UC Riverside

Traffic congestion has always been a serious problem in metropolitan areas. Fortunately, the emergence of autonomous vehicles (AVs), vehicle-to-everything (V2X) communications, and advanced machine learning algorithms have unlocked uncountable opportunities to improve transportation system management and vehicle operations in terms of safety, mobility, efficiency, and environmental sustainability. In this dissertation, agent-based traffic management strategies have been developed to mitigate traffic congestion for three representative scenarios, including: 1) signalized intersection control using deep reinforcement learning (DRL); 2) signal-free intersection management based on optimal scheduling; and c) Reservation-based Network Traffic Management with a multi-agent system (MAS) approach.

As for the first part, we focused on the signal timing problem at an isolated intersection by using deep reinforcement learning approach. The system is designed and test in a realistic transportation simulation platform (SUMO) and the intersection configuration is also developed from a real-world scenario (University Ave @ Chicago Ave). The method significantly achieved improvement at perspective of both congestion reducing and energy saving when comparing with traditional signalized intersection management methods. As for the second part, a signal free intersection management module was installed in multiple Raspberry Pi based vehicle models to solve the passing sequence at signal-free intersections in an artificial urban network. In the last part, we moved from small scale scenario(intersection) to large scale scenario (network). A reservation – based network traffic management method using MAS is proposed to route individual CAV traversing a given network in terms of minimizing its arrival time. The results show that our system can reduce travel time in the range of 8 - 12%, compared with the state-of-the-practice strategy.