The University of Sussex

Evolving neural network controllers for task defined robots

Kyran Dale

Some recent attention in Artificial Intelligence (AI) research (specifically the sub-discipline known as Artificial Life) has been focussed on the possibility of using genetic algorithms to vevolve neural network controllers for task-defined robots. Employing techniques formalised by Holland (1975), the hope is that by using various encoding methods for representing a neural network on a 'genome' - commonly a binary string - and then manipulating a population of these genomes using, primarily, cross-over and mutation operators according to fitness-preferential dictates, one may efficiently search a large parametric state-space for useful networks. This paper deals with my attempt to evolve a neural network that, by mediating between a simulated robot's actions and its environmental input leads to a 'guard-dog' behaviour.

Download compressed postscript file