Abstract: | In this paper we argue for a division of the evaluation of robot learning along several lines. First, both external and internal measures of a robot's performance are required. Second, both quantitative and qualitative methods of description should ideally be employed. Four methods for agent analysis are then presented. These are drawn from a variety of fields: psychology, ethology, machine learning, statistics and engineering. We describe the application of these techniques to the construction and analysis of a robot that learns to push boxes from reinforcement. the paper concludes with general remarks about the efficacy of each of these techniques. We also emphasize the importance of careful experimental design in order to integrate the various forms of evaluation.
|