Autonomic Learning Framework Based on Behavior Policy

Abstract

In this paper, a novel autonomous learning framework is proposed. Different form the traditional learning framework, which maps environment to the action space directly, the new method abstracts behavior policies from the demonstrations and expresses it with a parameterized model. Based on the parameterized policies, cost function and environment constraints are generated, which together with the dynamic constraints of the robot are used to optimize the decision actions. A simulation on the grid world proved the effectiveness of the proposed method.