Temporally adaptive networks: Analysis of GasNet robot ontrollers

There are immense problems in developing arti ial nervous systems for autonomous ma hines operating in non-trivial environments. In parti ular, no prin ipled methodology is in pla e to de ide between solution lasses and representations, and between methods by whi h solutions might be developed using hand-design or sear h te hniques. In this paper we apply the te hniques of dynami al systems theory to the analysis of su essfully evolved robot ontrol systems, in order to identify useful properties of the underlying ontrol ar hite ture. We investigate the suitability of two di erent neural network lasses for a roboti visual dis rimination task, through analysis of both su essful ontroller behaviour and ontinued evolution of su essful solutions in environments with modi ed hara teristi s. We argue that the temporally adaptable properties of the GasNet lass identi ed through dynami al systems analysis, and found to be useful in order to re-evolve in modi ed environments, are ru ial to the evolution of su essful ontrollers for the original environment.