Decentralized Model-Based Policy Optimization for Cooperative Intersection Management

Traffic management at road intersections remains a challenge, even with the evolving technologies that enable autonomous vehicles to communicate and cooperate. The challenge switched to managing large amounts of data from driving agents and their surrounding environment. In this paper, we adapted the Decentralized Model-Based Policy Optimization (DMPO) [1], developed for large-scale network control, to address this challenge. DMPO, with its great scalability guarantees, is designed for large-scale static network topologies, while the traffic management problem has a large scale of autonomous vehicle nodes; each node or agent’s action decision alters the graph’s architecture, making it a dynamic graph. Our implementation fixes this by rebuilding the κ-hop communication graph at each time step based on the nodes’ proximity. We test the method in a SUMO simulation of a 4-way signalized intersection with 2 lanes per approach, 30 cars, and mixed traffic (60% CAV penetration). The results show that there were no collisions during the evaluation episodes, that the average wait time was only 0.26 seconds, and that fuel use decreased by 27.4% and emissions decreased by 24.8% compared to driving in a city with stop-and-go traffic. These findings show that DMPO is applicable beyond static networks to transportation scenarios defined by time-varying agent populations.