A Game-Theoretic Reinforcement Learning Approach for Adaptive Interaction at Intersections

In this paper, we propose a bi-level algorithm for motion planning at intersections based on a scheme of reasoning game theory and heuristic reinforcement learning. In the upper level, a recurrent neural network is introduced to estimate the type of opponent agent. In the lower level, Q-networks are selectively connected to implement the game with different type. Then the ego agent could update its estimation step-by-step and conclude correspond action from historical joint state. The simulation results show that the bi-level controller improves pass times and collision avoidance performance.