Learning to Box: Reinforcement Learning using Heuristic Three-step Curriculum Learning

The reinforcement learning paradigm is a widely used approach to solving sequential problems. In this paper, we utilized reinforcement learning with curriculum learning to train the agent to play boxing without any human demonstration. We presented as a curriculum three steps for a single agent to learn boxing: standing, walking and punching. The agent who learned through the proposed curriculum successfully threw a jab into a punching bag, but the agent who trained to punch from the beginning did not. Not only did we show the effectiveness of curriculum learning in single-agent boxing, but we also showed that the agent could play boxing without human demonstrations.