Graph Convolution Reinforcement Learning for Decision-Making in Highway Overtaking Scenario

Abstract

Overtaking of autonomous vehicles (AVs) is an extremely complex process, which involves many factors and poses great safety hazards. However, most of the current research does not consider the impact of the dynamic environment on autonomous vehicles. In order to solve the multi-agent overtaking problem on the highway, this paper proposes a decision-making algorithm for AVs. The algorithm is based on graph neural network (GNN) and deep reinforcement learning (DRL), and adopts different training methods including as deep Q network (DQN), double DQN, dueling DQN, and D3QN for simulation. Firstly, the simulation environment is a 3-lane highway constructed in sumo. Secondly, there are both human-driven vehicles (HDVs) and AVs with maximum speeds of 10km/h and 20km/h on the highway. Finally, these two kinds of vehicles will appear in the right lane with different probabilities. The training effect is evaluated by the time it takes for the vehicle to enter and exit the current environment and the average speed of the AV. The simulation results show that the algorithm improves the efficiency of the overtaking process and reduces the accident rate.