Intrinsic Rewards for Maintenance, Approach, Avoidance, and Achievement Goal Types

摘要

In reinforcement learning, reward is used to guide the learning process. The reward is often designed to be task-dependent, and it may require significant domain knowledge to design a good reward function. This paper proposes general reward functions for maintenance, approach, avoidance, and achievement goal types. These reward functions exploit the inherent property of each type of goal and are thus task-independent. We also propose metrics to measure an agent's performance for learning each type of goal. We evaluate the intrinsic reward functions in a framework that can autonomously generate goals and learn solutions to those goals using a standard reinforcement learning algorithm. We show empirically how the proposed reward functions lead to learning in a mobile robot application. Finally, using the proposed reward functions as building blocks, we demonstrate how compound reward functions, reward functions to generate sequences of tasks, can be created that allow the mobile robot to learn more complex behaviors.