Signal Optimization via Heuristic Search and Traffic Simulation

Abstract

Traffic congestion is a severe problem in many cities. One way to reduce it is by optimizing traffic signal timings. Experts spend a lot of time analyzing traffic patterns to produce good handcrafted timing schedules. However, these timing schedules can be less responsive when there is a sudden change in traffic flow. In this thesis, a novel way to formulate the traffic signal optimization problem as a signal-player game is proposed. The model uses a heuristic search algorithm called Monte Carlo Tree Search (MCTS) which is incorporated with a traffic simulator called Simulation of Urban MObility (SUMO) to approximate the optimal traffic signal timings. The model is tested against handcrafted timing schedules across different types of road networks, such as interconnected intersections, a long corridor of intersections, and intersections with Light Rail Transit (LRT) crossings. Experimental results show that our model performs significantly better in most cases when compared to our handcrafted policies. For instance, in one of the networks, our search model outperforms the handcrafted policy by 29% in all performance measures we considered. Moreover, in a real-world scenario with LRT crossings, MCTS surpassed the handcrafted policy by 18%. The strength of our model is that it can foresee changes in traffic flow patterns through simulations and react accordingly. Therefore, MCTS along with simulations is a viable alternative to experts handcrafting traffic light timing policies manually.