On the performance of Scatter Search for post-enrolment course timetabling problems

This study presents an investigation of enhancing the capability of the Scatter Search (SS) metaheuristic in guiding the search effectively toward elite solutions. Generally, SS generates a population of random initial solutions and systematically selects a set of diverse and elite solutions as a reference set for guiding the search. The work focuses on three strategies that may have an impact on the performance of SS. These are: explicit solutions combination, dynamic memory update, and systematic search re-initialization. First, the original SS is applied. Second, we propose two versions of the SS (V1 and V2) with different strategies. In contrast to the original SS, SSV1 and SSV2 use the quality and diversity of solutions to create and update the memory, to perform solutions combinations, and to update the search. The differences between SSV1 and SSV2 is that SSV1 employs the hill climbing routine twice whilst SSV2 employs hill climbing and iterated local search method. In addition, SSV1 combines all pairs (of quality and diverse solutions) from the RefSet whilst SSV2 combines only one pair. Both SSV1 and SSV2 update the RefSet dynamically rather than static (as in the original SS), where, whenever a better quality or more diverse solution is found, the worst solution in RefSet is replaced by the new solution. SSV1 and SSV2 employ diversification generation method twice to re-initialize the search. The performance of the SS is tested on three benchmark post-enrolment course timetabling problems. The results had shown that SSV2 performs better than the original SS and SSV1 (in terms of solution's quality and computational time). It clearly demonstrates the effectiveness of using dynamic memory update, systematic search re-initialization, and combining only one pair of elite solutions. Apart from that, SSV1 and SSV2 can produce good quality solutions (comparable with other approaches), and outperforms some approaches reported in the literature (on some instances with regards to the tested datasets). Moreover, the study shows that by combining (simple crossover) only one pair of elite solutions in each RefSet update, and updating the memory dynamically, the computational time is reduced.