters are determined. The GA
developed is replicated 20 times for each problem. For all sets of problems, each replication of the GA
is run for 100 generations. The objective function of each new chromosome is compared with the lower
bound for the problem under consideration. If they are equal, then, as an optimal solution is found, the
runs for that replication are terminated before completing all 100 generations. The population size is
selected to be 140. For all the problems, mating pool formation rate C and generation gap G are taken
to be 1 and 0.04, respectively. Mating pool formation strategy is fitness-based, while gap generation
strategy is elitist where all the best chromosomes (N'G) are preserved from one generation to the next.
Crossover rate is selected to be 0.8. The rate of operation swap mutation per chromosome is 0.4 and the
bitwise AGV mutation rate is 0.01.

5.3. Computation of a lower bound


The problem of simultaneous scheduling of machines and AGVs is a two dimensional problem where
both of the dimensions are NP-hard. The problem is also highly nonlinear. It is therefore quite difficult
to come up with a good suggestion for a tight lower bound (LB). A suggestion for an LB is given
below.
It has been assumed that an AGV is always available when needed and that no idle time need to be
inserted inbetween operations so that on each machine all the operations can be processed one after the
other. The only idle time on a machine is the time span necessary for the arrival of the first workpiece
to be processed on that machine. This time span includes the travel times and the nondelayed processing
of the predecessor operations. If there is an operation within the set of operations to be processed on a
machine which is the first operation of its part, then only the travel time from the L/U unit to that
particular machine need to be accounted for. If not, then, from among alternative choices for the first
operation, the one is chosen which gives rise to the smallest idle time. LB is then taken to be the
maximum of all LBs for completion times. Note that the method proposed here is not only easy to apply
but also computationally feasible. Equations (1)-(4) detail the formulation of the lower bound.

f L=t[L/U,M(j)] + U, if P(j)=~
PT
J=l E (L+t[g(pred(k)), M(k)]+pk+U)+L+t[M(pred(j)),M(j)]+U, o/w (1)
ITI~= min {PT~}, (2)
JeOui
Cui=ITli+ Y~ pj (3)
j~o,~
LB= min {CMA (4)
344 Gttnd0z Ulusoy et al.

where
P(j) set o f predecessors o f operationj
OMi set o f operations to be processed on machine i
pred(j) predecessor o f operation j; the first operations o f each part are preceded by a dummy operation
d where M(d)=L/U
M(j) machine index upon which operationj is performed
L,U loading and unloading times
P5 time needed for the process and transport of all predeccessors o f operationj
ITli the first idle time on machine i
CM, completion time o f the last operation processed on machine i.
The computation o f the lower bound is demonstrated at the hand o f an example problem illustrated
in Fig. 8. In the example problem, the processing o f parts P1 and P2 start on M I ; P3 and P5 start on
M3 and P4 starts on M4. Therefore these machines can start processing as soon as an A G V transfers a
part from the L/U station to the respective machine. This means, a lower bound on the completion times
for M 1, M3 and M4 can be computed as the sum o f the processing times o f the operations which are to
be processed on these machines and the transfer time from the L/U station. They are calculated and
are shown in Fig. 6. As no part has its first operation processed on M2, a lower bound on M2's comple-
tion time is determined by considering all the operations to be performed on M2 (namely, b,f, and k) one
by one as the first operation to be processed by M2. Starting with operation b leads to an idle time o f

t(L/U,M1)

M1
4 12 32 47 62
ITmin I
M2 I b I ~' I k 1
t ( t ~ . M 3 ) 16 32 50 68

I
M3 I 8 I el l I
t(L/U~I4) 22 32 42
/
M4 I O l cl"l
0 14 28 40 48

(a) The Gantt chin for the lowerbound

machine ~ MI M2 M3 M4
I./U 0 4 8 10 14
MI 18 0 4 6 10
M2 20 14 0 8 6
M3 12 8 6 0 6
M4 14 14 12 6 0

(b) Travel lime (t) matrix

P1 P5
op P p roach p mach op p roach op p mach
a 8 M1 20 MI 12 M3 14 M4 I 10 M3
b 16 M2 e I0 M3[h 8 M4 18 M2 m 15 MI
C 12 M4 f 18 M2 15 MI

(c)ProcessingTimes
Fig. 8. Example problem for the computation of the lower bound
A genetic algorithm approach 345

Table 1. Results: Job Set l--Layout I

GA
Problem (tip) LB %Deviation I00
mean st.dev, best (best-LByLB

Jl 0.08 89.83 1.28 86.20 84.75 1.71


J2 0.07 83.04 0.22 82.75 82.75 0.00
J3 0.07 89.19 2.60 81.35 80.75 0.74
J4 0.08 74.77 0.31 74.60 74.60 0.00
J5 0.06 78.35 0.48 78.20 78.20 0.00
J6 0.08 80.57 1.47 78.10 68. I0 14.68
J7 0.07 81.07 1.07 80.20 78.05 2.75
J8 0.07 75.93 O.I I 75.90 75.90 0.00
J9 0.08 67.64 0.09 67.60 67.60 0.00
J10 0.08 81.59 2.14 79.35 74.90 5.94
J 11 0.07 78.05 0.00 78.05 78.05 0.00
J 12 0.07 82.59 0.82 82.20 77.75 5.72
J 13 0.07 70.05 0.00 70.05 70.05 0.00
J14 0,07 96.15 0.29 96.05 96.05 0.00
J 15 0,06 103.05 0.00 103.05 103.05 0.00
J16 0.08 85.05 0.00 85.05 85.05 0.00
J 17 0.08 86.87 0.20 86.75 86.75 0.00
J 18 0.07 7 I. 10 0.45 69.20 63.75 8.55
JI9 0.07 69.13 1.27 67.90 67.90 0.00
J20 0.08 69.10 0.24 69.05 69.05 0.00
J21 O.12 43.46 0.80 42.95 33.75 27.26
J22 0.12 47.33 0.00 47.33 47.33 0.00
J23 0.12 42.47 0.51 40.83 36.83 10.86
J24 0.12 48.66 0.79 47.35 46.70 1.39
J25 0.12 50.33 0.08 50.31 47.42 6.09
J26 0.11 44.35 0.17 44.18 44.18 0.00
J27 0.12 48.78 0.65 47.85 47.85 0.00
J28 0.12 59.22 0.88 57.27 55.04 4.05
J29 0.11 41.17 0.48 40.99 40.68 0.76
J30 0.13 48.73 0.88 47.68 47.68 0.00

Average percent deviation from the lower bound =3.02.

t[L/U, M 1] +p(a) + t[M 1,M2] = 4 + 8 + 4 = 16, which is the processing time of the predecessor operation
a plus the transfers. Similarly, starting with the operation f gives rise to an idle time of
t[L/U, M1]+p(d)+t[M1,M3]+p(e)+t[M3,M2]=46 and with operation k to an idle time of

Table 2. Results: Job Set I - - ~ y o u t 2

GA
Problem (t/p) LB %Deviation
mean st.dev, best 100 (best-LB)/LB

J1 0.10 90.73 0.72 90.48 84.80 6.70


J2 0.10 83.22 0.36 82.80 82.80 0.00
J3 0,09 90.48 0.00 90.48 80.80 11.98
J4 O. lO 74.91 0.93 74.50 74.50 0.00
J5 0.09 78.20 0.08 78.18 78.18 0.00
J6 0.11 81.31 2.53 76.48 69.85 9.49
J7 0.10 83.03 1.62 80.18 77.88 2.95
J8 0.10 76.27 0.14 76.18 76.18 0.00
J9 0. I I 67.97 I. 11 67.50 67.50 0.00
Jl0 0.10 81.77 2.63 79.48 75.18 5.72
JI I 0.10 77.88 0.00 77.88 77,88 0.00
J12 O.lO 83.55 1.24 79.50 77.80 2.19
JI3 0.09 69.88 0.00 69.88 69.88 0.00
J 14 O.10 96.20 0.78 95.88 95.88 0.00
Jl5 0.08 102.88 0.00 102.88 102.88 0.00
J16 0.10 84.88 0.00 84.88 84.88 0.00
J17 0.10 87.49 0.81 86.80 86.80 0.00
JIB 0.10 72.56 0.38 72.48 63.80 13.61
JI9 0.09 69.36 1.94 68.18 68.18 0.00
J20 O.I I 69.06 0.55 68.88 68.88 0.00
J21 0.16 43.01 0.69 41.63 33.65 23.71
J22 O.16 47.50 O.14 47.46 47.46 0.00
J23 O.17 42.98 0.96 41.64 36.65 13.62
J24 O.16 49.68 0.70 48.40 46.97 3.04
J25 0.17 51.39 0.12 51.36 47.24 8.72
J26 0.15 45.46 0.58 45.15 45.15 0.00
J27 0.16 48.97 0.89 47.92 47.60 0.67
J28 0.16 60.46 1.50 58.86 56.39 4.38
J29 0.15 42.10 0.27 41.97 40.95 2.49
J30 0.17 50.26 0.71 48.56 47.58 2.06

Average percent deviation from the lower bound =3.71


346 Giindtlz Ulusoy et al.

Table 3. Results : Job Set I---Layout 3

GA
Problem (7//3) LB %Deviation
mean st.dev, best 100 (best-LB)/LB

JI 0.11 89.37 1.95 85.15 84.50 0.77


J2 0.11 83.61 0.56 82,50 82.50 0.00
J3 0.10 88.70 2.41 83.15 80.50 3.29
J4 0.12 74.73 1.23 74.15 74.15 0.00
J5 0.10 78.95 0.95 78.50 78.50 0.00
J6 0.12 81.24 3.28 76.15 70.15 8.55
J7 0.11 83.25 0.92 81.15 78.85 2.92
J8 0.11 76.50 0.00 76.50 76.50 0.00
J9 0.12 68.25 1.08 67.50 67.15 0.52
Jl0 0.11 81.98 2.49 78.50 75.50 3.97
JIl 0.11 78.85 0.00 78.85 78.85 0.00
J12 0.11 82.74 1.82 77.50 77.50 0.00
JI3 0.10 70.85 0.00 70.85 70.85 0.00
JI4 0.11 97.40 1.09 96.85 96.85 0.00
JI5 0.09 103.85 0.00 103.85 103.85 0.00
JI6 0.12 85.85 0.00 85.85 85.85 0.00
J17 0.11 86.80 0.52 86.50 86.50 0.00
J18 0.11 72.15 0.00 72.15 63.50 13.62
J19 0.10 70.48 2.13 68.50 68.50 0.00
J20 0.12 70.26 1.08 69.85 69.85 0.00
J21 0.18 44.06 0.95 43.15 33.64 28.27
J22 0.18 48.40 0.79 47.83 47.83 0.00
J23 0.19 43.84 0.81 42.15 37.63 12.01
J24 0.18 50.74 0.84 48.95 47.30 3.49
J25 0.18 52.31 0.00 52.31 48.22 8.48
J26 0.17 45.35 0.67 44.83 44.83 0.00
J27 0.18 49.03 1.12 47.79 47.30 1.04
J28 0.18 60.99 1.01 58.97 56.74 3.93
J29 0.17 42.68 0.42 42.33 41.28 2.54
J30 0.19 51.89 0.94 50.12 47.23 6.12

Average percent deviation from the lower bound =3.32

t[LAJ,M4] +p(/)+t[M4,M2] =40. The minimum of these is 16 for operation b (denoted as ITmin in Fig. 6),
thus b is processed first on M2. The lower bound on the completion time of M2 becomes the sum of the
processing times of b, k, andfplus ITmi n. Lower bound LB is determined as the maximum of the lower
bounds on the completion times of all the machines. Here, it corresponds to that of M2 and has a value
of 68.

6. N U M E R I C A L RESULTS

6.1. The set of 180 problems


The solutions and related statistics for all four .job sets and the three layouts obtained using the
proposed GA are presented in Tables 1 respectively. For each layout and job set, the best GA solution
among the 20 replications is given together with the mean and standard deviation of the best solutions
of these 20 replications. One interesting observation is that compared to the magnitude of the mean

Table 4. Results : Job Set 2

GA
Layout Problem (i/p) LB %Deviation
mean st.dev, best from LB

I J31 0.07 150.05 0.00 150.05 150.05 0.00


J32 0.07 168.90 0.00 168.90 168.90 0.00
J33 0.07 130.05 0.00 130.05 130.05 0.00
J34 0.07 158.75 0.00 158.75 158.75 0.00
J35 0.08 151.05 0.00 151.05 151.05 0.00
2 J31 0.I0 149.88 0.00 149.88 149.88 0.00
J32 0.I0 169.18 0.00 169.18 169.18 0.00
J33 0.10 129.88 0.00 129.88 129.88 0.00
J34 0.I0 158.80 0.00 158.80 158.80 0.00
J35 0.II 150.88 0.00 150.88 150.88 0.00
3 J31 0.11 151.00 0.67 150.85 150.85 0.00
J32 0.II 169.50 0.00 169.50 169.50 0.00
J33 0.11 130.85 0.00 130.85 130.85 0.00
J34 0.11 158.50 0.00 158.50 158.50 0.00
J35 0.12 151.85 0.00 151.85 151.85 0.00

Average deviation of best GA from LB: 0.00%.


A genetic algorithm approach 347

Table 5. Results : Job Set 3--Layout l

GA
Problem (i//~) LB %Deviation
mean st.dev, best 100 (best-LB)/LB

FI 0.08 53.80 1.47 53.20 51.20 3.9 I


F2 0.08 75.73 2.01 72.05 65.60 9.83
F'3 0.07 59.14 1.94 58.05 55.20 5.16
1:4 0.07 93.35 0.00 93.35 93.35 0.00
1:5 0.07 78.75 0.00 78.75 78.75 0.00
F6 0.08 55.40 0.62 55.20 50.75 8.77
F7 0.08 67.35 0.00 67.35 67.35 0.00
F8 0.08 76.70 1.13 75.35 74.90 0.60
F9 0.08 64.90 0.00 64.90 64.90 0.00
FIO 0.07 75.75 0.00 75.75 75.75 0.00
FI l 0.08 86.90 0.00 86.90 86.90 0.00
FI2 0.08 76.09 2.81 73.90 73.9O 0.00
FI3 0.08 66.35 0.00 66.35 66.35 0.00
FI4 0.07 108.58 0.72 108.35 107.35 0.93
FI5 0.08 67.80 1.21 66.75 66.75 0.00
F 16 0.07 89.65 2.18 85.75 85.75 0.00
FI7 0.07 72.45 2.82 69.20 59.35 16.60
F 18 0.07 84.98 0.71 84.75 84.75 0.00
FI9 0.08 73.80 1.35 73.20 70.75 3.46
F20 0.08 51.18 0.44 51.05 49.35 3.44

Average percent deviation from the lower bound=2.64

values, the standard deviation values are extremely small. Indeed, for problems with a nonzero standard
deviation the coefficient of variation varies in the range [0.10, 4.28]. Further, for 61 out of 180 problems
the standard deviation is zero. When one looks more closely to the final pools of the 20 replications for
these problems, one observes that various distinctly different solutions exist with the same Cm~ value.
This implies that there are many alternative optima and the proposed GA is capable of finding them.
The comparison of the proposed GA with the lower bound is also given in Tables 1-8. In 108 out of
180 problems (60%), the GA solution coincides with the LB implying that the solution is optimal. The
average deviation of GA solutions from LB over all problems is calculated to be 2.53%. The average
deviation varies between 0.00% and 3.71% among the twelve problem sets. Note that these values would
be lower except for a few problems where the deviation from the lower bound is much larger than the
average. The performance of the lower bound proposed is satisfactory in general.
Another observation is that the means are relatively close to each other over the different layouts. This
observation is tested through a t-test the results of which are presented in Table 9. These results confirm
the observation for all problem sets at an a-level of 0.005. A major reason for this result appears to be
the relatively low (tip) ratio.

Table 6. Results : Job Set 3--Layout 2

GA
Problem (i/p) LB %Deviation
mean st,dev, best 100 (best-LB)/LB

FI 0.11 53.38 0.89 53.18 51.17 3.93


F2 O.11 73.92 2.44 71.80 65.50 9.62
F3 0.10 58.73 1.78 57.80 55.17 4.77
F4 0.09 93.48 0.00 93.48 93.48 0.00
F5 0.10 78.93 0.60 78.80 78.80 0.00
1::6 0.11 55.18 0.00 55.18 50.80 8.62
F7 O.I 1 67.48 0.00 67.48 67.48 0.00
F8 0.11 76.71 1.06 75.18 75.18 0.00
F9 0.11 65.18 0.00 65.18 65.18 0.00
FIO 0.09 75.80 0.00 75.80 75.80 0.00
FII 0.11 87.18 0.00 87.18 87.18 0.00
FI2 0.11 76.33 3.07 74.18 74.18 0.00
FI3 0.10 65.88 0.00 65.88 65.88 0.00
FI4 0.09 108.48 0.00 108.48 106.88 1.50
FI5 0.11 67.60 1.12 66.80 66.80 0.00
FI6 0.10 90.73 1.31 86.73 85.80 1.08
FIT 0.10 71.99 2.92 69.18 59.48 16.31
FI 8 0.09 84.80 0.00 84.80 84.80 0.00
FI9 O.ll 74.01 1.63 73.18 70.80 3.36
F'20 0.10 50.97 0.52 50.80 49.48 2.67

Average percent deviation from the lower bound = 2.59


348 Gtind~ Ulusoy et al.

Table 7. Results : Job Set 3--Layout 3

GA
Problem (j/p) LB %Deviation
mean st.dev, best 100 (best-LB)/LB

FI 0.12 53.30 1.22 51.50 51.50 0.00


F2 0.12 74.34 2.19 71.50 65.15 9.75
F3 0. I I 58.77 2.36 55.85 55.50 0.63
1:4 0.10 93.85 0.00 93.85 93.85 0.00
F5 0. I I 78.50 0.00 93.85 78.50 0.00
F6 0.12 54.96 1.21 52.85 50.50 4.65
F7 0.12 67.85 0.00 67.85 67.85 0.00
F8 0.12 77.90 1.130 75.50 75.50 0.00
F9 O.13 65.50 0.00 65.50 65.50 0.00
FIO 0.10 75.50 0.00 75.50 75.50 0.00
FI1 O.12 87.50 0.00 87.50 87.50 0.00
F 12 0.12 76.24 2.63 74.50 74.50 0.00
FI3 0.12 66.85 0.00 66.85 66.85 0.00
FI4 0.10 109.35 1.05 108.85 107.85 0.93
FI5 0.12 67.52 1.38 66.50 66.50 0.00
FI 6 O. 11 86.29 1.36 85.50 85.50 0.00
FI7 0.11 70.13 2.20 66.85 59.85 11.70
FI8 0.10 84.50 0.00 84.50 84.50 0.00
FI9 0.12 73.20 1.39 70.50 70.50 0.00
F20 0.12 50.77 0.64 50.50 49.85 1.30

Average percent deviation from the lower bound= 1.45

6.2. The set of 82 problems


A heuristic solution procedure suggested for this same problem is the time window approach which is
tested on 82 problems reported [23]. This set of 82 problems is also solved using the GA proposed here
and the results are displayed in Tables 10 and 11 together with the corresponding results of the time
window approach and the lower bound. The problems are grouped into two groups; one with relatively
low (t//~) ratio (Table 10 ) and one with relatively high (t//3) ratio (Table 11). In 59% of the problems GA
Table 8. Results : Job Set 4

GA
Layout Problem 0//~) LB %Deviation
mean st.dev, best from LB

I F21 0.08 110.32 1.15 110.05 105.60 4.21


F22 0.08 132.90 0.00 132.90 132.90 0.00
F23 0.07 147.39 6.04 143.00 139.35 2.62
F24 0.08 131.45 0.35 130.90 130.90 0.00
F25 0.07 130.03 0.35 129.90 129.90 0.00
2 F21 0.11 109.80 0.00 109.80 105.50 4.08
F22 0.10 133.18 0.00 133.18 133.18 0.00
F23 0.10 148.36 3.55 144.48 138.88 4.03
F24 0.10 131.29 0.28 131.18 131.18 0.00
F25 0.09 130.18 0.00 130.18 130.18 0.00
3 F21 0.12 107.10 2.80 105.15 105.15 0.00
F22 O. 12 133.50 0.00 133.50 133.50 0.00
I=23 O. I I 150.15 6.43 146.85 139.85 5.01
F24 0.11 131.79 1.01 131.50 131.50 0.00
F25 0.10 130.50 0.00 130.50 130.50 0.00

Average deviation of best GA from LB: 1.33%

Table 9. t-Tests for the effect of layout differences


Job set Layouts Mean difference D st.dev. (b) t-test dof

1 I vs 2 -0.601 0.374 - 1.608 29


1 vs 3 - 0.469 0.260 - 1.808
2 vs 3 0.132 0.340 0.387
2 1 vs 2 0.036 0.089 0.404 4
I vs 3 - 0.580 0.215 2.700
2 vs 3 -0.616 0.268 2.303
3 1 vs 2 -0.044 0.066 -0.665 19
1 vs 3 0.485 0.253 1.915
2 vs 3 0.529 0.245 2.158
4 I vs 2 -0.144 0.248 0.581 4
1 vs 3 -0.190 0.962 0.198
2 vs 3 - 0.046 0.740 0.062

t4~=8.610; tig.o.eoos=3.883:t~es= 3.659


A geneticalgorithmapproach 349

outperforms the time window approach where the reverse is true only in 6% of the problems. GA
performs up to 7.19% better than the time window approach in the first group and up to 6.54% better in
the second group. The relative performance of GA is even better in problems with relatively high (t//5)
ratio. Average improvement of GA over the time window approach increases from 1.16% in Table 10 to
2.16% in Table ll.
Another observation is with regard to the performance of the lower bound proposed. The deviation of
GA from the lower bound is very different for the two groups of problems. For the 42 problems in Table
10, the deviation is 5.30%. For the 40 problems in Table 11, the deviation is 26.68%. The results reported
in Table l0 and Table 1l, as well as those of the previous 180 problems indicate that with increasing (t/p)
ratios the quality of the lower bound deteriorates. This is an expected result, as the lower bound procedure
ignores the travelling times except for the first operation on each machine.

7. C O N C L U S I O N AND EXTENSIONS

A GA is proposed here for the solution of the simultaneous scheduling of the machines and a number
of identical AGVs in an FMS. The algorithm has been tested extensively. The comparison of the solutions
obtained with the lower bound proposed here indicates that the algorithm performs well. The lower bound
proposed is easy to apply and is computationally feasible.
The objective of makespan minimization belongs to the class of performance criteria based upon
completion times. A second class of performance criteria is based upon due dates and includes
performance criteria such as mean tardiness and maximum tardiness. Recently, the performance criterion
of 'sum of weighted earliness and weighted tardiness' is used more frequently mainly as it reflects closely
the just-in-time manufacturing environment [36]. The GA proposed here can easily be modified to

Table 10. Summary of the results for problems with (rib)<0.25


Problem (i/p) Lower Time GA %Improvement
Bound Window (TW) Best Mean Std Devn 100 (TW-best)/TW

EXII0 0.15 126 126 126 126.00 0.00 0.00


EX210 0.15 148 148 148 148.00 0.00 0.00
EX310 0.15 138 150 148 149.75 0.64 1.33
EX410 0.15 112 121 119 120.00 0.97 1.65
EXSI0 0.21 102 102 102 102.00 0.00 0.00
EX610 0.16 163 186 186 186.90 2.20 0.00
EX710 0.19 137 137 137 137.00 0.00 0.00
EX810 0.14 271 292 271 273.75 0.79 7.19
EX910 0.15 150 176 176 176.40 1.39 0.00
EXI010 0.14 218 238 236 244.50 4.44 0.84
EXI20 0.12 123 123 123 123.00 0.00 0.00
EX220 0.12 143 143 143 143.00 0.00 0.00
EX320 0.12 135 148 145 145.00 0.00 2.03
EX420 0.12 Ill ll6 ll4 115.05 0.89 1.72
EX520 0.17 99 100 100 ! 00.00 0.00 0.00
EX620 0.12 160 183 181 181.00 0.00 1.09
EX720 0.15 136 136 136 136.00 0.00 0.00
EX820 0. I l 268 287 268 268.95 0.22 6.62
EX920 0.12 150 174 173 173.35 1.18 0.57
EX 1020 0. I l 216 236 238 243.55 3.33 - 0.85
EXI30 0.13 122 122 122 122.00 0.00 0.00
EX230 0.13 146 146 146 146.00 0.00 0.00
EX330 0.13 136 149 146 146.00 0.00 2.01
EX430 0.13 ll0 116 ll4 114.15 0.37 1.72
EX530 0.18 98 99 99 99.00 0.00 0.00
EX630 O.14 161 184 182 183.20 2.46 1.09
EX730 0.17 137 137 137 137.00 0.00 0.00
EX830 0. ! 3 269 288 270 270.35 1.57 6.25
EX930 0.13 151 176 174 175.00 2.18 1.14
EXI030 0.12 217 237 241 247.05 3.90 - 1.69
EXI40 0.18 124 124 124 124.00 0.00 0.00
EX241 0.13 217 217 217 217.00 0.00 0.00
EX340 0.18 138 151 151 151.00 0.00 0.00
EX341 0.12 203 222 221 221.00 0.00 0.45
EX441 0.19 166 179 172 173.05 1.05 3.91
EX541 0.18 148 154 148 148.00 0.00 3.90
EX640 0.19 161 185 184 185.45 2.54 0.54
EX740 0.24 137 138 137 137.00 0.00 0.72
EX741 0.16 203 203 203 203.00 0.00 0.00
EX840 0.18 272 293 273 282.85 8.44 6.83
EX940 0.19 149 177 175 175.75 1.83 1.13
EXI040 0.17 219 240 244 251.25 3.80 - 1.67

Average improvement of GA over time window approach: 1.16%


350 Giindiiz Ulusoy et al.

Table 11. Summary of the results for problems with (t//~)>0.25


Problem (t/#) Lower Time GA %Deviation
Bound Window (TW) Best Mean Std Devn 100 (TW-best)fl'W

EX I I 0.59 72 96 96 96.50 1.00 0.00


EX21 0.61 86 105 104 106.35 1.76 0.95
EX31 0.59 81 105 105 109.45 2.56 0.00
EX41 0.91 62 118 116 120.95 3.14 1.69
EX51 0.85 60 89 87 89.15 0.75 2.25
EX61 0.78 96 120 121 125.00 2.90 - 0.83
EX71 0.78 76 119 118 123.70 3.26 0.84
EX81 0.58 146 161 152 158.40 2.85 5.59
EX91 0.61 93 120 117 120.60 1.88 2.50
EXI01 0.55 124 153 150 157.40 3.05 1.96
EX 12 0.47 66 82 82 82.00 0.00 0.00
EX22 0.49 76 80 76 79.50 1.82 5.00
EX32 0.47 75 88 85 86.30 1.53 3.41
EX42 0.73 60 93 88 93.10 2.47 5.38
EX52 0.68 54 69 69 70.55 1.88 0.00
EX62 0.54 86 100 98 102.95 2.24 2.00
EX72 0.62 74 90 85 87.30 2.18 5.56
EX82 0.46 140 151 142 143.05 2.80 5.96
EX92 0.49 91 104 102 104.85 2.16 1.92
EXI02 0.44 114 139 137 142.50 3.10 1.44 ~
EX 13 0.52 64 84 84 85.80 0.89 0.00
EX23 0.54 82 86 86 86.75 1.86 0.00
EX33 0.51 77 86 86 87.45 1.73 0.00
EX43 0.80 58 95 91 97.10 3.55 4.21
EX53 0.74 52 76 75 75.85 0.37 1.32
EX63 0.54 88 104 104 106.45 1.82 0.00
EX73 0.68 76 91 88 91.80 3.07 3.30
EX83 0.50 142 153 143 144.60 2.09 6.54
EX93 0.53 93 110 105 107.30 1.59 4.55
EX 103 0.49 116 143 143 147.40 2.78 0.00
EXI4 0.74 68 108 103 105.80 2.26 4.63
EX24 0.77 84 116 113 116.60 2.95 2.59
EX34 0.74 81 116 113 119.85 3.82 2.59
EX44 1.14 62 126 126 132.85 2.94 0.00
EX54 1.06 56 99 97 98.30 1.63 2.02
EX64 0.78 90 120 123 130.00 3.39 - 2.50
EX74 0.97 76 136 128 136.80 4.67 5.88
EX84 0.72 148 163 163 167.10 3.34 0.00
EX94 0.76 91 125 123 124.75 1.92 1.60
EXI04 0.69 120 171 164 171.30 4.49 4.09

Average improvement of GA over time window approach: 2.16%

incorporate different due date based performance criteria by changing the evaluation routine of the
schedule accordingly.
It should be emphasized that although this problem has been stated in an FMS environment with the
material handling vehicles being AGVs, it is also a valid problem formulation in a job-shop environment
where the transfer of parts between machines might not be assumed to be instantaneous due to limited
material handling resources.
The problem and the solution procedure are not only of interest to the scheduler for off-line scheduling
but also to the designer. A series of problems can be defined to represent different operating conditions
and solved to understand better the design situation. As a result of such an exercise the designer can
extract some clues regarding the machine layout, path layout and the number of AGVs needed.
An interesting extension would be to introduce into the problem routing flexibility and scheduling
flexibility in the sense defined in [37] and [38]. This would require a redefinition of the chromosome,
but it should be handled without a major revision of the GA.
REFERENCES
1. R a m a n , N., Talbot, E B. a n d R a c h a m a d g u , R. V., Simultaneous scheduling o f machines a n d material handling devices in
a u t o m a t e d manufacturing. In Proc. of the Second ORSA/TIMS Conf. on Flexible Manufacturing Systems, eds K. E. Stecke a n d
R. Suri, pp. 455-466. Elsevier, A m s t e r d a m , 1986.
2. G r e e n w o o d , N. R., Implementing Flexible Manufacturing Systems. Macmillan, London, 1988.
3. Bozer, Y. A., G u i d e d vehicle systems: information/control system implication o f alternative design a n d operation strategies. In
Advanced Information Technologies for Industrial Material Flow Systems, eds S. Y. N o f and C. L. Moodie, N A T O ASI Series,
pp. 417--436. Springer-Verlag, Berlin, 1989.
4. Han, M. a n d McGinnis, L. E , Control o f material handling transporter in automated manufacturing, liE Trans., 1989, 21,
184--190.
5. Kouvelis, P., Design and planning problems in flexible manufacturing systems: a critical review. J. Intelligent Manufacturing,
1992, 3, 7 5 - 9 9 .
6. Noble, J. S. a n d Tanchoco, J. M. A., Design justification o f material handling systems. In Material Flow Systems in
A genetic algorithm approach 351

Manufacturing, ed. J. M. A. Tancboco, pp. 54--72. Chapman and Hall, 1994.


7. Maxwell, W. L. and Muchstadt, J. A., Design of antomated guided vehicles, lie Trans., 1982, 14, 114--124.
8. Grosseschallan, W. and Kusiak, A., An expert system for design of automated material handling systems (INSIMAS). Mater.
Flow, 1985, 2, 157-166.
9. Chang, Y., Sullivan, R. S. and Wilson, J. R., Using SLAM to design the material handling system of an FMS. Int. J. Prod. Res.,
1986, 24, 15-26.
10. Mahadevan, B. and Narendran, T. T., Design of an automated guided vehicle design-based material handling system for an
FMS. Int. J. Prod. Res., 1990, 28, 1611-1622.
11. Malmborg, C. J., A model for the design of zone control automated guided vehicle systems, b~t. J. Prod. Res., 1990, 28,
1741-1758.
12. Gobal, S. L. and Kasilingam, R. G., A simulation model for estimating vehicle requirements in automated guided vehicle
systems. Computers Industrial Engg, 1991, 21, 623--627.
13. Rembold, B. E, and Tanchoco, J. M. A., An integrated framework for the design of material flow systems. In Material Flow
Systems in Manufacturing, ed. J. M. A. Tanchoco. Chapman and Hull, 1994, pp. 3-53.
14. Wu, S. D. and Wysk, R. A., An application of discrete-event simulation to on-line control and scheduling in flexible
manufacturing. Int. J. Prod. Res., 1989, 27, 1603-1623.
15. Ro, I. and Kim, J., Multi-criteria operational control rules in flexible manufacturing systems. Int. J. Prod. Res., 1990, 28,
47-63.
16. Sabuncuo~,lu, I. and Hommertzheim, D. L., Dynamic dispatching algorithm for scheduling machines and automated guided
vehicles in a flexible manufacturing system. Int. J, Prod. Res., 1992, 30, 1059-1079.
17. Sawik, T., Scheduling of machines and vehicles in an FMS: single-level versus multi-level approach. Automatyka, 1993, 64,
541-568.
18. Taghaboni, F. and Tanchoco, J. M. A., A LISP based controller for free-ranging automated guided vehicle systems. Int. J. Prod.
Res., 1988, 26, 173-188.
19. Tanchoco, J. M. A. and Co, C., Real-time control strategies for multiple-load AGVs. In Material Flow Systems in
Manufacturing, ed. J. M. A. Tanchoco. Chapman and Hull, 1994, pp. 300-331.
20. Karabtlk, S. and Sabuncuo~,lu, I., A beam search based algorithm for scheduling machines and AGVs in an FMS. In Proc.
Second Industrial Engg Res. Conf., 1993, pp. 308-312.
21. Blazewicz, J., Eiselt, H. A., Finke, G., Laporte, G. and Weglarz, J., Scheduling tasks and vehicles in a flexible manufacturing
system. Int. J. Flexible Manufacturing Sys., 1991,4, 5-16.
22. Pandit, R. and Palekar, U. S., Job shop scheduling with explicit material handling considerations. Working Paper, Iowa State
University, Ames, Iowa, 1993.
23. Bilge, O. and Ulusoy, G., A time window approach to simultaneous scheduling of machines and material handling system in
an FMS. Opns Res., 1995, 43, 1058-1070.
24. Stecke, K. E., Design, planning, scheduling, and control problems of flexible manufacturing systems. Annals of Opns Res.,
1985, 3, 3-12.
25. Kiran, A. S. and Tansel, B., Mathematical programming models for FMSs. Working Paper 86-01, University of Southern
California, Department of Industrial and Systems Engineering, Los Angeles, California, 1986.
26. Holland, J. H., Adaptation in Natural and Artificial Systems, The University of Michigan Press, Ann Arbor, Michigan, 1975.
27. Syswerda, G., Schedule optimization using genetic algorithms. In Handbook of Genetic Algorithms, ed. L. Davis. 1991, pp.
332-349.
28. Holsapple, G. W., Jacop, V. S., Pakath, R. and Zaveri, J. S., A genetics based hybrid scheduler for generating static schedules
in flexible manufacturing contexts. IEEE Trans. Systems Man and Cybernetics, 1993, SMC-23, 953-972.
29. Uckun, S., Bagchi, S., Kawamura, K. and Miyabe, Y., Managing genetic search in job shop scheduling, 1EEE Expert, 1993,
pp. 15--24.
30. Kanet, L J. and Sridharan, V., PROGENITOR: A GA for production scheduling. Wirtschafisinformatik, 1991,33, 4 332-336.
31. Bruns, R., Direct chromosome representation and advanced genetic operators for production scheduling. In Proc. 5th Int. Conf.
on Genetic Algorithms, ed. S. Forest. Morgan Kaufmann, Fairfax, Virginia, 1993, pp. 352-359.
32. Fang, H.-L., P. Ross, P. and Come, D., A promising GA approach to job-shop scheduling, rescheduling and open-shop
scheduling problems. In Proc. 5th Int. Conf. on GeneticAIgorithms, ed. S. Forest. Morgan Kaufmann, Fairfax, Virginia, 1993,
pp. 375-382.
33. Lee, I., Sikora, R. and Shaw, M. J., Joint lot sizing and sequencing with GAs for scheduling: Evolving the chromosome
structure. In Proc. 5th Int. Conf. on Genetic Algorithms, ed. S. Forest. Morgan Kaufmann. Fairfax, Virginia, 1993, pp.
383-389.
34. Grefenstette, J. J., Optimization of control parameters for genetic algorithms. IEEE Trans. Systems Man and Cybernetics, 1986,
SMC-16, 122-128.
35. Goldberg, D. E., Sizing populations for serial and parallel genetic algorithms. In Proc. 3rd Int. Conf. on Genetic Algorithms,
ed. D. J. Schaffer. Morgan Kaufmann, Fairfax, Virginia, 1989, pp. 70-79.
36. Baker, K. R. and Scudder, G. D., Sequencing with earliness and tardiness penalties: A review. Opns Res., 1990, 38, 22-36.
37. Chang, Y. L., Matsua, H. and Sullivan, R., A bottleneck based beam search for job scheduling in a flexible manufacturing
system. Int. J. Prod. Res., 1989, 27, 1949-1963.
38. Rachamadugu, R. and Schriber, T. J., Performance of dispatch rules under perfect sequence flexibility. In Proc. of the Winter
Simulation Conf., eds O. Balci, R. Sadowski and R. E. Nance. 1990, pp. 653-658.

Вам также может понравиться