Design, analysis and comparison of robot learners

This paper outlines some ideas as to how robot learning experiments might best be designed. There are three principal ndings: (i) in order to evaluate robot learners we must employ multiple evaluation methods together; (ii) in order to measure in any absolute way the performance of a learning algorithm we must characterise the complexity of the underlying decision task formed by the interaction of the agent, task and environment; (iii) that in fact this goal is too diicult to attain in practice, and progress in robot learning must rely on comparative work. Four methods for agent analysis are presented. These are used to analyse a robot that learns to push boxes from reinforcement. Using these techniques we have been able to show that Q() learning outperforms one step Q-learning on a typical robot task. The diierences are statistically signiicant. We emphasize the importance of experimental design in order to integrate the various forms of evaluation.