Decision Making in Autonomous Driving by Integrating Rules with Deep Reinforcement Learning

Abstract

Human error is the main contributing factor to traffic accidents. The advancement of autonomous driving has a great potential to improve road safety. As a promising decision-making technique, reinforcement learning has been research extensively in the autonomous driving domain. However, reinforcement learning suffers from safety concerns arising from exploration during training and unpredictable behavior when testing in unknown environments. This thesis combines reinforcement learning with a well-defined rule-based method, which assists a vehicle prior to a potential collision in a pedestrian crossing scenario. The proposed algorithm takes into consideration of safety, efficiency, and comfort simultaneously by expressing these requirements as reward functions. The proposed approach was studied experimentally in four different training scenarios in a simulated environment. The experimental results showed that the algorithm has learned to execute longitudinal control when uncertainty is introduced to the environment. In addition, the proposed algorithm learns to prevent collisions both during training and testing.