A ‘Fitness Landscaping’ Comparison of Evolved Robot Control Systems

Sign In

Full text access may be available.

To access full text, please use your member or institutional sign in.

The purchase and pricing options for this item are unavailable. Select items are only available as part of a subscription package. You may try again later or contact us for more information.
4 Author(s)

Most work in evolutionary robotics has focussed on evolving neural network based control systems, rather than high level control programs using genetic programming. One reason for this is that the fitness landscapes generated by a genetic programming framework may be poorly suited to the evolutionary techniques employed. The aim of this paper is to investigate this claim. The first part of the paper demonstrates two simulations of evolving populations of simple robots, one based on genetic programming, the other based on evolved neural nets. The latter part of the paper then introduces a technique for ‘fitness landscaping’, that is for constructing 3-D visualisations of the sorts of complex, high dimensionality fitness landscapes involved. This technique is applied to the results from the two siniulations, and the resulting representations are discussed