Parallel genetic algorithms with local search

This paper presents methods of applying local search to global optimization problems. The most common approach, multistart, selects the best solution from restarts of local search from random starting points. Partitional methods augment local search with general principles concerning the location of global optima in real space, significantly improving the effectiveness of local search in function optimization problems. Standard partitional methods, however, are not directly applicable to combinatorial optimization problems. We describe a genetic algorithm, GALO, that is similar to the partitional methods, but can be applied to combinatorial problems. Empirical results are presented for a parallel implementation of GALO that show it to be effective for the quadratic assignment problem.