Genetic algorithms in timetabling and scheduling

This thesis investigates the use of genetic algorithms (GAs) for solving a range of timetabling and scheduling problems. Such problems are very hard in general, and GAs o er a useful and successful alternative to existing techniques. A framework is presented for GAs to solve modular timetabling problems in educational institutions. The approach involves three components: declaring problemspeci c constraints, constructing a problem-speci c evaluation function and using a problem-independent GA to attempt to solve the problem. Successful results are demonstrated and a general analysis of the reliability and robustness of the approach is conducted. The basic approach can readily handle a wide variety of general timetabling problem constraints, and is therefore likely to be of great practical usefulness (indeed, an earlier version is already in use). The approach relies for its success on the use of specially designed mutation operators which greatly improve upon the performance of a GA with standard operators. A framework for GAs in job-shop and open-shop scheduling is also presented. One of the key aspects of this approach is the use of specially designed representations for such scheduling problems. The representations implicitly encode a schedule by encoding instructions for a schedule builder. The general robustness of this approach is demonstrated with respect to experiments on a range of widely-used benchmark problems involving many di erent schedule quality criteria. When compared against a variety of common heuristic search approaches, the GA approach is clearly the most successful method overall. An extension to the representation, in which choices of heuristic for the schedule builder are also incorporated in the chromosome, is found to lead to new best results on the makespan for some well known benchmark open-shop scheduling problems. The general approach is also shown to be readily extendable to rescheduling and dynamic scheduling. ii Acknowledgements Mostly, I would like to thank my rst supervisor Dr Peter Ross for his boundless and instructive help, criticism and patience. Without his supervision and expert knowledge in Arti cial Intelligence and especially Genetic Algorithms, this thesis would not have been possible. I would also like to thank my second supervisor Dave Corne, who introduced me to Genetic Algorithms and gave me much invaluable help, comments and also careful proof-reading. Thanks also to my former second supervisor Dr Chris Mellish for the supervison of the early parts of my project. Thanks to Drs Robert Fisher and Alan Smaill for providing the MSc examinations and lectures information. The former also provided much invaluable help throughout the timetabling part of my project and comments on parts of this dissertation. Thanks also to Howard Beck, Tim Duncan and Dr Pauline Berry (who has since left) in the Arti cial Intelligence Applications Institute who taught me about the scheduling domain through seminars and discussions, and proof-read the early chapters of this dissertation. I would also like to thank many of the colleagues and friends who have provided help, advice and companionship during my PhD study in the Department of Arti cial Intelligence. In no particular order, thanks to: Chris Gathercole (for proof-reading chapters of this dissertation), W W Vasconcelos, Rolando S. Carrera-S., Dr Kee Yin How, Albert Burger, Jessica Chen-Burger, Simon Liang, Chun Chi Wang, Ching Long Yeh and Cheng Hsiang Pan. Finally, I thank my parents, my wife Hui-Chen and my daughter Shyuan-Yih for their encouragement and company, and also thank my employer, China Steel Corporation, Taiwan, R.O.C., especially the Arti cial Intelligence/Expert Systems Projects and those who gave me the support during these last few years. iii Declaration I hereby declare that I composed this thesis entirely myself and that it describes my own research unless otherwise indicated. Hsiao-Lan Fang Edinburgh September 30, 1994 iv