A Comparison of Evolutionary Methods for the Discovery of Local Search Heuristics

Methods of adaptive constraint satisfaction have recently become of interest to overcome the limitations imposed on “black-box” search algorithms by the no free lunch theorems. Two methods that each use an evolutionary algorithm to adapt to particular classes of problem are the CLASS system of Fukunaga and the evolutionary constraint algorithm work of Bain et al. We directly compare these methods, demonstrating that although special purpose methods can learn excellent algorithms, on average standard evolutionary operators perform even better, and are less susceptible to the problems of bloat and redundancy.