Deep reinforcement learning for table tennis

Abstract

Deep Reinforcement Learning (DRL) is a powerful branch of machine learning that combines deep learning with reinforcement learning (RL) to enable intelligent agents to learn complex behaviors. DRL has achieved remarkable success in areas such as robotics, game playing (e.g., AlphaGo, OpenAI Five), autonomous vehicles, and finance. This report compares various configurations of DRL algorithms to train agents in games. This report also continues with the work by Professor Seah Hock Soon and his team, who built a novel table tennis game in the Unity environment and successfully trained agents of different proportions to perform forehand shots using DRL. We attempted to teach an agent more complicated tasks. Firstly, we experimented with backhand shots before attempting to teach an agent to perform both forehand and backhand shots appropriately. Through these attempts, we discovered more methods to improve results and training with DRL. The key contribution of this report is the discovery of methods to improve training speed and effectiveness DRL training, especially when rewards are sparse, like in the table tennis environment, which can be valuable when DRL is applied to other areas.