Successful Lecture Timetabling with

Evolutionary Algorithms
Peter Ross, Dave Corne, Hsiao-Lan Fang
Department of Articial Intelligence
University of Edinburgh
80 South Bridge, Edinburgh EH1 1FN
email: [email protected]

Abstract
Arranging a lecture/tutorial/lab timetable in a large university department or school is
a hard problem faced continually in educational establishments. We describe how this prob-
lem has been solved in one institution via the use of evolutionary algorithms. The technique
extends easily and straightforwardly to any lecture timetabling problem. Although there
may be more eective ways to handle particular instances of the general lecture timetabling
problem, we note that the combination of speedy, good results and ease of development for
the particular application in hand make the EA-based technique we present potentially
widely useful in general.

1 Introduction
Lecture timetabling is the problem of assigning times and places to a many separate lectures,
tutorials, etc : : : , to satisfy several constraints concerning capacities and locations of available
rooms, free-time needs and other such considerations for lecturers, and relationships between
particular courses. The most prominent overall constraint (central to all timetabling problems)
is that there should be no clashes; that is: any pair of lectures (or tutorials, etc : : : ) which are
expected to share common students or teachers should not be scheduled simultaneously.
Typically, this is addressed by drawing up an initial draft timetable, followed by perhaps
weeks of redrafting as complaints about the most recent draft ow in from various sources. The
space of possible timetables can be nightmarish to traverse, and there have been several attempts
to nd useful AI/OR approaches to aid the process [2, 15, 4, 11, 3]. Success has been recently
reported for using evolutionary algorithms (EAs) for timetabling [1, 6, 14, 7, 8, 17, 16]. Here we
describe one such EA approach, and present illustrative results on some real lecture timetabling
problems.
In presenting results on some real problems, we augment similar work which also reports
results on real problems [6, 7, 8, 17, 16], but we also present a fuller and deeper discussion of the
general approach, clarifying how it may be used on a much wider range of problems than that
studied. Some comparison is also made between the use of dierent EA selection schemes. Also,
1
in reporting EA-based results in comparison with the independently and `expertly' calculated
timetables for four real lecture timetabling problems we show clearly how this approach can yield
very benecial improvements.

Overview
We rst describe the kind of problem addressed in more detail in section 2. Description of our
EA based approach follows in section 3, and notes on implementing the approach then appear
in section 4. Illustrative experiments on real problems appear in section 5, followed by general
discussion in section 6.

2 Lecture Timetabling Problems

The basic element of a lecture timetabling problem is a set of events E = fe1; e2; : : : ; ev g. Each
member of E is a unique event requiring assignment of a time and a place. That is, it may be a
lecture, a tutorial, a lab session, or some other event which plays a part in the term timetable.
We could alternatively formulate this, for example, in terms of a set of subjects S , each of which
has associated numbers of lectures, tutorials, lab sessions, etc : : : per week. However it is simpler
to take as our starting point the set E as described, which is easily generated from the latter
kind of data. Eg, two separate Lisp Programming lectures and ve Lisp Programming tutorials
will constitute seven members of E .
Each event ei has an associated length li (how long the event is in, say, minutes), and an
associated size si , which is either known or an estimate of the number of students expected to
attend that event. There is also a set of `agents' A = fa1; a2; : : :; atg; these are lecturers, tutors,
technicians, etc : : : | people with some kind of distinguished role to play in an event. Finally,
there is a set of places P = fp1; p2; : : :; pq g, and a set of times T = ft1; t2; : : :; tsg. An assignment
is a four-tuple (a; b; c; d), in which a 2 E; b 2 T; c 2 P; d 2 A, with the interpretation "event a
starts at time b in place c and is taught (lectured, tutored, : : : ) by agent d. A lecture timetable
is simply a collection of n assignments, one for each event.
Such problems are beset by many kinds of constraint. A fuller presentation of these appears
in [18], and various more simplied treatments have been presented elsewhere [2, 1, 3]. The
following brie y notes the necessary aspects of the approach discussed in [18] which are needed
for understanding the rest of this paper. Following this we describe the constraints which need to
be met in a specic series of lecture timetabling problems; this serves as a background case-study
against which we can illutrate some general aspects of the implementation of the approach to an
arbitrary timetabling problem.

3 Evolutionary Timetabling
Assuming familiarity with the basic processes in EAs, it suces to describe our approach by
reference only to the chromosome representation and the tness function. Those unfamiliar with
the basics of EAs can consult good texts such as [12, 10].
2
The Chromosome Representation
A `timetable' chromosome is a vector of symbols of total length 3v (recall: v is the number of
events), divided into contiguous three-gene chunks. The three alleles in the ith chunk, where
1 <= i <= v, represent the time, place, and agent assignments of ith event. Naturally, the sets
of possible alleles at time, place, and agent genes are respectively identied with the sets T , P ,
and A. The simple example chromosome \abcdef" represents a timetable in which event e1 starts
at time a in place b, involving agent c, and event e2 starts at time d in place e, involving agent
f.
This constitutes a `direct' representation, as opposed to the more indirect style common in EA-
based job shop scheduling work, and recently implemented for timetabling in =citePaechter+94.
Relative advantages and disadvantages of these two styles are beyond the scope of this paper,
but are a central point of interest. Suce to say here that comparison of the two is beset by
complications, but it so far seems that the direct approach enhanced with the introduction of
intelligent mutation operators (which take great advantage of the directness of the representation,
and hence cannot be feasibly constructed for use with the indirect style) [8], vies on equal terms
with an ingenious version of the indirect style [16]; these observations are yet to be properly
backed up empirically.

3.1 The Fitness Function

A maximally t timetable is clearly one which satises all of the imposed constraints. Also, it
seems reasonable to distinguish between timetables in terms of tness based on the numbers and
kinds of dierent constraints violated. A choice of tness function which meets this behaviour
is as follows, where C is the set of constraints in the problem, Pi is a penalty associated with
constraint i, and vi(g) = 1 if timetable g violates constraint i, and 0 otherwise:
f (g) = 1=(1 +
X Pivi(g)) (1)
i2C
The relative penalty values may be chosen to re ect intuitive judgement of the relative im-
portance of satisfying dierent kinds of constraint. Further discussion of this kind of objective
function and other possibilities is beyond the scope of this paper, but appears elsewhere, eg:
[8, 19, 13]. In general, however, a penalty function as above, using a rough choice of penalty
settings derived from the course organisers' notion of relative importance of dierent constraints,
appears adequately robust for many problems.

4 A Specic Lecture Timetabling Problem

An MSc course in the Department of Articial Intelligence, University of Edinburgh (EDAI)
involves eight taught course modules spread over two terms. The course is organised into `themes',
each involving a particular combination of 8 modules, of which some are compulsory and some
optional. As well as choices from among the 30+ modules available in the AI Department,

3
students may also choose modules from the Computer Science (CS) Department and others. A
complicating factor here is that the CS Dept is an inconvenient bus ride away from the AI Dept.
T comprises 80 start times, 16 per day on each day of a ve day week. Each day's slots are
at half-hourly intervals from 9am to 4:30pm. E comprises a large collection of lectures, tutorials,
and lab sessions, mostly an hour long, but sometimes two hours long. A student enrolled on a
course must attend all the lectures in E involving the course, but only one of the tutorials or
labs (E will include several tutorials or labs for each course). Separate courses are pre-assigned
to either term 1 or term 2. Hence, in one academic year there is a separate lecture timetabling
problem for each term involving roughly half of the modules available on the course as a whole.
The full set of constraints which need to be faced are as follow:
Options : Student's options should be kept open as far as possible. No pair of lectures in the
same theme should overlap in the timetable. More generally, lectures x and y should not
overlap if there is expectation that one or more students may wish to take both courses x
and y.
Event Spread : The individual timetable for any student must be spread out fairly evenly. Eg:
A student should not have to sit through four lectures in a single day. Rather, the events
an individual student must attend should be evenly spaced out during the week. Also,
dierent lectures on the same topic (eg: there may be 2 Prolog lectures per week) should
occur on dierent days.
Travel Time : A student should have at least 30 minutes free for travel between events in the
CS Dept and events in the AI Dept.
Slot Exclusions : CS lectures should occur in morning slots, and AI lectures in afternoon slots
(this arises from an inter-departmental agreement). Also, lectures at lunchtime (starting
at 1:00pm or 1:30pm) should be avoided if possible. In a similar vein, various constraints
of the form \event e cannot start at time t" arise owing to other commitments of the sta
involved.
Slot Specications : Various constraints are given in the form \event e must start at time t",
arising for various reasons.
Capacity : The size of a lecture or tutorial should not exceed the capacity of the room it occurs
in. Also, a room can only cope with one event (lecture, tutorial, or lab) at a time (this is
the main dierence between lecture and examination timetabling).
Room Exclusions : Many constraints on room assignments for particular events can easily be
derived from the Capacity constraints, along with information about the expected sizes of
events. In addition however, there are other considerations which lead to several a priori
constraints of the form \event e cannot occur in room r". For example, event e may demand
disabled access, or certain audio-visual requirements unavailable in room r.
Room Specications : Similarly, several constraints are apparent of the form \event e must
occur in room r".
4
Juxtaposition : Preferably, all tutorials or laboratory sessions for any course should occur later
in the week than the week's rst lecture on that course. Sometimes this is particularly
necessary, since a tutorial or lab session may be based on the lectures which were held
(hopefully) earlier in the week. In other cases this is desirable but not vital.

Translating the Constraints into a Fitness Function

In this case, it so happens that every lecture's agent (ie: lecturer) is predetermined, and individual
lecturers have already decided in advance (providing details in the style of exclusion constraints)
which slots they are not available for. Tutors for tutorials and lab sessions are not pre-determined
in this way, but for these there is no point in incorporating them into the timetable at this stage
(usually well in advance of term), since we simply do not know for sure who will be available and
when. For this problem we therefore need not consider the set A, and hence can use chromosomes
of length 2v. The above constraints can be dealt with as follows:
Keeping Options Open
The Options constraint is handled by interpreting it as a large collection of binary constraints,
each involving a distinct pair of events expected to share students. For convenience, we derive
a set of `virtual' students, each of whom takes a distinct one of the set of possible four or ve-
module options for the term. This set of virtual students is then used to derive the collection
of distinct binary constraints between events. Here, such a binary constraint occurs between
every pair of distinct lectures taken by some virtual student. A similar constraint also holds
between distinct lectures on the same module, and between lectures and tutorials on the same
module. No such constraint is needed between dierent tutorials or labs for the same course, or
even between tutorials and labs on dierent courses. Such may be scheduled simultaneously, and
often are; in due course this gives rise to constraints which aect each students' choice from the
set of tutorial and/or lab sessions available for each module.
This amounts to a collection of binary constraints of the form \e1 must not overlap in time
with e2"; the tness function must check for violation of each of these in turn. Similarly, it should
be clear how the basic problem of avoiding clashes in any other lecture or exam timetabling
problem can be dealt with. Notice too that we can incorporate the Travel Time constraint here.
If e1 and e2 are timetabled with a break of, say, k minutes between them, but are assigned to
rooms which take more than k minutes to travel between, then any Options constraint between
them is eectively violated. Hence, the violation check for Options constraints can, simply via
accessing the `place' genes for e1 and e2 and a given travel-time matrix for the places, also account
for Travel Time constraints.
Event Spread constraints
Event spread constraints can be handled in a number of ways. Eg, to even out the event spread
for individual students we might explicitly calculate some measure or measures of event spread
for each virtual student. Alternatively, we could reasonably approximate this by examining some

5
measure or measures of the spread of the timetable as a whole. Both would seem to oer the
same overall eect; the latter will typically be computationally cheaper, but the former approach
would seem to oer more potential for control and tradeo of dierent aspects of the event spread
for individual students.
The method used in the experiments detailed later is as follows: the tness function notes, for
each virtual student, the number of instances of the following two `oenses': a) four events are
scheduled in one day for this virtual student; b) ve or more events are scheduled in one day for
this virtual student. A dierent penalty term is associated with each, and the penalty weighted
sum of instances of these oenses, summed over virtual students, makes up the contribution to
(or, rather, detraction from : : : ) tness of the overall event spread constraint.
Finally, `dierent-day' constraints can clearly be handled in the same way as Options con-
straints. For any pair of lectures on the same module, we simply check directly from the chromo-
some whether or not their assigned slots are on the same day. If they are, then an appropriate
penalty is added.
Exclusions and Specications
It is easy to see how exclusion constraints can be directly translated into one or more simple
violation-check functions, given the chromosome representation in use. Notice however that we
can just as simply pre-arrange it so that chromosomes never violate these constraints in the rst
place. We can doctor the allele range of each gene so that it is always the specied allele (if any),
or only ranges over the non-excluded alleles.
Choosing between such pre-satisfaction of exclusion and specication constraints, and the
option of penalising violations of them, is not always straightforward. If many such constraints
exist, the `pre-satised' space may well lack excellent timetables which violate a few exclusions,
for example, but make up for this in other ways. On the other hand, pre-satisfaction speeds up
evaluation and promises to speed up search via reducing the search space. The full ramications
of this choice are beyond the scope of this paper, but it suces to point out here that either option
should be available in a system which implements this technique. In the experiments discussed
later, most exclusion and specication constraints were prespecied. The only `penalised' such
constraint was that for lunchtime lectures. According to the EDAI MSc course organisers, it is
preferable to avoid these, but acceptable to trade these o against other constraint violations.
Capacity constraints
To check that a room's capacity isn't exceeded, we must rst translate the overall room capacity
constraint into constraints of the form \room r should contain no more than rcap students in
timeslot t", for each room r and timeslot t, where rcap is the student capacity of room r. During
evaluation, the system simply precomputes from the current candidate timetable the student
load for each room in each slot, and then runs through this list of constraints checking each in
turn and accumulating penalties for violations. Evidently, the same technique can be applied for
a very wide range of similar problems, and also applies to constraints concerned with `teaching
loads' constraints in problems for which agents must be considered in the representation.

6
Juxtaposition Constraints
Finally, it is evident how the ordering constraints between given groups of events can be incorpo-
rated. We rst derive a collection of binary constraints from those given. Eg, \all lisp tutorials
should occur later than the rst lisp lecture of the week" is translated into a collection of binary
constraints of the form \lisp 1 must be before lisp t3". Checking for violations of such con-
straints is then straightforward . Similarly, it should be clear how any juxtaposition constraint
(eg: \there should be at least two days between event e1 and event e2") can be similarly handled.

5 Experiments
Experimental Setup
We address the EDAI MSc lecture/tutorial problems for both terms of the academic years 92/93
and 93/94, respectively involving 76, 73, 82, and 73 events. Other features of these problems are
as described in section 4, and full details are available from the authors.
In all cases, the EA used a population size of 50, uniform crossover, gene-by-gene mutation,
and elitist generational reproduction. The crossover and mutation rates pC and pM were dynam-
ically altered as follows. pC started at 0.8 and was decreased by 0.001 after each generation (ie:
after every 50 evaluations), with a lower limit of 0.6, while pM started at 0.003 and increased
by 0.0003 each generation, with an upper limit of 0.02. Each trial was run for 200 generations
(10000 evaluations). Separate experiments are recorded for each problem for each of three dif-
ferent selection strategies: tness-proportionate (FIT), GENITOR-style rank-based with bias 2
(RANK) [20], and tournament selection with tournament size 10 (TOUR).
The result of a trial was a maximally t timetable found during the trial. From this we record
a vector of violations V = fc; j; p; l; e; d5+; d4; tg, which respectively denote violations of Options
constraints (clashes), important juxtaposition constraints, desirable juxtaposition constraints,
lunchtime constraints, slot-exclusion constraints, cases where a `virtual student' faced more than
four events in a a day, cases where a `virtual student' faced four events in a single day, and,
nally, travel time constraints. Violations of all other constraints mentioned above (eg: room
exclusions, capacity constraints) are not recorded, since they were fully satised in all cases.
Ten trials were run for each experiment, and results for each problem record the best V found
overall, and the mean of V over the ten trials, for each of three selection schemes. We also present
V for the timetables produced by the course organisers for each problem, and which were the
actual timetables used (or in use), since the EA system itself was not yet in regular use. `Best'
means relative to the penalty-weighted sum of violations. The xed penalty values used in these
trials for the various violations were, in the order in which they appear in the tables: 500, 300,
30, 30, 10, 5, 1, 1. Hence, violations are listed in order of decreasing importance, as judged by
the course organisers.

5.1 Results

7
 Problem c j p l e d5+ d4 t
92/93 term1
Course Organisers 0 0 4 4 14 0 2 65
FIT / Best of 10 0 0 0 0 0 0 0 8
FIT / Mean of 10 0 0.1 0.2 1.2 0 0 0.9 16.6
RANK / Best of 10 0 0 0 0 0 0 0 17
RANK / Mean of 10 0 0 0 1.6 0 0 0.6 17.2
TOUR / Best of 10 0 0 0 0 0 0 0 0
TOUR / Mean of 10 0 0 0 1 0 0 0.4 0.7
92/93 term2
Course Organisers 0 1 9 2 0 1 4 49
FIT / Best of 10 0 0 0 0 0 0 1 0
FIT / Mean of 10 0 0 0.1 0.5 0 0 2.1 9.2
RANK / Best of 10 0 0 0 0 0 0 3 9
RANK / Mean of 10 0 0 0.5 1 0 0 2.4 13.4
TOUR / Best of 10 0 0 0 0 0 0 0 0
TOUR / Mean of 10 0 0 0.3 0.6 0 0 0.7 0.1
Table 1: Comparative performance on the 93/94 problems

Tables 1 and 2 clearly show that the EA approach leads to much better timetables in each
case. One course-organiser produced timetable failed to keep all reasonable options open (ie: had
clashes), while many failed to fully keep lectures and tutorials away from various restricted slots,
and all failed constraints at least as important as the need to avoid lunchtime events. Problems
caused involve lecturers and tutors being forced to work during timeslots previously designated
for other things (eg: regular weekly seminars), forced to give up free afternoons or mornings
designated for research, students facing excessively demanding days, and so on.
On the other hand, for each problem, tournament selection found either a perfect timetable
or one with a single travel-time violation in at least one of ten trial runs. For each selection
scheme, mean and best results compared very favourably with the experts' eorts. Each EA
trial was completed within 5 minutes on a sun SPARC. Occasionally, a EA solution, was worse
in terms of some attribute (eg: violations of desirable juxtaposition constraints) than the course
organisers' solution, but better overall in terms of the penalty-weighted sum of violations. This
suggests that the EA was more successful at trading o the relative occurrences of violations of
dierent importance.
Tournament selection appears to be the best choice, with rank-based selection of the style
used in [20] and tness proportionate selection, in that order, being next best. This relative
performance of dierent selection schemes cannot strictly be taken as read from these results
without further experiments using dierent tournament sizes, biases, and so on; however, much
EA literature backs up this ordering of relative performance.

8
 Problem c j p l e d5+ d4 t
93/94 term1
Course Organisers 0 0 0 5 7 0 2 84
FIT / Best of 10 0 0 0 2 0 0 0 15
FIT / Mean of 10 0 0 2.6 2.3 0 0 0.7 28.6
RANK / Best of 10 0 0 1 2 0 0 0 23
RANK / Mean of 10 0 0 2.2 2.5 0 0.1 0.3 31.5
TOUR / Best of 10 0 0 0 0 0 0 0 1
TOUR / Mean of 10 0 0 0.6 1.4 0 0 0.2 0.6
93/94 term2
Course Organisers 2 0 3 3 1 0 0 50
FIT / Best of 10 0 0 1 0 0 0 0 11
FIT / Mean of 10 0 0 0.2 1.4 0 0 1.4 14
RANK / Best of 10 0 0 0 1 0 0 1 13
RANK / Mean of 10 0 0 0.6 1.7 0 0 1.3 13.6
TOUR / Best of 10 0 0 0 0 0 0 0 0
TOUR / Mean of 10 0 0 0 0.7 0 0 0.2 0.2
Table 2: Comparative performance on the 93/94 problems

6 Discussion
The results clearly indicate the benets of using a penalty-function based EA approach on this
problem, and by implication suggest similar utility for the same approach on similar problems. In
designing the penalty-weighted tness function itself for these experiments, several of the design
decisions were ad hoc. For example, penalty values were chosen according to a rough judgement
of relative importance. Also, there were several other possibilities, as discussed earlier, for dealing
with the event-spread constraints. Even the underlying EA itself was far from optimal in terms
of parameter settings and general conguration. Better choices of selection scheme, for example,
are spatially-oriented schemes as presented in [5] and [9], while better overall choices for the EA
are certainly possible.
The `rough-and-ready' aspect of the experimental congurations used in this paper, coupled
with the good results reported and the ease of implementing the approach strongly suggests a
promising future for both further research and also practical use of EAs on general timetabling
problems. Naturally, there are a considerable number of theoretical and practical issues that
need to be answered. An illustrative collection of these follow:
Scaling Up
How does this approach scale up to larger and more tightly constrained problems? The real
problems addressed here, and similarly those addressed in [6, 14, 7], are similar in size or larger
than a large proportion of the timetabling problems faced in many institutions. Hence the useful-
ness of this approach seems justied, inasmuch as we can expect the benecial results displayed

9
here to carry over to dierent timetabling problems of similar or smaller size. How the approach
scales with increasing size and/or complexity is a harder question, which is the subject of con-
tinuing research. Initial indications in unpublished work are that the basic approach scales well,
but suers from a problem common to EA-based optimisation: that is, solutions near optimal
regions are rapidly found on large complex problems, but further evolution towards optima be-
comes considerably slow, and may stop altogether. Fortunately this diculty is readily aided by
the use of smart hillclimbing mutation operators. As detailed in [8], use of such operators helps
to vastly increase the scope of the approach in terms of problem size. Similarly, an alternative
chromosome representation used in [16] is also found to signicantly improve on solution quality
when compared with the basic approach as presented here.
Generalising Across
How does the approach perform on other timetabling problems? The general nature of the
approach suggests that it would be just as well employed on many similar problems. A key
aspect which matters here is speed of evaluation. As long as the numbers of constraints which
need checking coupled with the computational ease of checking them make for a relatively speedy
evaluation function, it seems safe to suggest that useful performance is promised. The kinds of
individual constraints that usually occur in timetabling problems are computationally quick to
check when using the direct chromosome representation. The general prospects for EA-based
timetabling in this respect are in any case illustrated by the variety of real problems so far
successfully addressed.
Dierent Approaches
Dierent representations, use of domain specic recombination operators, and hybridisation
of the EA with other techniques are all candidates for renement of this approach. Much further
research in this vein is in order. It is also interesting and important to compare EA approaches
with other methods such as branch & bound search, simulated annealing, and so on. This
endeavour is complicated by the dierences between the techniques themselves. Eg, the promise
of the EA-based approach is most strongly manifest in its robustness across a very wide range of
dierent timetabling problems. Comparison with rule-based approaches to test this claim on the
same variety of problems would then necessitate the lengthy and dicult development process
of building rule-based systems with similarly wide applicability. Comparison with simulated
annealing is a more likely prospect, and such is planned in due course.
Practice
Some common needs are not met by the approach as discussed here. Eg: timetablers may wish
to generate several distinct timetables to choose from. Such considerations require renements
and extensions, although the basic approach we have discussed remains a useful partial tool for
such requirements. More relevantly, space prevents us from properly covering here the general
process and the many possible choices involved in interpreting the constraints of a problem into
a particular choice of tness function. It is rarely apparent how best to do this, and further work
is required to assess the various possibilities. Experience shows, however, that there is unlikely
to be a major dierence in performance between dierent such choices for problems of the size
and type addressed here; hence, we feel that natural and/or arbitrary choices may be made with
impunity for penalty settings, pre-specications, event-spread constraint handling, and so on : : : ,
at least for problems of the size and type found in small or medium sized university departments.
10
Acknowledgements
We would like to thank Bob Fisher and Alan Smaill for their help in describing the lecture
timetabling problems used here and providing data. Thanks also to the UK Science and Engineer-
ing Research Council for support of Dave Corne via a grant with reference number GR/J44513.

References
[1] D. Abramson and J. Abela, `A parallel genetic algorithm for solving the school timetabling
problem', Technical report, Division of Information Technology, C.S.I.R.O., (April 1991).
[2] Alan M. Barham and John B. Westwood, `A simple heuristic to facilitate course timetabling',
Journal of the Operational Research Society, 29, 1055{1060, (1978).
[3] Mirjana Cangalovie and Jan A.M. Schreuder, `Exact colouring algorithm for weighted graph
applied to timetabling problems with lectures of dierent lengths', European Journal of
operations research, 51, 248{258, (1991).
[4] Michael W. Carter, `A survey of pratical applications of examination timetabling algo-
rithms', Operations Research, 34(2), 193{202, (March-April 1986).
[5] Robert J. Collins and David R. Jeerson, `Selection in massively parallel genetic algorithms',
in Proceedings of the Fourth International Conference on Genetic Algorithms, eds., R.K.
Belew and L.B. Booker, pp. 249{256. San Mateo: Morgan Kaufmann, (1991).
[6] Alberto Colorni, Marco Dorigo, and Vittorio Maniezzo, `Genetic algorithms and highly
constrained problems: The time-table case', in Parallel Problem Solving from Nature, eds.,
G. Goos and J. Hartmanis, 55{59, Springer-Verlag, (1990).
[7] Dave Corne, Hsiao-Lan Fang, and Chris Mellish, `Solving the module exam scheduling
problem with genetic algorithms', in Proceedings of the Sixth International Conference in
Industrial and Engineering Applications of Articial Intelligence and Expert Systems, eds.,
Paul W.H. Chung, Gillian Lovegrove, and Moonis Ali, 370{373, Gordon and Breach Science
Publishers, (1993).
[8] Dave Corne, Peter Ross, and Hsiao-Lan Fang, `Fast practical evaolutionary timetabling', in
Proceedings of the AISB Workshop on Evolutionary Computation, (1994).
[9] Yural Davidor, `A naturally occuring niche & species phenomenon: The model and rst
results', in Proceedings of the Fourth International Conference on Genetic Algorithms, eds.,
R.K. Belew and L.B. Booker, pp. 257{263. San Mateo: Morgan Kaufmann, (1991).
[10] Handbook of Genetic Algorithms, ed., L. Davis, New York: Van Nostrand Reinhold, 1991.
[11] K. A. Dowsland, `A timetabling problem in which clashes are inevitable', Journal of opera-
tions research society, 41, 907{918, (1990).
11
[12] David E. Goldberg, Genetic Algorithms in Search, Optimization & Machine Learning, Read-
ing: Addison Wesley, 1989.
[13] Sami Khuri, Thomas Back, and Jorg Heitkotter, `An evolutionary approach to combinatorial
optimization problems', in Proceedings of the 1994 Computer Science Conference (CSC94),
Phoenix, Arizona, (March 1994). ACM Press. to appear.
[14] Si-Eng Ling, `Intergating genetic algorithms with a prolog assignment problem as a hybrid
solution for a polytechnic timetable problem', in Parallel Problem Solving from Nature, 2,
eds., R. Manner and B. Manderick, 321{329, Elsevier Science Publisher B.V., (1992).
[15] N. K. Mehta, `The application of a graph coloring method to an examination scheduling
problem', Interfaces, 11, 57{64, (1981).
[16] B. Paechter, H. Luchian, A. Cumming, and M. Petruic, `Two solutions to the general
timetable poblem using evolutionary methods', in Proceedings of the IEEE Conference on
Evolutionary Computation, (1994).
[17] Ben Paechter, `Optimising a presentation timetable using evolutionary algorithms', in Pro-
ceedings of the AISB Workshop on Evolutionary Computation, (1994).
[18] Peter Ross, Dave Corne, and Hsiao-Lan Fang, `Timetabling by genetic algorithms: Issues
and approaches', Technical Report AIGA-006-94, Department of Articial Intelligence, Uni-
versity of Edinburgh, (1994). revised version to appear in Applied Intelligence.
[19] Alice E. Smith and David M. Tate, `Genetic optimisation using a penalty function', in
Proceedings of the Fifth International Conference on Genetic Algorithms, ed., S. Forrest,
pp. 499{503. San Mateo: Morgan Kaufmann, (1993).
[20] Darrell Whitley, `The genitor algorithm and selection pressure', in Proceedings of the Third
International Conference on Genetic Algorithms, ed., J. D. Schaer, 116{121, San Mateo:
Morgan Kaufmann, (1989).

12