Behavior evaluation and learning from an internal point of view.

摘要

This paper describes a new approach that makes a robot learn by evaluating its own performance based on the use of its resources. For a behavior-based robot, this means that learning is accomplished from the observation of behavior use over time. The acquired knowledge can then be exploited for future selection of behaviors. When applied to the multirobot domain, this approach will make the robot find regularities in its interactions with its environment and exploit them efficiently, eventually resulting in specialization within the group. 1 INTRODUCTION Learning can be used in many ways to improve the ability of a robot to behave in its environment, such as by acquiring knowledge about the world (e.g., using a topological representation) [6, 9], by acquiring special skills (behaviors or control policies) [4, 10] or by learning when to use such skills according to sensed conditions [3, 7]. Learning in the mobile robot domain is challenging. First, there are difficulties intrinsic to t...