Towards Behavioral Consistency in Neuroevolution
Charles Ollion, Stéphane Doncieux · 2012
Abstract. To survive in its environment, an animat must have a be-havior that is not too disturbed by noise or any other distractor. Its behavior is supposed to be relatively unchanged when tested on similar situations. Evolving controllers that are robust and generalize well over similar contexts remains a challenge for several reasons. One of them comes from the evaluation: how to check a controller for such proper-ties? The fitness may evaluate a distance towards a behavior known to be robust, but such an example is not always available. An alternative is to test the behavior in multiple conditions, actually as many as possible, to avoid overfitting, but this significantly slows down the search pro-cess. This issue is expected to become even more critical when evolving behaviors of increasing complexity. To tackle this issue, we propose to formulate it as a problem of behavioral consistency in different contexts. We then propose a fitness objective aimed at explicitly rewarding behav-ioral consistency. Its principle is to define different sets of contexts and compare the evolved system behavior on each of them. The fitness func-tion thus defined aims at rewarding individuals that exhibit the expected consistency. We apply it to the evolution of two simple computational neuroscience models. 1