A decentralized control method in which each autonomous vehicle acts as an independent reinforcement-learning agent, trained with Multi-Agent Deep Deterministic Policy Gradient (MADDPG) to decide when to change lanes or keep its lane in order to avoid dynamic roadblocks. The decision problem is cast as a Markov game whose reward maximizes the mean harmonic speed of traffic subject to minimum-speed and lane-change-frequency constraints, and vehicles exchange state over 6G-V2X links so that centralized training yields policies executed in a decentralized way.
This paper introduces a multi-agent deep reinforcement learning controller for autonomous vehicles navigating highways with dynamic roadblocks such as accidents or maintenance zones. Each vehicle is …