A new traffic signaling model based on graph and deep reinforcement learning

Abstract

In this study, a new traffic signaling methodology based on Graph and Deep Reinforcement Learning is proposed to provide solutions to traffic congestion at intersections. A distributed signaling system has been developed by combining the fast and accurate calculation ability of the Ford-Fulkerson graph algorithm with the dimensionality and generalization capabilities of DRL. The developed signaling system was modeled on SUMO using the real map model and real vehicle data. It was trained and tested on SUMO. The flow chart of the proposed method is given in Figure A. Figure A. Proposed Graph based DRL method Purpose: It is aimed to increase the transit efficiency at intersections by reducing the average queue length and total waiting times at intersections. Theory and Methods: Within the scope of the study, a new traffic signaling based on Graph and Deep Reinforcement Learning is proposed. This method calculates the phase sequence by DRL and the duration by the Ford Fulkerson graph method. Results: The signaling model obtained as a result of training using the real map and real data reduced the queue length by 45% compared to the fixed time and sequential phase method. Conclusion: In this study, a new signaling method has been developed based on distributed architecture without the need for a central management system, increasing signaling efficiency.