How to select a winner in evolutionary optimization?

In many evolutionary optimization domains evaluations are noisy. The candidates are tested on a number of randomly drawn samples, such as different games played, different physical simulations, or different user interactions. As a result, selecting the winner is a multiple hypothesis problem: The candidate that evaluated the best most likely received a lucky selection of samples, and will not perform as well in the future. This paper proposes a technique for selecting the winner and estimating its true performance based on the smoothness assumption: Candidates that are similar perform similarly. Estimated fitness is replaced by the average fitness of candidate's neighbors, making the selection and estimation more reliable. Simulated experiments in the multiplexer domain show that this technique is reliable, making it likely that the true winner is selected and its future performance is accurately estimated.