Evaluation of learning performance of situated embodied agents

Abstract

This paper discusses the complexities of designing and evaluating Alife learning systems using physical robots interacting in complex, dynamic environments. We use the learning data from two different implemented multi-robot learning systems to illustrate the difficulties with traditional "objective" methods of evaluation. We then describe the methods we used to evaluate the behavior of such learning systems in order to preserve its dynamics and demonstrate their effect on the performance.