In a Simulation for Multi-Intersection Traffic Light Control, Average Waiting Time was Reduced by 6.42 Seconds and Average Speed Increased by 1.57 m/s for the Trained Controller versus the Fixed-Timing Controller.
Traffic Controllers were Trained with a Deep Q-Network with an Auto-Encoder to Evaluate its Effectiveness in Managing Traffic Congestion at Intersections in California.
California
Addressing Urban Traffic Congestion: A Deep Reinforcement Learning-Based Approach
Summary Information
Urban traffic congestion profoundly impacts the economy public health, and overall quality of urban living. The dynamic nature of traffic requires traffic management solutions that can monitor, analyze, and intervene in real time.
Traffic Management Systems (TMS) present a modern approach to resolving the long-existing problem of urban traffic congestion. TMS integrates state-of-the-art technology, real-time data analytics, and strategic traffic engineer to offer a comprehensive solution to improve traffic flow, reduce bottlenecks, and enhance the overall efficiency of urban transportation networks.
The goal of the research was to explore and quantify effectiveness of TMS and Vehicle-to-Everything (V2X) communication technologies mitigating urban traffic congestion in California. The research focuses on using reinforcement learning algorithms to optimize traffic signal timings. HyperOPT was also utilized to navigate the hyperparameter space effectively to find the optimal solution. Simulations were conducted for the different case studies using Simulation of Urban MObility (SUMO) to model the intersections.
METHODOLOGY
For single-intersection traffic light control, extensive analysis, conducted over 400 episodes using the SUMO traffic simulation, focused on evaluating the effectiveness of three distinct methodologies in managing traffic congestion measured by the average waiting time of vehicles. These methodologies include a pre-existing implementation of traffic management, Deep Q-Network with an auto-encoder, and a naive method devoid of deep learning techniques, where the traffic signals have been programmed with a fixed time interval for each of the phases.
FINDINGS
- The proposed method was tested on a multi-intersection scenario, and its effectiveness was validated. The performance of the trained controller and a fixed timing controller (a pre-timed traffic signal controller where each phase has a predetermined length) was compared in terms of vehicles’ average waiting time and speed.
- With the trained controller, the vehicles’ average waiting time was reduced to 0.59 seconds and the average speed was increased to 7.77 m/s compared to 7.01 seconds and 6.20 m/s respectively with the fixed-timing controller.
- The results indicate that the proposed method works for multiple intersection scenarios and can achieve better performance compared to the traditional fixed-timing traffic signal controllers.
