You are on page 1of 5

2009 2nd Conference on Data Mining and Optimization

27-28 October 2009, Selangor, Malaysia

Examination Timetabling Using Scatter Search


Hyper-Heuristic
Nasser R. Sabar1 and Masri Ayob2
1, 2
Data Mining and Optimisation Research Group (DMO), Center for Artificial Intelligence Technology (CAIT),
Universiti Kebangsaan Malaysia, 43600 UKM, Bangi Selangor, Malaysia.
naserdolayme@yahoo.com, masri@ftsm.ukm.my

Abstract—Hyper-heuristic can be defined as a “heuristics to Moreover, it’s difficult to apply them to the new problem
choose heuristics” that intends to increase the level of generality and able to produce the good results for one or more
in which optimization methodologies can operate. In this work, datasets.
we propose a scatter search based hyper-heuristic (SS-HH) However, recently, there has been a new trend in search
approach for solving examination timetabling problems. The methodology with aim to develop general methods; hyper-
scatter search operates at high level of abstraction which heuristics (Burke et al. [2]). The term hyper-heuristic refers
intelligently evolves a sequence of low level heuristics to use for
to an approach that aims to explore the search space of
a given problem. Each low level heuristic represents a single
neighborhood structure. We test our proposed approach on the heuristics rather than a search space of solutions (Burke et al.
un-capacitated Carter benchmarks datasets. Experimental [2]; Qu and Burke [4]). Burke et al. [2] define a hyper-
results show the proposed SS-HH is capable of producing good heuristic as: “The process of using meta-heuristics to choose
quality solutions which are comparable to other hyper- (meta-) heuristics to solve the problem in hand”. Usually,
heuristics approaches (with regarding to Carter benchmark low level heuristics can be a simple local search; several
datasets). neighbourhood functions or neighbourhood exploration that
could be used to improve the solution in hand.
Keywords: Educational Timetabling, Scatter search, Hyper- Some hyper-heuristics approaches for examination
Heuristic timetabling problems include tabu search (Burke et al. [5],
[2]), case-based reasoning (Burke et al.[6] ), variable
I. INTRODUCTION neighborhood search (Qu and Burke[4]; Qu and Burke [7]),
The examination timetabling can be defined as the process of graph based methods (Asmuni et al. [8]) and memetic
scheduling a given number of exams into a limited number algorithms (Ersoy et al. [9]).
of timeslots so as satisfy all hard constraints and to minimize Therefore, in this work, we propose a scatter search based
the soft constraints violation as much as possible (Qu et hyper-heuristic (SS-HH) for solving examination timetabling
al.[1]). Hard constraints have to be satisfied under any problem. The SS-HH employs scatter search algorithm at
circumstances; otherwise the generated timetable is high level to intelligently produce a sequence of low-level
infeasible. As an example, no student can sit two exams at heuristics. The goal of the SS-HH is to evolve a sequence of
the same time. Whilst, on the other hand, soft constraint low-level heuristics for each given instance of the problem
should be satisfied as far as possible but the timetable is still and use these sequences to improve the solution quality by
called feasible if any soft constraints are violated. However, applying them sequentially. We tested our proposed SS-HH
the quality of the timetable is measured based on soft on the un-capacitated (where the size of the room is
constraint violation. An example of a soft constraint is the disregarded) Carter benchmarks [10] (Toronto b, see Qu et
spreading of exams, for a given student, as much as possible. al. [1]) examination timetabling datasets and evaluated the
During the last four decades, there have been a numerous quality of the timetable using proximity cost [10].
heuristics and methodologies have been developed for
examination timetabling. For example, graph based
sequential techniques, constraint based techniques, local II. SCATTER SEARCH –HYPER HEURISTICS (SS-HH)
search methods (e.g. tabu search, simulated annealing), Scatter search (SS) is a population-based metaheuristic
population based algorithms (e.g. genetic algorithms, ant introduced by Glover [11] in 1977. The main idea of SS is to
colony optimization) etc. (Qu et al. [1]). keep a small population of reference solutions, called a
Most of previous methods are known as problem specific reference set, and to combine them to create new solutions.
domain which require human expert to fine-tune parameters. The SS procedure can be as follows. It begins by creating
a reference set from population of solutions. This is done by
first creating a large number of solutions (i.e. population)

978-1-4244-4944-6/09/$25.00 ©2009 IEEE 127


using a diversification generation method. Then, these directly to the problem itself. Figure I present the pseudo-
solutions (i.e. population) are improved by conducting local code of the SS-HH approach that we have developed.
search method. From these improved solutions (i.e.
population), the reference set is created. The reference set
solutions are selected based on the solution quality and Procedure SS-HH-
diversity. Then, the solutions in reference set are used to
produce a subset of solutions. The solutions in each subset 1- Initialization
are candidates for combination. Solutions within each subset - Generate initial solutions S
are combined to generate new solutions. Each newly -Set population size, ReferenceSet size, subset- size,
generated solution is improved by local search and possibly steepest descent maximum idle iteration.
replaces one solution in the reference set. The process of 2- GenerateDiverseSolutions(pop)
subset generation, solution combination, and local search is -for i=1 to population size do
interacting iteratively until the reference set does not change Generate diverse solutions soli, and add it to
anymore. solution
In this work, at the higher level, a SS is used to population (pop)
intelligently evolve a sequence of low level heuristics. -end for
Whilst at the lower level, ten different neighborhood 3- Improvement stage
structures are used to generate sequence ordering of low -Apply steepest descent local search at current
level heuristics. The ten different neighborhoods structures population (pop).
can be summarized as follows (adapted from Burke et al. -Sbest ← BestOfPopulation (pop).
[12]): 4- GenerateReferenceSet(rs)
-while rs < ReferenceSet size
LLH1: Select one exam at random and move it to feasible Select the best solution from the diverse
timeslots selected randomly. solutions (pop) and add it to rs.
Remove the selected solutions from diverse
LLH 2: Select two exams at random and swap their solutions (pop)
timeslots with ensuring feasibility. -end while
5- While stopping criterion is not satisfied do
LLH 3: Move two exams at random to a feasible timeslots. - Generate a subset (S_set)
For i=1 to subset-size do
LLH 4: Select two timeslots at random and swap their Randomly select one solution from rs
exams. and add it to S_set.
end for
LLH 5: Move the highest penalty exam to a random feasible - while S_set not empty do
timeslots. Randomly select two solutions from S_set
and combine them using one-point
LLH 6: Make random kempe chain (The advantages of this crossover (sol*).Apply steepest descent
technique for exam timetabling are highlighted by Thomson local search at the new generated solution
and Dowsland in [13].). (sol**).
-Update rs
LLH 7: Move a random exam into a random timeslot (using Sworst ←WorstOfPopulation (rs)
Kempe chains in the case of infeasibility). If (sol**) ∉ rs and f(sol**) >= f(Sworst)
then
LLH 8: Randomly order timeslots. Replace sol** with Sworst
If f(sol**) >f(Sbest) then
LLH 9: Make two random kempe chains. Sbest=sol** and f(Sbest) = f(sol**)
end if .
LLH10: Move the highest penalty exams from a random 5% end if .
selection of the exams to a random feasible timeslots. -S_set= S_set-2;
- end while
In this work, the SS is implemented at a higher level of end while
abstraction rather than being applied directly to the solution
space. This is similar to the work implemented by Terashima Return the best solution Sbest.
et al. [14], Ross et al. [15], Han et al. [16] and Burke et al.
[12] where they successfully employs a Genetic algorithm at Figure I SS-HH pseudo-code
a higher level of abstraction rather than being applied

978-1-4244-4944-6/09/$25.00 ©2009 IEEE 128


The SS-HH begins by first generating initial feasible solutions by procedure is repeated until the stopping condition is
using hybrid graph coloring methods (see Ayob et al. [17]). Then satisfied.
we set the population size, reference set size, subset size, and
steepest descent maximum idle iteration. In this work, the SS-HH III. EXPERIMENTS AND RESULTS
uses an indirect chromosome representation where each The SS-HH has been tested on Carter’s un-capacitated exam
chromosome represents sequences of low level heuristics rather timetabling benchmark datasets [10] (type I, [1]). The
than actual solution. A simple chromosome representation is used characteristics of the datasets are presented in Table I. For
with fixed length (i.e. length= 5). Each low level heuristics (i.e. each dataset, the SS-HH was run 10 times with different
single neighborhood) is represented by an integer number (Han et initial solutions. Each run take about 2-4 hours based on
al. [16]). Figure II is an example of SS-HH indirect chromosome datasets complexity.
representation, where integer 1 in the chromosome represents the The SS-HH algorithm coded using Visual C++ 6.0 on a
low-level heuristic LLH1, 2 represent LLH2, 5 represent LLH5 PC AMD Athlon with a 1.92 GHz processor and 512 RAM
and so on. running Windows XP 2002. Table II summarize all
parameters used in SS-HH with their assigned values. we
have tested number of different parameters setting and the
3 1 4 2 5 chosen values are based on our preliminary numerical
TABLE I
Figure II shows example of SS-HH chromosome CARTER BENCHMARK EXAM TIMETABLING DATASET
representations
Data sets Number of Number of Number
Then, we generate a population of diverse solutions (pop) by timeslots examinations of
Students
randomly select one low level heuristic (i.e. single neighborhood)
and inserting it to the chromosome with ensuring no one low Car-f-92 32 543 18419
level heuristic selected more than one times. Afterward, we apply Car-s-91 35 682 16925
steepest descent algorithm. The steepest descent algorithms Ear-f-83 24 190 1125
implemented by sequentially select one low-level heuristic from Hec-s-92 18 81 2823
Kfu-s-93 20 461 5349
chromosome and apply it to the initial solution S. The steepest
Lse-f-91 18 381 2726
descent algorithms apply the next low level heuristic in the
Rye-s-93 23 486 11483
chromosome ordering as soon as the maximum number of idle Sta-f-83 13 139 611
iteration reached. Then, the chromosome that has made a big Tre-s-92 23 261 4360
improvement to the solution S compared to other chromosome is Uta-s-92 35 622 21267
selected as the best one (i.e. Sbest). Next, we generate a reference Ute-s-92 10 184 2750
sets by select the best chromosome (i.e. solution) from the Yor-f-83 21 181 941
diverse solution (i.e. pop). Then, we generate a subset (s_set)
from the reference sets rs by randomly selecting one TABLE II
chromosome (i.e. solution) from the reference sets rs and add it SS-HH PARAMETERS
into the subset s_set. Two chromosomes are selected from subset No Parameters Value
s_set and combined using one-point crossover to generate new
chromosomes. Figure III shows an example of crossover for two 1- Population size 80
chromosomes parent A and parent B. 2- Reference set size 10
3- Max. No. of idle iteration 5
4- Subset size 4
5- Number of iteration 100
6- Crossover type one point
experiments. These parameters will be tuned in our future
work to be more general for complex datasets.

The SS-HH best results are presented in Table III. Also, this
table contains the comparison of our best results with a range
of published hyper-heuristics method (the best results are
Then we improve the new generated chromosomes by shown in bold). SS-HH was compared with five hyper-
implementing steepest descent algorithms. If there is no heuristics methods;
chromosome inside rs has same sequence of low level
heuristic and the quality of the improved solution is better • FZLO: fuzzy multiple ordering criteria for examination
than the worse one, replace it with the worst one in the rs timetabling problem applied by Asmuni et al [8].
and update the best chromosome (i.e. solution). The

978-1-4244-4944-6/09/$25.00 ©2009 IEEE 129


• VNS-Hybrid neighborhood hyper-heuristics for exam REFERENCES
timetabling problems applied by Qu and Burke [7]. [1] Qu, R., Burke E., K., McCollum B., Merlot L.T.G., Lee S.Y.
• TS*- Tabu search hyper-heuristics applied by Burke et al. (2009), A survey of search methodologies and automated
[5]. system development for examination timetabling, Journal of
• TS- Tabu search hyper-heuristics applied by Burke et al. Scheduling, 12(1): 55-89,
[3]. [2] Burke, E., Hart, E., Kendall, G., Newall, J., Ross, P.,
Schulenburg, S., (2003). Hyperheuristics: An emerging
• GHH- hybridizations with a graph based hyper-heuristics direction in modern research technology. In: Handbook of
applied by Qu and Burke [4]. Metaheuristics. Kluwer Academic Publishers, pp. 457-474
(Chapter 16).
[3] Burke, E. K., McCollum, B., Meisels, A., Petrovic, S. & Qu,
R. (2007). A graph based hyper-heuristic for exam timetabling
Table III SS-HH Results problems. European Journal of Operational Research, 176,
177–192.
Data set Our best FZLO VNS TS* TS GHH
[4] Qu, R., & Burke, E. K. (2009). Hybridisations within a graph
Results
Car-f-92 4.41 4.52 4.7 - 4.53
based hyper-heuristic framework for university timetabling
4.16
Car-s-91 5.36 5.2 5.4 - 5.36 5.16
problems. To appear in Journal of Operational Research
Ear-f-83 37.74 37.02 37.29 45.60 37.92 35.86 Society, doi: 10.1057/jors.2008.102.
Hec-s-92 11.87 11.78 12.23 - 12.25 11.94 [5] Burke, E. K., Dror, M., Petrovic, S., & Qu, R. (2005). Hybrid
Kfu-s-93 15.70 15.81 15.11 - 15.2 14.79 graph heuristics in hyper-heuristics applied to exam
Lse-f-91 12.53 12.09 12.71 - 11.33 11.15 timetabling problems. In B. L. Golden, S. Raghavan, & E. A.
Pur-s-93 6.27 - - - - - Wasil (Eds.), the next wave in computing, optimisation, and
Rye-s-93 9.88 10.35 - - - - decision technologies (pp. 79–91). Maryland: Springer.
Sta-f-83 160.77 160.42 158.8 158.2 158.19 159.00 [6] Burke, E. K., Petrovic, S., & Qu, R. (2006). Case-based
Tre-s-92 8.90 8.67 8.67 - 8.92 8.6 heuristic selection for timetabling problems. Journal of
Uta-s-92 3.87 3.57 3.54 4.52 3.88 3.59 Scheduling, 9, 115– 132.
Ute-s-92 32.67 27.78 29.68 35.40 28.01 28.3 [7] Qu, R., Burke, E.K., (2005). Hybrid variable neighbourhood
Yor-f-83 39.11 40.66 43.0 - 41.37 41.81 hyper-heuristics for exam timetabling problems In:
Proceedings of the MIC2005: The Sixth Metaheuristics
International Conference, Vienna, Austria, August 2005.
SS-HH could obtain comparable results when comparing to [8] Asmuni, H., Burke, E. K., Garibaldi, J., & McCollum, B.
all hyper-heuristics method for all tested datasets. Moreover, (2005). Fuzzy multiple ordering criteria for examination
timetabling. In E. K. Burke & M. Trick (Eds.), Lecture notes
for 3 datasets (Pur-s-9, Rye-s-93 and Yor-f-83) among five
in computer science: Vol. 3616. Practice and theory of
hyper-heuristics, SS-HH could obtain best results. Also, it is automated timetabling V: selected papers from the 5th
shown from Table III, SS-HH outperforms FZLO and TS* international conference (pp. 334–353). Berlin: Springer.
for five tested datasets and VNS , TS GHH for 6, 9 and 4 [9] Ersoy, E., Özcan, E., & Etaner, A. S. (2007). Memetic
tested datasets respectively. Note that TS* has been tested algorithms and hyperhill-climbers. In Proceedings of the 3rd
for four datasets only. Moreover, there are no reported multidisciplinary international conference on scheduling:
results for Pur-s-9 and Rye-s-93 datasets among all theory and applications (pp. 159–166), Paris, France, August
compared methods except FZLO reported only Rye-s-93. 2007.
Therefore, we may conclude that our SS-HH is generally [10] Carter, M. W., Laporte, G., & Lee, S. Y. (1996). Examination
timetabling: algorithmic strategies and applications. Journal of
able to produce good results when compared against other
Operational Research Society, 47(3), 373–383.
hyper-heuristics. We belief SS-HH might produce a good [11] Glover F (1977), ‘Heuristics for integer programming using
quality results if the number of low level heuristic are surrogate constraints’, Decision Science, vol.8, pp. 156-166.
increased or mutation operator is used. [12] Burke E.K., Eckersley A.J., McCollum B., Petrovic S. and Qu.
R. (2009). Hybrid variable neighbourhood approaches to
university exam timetabling. Accepted by European Journal of
IV. CONCLUSIONS Operational Research, Technical Report NOTTCS-TR-2006-2,
School of Computer Science, University of Nottingham,
In this work, Scatter search based hyper-heuristics (SS-HH) United Kingdom.
for solving examination timetabling problems has been [13] Thompson, J.M., Dowsland, K.A.: Variants of simulated
presented. The SS-HH algorithm employs SS at high level to annealing for the examination timetabling problem. Annals of
evolve a sequence of low level heuristics. At the low level, Operations research, Vol. 63 (1996) 105-128.
ten different neighborhoods structure are used as low level [14] Terashima-Mar´ın H., Ross P. and Valenzuela-Rend´on M.
heuristics. SS-HH used indirect representation where each Evolution of constraint satisfaction strategies in examination
chromosome represents a sequence of low level heuristics timetabling. In: Proceedings of the genetic and evolutionary
conference, Orlando, Florida, July 13-17 1999, pp 635-642.
rather than solutions. SS-HH tested on Carter un-capacitated
998, 25, pp 637-648.
examination benchmark datasets. Experimental results show [15] Ross P., Hart E. and Corne D. Some observations about GA-
that SS-HH able to produce a comparable results when based exam timetabling. In: E. K. Burke and M. W. Carter
compared to other hyper-heuristic. (eds). Practice and theory of automated timetabling: Selected

978-1-4244-4944-6/09/$25.00 ©2009 IEEE 130


papers from the second international conference. 27 Volume
1408 of Lecture notes in computer science. Springer-Verlag,
Berlin, Heidelberg, 1998, pp 115-129.
[16] Han L., Kendall G. and Cowling P. An adaptive length
chromosome hyperheuristic genetic algorithm for a trainer
scheduling problem. In: Proceedings of the 4th Asia-Pacific
conference on simulated evolution and learning (SEAL’02),
Singapore, November 18-22, 2002, pp 267-271.
[17] Ayob M., Malik A. M. A., Abdullah S., Hamdan A. R.,
Kendall G., and Qu R. (2007). Solving a Practical
Examination Timetabling Problem: A Case Study”. In O.
Gervasi and M. Gavrilova (Eds.): ICCSA 2007, Lecture Notes
In Computer Science, Vol. 4707, Part III, Springer-Verlag
Berlin Heidelberg, 611–624.

978-1-4244-4944-6/09/$25.00 ©2009 IEEE 131

You might also like