Simulating exploratory behavior in evolving Artificial Neural Networks

Abstract

Animals released into unfamiliar environments will often engage in roaming behavior, apparently for exporatory purposes. It is likely that this behavior constitutes a behavioral primitive which can be used in the construction of more complex behaviors. This paper reports a series of experiments in which a Genetic Algorithm is successfully used to evolve efficient exporation strategies in a population of software simulated Khepera robots, controlled by Artificial Neural Networks. Robots based on simple perceptrons with no hidden neurons outperformed those with more complex contro networks. These robots tended however to adapt to the specific environments where they had evolved. More robust behavior was obtained from robots where input from the external environment was enriched with data from cyclical time sensors. It is suggested that control-networks based on this architecture could become a useful component in more complex systems.