Hard test generation for maximum flow algorithms with the fast crossover-based evolutionary algorithm

Abstract

Most evolutionary algorithms not only throw out insufficiently good solutions, but forget all information they obtained from their evaluation, which reduces their speed from the information theory point of view. An evolutionary algorithm which does not do that, the (1 + (λ; λ)) EA was recently proposed by Doerr, Doerr and Ebel. We evaluate this algorithm on the problem of finding hard tests for maximum flow algorithms. Experiments show that the (1 + (λ; λ)) EA is never the best, but is quite stable. However, its adaptive version, known for being superior for the OneMax problem, is shown to be one of the worst.