Comparing genetic algorithms, simulated annealing, and stochastic hillclimbing on timetabling problems

Abstract

Much recent research has focussed on applying genetic algorithms (GAs) to real educational institution timetabling problems. This work is generally successful, but it is as yet unclear whether a simpler stochastic hillclimbing (SH) strategy would generally do just as well, and how both GA and SH might compare with the use of simulated annealing (SA) on timetabling problems. We begin to investigate these concerns by comparing GA, SH, and SA on a collection of real timetabling problems. Comparisons are done in terms of final solution quality, and number of distinct solutions obtained. When considering the latter criterion, we necessarily compare the GA with modified SH and SA algorithms which continually restart to look for new solutions. The main conclusions are that SH and SA are generally the best strategy as far as solution quality is concerned. For a certain fairly small range of problems though, the GA either betters or equals the performance of SA and SH, but delivers the added value of a large number of usefully distinct, equally good solutions. Finally, we note that our results are to be taken in the context of particular implementations of SA, SH, and GA; although steps are taken to optimise parameters and such for each implementation, different conclusions may have been reached if, in particular, we had used more sophisticated SA cooling schedules, and/or more sophisticated GA operators. Such complexities concerning GA/SA comparisons in general are discussed.