Multi-Point Constructive Search: Extended Remix

Multi-Point Constructive Search maintains a small set of “elite solutions” that are used to heuristically guide constructive search through periodically restarting search from an elite solution. Empirical results indicate that for job shop scheduling optimization problems, multi-point constructive search leads to significantly better solutions for equivalent search effort when compared to chronological backtracking and bounded backtracking with random restart. For satisfaction problems (quasigroup with holes completion), significant reduction in the magnitude of mean search effort (the number of fails and run-time) is also achieved versus chronological backtracking and bounded backtracking with random restart. Two conjectures about the relationship between the clustering of good solutions in a search tree and the performance of multi-point constructive search are made. Preliminary empirical results are consistent with the conjectures, suggesting directions for future work to develop a deeper understanding of the observed performance.