An Emotion-Based Approach to Reinforcement Learning Reward Design

Reward is one of the crucial factors in reinforcement learning, which affects the improvement of control strategies. However, the role of reward design has received relatively little attention. In this paper, an emotion-based target reward function is proposed which requires the agent to possess the ability to reflect. In this approach, the learning process information of the agent is mapped to its internal changes in any episode. The difference in internal values of adjacent episodes induces the generation of the agent's emotions, which is a key way to assist the agent to internally measure preset external target reward. Our proposed approach is combined with traditional RL algorithms (i.e., Q-learning, Sarsa and Q(λ)-learning) to test its effectiveness. All experimental results show that emotion-based target reward can accelerate the learning process.