Item

Abstract

Rigorous runtime analyses of evolutionary algorithms (EAs) mainly investigate algorithms that use elitist selection methods. Two algorithms commonly studied are Randomized Local Search (RLS) and the (1+1)~EA and it is well known that both optimize any linear pseudo-Boolean function on $n$ bits within an expected number of $\ensuremath{{O}}(n \log n)$ fitness evaluations. In this paper, we analyze variants of these algorithms that use fitness proportional selection. A well-known method in analyzing the local changes in the solutions of RLS is a reduction to the gambler's ruin problem. We extend this method in order to analyze the global changes imposed by the (1+1)~EA. By applying this new technique we show that with high probability using fitness proportional selection leads to an exponential optimization time for any linear pseudo-Boolean function with non-zero weights. Even worse, all solutions of the algorithms during an exponential number of fitness evaluations differ with high probability in linearly many bits from the optimal solution. Our theoretical studies are complemented by experimental investigations which confirm the asymptotic results on realistic input sizes.