Simulation of Artificial Intelligence-Enabled Traffic Signal Control in Toronto, Canada, Decreased Vehicles’ Total Time Spent by Up to 19 Percent from the City's Signal Plan.
Simulation Study of Signalized Intersections along Toronto’s Sheppard Avenue Reduced Overall Network Delay by Considering Upstream Congestion.
Toronto, Ontario, Canada
Multi-Hop Upstream Anticipatory Traffic Signal Control With Deep Reinforcement Learning
Summary Information
Advanced traffic signal control is an ITS solution designed to optimize traffic flow at intersections, reduce congestion, and minimize delays. However, many of these traffic signal control strategies primarily focus on the immediate links at individual intersections and not the broader network context, which may limit their effectiveness in improving overall network efficiency. Using simulations of both simplified traffic networks and a segment of Sheppard Avenue in Toronto, this study used artificial intelligence (AI) to integrate upstream conditions, allowing controllers to prioritize traffic movements that more effectively alleviate congestion.
METHODOLOGY
This simulation study modeled traffic signal control using an AI algorithm with reinforcement learning capability. The proposed control method captured the diminishing influence of distant congestion while still accounting for its cumulative effects. The method was evaluated in a traffic simulator for both synthetic and realistic scenarios. The former were two different traffic networks at three levels of traffic demand. The latter was a simulation of 12 signalized intersections along Sheppard Avenue in Toronto, Canada. For the Toronto scenario, traffic profiles and demand were chosen and calibrated using publicly available turning movement counts and volume data. The study considered three performance evaluation metrics: total time spent, total queue time, and total virtual queue time (vehicles’ waiting time as if they are in line, even if they haven't physically joined a visible queue yet) to measure network efficiency. Results were dependent on simulated demand levels and network topology.
FINDINGS
- The proposed traffic signal control method (“3-hop”) improved total time spent by approximately 19 percent compared to the city’s signal plan (~2,303 hours versus 1,878 hours).
- Total queue time and total virtual queue time also decreased compared to the city’s plan (1,171 hours versus 901 hours and 202 hours versus 1.8 hours).
