Artificial Intelligence for Engineering Design, Analysis and Manufacturing ~2001!, 15, 251–263. Printed in the USA.

Copyright © 2001 Cambridge University Press 0890-0604001 $12.50

A review of machine learning in dynamic scheduling

of flexible manufacturing systems

PAOLO PRIORE, DAVID DE LA FUENTE, ALBERTO GOMEZ, and JAVIER PUENTE

ETSII e II, Campus de Viesques, 33204 Gijón, Spain
(Received January 25, 1999; Accepted February 1, 2001!

Abstract
A common way of dynamically scheduling jobs in a flexible manufacturing system ~FMS! is by means of dispatching
rules. The problem of this method is that the performance of these rules depends on the state the system is in at each
moment, and no single rule exists that is better than the rest in all the possible states that the system may be in. It would
therefore be interesting to use the most appropriate dispatching rule at each moment. To achieve this goal, a scheduling
approach which uses machine learning can be used. Analyzing the previous performance of the system ~training
examples! by means of this technique, knowledge is obtained that can be used to decide which is the most appropriate
dispatching rule at each moment in time. In this paper, a review of the main machine learning-based scheduling
approaches described in the literature is presented.
Keywords: Discrete Simulation; Dispatching Rules; Dynamic Scheduling; Flexible Manufacturing Systems;
Machine Learning

1. INTRODUCTION conventional manufacturing systems to disturbances, as their

Scheduling, a part of any manufacturing system’s control components are more synchronized, more integrated, and
process, is necessary when a common set of resources needs more interdependent. They therefore require immediate
to be shared to manufacture several different products dur- response to changes in system states, using a real-time sched-
ing the same time period. The goal of scheduling is to as- uling method. If system states change dynamically, sched-
sign machines and other resources to jobs, or operations uling of parts should be done as a function of the current
within jobs, in an efficient manner, as well as to determine state of the system ~Jeong & Kim, 1998!.
the moment when each of the jobs is processed ~Shaw et al., The rest of this paper is organized as follows. The differ-
1992! ent techniques described in the literature to schedule FMS
Processing times in flexible manufacturing systems jobs are first described. Then, there is a description of two
~FMSs! are almost deterministic, as operations are computer- ways of modifying the dispatching rules dynamically so as
controlled and mainly processed by numerically controlled to overcome the problem that dispatching rules cause when
machines, and setups between consecutive operations are applied statically. One of these two ways is based on the use
automated. As a result, providing the system is not dis- of a simulation model, whilst the other utilizes “scheduling
turbed in some way, results can be predicted and a fixed knowledge” of the manufacturing system. A review of the
off-line scheduling system is sufficient. work done on the two approaches is then provided, and
However, the actual states of FMSs may not be predict- their main characteristics are described. Finally, the paper
able because of part arrivals, machine states ~up or down!, ends with a consideration of a series of generalized short-
tool breakages, rushed jobs, and many other system distur- comings of knowledge-based systems that need to be dealt
bances. This dynamic, uncertain nature of the FMS sug- with in future research.
gests that an off-line scheduling system is not really the
most adequate. Moreover, FMSs are more sensitive than
2. APPROACHES TO SCHEDULING IN FMSs

Address correspondence and reprint requests to: Paolo Priore, ETSII e The different approaches available to solve the problem of
II, Campus de Viesques, 33203 Gijón, Spain. E-mail: [email protected] FMS scheduling can be divided into the following categories:
251
252 P. Priore et al.

1. The analytical approach. at least as well as the best of the candidate dispatching rules
2. The heuristic approach. being considered. Secondly, this approach can adapt its
choices dynamically to changing scenarios. This adaptabil-
3. The simulation-based approach. ity should result in job scheduling of a higher quality than
4. The artificial intelligence-based approach. even the best dispatching rules ~Shaw et al., 1992!.
Basically, two approaches to modifying dispatching rules
The analytical approach interprets an FMS scheduling dynamically can be found in the literature. Firstly, the rule
problem as an optimization model with certain constraints, is selected at the appropriate moment by simulating a set of
in terms of an objective function and explicit constraints. preestablished dispatching rules and choosing the one that
An appropriate algorithm is then used to resolve the model provides the best performance. In the second approach, be-
~see, e.g., Stecke, 1983; Kimemia & Gershwin, 1985; longing to the field of artificial intelligence, a set of earlier
Shanker & Tzen, 1985; Lashkari et al., 1987; Han et al., system simulations ~training examples! is used to deter-
1989; Hutchison et al., 1989; Shanker & Rajamarthandan, mine which is the best rule for each possible system state.
1989; Wilson, 1989!. These training cases are used to train a machine learning
In general, these problems are of a NP-complete type module to acquire knowledge about the manufacturing sys-
~Garey & Johnson, 1979!. Heuristic and off-line type algo- tem. Such knowledge is then used to make intelligent deci-
rithms are therefore usually proposed to resolve this kind of sions in real time. These scheduling systems are normally
problem ~Cho & Wysk, 1993; Chen & Yih, 1996!. How- said to be knowledge-based.
ever, these analytical models contain simplifications that In contrast, there are other scheduling schemes within
are not always valid in practice. Indeed, Basnet and Mize the artificial intelligence approach in which dynamic mod-
~1994! state that some models are so singular that one has ification of the dispatching rule does not take place ~see,
the impression that the problems are invented to fit the e.g., Fox & Smith, 1984; Maimon, 1987; Maley et al., 1988;
model rather than vice versa. They are not efficient for Shen & Chang, 1988; Shaw & Whinston, 1989; Chaturvedi
reasonably large-scale problems either. et al., 1993; Dong & Kitaoka, 1994; De & Lee, 1998!.
The above-mentioned difficulties of applying the analyt- Kanet and Adelsberger ~1987!, as well as Kusiak and Chen
ical approach to scheduling problems led to research into ~1988!, both present reviews of expert systems as applied
many heuristic approaches. These are usually dispatching to scheduling. Jain and Meeran ~1998! review scheduling
rules, although they may be more complicated than that, systems that use neural networks. Aytug et al. ~1994! and
and they are generally used to schedule the jobs in a man- Minton ~1993! present a review of work in which machine
ufacturing system dynamically. These heuristics use differ- learning is applied to solving scheduling and planning prob-
ent priority schemes to order the different jobs competing lems. Zweben and Fox ~1994! give different scheduling
for the use of a given machine. Each job is assigned a pri- systems that use artificial intelligence, including real sys-
ority index and the one with the lowest index is selected tems used in different industrial fields ~aerospace, defence,
first. heavy industry, and semiconductor manufacturing!.
Many researchers ~see, e.g., Panwalkar & Iskander, 1977; Finally, in the literature on FMSs scheduling, there are
Blackstone et al., 1982; Baker, 1984; Russel et al., 1987; many reviews of a general nature that apply any of the four
Vepsalainen & Morton, 1987; Ramasesh, 1990; Kim, 1990! above-mentioned approaches ~see, e.g., Harmonosky &
have evaluated the performance of these dispatching rules Robohn, 1991; Kouvelis, 1992; Gunasekaran et al., 1993;
on manufacturing systems using simulation. The conclu- Basnet & Mize, 1994!.
sion to be drawn from such studies is that their performance
depends on many factors, such as the criteria that are se-
3. SIMULATION-BASED SYSTEMS
lected, the system’s configuration, the work load, and so on
~Cho & Wysk, 1993!. With the advent of FMSs came many The general scheme of simulation-based scheduling sys-
studies analyzing the performance of dispatching rules in tems is shown in Figure 1. When the simulator receives a
these systems ~see, e.g., Stecke & Stolberg, 1981; Egbelu & request to select a rule, it carries out a series of simulations
Tanchoco, 1984; Denzler & Boe, 1987; Choi & Malstrom, with each of the a priori selected rules. From amongst the
1988; Henneke & Choi, 1990; Montazeri & Van Wassen- results of the simulations carried out, the selector chooses
hove, 1990; Tang et al., 1993!. the best dispatching rule to use to schedule the manufactur-
In view of the variable performance of dispatching rules, ing system’s jobs. Finally, if an anomaly occurs in the sys-
it would be interesting to modify these rules dynamically tem, the control system sends a signal to the rule selector.
and at the right moment according to the system’s condi- This can then send the simulator a new request according to
tions. That this approach will be better than the conven- the type of anomaly observed. Some simulation-based sys-
tional system of using a dispatching rule constantly is an a tems that vary the dispatching rule applied at each particu-
priori assumption, for two reasons. Firstly, because it can lar moment dynamically will next be reviewed.
identify the best rule for a given manufacturing scenario. Wu and Wysk ~1989! put forward a scheduling system
Given such a selection capacity, the system should perform and on-line control that selects the best rule via simulation
FMS scheduling by machine learning 253

Fig. 1. General overview of a simulation-based scheduling system.

for each time period; they call this the scheduling interval. problem of “censored data” and the constant scheduling
Although they do point out that the duration of this interval interval. Moreover, the average improvement according to
is an important factor insofar as it determines the system’s the six performance criteria that are considered is 5.17% if
performance, they do not provide a general procedure by the third strategy is used to lessen the problem of “censored
which it might be defined. Simply by using multiples of the data”. Curiously, the authors claim that the algorithm pro-
average total processing time, they make the claim and ob- posed by Wu and Wysk ~1989! is 4.12% inferior to the best
servation that three times this amount is the best possible of the rules used constantly.
scheduling interval. However, a constant interval cannot Kim and Kim ~1994! suggest a real-time simulation-
keep up with the changes of state in a system as dynamic as based scheduling method whose dispatching rules vary dy-
the FMS is. They likewise define a simulation window, to namically. The main elements of the proposed method are
simulate a model and evaluate the performance of the can- the simulator and the real-time control system. The func-
didate rules, which is the same as the scheduling interval. tion of the former is to evaluate the rules and select the best
Because of the parts that remain in the system at the end of one for a given performance criterion. The latter compo-
each simulation period, this window affects each of the nent supervises the manufacturing system and checks its
rules differently ~this problem is called “censored data”!. performance periodically.
Despite all this, the authors achieve an improvement of The selected dispatching rule is applied until the differ-
7.7% and 21.11%, respectively, for the FMS they used, com- ence between actual performance and performance as cal-
pared to employing one rule constantly, using mean tardi- culated by the simulator exceeds a given limit, or until there
ness and mean flow time as performance measurements. is a major disturbance. When either of these occur, a new
Ishii and Talavage ~1991! present an approach that at- rule is selected by the simulator with the jobs that remain to
tempts to solve the above problems which has three main be carried out. The authors study the methodology in rela-
components. First, it calculates the scheduling intervals as tion to the monitoring period of the control system and the
a function of an index that measures the system’s state. The limits in performance differences. The improvement with
second element determines the dispatching rule to be used. respect to mean tardiness and mean flow time is 6.10% and
The authors propose four strategies that define different 2.08%, respectively.
simulation windows to reduce the problem of “censored Jeong and Kim ~1998! use a scheme based on the work of
data”. The final element is the FMS simulator. Kim and Kim ~1994! and analyze two factors that can in-
When the authors compared this approach with the one fluence the scheduling system. First, the type of simulation
proposed by Wu and Wysk ~1989!, they observed that the model that is used; this can be dynamic or static, depending
latter displayed extremely variable behavior, owing to the on whether it includes probability distributions of system
254 P. Priore et al.

disturbances. Second, the authors study the right moment to One way of achieving these characteristics is to utilize
select a new rule, for which there are four options: some class of knowledge about the relationship between
the manufacturing system’s state and the rule to be applied
1. Once only, at the beginning of the planning horizon. at that moment. It is therefore useful to use “scheduling
2. When there is a major disturbance. knowledge” of the manufacturing system to save time and
get a rapid response in a dynamically changing environ-
3. If there is a major disturbance or if the difference ment ~as are FMS environments!. However, one of the most
between the actual performance value and the one difficult problems to solve in a knowledge-based system is
estimated by the simulator goes over a given limit. precisely how this knowledge is to be acquired.
4. When any kind of disturbance occurs. To acquire knowledge, machine learning techniques, such
as inductive learning or neural networks, are used. These
Their results show that modifying the rules in response reduce the effort involved in determining the knowledge
to changes in the manufacturing system reduces the mean required to make scheduling decisions. However, the train-
flow time between 1.77% and 3.20%. Moreover, mean tar- ing examples and the learning algorithm must be right for
diness goes down between 8.34% and 12.05%. Further- this knowledge to be useful. Moreover, in order to get the
more, ANOVA analysis shows that the simulation model training examples, the attributes that are selected are cru-
type is not relevant. The authors conclude by calculating cial to the performance of the scheduling system that is
the percentage use of each of the rules, and they come to the generated ~Chen & Yih, 1996!.
conclusion that no rule predominates over the others, and There are at least four reasons why a knowledge-based
that it is therefore useful to substitute rules when the state approach might perform worse than the best rules used
of the system changes. individually:
The main generalized drawbacks to these simulation-
based systems are the following: 1. The training set is a subset of the universe of all pos-
sible cases. However, situations in which the sched-
1. The time required to examine the performance of the uling system does not work properly can always be
set of candidate rules, which can make real-time sched- observed and added as training examples.
uling difficult. 2. The system’s performance depends on the number and
range of control attributes taken into account in the
2. Very frequent changes in the system. As the evalua-
design of the training examples.
tion of each of the rules is carried out until the end of
the period considered, there may not be a match be- 3. A rule may perform well in a simulation over a long
tween the rule that is proposed and the one that is time period for a set of given attributes, but will per-
really required, as the one that is chosen is used for a form poorly when applied dynamically.
period of time that is less than the time used during its 4. The system can be prone to inadequate generaliza-
evaluation. tions in extremely imprecise situations.
3. Methods to avoid unnecessary modifications of the An overview of a knowledge-based scheduling system is
dispatching rules during transitory changes are not shown in Figure 2. The examples generator uses a simula-
available. tion model to generate different manufacturing system states
4. No knowledge is acquired about the system. and search for the best dispatching rule for that state. The
training examples that the machine learning module needs
5. There are no methods to determine a reasonable sim-
are generated by an information processor, based on the
ulation window.
simulation results. The machine learning module acquires
6. The scheduling interval must be defined at machine, the knowledge that is necessary to make future scheduling
not system, level ~Chiu & Yih, 1995!. decisions by using the training cases. The knowledge may
need to be refined, depending on the manufacturing sys-
tem’s performance, by generating further training exam-
4. KNOWLEDGE-BASED SYSTEMS ples. The remaining elements of the system have similar
A real-time scheduling system that modifies dispatching functions to those described in Figure 1.
rules dynamically should fulfil two contradictory character- Several knowledge-based approaches that dynamically
istics to work adequately ~Nakasuka & Yoshida, 1992!: modify the dispatching rule being used at a specific in-
stance are reviewed next. According to the type of machine
1. Rule selection must contemplate a variety of informa- learning algorithm used, these approaches can be divided
tion about the manufacturing system in real time. into the following categories:

2. Rule selection must be completed in such a short time 1. Approaches that do not use knowledge-acquisition
that real operations are not delayed. algorithms.
FMS scheduling by machine learning 255

Fig. 2. General overview of a knowledge-based scheduling system.

2. Inductive learning-based approaches. system was not acquired by any machine learning proce-
3. Neural network-based approaches. dure, but rather by inspection carried out directly on the
simulation results. The authors observe that the manufac-
4. Mixed approaches. Here a combination of different turing system increases the number of parts produced in
types of learning algorithms is applied. percentages ranging between 7% and 30%.
5. Other approaches based on machine learning Sarin and Salgame ~1990! define an expert system to
algorithms. schedule dynamically. At the beginning of a given time
period, the system has a known schedule of jobs that is
followed during this period. This system reacts when a
4.1. Approaches that do not use
change occurs. Changes are classified into different groups:
knowledge-acquisition algorithms
machine breakdown, rush jobs, new batch of jobs, material
Thesen and Lei ~1986! propose an expert system for sched- shortage, labor absenteeism, job completion at a machine,
uling robots in a flexible electroplating system. The authors and change in shift. The system they propose has the fol-
carry out a series of a priori simulations using different lowing parts to it: a scheduling knowledge, a global data-
dispatching rules to study the performance of the manufac- base, a user interface, and a control block. The knowledge
turing system in different situations; 38 training examples is divided into several groups, each of which has rules to
were acquired in the process. However, knowledge of the solve different types of problems depending on the changes
256 P. Priore et al.

that occur in the system. The rules represent the heuristics The authors compare the algorithm that is proposed, using
of a human expert. mean flow time and mean tardiness as performance criteria,
The global database has information on the different jobs with the two best dispatching rule combinations for ma-
and shifts that exist at a specific point in time. The group of chines and AGVs. To do this they proposed different sce-
rules that needs information goes to the global database. narios, varying the load level, the queue capacity, F ~flow
Finally, the control block, which has the form of a tree allowance; Baker, 1984!, the type of processing time distri-
~“meta-rules” or “knowledge about knowledge”!, chooses bution, and the performance criterion. The algorithm was
the group of rules that fits the new problem that originated found to perform better than the best rule combinations
the change in the system. when machine load is high ~if it is low there are hardly
The same authors likewise present an integrated system parts in the queue so the rule selected has no influence! and
made up of two modules. The first, which is backed up by queue size is small. In such conditions an improvement of
mathematical programming, determines a predictive sched- over 12% is achieved.
ule as a starting point. The second, the expert system, takes Pierreval and Mebarki ~1997! introduce a heuristic meth-
over control of executing dynamic or reactive scheduling as odology, called SFSR ~shift from standard rules!, to dynam-
the new situation demands, whenever a change occurs. Fi- ically modify the dispatching rules according to two
nally, they point out that this integrated approach has still performance criteria ~one primary and the other second-
not been implemented in a real case. ary!. The SFSR heuristic checks the manufacturing sys-
Chandra and Talavage ~1991! present a system called tem’s state when a resource becomes available or a new job
EXPERT, made up of a set of decision rules. The informa- arrives. By using rules that were defined beforehand that
tion that is used in the decision process is the congestion are functions of parameters to be optimized, the presence of
level of the manufacturing system, the preference of a part certain symptoms in the manufacturing system can be de-
for a machine, how critical the part is ~it indicates the part’s tected ~for example, when the system is congested or when
ability to meet its due date!, and the objective of the man- there is a job which is waiting too long, etc.!. The optimal
ufacturing system at that specific moment. The authors state values of the parameters are calculated by the Hooke-
that in principle the aim of maximizing the work progress Jeeves method ~Hooke & Jeeves, 1961!. If there are no
rate is an interesting one, even though there is the risk that symptoms in the system, standard rules taken from the lit-
some jobs are delayed ~especially if the system is over- erature depending on the criterion to be optimized are used.
loaded and there are a lot of critical jobs!. Excessive pre- The dispatching rule to be applied can thus be calculated.
occupation with critical jobs can make the system worse; If the opposite is the case, rules defined by the authors
for this reason the objective of maximizing the work progress are used; these depend on the criteria to be optimized, on
rate is chosen as the first criterion. the symptom detected, and on the state of the system. In
Jobs are furthermore divided into groups ~high, medium, general, the methodology that is proposed improves upon
and low preference! instead of them being classified indi- the alternative of using a rule constantly according to the
vidually. The system that is proposed selects the job as- primary criterion; if this is not the case, it compensates with
signed to a machine, beginning with the high preference the secondary criterion. The improvements vary between
ones, pursuing the primary objective whilst also searching 12.3% and 33.8%. The greatest defect of the methodology
for opportunities to improve the secondary one ~minimiz- is that the standard rules are defined according to research
ing the number of tardy jobs!. In certain cases, jobs avail- results already presented in the literature. An alternative
able in the near future are inspected. If there is a tie, or approach would be for them to be generated by inductive
when a clear decision is not taken, then the shortest pro- learning, so as to take account of the peculiarities of the
cessing time ~STP! rule is applied. At the experimental stage, system under study. Inductive learning could also be used
it was shown that the EXPERT system is superior to con- to generate other types of rules used in SFSR.
ventional dispatching rules.
Sabuncuoglu and Hommertzheim ~1992! suggest a dy-
4.2. Inductive learning-based approaches
namic algorithm to schedule jobs in machines and auto-
matic guided vehicles ~AGVs!. The algorithm they propose Pierreval and Ralambondrainy ~1990! suggest an inductive
is based on the idea that a job should not be assigned to a learning algorithm called GENREG, to obtain heuristic rules
machine if it has to wait for an AGV in the following oper- to know system performance with different dispatching rules
ation and vice versa. It uses several priority schemes ~or and states of the manufacturing system. In the methodology
rules! and information about the system ~queue levels, the that they propose, a rule is obtained with each training ex-
number of parts in the system, machine state, etc.! and about ample. As the number of rules is extremely large, GENREG
the jobs ~processing times, number of operations, etc.!. The is then used to generalize them, and thereby reduce this
algorithm has two fundamental parts: a set of procedures number. The approach they propose is applied in a simpli-
for scheduling jobs in the machines, and others to schedule fied flow-shop configuration with two machines, using 198
the jobs in the AGVs. The latter check whether there are training examples. However, the rules obtained using GEN-
blocked or empty stations or parts in the central buffer. REG are not used dynamically.
FMS scheduling by machine learning 257

Shaw et al. ~1992! present a system called PDS ~Pattern- 4.3. Neural network-based approaches
Directed Scheduling! to schedule jobs in an FMS that uses
inductive learning. Here, the learning algorithm that is used Chen and Yih ~1996! define an approach to determine the
is ID3. During the knowledge-acquisition stage 130 train- most important attributes as a first step to constructing
ing examples are used, applying mean tardiness as perfor- knowledge-based scheduling systems. The approach for iden-
mance criterion. This system provides a mean tardiness tifying attributes has three steps to it:
reduction of 11.5%.
1. Data collection via manufacturing system simulation.
The authors observe that the maximum effectiveness of
the approach is obtained when the number of changes in the 2. Building of attributes0performance mapping func-
manufacturing system’s states ~patterns! is between me- tions of the manufacturing system using back-
dium and reasonably high. Moreover, the number of alter- propagation neural networks.
native machines to process a given operation does not need 3. Selection of essential attributes.
to be very high. The authors confirm that all these charac-
teristics occur in most real FMSs. To do this, an attribute is omitted and the difference be-
Nakasuka and Yoshida ~1992! propose a scheduling tween the original output and the output obtained with the
scheme called LADS ~Learning-Aided Dynamic Sched- attribute omitted is measured. More important attributes
uler! that incorporates an inductive learning algorithm with have a greater difference than those with a lower signifi-
two characteristics that differentiate it from conventional cance level. Twenty candidate attributes taken from the lit-
algorithms. Firstly, there is a new criterion to decide how to erature ~Nakasuka & Yoshida, 1992; Cho & Wysk, 1993;
separate data groups; owing to data noise in scheduling Chiu, 1994! are used in the experimental study ~using
problems, dividing them so that they belong to a single attribute variability defined as the variance of the attribute
class is of no interest, as the number of data in each group divided by its mean!, along with six performance measure-
would be very small. The second characteristic of the algo- ments and ten dispatching rules. One thousand three hun-
rithm that is proposed is the generation of linear combina- dred training examples are generated for each type of rule
tions of attributes fed into the system. The scheduling scheme and the corresponding neural network, and the ten most
being proposed is used in a simplified flow-shop system important attributes are selected.
with three machines so as to minimize the makespan and Finally, the authors compare three neural networks with
keep mean tardiness below a set level. The authors point input nodes formed by alternative groups of attributes ~with
out that the system is superior ~by both criteria! to using the ten selected attributes, with the first twenty, and with
one rule constantly. the other ten! and output nodes corresponding to the rules
Piramuthu et al. ~1993! define an approach to scheduling that are selected. There is verification of the fact that the
jobs dynamically. The approach they propose, along with network’s capacity for generalization obtained with the ten
the examples they apply it to, is similar to the one the au- significant attributes is 9% superior with respect to the net-
thors present in others works ~see for example, Shaw et al., work formed by the first twenty attributes. The authors claim
1992; Piramuthu et al., 1994!. What this paper does con- that using as many attributes as possible to build up a knowl-
tribute, compared with others that the same authors have edge base does not improve generalization or prediction
published, is a more elaborate theoretical framework for capacity. The more attributes that are included, the greater
each of the parts that make up the job scheduling system. the effort required to develop the knowledge base, and the
Piramuthu et al. ~1994! define a methodology to sched- more complex its structure becomes. The main defects of
ule jobs using inductive learning in a flexible flow-shop the approach being proposed are that it does not identify
manufacturing system with mean flow time as the perfor- important attributes if they are not considered initially and
mance criterion. By using C4.5 as the learning algorithm, the process must be repeated if performance measurements
two decision trees are generated via 66 training examples. change.
The first is to schedule jobs in the machines themselves, Sun and Yih ~1996! apply an approach that uses a back-
and the second is for part-release decision. They also present propagation neural network on each machine to select the
a refinement procedure for the decision trees, which con- most adequate dispatching rule in a multiple criterion envi-
sists of including cases that the system misclassifies in the ronment. At each point of decision, when a machine has to
training set. select a new job, an adjustment module determines the rel-
The authors observe that incorporating a decision tree to ative importance of each performance criterion as a func-
select the dispatching rule does not improve results signif- tion of the desired and current values. Taking the value
icantly with respect to the alternative of using the decision provided by the adjustment module and the current state of
tree only for part-release decision using a dispatching rule the machine as the input value, the neural network provides
constantly. Moreover, this methodology is particularly use- the most adequate dispatching rule.
ful when input buffer size is limited and small, and there is The authors used around 1000 training examples for each
a great variation in processing times for parts in the bottle- neural network, and show that the proposed approach has
neck machines. an average performance 4.2% better than the best rules used
258 P. Priore et al.

constantly. Moreover, it is very adaptable to changes in the The intelligent scheduling module is activated when a
choice of performance criteria that are given priority. The new job arrives, or if there is an anomaly in the system. The
greatest defect of this approach is that the manufacturing task of the simulator is to examine the performance of the
system being studied does not have flexible routes and the dispatching rules suggested by the intelligent scheduling
number of parts is limited. module and to select the best. Finally, the control module
Min et al. ~1998! propose a methodology that uses com- allows scheduling in the physical cell. Moreover, most in-
petitive neural networks. Here, differences between the val- formation about the cell, which is essential to control it, is
ues of performance criteria and system state variables in obtained and manipulated through this module. The authors
different time intervals are used as attributes. By simulating point out that the proposed scheme produces a performance
the system for a long time period and modifying the dis- improvement in the manufacturing system of between 2.3%
patching rules randomly, the training examples are ob- and 29.3%, when compared to the alternative of using a
tained with these differences. Three thousand five hundred dispatching rule constantly.
examples are used in the training stage and 40 network Rabelo and Alptekin ~1989! define an approach called
nodes or classes are defined. Neural networks are then used ISS0FMS ~Intelligent Scheduling System for FMS! to sched-
to obtain classes from the training examples. ule jobs, made up of three basic modules. The first of them
Their scheduling system works in real time in the follow- is an expert system which decides the heuristic rule to use
ing way. First, the user sets differences as an objective and based on certain information ~data on work to be done,
the class is identified using the neural network. Then, from constraints imposed by the workshop, cell state, etc.!. It
amongst all the training examples, those that have the same takes into account data provided by a neural network and a
class and the same decision variables ~dispatching rules! statistical analysis model that study past cases obtained via
from the previous interval are sought. If this example is not a simulation study. The second module carries out a heuris-
found, which is the likely case, the most used rule within tic process ~Kiran & Alptekin, 1989! that depends on two
this class is chosen for each decision variable. coefficients determined by a neural network as a function
An interesting characteristic of this approach is that it of the characteristics of the scheduling problem being solved.
uses earlier dispatching rules to find the new ones. The The third component chooses the best of the solutions cal-
system proposed is compared to another one that chooses culated by the previous two modules.
rules randomly ~ten replicas of the random system are in Cho and Wysk ~1993! present a system called IWC ~In-
fact made and the best from amongst them is chosen! and it telligent Workstation Controller! that uses neural networks
is shown to have superior performance. The drawbacks of and a simulator, based on the work of Wu and Wysk ~1989!.
the approach are the lack of a method to systematically The neural network has seven input nodes, corresponding
search for an optimum number of output nodes to the neural to the state of the system, and nine output nodes, one for
network, and that it is compared to a random system rather each of the dispatching rules considered. The network is
than the best possible combination of the proposed dispatch- trained with 90 examples taken from the literature, taking
ing rules. into account different configurations of the hidden layers
and different learning rates.
Using the two best rules provided by the network, the
4.4. Mixed approaches simulator selects the better of them as a function of the state
of the manufacturing system. Moreover, the authors exper-
Wu and Wysk ~1988! propose a control and scheduling imentally calculate the most adequate simulation window
scheme called MPECS ~Multipass Expert Control System! for the performance criterion chosen using a set of simula-
that combines expert systems, simulation, and inductive tions. It is observed that IWC is superior to the use of one
learning. The system they propose has three modules: an rule constantly, although percentage improvement never goes
intelligent scheduling module, a manufacturing system sim- beyond 3%.
ulator, and a cell control module. The first element is in turn Li and She ~1994! use an approach that utilizes cluster
made up of a knowledge base, an inference engine, and a analysis ~Evert, 1980! and inductive learning. Six hundred
learning module. examples are uniformly generated from the decision spec-
The base has declarative knowledge ~information about trum. By using cluster analysis, seven classes with similar
the system state, scheduling heuristics, and rules!, and pro- performance values are established. Then an algorithm sim-
cedural information ~general criteria, in rule form, to select ilar to C4.5 establishes two sets of rules that determine the
dispatching rules!. The inference engine is a search mech- class as a function of the decision and performance attributes.
anism to select the right rules of procedural knowledge. One way that the authors suggest of using this scheduling
Finally, the learning module generates a set of rules from knowledge is to set performance conditions and determine
the training examples that associate dispatching rules, per- the class they correspond to. Once the class is known, the
formance measurements, and the system’s characteristics. decision variables that will be taken are determined using
The rules created by the learning module are sent to proce- the other set of rules. However, this methodology is not
dural knowledge. compared with any other to check how it works.
FMS scheduling by machine learning 259

Chiu and Yih ~1995! put forward a system that uses in- incomplete data, and have the lowest level of test error
ductive learning and genetic algorithms. The latter are used ~1.2%!.
to search for a set of good-quality training examples. To do Bowden and Bullington ~1996! suggest a scheme called
this, at each point of decision the best dispatching rule is GARDS ~Genetic Algorithm Rule Discovery System! to
chosen and, this rule forms a training case along with the determine control strategies using genetic algorithms.
state of the system. Furthermore, the learning algorithm GARDS has three fundamental modules:
can modify the decision tree when new examples are pre-
sented only if the change is significant. The authors show 1. A simulation model to analyze the performance of the
that the proposed approach is superior to using a dispatch- different strategies that are generated.
ing rule constantly. The greatest defect of the approach lies 2. An algorithm that determines the most adequate rule
in the need to change the induced scheduling knowledge within a strategy or plan for the current state of the
when there are small modifications in the manufacturing manufacturing system.
system.
3. A genetic algorithm that uses traditional crossover and
Lee et al. ~1997! propose a scheme that also uses induc-
mutation operators to improve the initial plans by
tive learning and genetic algorithms. The first technique
choosing the best for the control system.
serves to generate a decision tree using C4.5 to select the
best rule to control the input flow of jobs to the system. The
The system that is proposed is tested on two configura-
genetic algorithms are used to select the most appropriate
tions of different complexity with the aim of minimizing
dispatching rules for each of the system’s machines. The
the number of tardy jobs. It was observed that GARDS
authors verify the approach they propose with two job shop
improves the performance of manufacturing systems with
systems ~one of them with a bottleneck machine!, using
respect to several classical heuristic methods ~e.g., sending
mean tardiness as performance criterion, and show that it
jobs to the queue of the machine with fewest jobs!.
beats the best combination of rules used constantly, at a rate
Table 1 recapitulates and summarizes the different ap-
of between 20.34% and 25.28%. However, times required
proaches to be found in the literature, classified by the meth-
~26 and 168 min for the first and second cases, respec-
odology used. Table 2 shows a collection of different
tively! are rather high for this system to work in real time.
approaches that dynamically modify dispatching rules, clas-
Kim et al. ~1998! suggest a scheme that broadens the
sified by the type of machine learning algorithm used.
scope of earlier work ~Min et al., 1998! and uses competi-
tive neural networks and inductive learning. Once the classes
are obtained from the neural networks, inductive learning is 5. LIMITATIONS OF THE SCHEDULING
applied to express knowledge in tree form and production APPROACHES AND FUTURE RESEARCH
rules. The authors use 99.999 training cases and establish a DIRECTIONS
network with 100 groups or classes.
The scheduling system works in real time in the same A number of limitations that would be desirable character-
way as the previously described approach ~Min et al., 1998!. istics can be detected across the board in the knowledge-
The only difference is that the class is identified by the based approaches using machine learning algorithms that
production rules obtained from the C4.5 inductive learning have been considered above. These limitations point to fu-
algorithm. The authors compare this system with another ture directions of research in the field of dynamic schedul-
that only uses a competitive neural network, and demon- ing of manufacturing systems, by modifying the dispatching
strate its superiority, due to the C4.5’s tree pruning algo- rule that is employed. Future research directions would in-
rithm, which deals with noise in the data more efficiently. clude the following:
This scheme displays the same defects as the approach sug-
gested in Min et al. ~1998!. 1. Comparison of the different machine learning meth-
odologies. The approaches described in the literature
employ a methodology or, in certain cases, a combi-
4.5. Other approaches based on nation of methodologies. However, there is no com-
machine learning algorithms parative study that determines which of them is the
best. Furthermore, because of the wide range and dis-
Quiroga and Rabelo ~1995! solve the problem of schedul- parity of the FMSs used in the literature that has been
ing jobs for a machine by inductive learning ~ID3!, back- reviewed, it is not possible to have even an inkling of
propagation neural networks, and fuzzy logic. They use 358 which of the methodologies described is the most ad-
training cases and 198 test cases, and the test error is lower equate for resolving this type of scheduling problem.
than 10% in the three methodologies. Inductive learning 2. The use of CBR ~Case-Based Reasoning! as a ma-
and fuzzy logic have the advantage of generating rules that chine learning methodology in scheduling systems.
are intelligible to humans, which is not the case for neural These algorithms are very efficient at classification,
networks. However, the latter are less sensitive to noise or despite their simplicity ~Rachlin et al., 1994!. Yet none
260 P. Priore et al.

Table 1. Classification of references according Table 2. Classification of references according

to methodology applied to the machine learning algorithm applied

Methodology References Machine Learning

Algorithm References
Han et al. ~1989!; Hutchison et al. ~1989!;
Kimemia & Gershwin ~1985!; Lashkari et al. Chandra & Talavage ~1991!; Pierreval & Mebarki
Analytical approach ~1987!; Shanker & Rajamarthandan ~1989!; Is not used ~1997!; Sabuncuoglu & Hommertzheim ~1992!;
Shanker & Tzen ~1985!; Stecke ~1983!; Wilson Sarin & Salgame ~1990!; Thesen & Lei ~1986!.
~1989!.
Nakasuka & Yoshida ~1992!; Pierreval &
Choi & Malstrom ~1988!; Denzler & Boe ~1987!; Inductive learning Ralambondrainy ~1990!; Piramuthu et al. ~1993!;
Egbelu & Tanchoco ~1984!; Henneke & Choi Piramuthu et al. ~1994!; Shaw et al. ~1992!.
Heuristic approach
~1990!; Montazeri & Van Wassenhove ~1990!;
Chen & Yih ~1996!; Min et al. ~1998!; Sun & Yih
Stecke & Solberg ~1981!; Tang et al. ~1993!. Neural Networks
~1996!.
Simulation-based Ishii & Talavage ~1991!; Jeong & Kim ~1998!;
Chiu & Yih ~1995!; Cho & Wysk ~1993!; Kim
approach Kim & Kim ~1994!; Wu & Wysk ~1989!.
Mixed et al. ~1998!; Lee et al. ~1997!; Li & She ~1994!;
Bowden & Bullington ~1996!; Chandra & Rabelo & Alptekin ~1989!; Wu & Wysk ~1988!.
Talavage ~1991!; Chaturvedi et al. ~1993!; Chen
Bowden & Bullington ~1996!; Quiroga & Rabelo
& Yih ~1996!; Chiu & Yih ~1995!; Cho & Wysk Others
~1995!.
~1993!; De & Lee ~1998!; Dong & Kitaoka
~1994!; Fox & Smith ~1984!; Kim et al. ~1998!;
Lee et al. ~1997!; Li & She ~1994!; Maimon
~1987!; Maley et al. ~1988!; Min et al. ~1998!;
Artificial
Nakasuka & Yoshida ~1992!; Pierreval &
intelligence-based
Mebarki ~1997!; Pierreval & Ralambondrainy
approach
~1990!; Piramuthu et al. ~1993!; Piramuthu et al. 5. Determination of a mechanism or filter to smooth tran-
~1994!; Quiroga & Rabelo ~1995!; Rabelo & sitory states. On certain occasions the manufacturing
Alptekin ~1989!; Sabuncuoglu & Hommertzheim
system running on “scheduling knowledge” does not
~1992!; Sarin & Salgame ~1990!; Shaw et al.
~1992!; Shaw & Whinston ~1989!; Shen & perform as it is expected to, and is worse than the
Chang ~1988!; Sun & Yih ~1996!; Thesen & Lei alternative of using the best combination of dispatch-
~1986!; Wu & Wysk ~1988!. ing rules constantly. This phenomenon is explained
by the fact that the system reacts hastily to changes in
control attributes that are only transitory in time. Thus,
one proposal would be to use digital filters to smooth
transitory scenarios in control attributes. This mecha-
of the systems reviewed uses CBR. It would therefore
nism is not considered in most of the approaches that
be interesting to test how good they are for scheduling
have been reviewed, or when it is considered, neither
problems.
the different kinds of digital filters available nor their
3. Determination of the optimum number of training ex- interrelationship with the monitoring period is
amples. None of the approaches reviewed calculate analyzed.
the number of examples required to optimally train
the machine learning algorithm. Nor do they specify 6. Generation of new control attributes using an algo-
whether the test examples are the same, similar or rithm that can create attributes that are a combination
very different to the training examples. Yet classifica- of the initial ones. In some cases, it is necessary to
tion error of “scheduling knowledge”, and therefore check relationships of the following type in order to
the performance of a manufacturing system, depends select the best dispatching rule: use of machine 1 is
to a great extent on the number of training examples less than that of machine 2. To achieve these relation-
that is considered. It is therefore necessary to study ships, the arithmetical combinations of the basic ini-
classification error as a function of the number of ex- tial attributes would need to be defined. However, these
amples considered, and an adequate size of the train- combinations are often not known at the outset, and
ing set must be chosen. can only be discovered in simple manufacturing sys-
tems after detailed examination of simulation results.
4. Selection of an adequate monitoring period. A study
to determine the right monitoring period for each per- 7. Incorporation of a simulator. The performance of the
formance criterion is not generally done in the exist- scheduling system could be enhanced if a simulator
ing literature. However, the frequency of control was used to determine the best rule from amongst
attribute checking, to decide whether dispatching those that the machine learning system considered the
rules are to be changed or not, is a vitally important most important ones. On occasions, and given certain
question that determines manufacturing system’s control attribute values, “scheduling knowledge” de-
performance. termines that there are two or more dispatching rules
FMS scheduling by machine learning 261

Table 3. The characteristics of scheduling systems

Generation
Machine Learning Monitoring of New Incorporation Knowledge
References Algorithm Period Filter Attributes of a Simulator Refinement

Chandra & Talavage ~1991! No No No No No No

Pierreval & Mebarki ~1997! No No No No No No
Sabuncuoglu & Hommertzheim ~1992! No No No No No No
Sarin & Salgame ~1990! No No No No No No
Thesen & Lei ~1986! No No No No No No
Nakasuka & Yoshida ~1992! LADS No No Yes No No
Pierreval & Ralambondrainy ~1990! GENREG No No No No No
Piramuthu et al. ~1993! ID3; C4.5 No Yes No No Yes
Piramuthu et al. ~1994! C4.5 No Yes No No Yes
Shaw et al. ~1992! ID3 No Yes No No No
Chen & Yih ~1996! Backpropagation NN No No No No No
Min et al. ~1998! Competitive NN No No No No No
Sun & Yih ~1996! Backpropagation NN No No No No No
Chiu & Yih ~1995! Inductive Learning: No No No No Yes
Genetic Algorithm
Cho & Wysk ~1993! Backpropagation NN; Yes No No Yes No
Simulator
Kim et al. ~1998! C4.5; Competitive NN No No No No No
Lee et al. ~1997! C4.5; Genetic Algorithm No No No No Yes
Li & She ~1994! Inductive Learning; No No No No No
Cluster Analysis
Rabelo & Alptekin ~1989! Expert System; No No No No No
Backpropagation NN
Wu & Wysk ~1988! Expert System; Yes No No Yes Yes
Inductive Learning;
Simulator
Bowden & Bullington ~1996! Simulator; Genetic No No No Yes Yes
Algorithm
Quiroga & Rabelo ~1995! ID3; Fuzzy Logic; No No No No No
Backpropagation NN

that might in principle be the right one. In such cases, five characteristics and others include others. However, as
when “scheduling knowledge” decisions are not clear, none of the first three characteristics ~comparison of the
incorporating a simulator would be very useful. different methodologies for machine learning, use of CBR,
8. Refinement of the knowledge base. Once developed, and determination of the optimum number of training ex-
the knowledge base is not static, so it would be inter- amples! are found in any of the systems, they do not appear
esting to establish a procedure that would automati- in this table.
cally modify knowledge if important changes in the
manufacturing system occur. The main aim of the re- 6. CONCLUSIONS
finement module is to discover deficiencies in the
knowledge base and add training cases that cater to This paper provides a review of the literature on dynamic
them. These deficiencies can occur in certain ranges scheduling of FMSs using machine learning. A classifica-
of control attribute values. To solve this problem, such tion of general approaches to be found in the literature is
ranges have to be “covered” with new training cases, first provided. Then, two ways of dynamically modifying
so that the new “scheduling knowledge” obtained is dispatching rules in order to overcome their drawbacks when
able to deal with these situations. they are used statically are described. A review is then pro-
vided of the approaches available according to the machine
Table 3 provides a summary of the characteristics of each learning algorithm that is used. Next, we indicate a number
of the scheduling systems that have been reviewed above. of limitations that would be desirable characteristics, but
Only the last five characteristics ~of the eight listed above! which are lacking in the approaches we have reviewed.
are shown, as some of the systems include some of these Finally, the point is made that in future work it would be
262 P. Priore et al.

interesting for a scheduling module to be designed that in- Hooke, R., & Jeeves, T.A. ~1961!. Direct search solution of numerical and
corporates the eight characteristics that are listed, and for statistical problems. Journal of the Association of Computer Machines
8, 212–229.
the effect of each of them on the performance of scheduling Hutchison, J., Leong, K., Snyder, D., & Ward, F. ~1989!. Scheduling for
systems to be measured. random job shop flexible manufacturing systems. Proc. of the Third
ORSA0 TIMS Conf. on Flexible Manufacturing Systems, 161–166.
Ishii, N., & Talavage, J. ~1991!. A transient-based real-time scheduling
algorithm in FMS. International Journal of Production Research 29(12),
REFERENCES 2501–2520.
Jain, A.S., & Meeran, S. ~1998!. Job-shop scheduling using neural net-
Aytug, H., Bhattacharyya, S., Koehler, G.J., & Snowdon, J.L. ~1994!. A works. International Journal of Production Research 36(5), 1249–
review of machine learning in scheduling. IEEE Transactions on En- 1272.
gineering Management 41(2), 165–171. Jeong, K.-C., & Kim, Y.-D. ~1998!. A real-time scheduling mechanism
Baker, K.R. ~1984!. Sequencing rules and due-date assignments in a job for a flexible manufacturing system: Using simulation and dis-
shop. Management Science 30(9), 1093–1103. patching rules. International Journal of Production Research 36(9),
Basnet, C., & Mize, J.H. ~1994!. Scheduling and control of flexible man- 2609–2626.
ufacturing systems: A critical review. International Journal of Com- Kanet, J.J., & Adelsberger, H.H. ~1987!. Expert systems in production
puter Integrated Manufacturing 7(6), 340–355. scheduling. European Journal of Operational Research 29, 51–59.
Blackstone, J.H., Phillips, D.T., & Hogg, G.L. ~1982!. A state-of-the-art Kim, Y.-D. ~1990!. A comparison of dispatching rules for job shops with
survey of dispatching rules for manufacturing job shop operations. multiple identical jobs and alternative routings. International Journal
International Journal of Production Research 20(1), 27– 45. of Production Research 28(5), 953–962.
Bowden, R., & Bullington, S.F. ~1996!. Development of manufacturing Kim, M.H., & Kim, Y.-D. ~1994!. Simulation-based real-time scheduling
control strategies using unsupervised machine learning. IIE Transac- in a flexible manufacturing system. Journal of Manufacturing Systems
tions 28, 319–331. 13(2), 85–93.
Chandra, J., & Talavage, J. ~1991!. Intelligent dispatching for flexible man- Kim, C.-O., Min, H.-S., & Yih, Y. ~1998!. Integration of inductive learning
ufacturing. International Journal of Production Research 29(11), and neural networks for multi-objective FMS scheduling. Inter-
2259–2278. national Journal of Production Research 36(9), 2497–2509.
Chaturvedi, A.R., Hutchinson, G.K., & Nazareth, D.L. ~1993!. Supporting Kimemia, J., & Gershwin, S.B. ~1985!. Flow optimization in flexible man-
complex real-time decision making through machine learning. Deci- ufacturing systems. International Journal of Production Research 23,
sion Support Systems 10, 213–233. 81–96.
Chen, C.C., & Yih, Y. ~1996!. Identifying attributes for knowledge-based Kiran, A., & Alptekin, S. ~1989!. A tardiness heuristic for scheduling flex-
development in dynamic scheduling environments. International Jour- ible manufacturing systems. Fifteenth Conf. on Production Research
nal of Production Research 34(6), 1739–1755. and Technology: Advances in Manufacturing Systems Integration and
Chiu, C. ~1994!. A learning-based methodology for dynamic scheduling in Processes, 559–564.
distributed manufacturing systems. PhD Thesis. School of Industrial Kouvelis, P. ~1992!. Design and planning problems in flexible manufac-
Engineering, Purdue University, West Lafayette, Indiana. turing systems: A critical review. Journal of Intelligent Manufacturing
Chiu, C., & Yih, Y. ~1995!. A learning-based methodology for dynamic 3, 75–99.
scheduling in distributed manufacturing systems. International Jour- Kusiak, A., & Chen, M. ~1988!. Expert systems for planning and sched-
nal of Production Research 33(11), 3217–3232. uling manufacturing systems. European Journal of Operational Re-
Cho, H., & Wysk, R.A. ~1993!. A robust adaptive scheduler for an intelli- search 34, 113–130.
gent workstation controller. International Journal of Production Re- Lashkari, R.S., Dutta, S.P., & Padhye, A.M. ~1987!. A new formulation of
search 31(4), 771–789. operation allocation problem in flexible manufacturing systems: Math-
Choi, R.H., & Malstrom, E.M. ~1988!. Evaluation of traditional work ematical modelling and computational experience. International Jour-
scheduling rules in a flexible manufacturing system with a physical nal of Production Research 25, 1267–1283.
simulator. Journal of Manufacturing Systems 7(1), 33– 45. Lee, C.-Y., Piramuthu, S., & Tsai, Y.-K. ~1997!. Job shop scheduling with
De, S., & Lee, A. ~1998!. Towards a knowledge-based scheduling system a genetic algorithm and machine learning. International Journal of
for semiconductor testing. International Journal of Production Re- Production Research 35(4), 1171–1191.
search 36(4), 1045–1073. Li, D.-C., & She, I.-S. ~1994!. Using unsupervised learning technologies
Denzler, D.R., & Boe, W.J. ~1987!. Experimental investigation of flexible to induce scheduling knowledge for FMSs. International Journal of
manufacturing system scheduling rules. International Journal of Pro- Production Research 32(9), 2187–2199.
duction Research 25(7), 979–994. Maimon, O.Z. ~1987!. Real-time operational control of flexible manufac-
Dong, Y., & Kitaoka, M. ~1994!. Implementation of a case-based produc- turing systems. Journal of Manufacturing Systems 6(2), 125–136.
tion scheduling system in C language. Computers & Industrial Engi- Maley, J.G., Ruiz-Mier, S., & Solberg, J.J. ~1988!. Dynamic control in
neering 27(7), 265–268. automated manufacturing: A knowledge integrated approach. Inter-
Egbelu, P.J., & Tanchoco, J.M.A. ~1984!. Characterization of automated national Journal of Production Research 26(11), 1739–1748.
guided vehicle dispatching rules. International Journal of Production Min, H.-S., Yih, Y., & Kim, C.-O. ~1998!. A competitive neural network
Research 22(3), 359–374. approach to multi-objective FMS scheduling. International Journal of
Evert, B. ~1980!. Cluster Analysis. Heinemann, New York. Production Research 36(7), 1749–1765.
Fox, M.S., & Smith, S.F. ~1984!. ISIS-A knowledge-based system for Minton, S. ~1993!. Machine Learning Methods for Planning. Morgan Kauf-
factory scheduling. Expert Systems 1(1), 25– 49. mann, San Mateo, California.
Garey, M., & Johnson, D. ~1979!: Computers and Intractability: A Guide Montazeri, M., & Van Wassenhove, L.N.V. ~1990!. Analysis of scheduling
to the Theory of NP-Completeness. Freeman, New York. rules for an FMS. International Journal of Production Research 28(4),
Gunasekaran, A., Martikainen, T., & Yli-Olli, P. ~1993!. Flexible manu- 785–802.
facturing systems: An investigation for research and applications. Eu- Nakasuka, S., & Yoshida, T. ~1992!. Dynamic scheduling system utilizing
ropean Journal of Operational Research 66, 1–26. machine learning as a knowledge acquisition tool. International Jour-
Han, M., Na, Y.K., & Hogg, G.L. ~1989!. Real-time tool control and job nal of Production Research 30(2), 411– 431.
dispatching in flexible manufacturing systems. International Journal Panwalkar, S.S., & Iskander, W. ~1977!. A survey of scheduling rules.
of Production Research 27, 1257–1267. Operations Research 23(5), 961–973.
Harmonosky, C.M., & Robohn, S.F. ~1991!. Real-time scheduling in com- Pierreval, H., & Mebarki, N. ~1997!. Dynamic selection of dispatching
puter integrated manufacturing: A review of recent research. Inter- rules for manufacturing system scheduling. International Journal of
national Journal of Computer Integrated Manufacturing 4(6), 331–340. Production Research 35(6), 1575–1591.
Henneke, M.J., & Choi, R.H. ~1990!. Evaluation of FMS parameters on Pierreval, H., & Ralambondrainy, H. ~1990!. A simulation and learning
overall system performance. Computer Industrial Engineering 18(1), technique for generating knowledge about manufacturing systems be-
105–110. haviour. Journal of the Operational Research Society 41(6), 461– 474.
FMS scheduling by machine learning 263

Piramuthu, S., Raman, N., & Shaw, M.J. ~1994!. Learning-based schedul- Vepsalainen, A.P.J., & Morton, T.E. ~1987!. Priority rules for job shops
ing in a flexible manufacturing flow line. IEEE Transactions on Engi- with weighted tardiness costs. Management Science 33(8), 1035–1047.
neering Management 41(2), 172–182. Wilson, J.M. ~1989!. An alternative formulation of the operation-allocation
Piramuthu, S., Raman, N., Shaw, M.J., & Park, S. ~1993!. Integration of problem in flexible manufacturing systems. International Journal of
simulation modeling and inductive learning in an adaptive decision Production Research 27, 1405–1412.
support system. Decision Support Systems 9, 127–142. Wu, S.-Y.D., & Wysk, R.A. ~1988!. Multi-pass expert control system—a
Quiroga, L.A., & Rabelo, L.C. ~1995!. Learning from examples: A review control0scheduling structure for flexible manufacturing cells. Journal
of machine learning, neural networks and fuzzy logic paradigms. Com- of Manufacturing Systems 7(2), 107–120.
puters & Industrial Engineering 29, 561–565. Wu, S.-Y.D., & Wysk, R.A. ~1989!. An application of discrete-event sim-
Rabelo, L.C., & Alptekin, S. ~1989!. Integrating scheduling and control ulation to on-line control and scheduling in flexible manufacturing.
functions in computer integrated manufacturing using artificial intel- International Journal of Production Research 27(9), 1603–1623.
ligence. Computers & Industrial Engineering 17, 101–106. Zweben, M., & Fox, M.S. ~1994!. Intelligent Scheduling. Morgan Kauf-
Rachlin, J., Kasif, S., Salzberg, S., & Aha, D.W. ~1994!. Towards a better mann, San Mateo, California.
understanding of memory-based reasoning systems. Proc. of the Elev-
enth International Conf. on Machine Learning, 242–250.
Ramasesh, R. ~1990!. Dynamic job shop scheduling: A survey of simula-
tion studies. OMEGA: The International Journal of Management Sci-
ence 18(1), 43–57. Paolo Priore majored in industrial engineering at Univer-
Russel, R.S., Dar-El, E.M., & Taylor, B.W. ~1987!. A comparative analysis sity of Oviedo, Spain and graduated in 1991. He gained his
of the COVERT job sequencing rule using various shop performance
measures. International Journal of Production Research 25(10), Ph.D. from the University of Oviedo in 2001. He is cur-
1523–1540. rently an Assistant Professor at the School of Industrial
Sabuncuoglu, I., & Hommertzheim, D.L. ~1992!. Dynamic dispatching Engineering at Oviedo University. His research interests
algorithm for scheduling machines and automated guided vehicles in a
flexible manufacturing system. International Journal of Production include machine learning applications in production prob-
Research 30(5), 1059–1079. lems, simulation, and manufacturing. Dr. Priore is a mem-
Sarin, S.C., & Salgame, R.R. ~1990!. Development of a knowledge-based ber of EUROMA ~European Operations Management
system for dynamic scheduling. International Journal of Production
Research 28(8), 1499–1512. Association!.
Shanker, K., & Rajamarthandan, S. ~1989!. Loading problem in FMS: Part
movement minimization. Proc. of the Third ORSA0 TIMS Conf. on David de la Fuente majored in industrial engineering at
Flexible Manufacturing Systems, 99–104.
Shanker, K., & Tzen, Y.J. ~1985!. A loading and dispatching problem in a University of Madrid, Spain and graduated in 1983. He
random flexible manufacturing system. International Journal of Pro- gained his Ph.D. from the University of Oviedo in 1987.
duction Research 23, 579–595. Currently, he is a Professor at the School of Industrial Engi-
Shaw, M.J., Park, S., & Raman, N. ~1992!. Intelligent scheduling with
machine learning capabilities: The induction of scheduling knowl- neering at Oviedo Univesity. His present research interests
edge. IIE Transactions 24(2), 156–168. are in the areas of artificial intelligence, forecasting, and
Shaw, M.J., & Whinston, A.B. ~1989!. An artificial intelligence approach manufacturing. Dr. de la Fuente is a member of APICS.
to the scheduling of flexible manufacturing systems. IIE Transactions
21(2), 170–183.
Shen, S., & Chang, Y.-L. ~1988!. Schedule generation in a flexible manu- Alberto Gomez graduated from the University of Oviedo,
facturing system: A knowledge-based approach. Decision Support Sys- Spain in 1994, where he majored in the field of industrial
tems 4, 157–166.
Stecke, K.E. ~1983!. Formulation and solution of nonlinear integer pro- engineering. He is currently an Assistant Professor at the
duction planning problems for flexible manufacturing systems. Man- School of Industrial Engineering at Oviedo University. His
agement Science 29(3), 273–288. research interests include applications of genetic algo-
Stecke, K.E., & Solberg, J. ~1981!. Loading and control policies for a
flexible manufacturing system. International Journal of Production rithms in production problems, and information systems.
Research 19(5), 481– 490.
Sun, Y.-L., & Yih, Y. ~1996!. An intelligent controller for manufacturing Javier Puente graduated in industrial engineering in 1994
cells. International Journal of Production Research 34(8), 2353–2373.
Tang, L.-L., Yih, Y., & Liu, C.-Y. ~1993!. A study on decision rules of a at University of Oviedo, Spain. He is currently an Assistant
scheduling model in an FMS. Computer in Industry 22, 1–13. Professor at the School of Industrial Engineering at Oviedo
Thesen, A., & Lei, L. ~1986!. An expert system for scheduling robots in a University. His current research interests are in the areas of
flexible electroplating system with dynamically changing workloads.
Proc. of the Second ORSA0 TIMS Conf. on Flexible Manufacturing fuzzy logic in production problems, and total quality man-
Systems, pp. 555–566. Elsevier Science Publishers, Amsterdam. agement.