Multi-Agent Chance-Constrained Reinforcement Learning for Autonomous Vehicles

Abstract

Autonomous driving involves negotiating with other road users in various scenarios such as merging, overtaking, turning, and navigating unstructured urban roads. However, manually addressing all possible cases can result in a simplistic policy, and traditional methods such as optimization techniques or model-based solvers may not be feasible in real-time or may necessitate a comprehensive model to create a policy. Model-free deep reinforcement learning (DRL) offers a potential solution to these challenges. This study presents a multi-agent deep reinforcement learning approach with chance constraints for multiple autonomous vehicles in a simulated environment. The approach is formulated as a chance-constrained multiAgent Markov decision process (CC-MMDP) that limits the probability of being in a risky state by a specified threshold. The proposed approach holds promise for enhancing safety in autonomous driving while ensuring smooth traffic flow. Future work involves evaluating the approach in more complex scenarios with intersections and more autonomous vehicles.

Date of Award	Aug 2023
Original language	American English
Supervisor	Majid Khonji (Supervisor)