Tom Smith, Phil Husbands, Michael O'Shea
The majority of measures developed to predict the difficulty of finding good solutions to a given problem through search focus on epistasis, while other features which affect problem difficulty such as landscape neutrality, may be ignored completely. In this paper we develop a set of measures based on evolvability statistics of the fitness landscape surrounding sampled solutions. Averaging the measures over a sample of equal fitness solutions allows us to build up a fitness portrait of the fitness landscape, which we show can be used to predict the difficulty of finding good solutions in a set of tunably rugged and tunably neutral landscapes.
Download compressed postscript file