Not measuring evolvability: initial investigation of an evolutionary robotics search space

Abstract

Investigates the underlying search space of a difficult robotics problem. Previous work (P. Husbands et al., 1998) on the development of neural networks incorporating a model of gaseous neuromodulation (the GasNet) suggested that such networks are well-suited to evolutionary design for some problems. Networks that are allowed to use the gaseous signalling mechanism evolved significantly faster than networks with the mechanism disabled, implying a significant difference between the two search spaces. In this paper, we investigate this difference using a series of standard techniques for predicting the "difficulty" of searching in fitness landscapes. We show that, in this instance, measures based on random sampling do not discriminate between the two search spaces, due to the highly skewed nature of the fitness distributions, similar to those found in other difficult optimisation problems. It may be that such metrics are not useful as measures of difficulty for a class of complex problems.