Controller Module Planning Module Learning Module Controller Module Planning Module Learning Module Controller Module Planning Module Learning Module Environment

Most learning systems applied to problem solving have the goal of learning knowledge for solving problems more eeciently. However, very few systems have focused on the acquisition of planning operator descriptions. In this paper, we present a system that learns operator deenitions from the interaction with a completely unknown environment. In order to achieve better learning convergence, several agents that learn separately are allowed to interchange each learned set of planning operators. Learning is achieved by establishing plans, executing those plans in the environment, analyzing the results of the execution, and combining new evidence with prior evidence. Operators are generated incrementally by combining rote learning, induction, and a variant of reinforcement learning. The results clearly show that the integrated planning, learning, and executing system outperforms the basic planner with individual learning techniques in a robot environment. They also show how allowing the communication among individual learning (and planning) agents provides a much better percentage of successful plans, plus an improved convergence rate than the individual agents alone.