Multi-Agent Deep Reinforcement Learning for Cooperative Driving in Crowded Traffic Scenarios

Jongwon Park, Kyushik Min, Kunsoo Huh

Abstract: For autonomous vehicles, lane changes on crowded roads are difficult to be performed without interactions and cooperation between vehicles. This paper proposes a novel method to learn interaction and cooperate between the multiple vehicles to solve the complex traffic problem through Multi-Agent Reinforcement Learning (MARL). The proposed network is designed based on the interaction network to learn optimal control strategies considering interaction between vehicles. By applying the proposed algorithm, the network can control and train the agents regardless of the number of agents. It is a practical advantage because the number of the vehicles is constantly changed in the real environment. The proposed method is evaluated in the connected car environment where all vehicles can exchange information with each other.