Deep Reinforcement Learning Based Game-Theoretic Decision-Making for Autonomous Vehicles

This letter presents an approach for implementing game-theoretic decision-making in combination with deep reinforcement learning to allow vehicles to make decisions at an unsignalized intersection by use of 2D Lidar to obtain their observations of the environment. The main novelty of this work is modeling multiple vehicles in a complex interaction scenario simultaneously as decision-makers with conservative, aggressive, and adaptive driving behaviors. The game model allows anticipating reactions of additional vehicles to the movements of the ego-vehicle without using any specific coordination or vehicle-to-vehicle communication. The solution of the game is based on cognitive hierarchy reasoning and it uses a deep reinforcement learning algorithm to obtain a near-optimal policy towards a specific goal in a realistic simulator (ROS-Gazebo). The trained models have been successfully tested on the simulator after training. Experiments show that the performance of the lab cars in the real-world is consistent with it in the simulation environment, which may have great significance to improve the safety of self-driving cars, as well as may reduce their dependence on road tests.