Hyper-heuristics: Learning To Combine Simple Heuristics In Bin-packing Problems

Evolutionary algorithms (EAs) often appear to be a 'black box', neither offering worst-case bounds nor any guarantee of optimality when used to solve individual problems. They can also take much longer than non-evolutionary methods. We try to address these concerns by using an EA, in particular the learning classi£er system XCS, to learn a solution process rather than to solve individual problems. The process chooses one of various simple non-evolutionary heuristics to apply to each state of a problem, gradually transforming the problem from its initial state to a solved state. We test this on a large set of one-dimensional bin packing problems. For some of the problems, none of the heuristics used can £nd an optimal answer; however, the evolved solution process can £nd an optimal solution in over 78% of cases.