A study on evolutionary synthesis of classifier system architectures

摘要

We describe a general method to design architectures of reinforcement learning systems. The task of these systems is to create a stimulus-response pattern by which the expected long-term total reward is maximized. Reinforcement learning systems have high applicability to a broad task class of autonomous agents because of their flexibility and autonomy. However, it is difficult to determine the relevant set of learning parameters for a given task. These parameters dominate the system architecture and largely affect the learning performance. Therefore we propose a new approach involving evolutionary synthesis of simple classifier system architectures, which is known as a genetics-based machine learning system. This synthesis mechanism is realized using genetic algorithms. To examine the validity of our proposed method, the evolutionary synthesis technique is applied to motion planning tasks of a robot manipulator.