Learning to Race with Multi-Agent Deep Reinforcement Learning

Abstract

Autonomous driving cars are important due to improved safety and fuel efficiency. Various techniques have been described to use only a single task, for example, recognition, prediction, and planning with supervised learning techniques. However, most previous studies had limitations: (1) human bias from human demonstration; (2) multiple components such as localization, road mapping etc. to build a powerful autonomous car, with complicated logic; (3) for reinforcement learning, they focused on building a powerful model, but did determine, which perceptions or sensors were useful for an autonomous car, and how to design the best reward for driving a car. We describe end-to-end reinforcement learning for an autonomous car, which used only a single reinforcement learning model to create the autonomous car. Further, we designed a new efficient reward function to make the agent learn faster (18% improvement for all settings compared to the baseline reward function) and build the car with only the necessary perceptions and sensors. In addition, we also combined our propose for single agent into multi-agent system by modifies some reward function to make it work too. We show that it performed better with state-of-the-art off-policy reinforcement learning for continuous action (SAC, TD3).