The University of Sussex

Temporally adaptive networks: analysis of GasNet robot controllers

Tom Smith, Phil Husbands, Michael O'Shea

There are immense problems in developing artificial nervous systems for autonomous machines operating in non-trivial environments. In particular, no principled methodology is in place to decide between solution classes and representations, and between methods by which solutions might be developed using hand-design or search techniques. In this paper we apply the techniques of dynamical systems theory to the analysis of successfully evolved robot control systems, in order to identify useful properties of the underlying control architecture. We investigate the suitability of two different neural network classes for a robotic visual discrimination task, through analysis of both successful controller behaviour and continued evolution of successful solutions in environments with modified characteristics. We argue that the temporally adaptable properties of the GasNet class identified through dynamical systems analysis, and found to be useful in order to re-evolve in modified environments, are crucial to the evolution of successful controllers for the original environment.

Download compressed postscript file