Iterated Local Search: Framework and Applications

The importance of high performance algorithms for tackling difficult optimization problems cannot be understated, and in many cases the most effective methods are metaheuristics. When designing a metaheuristic, simplicity should be favored, both conceptually and in practice. Naturally, it must also lead to effective algorithms, and if possible, general purpose ones. If we think of a metaheuristic as simply a construction for guiding (problem-specific) heuristics, the ideal case is when the metaheuristic can be used without any problem-dependent knowledge. As metaheuristics have become more and more sophisticated, this ideal case has been pushed aside in the quest for greater performance. As a consequence, problem-specific knowledge (in addition to that built into the heuristic being guided) must now be incorporated into metaheuristics in order to reach state-of-the-art level. Unfortunately, this makes the boundary between heuristics and metaheuristics fuzzy, and we run the risk of loosing both simplicity and generality. To counter this, we appeal to modularity and try to decompose a metaheuristic algorithm into a few parts, each with its own specificity. In particular, we would like to have a totally general purpose part, so that any problem-specific knowledge built into the metaheuristic would be restricted to another part. Finally, to the extent possible, we prefer to leave untouched the embedded heuristic (which is to be “guided”) because of its potential complexity. One can also consider the case where this heuristic is