Partial Intention Encoding for Safe Multi-Agent Control via Deep Reinforcement Learning

Autonomous vehicles can improve urban traffic safety and efficiency. Nevertheless, safely and efficiently operating them is an ongoing research problem. This paper specifically studies the task of navigating autonomous vehicles through unsignalized intersections. We model the system as a Decentralized Partially Observable Markov Decision Process. Decentralized algorithms for each vehicle are trained using deep reinforcement learning through the QMIX method. Building upon the QMIX approach, we encode the intentions of a selected number of neighboring vehicles. Moreover, the reward function is designed to combine both individual performance (i.e., fuel consumption, speed, and trip completion) with global safety and fairness. This work shows improved performance compared to an existing conventional driving algorithm in reward, driving speed, and safety.