Interntional Study Evaluated Safety and Mobility Benefits Based on Real-time Safety Optimization Using Reinforcement Learning for Adaptive Traffic Signal Control Simulated in British Columbia.
British Columbia, Canada
Real-Time Crash-Risk Optimization at Signalized Intersections
Summary Information
Adaptive Traffic Signal Control (ATSC) algorithms have been considered popular solutions to alleviate traffic congestion and optimize traffic mobility using real-time traffic data, such as data from Connected Vehicles (CV). This study proposed an ATSC algorithm for real-time safety optimization utilizing a traditional reinforcement learning approach (i.e., Q learning) and extreme value theory real-time crash prediction models which estimate traffic crashes from the extreme value distribution of traffic conflicts to define the reward and evaluate safety. The algorithm was validated using real-world traffic video data collected in 2012 from two signalized intersections in Surrey, British Columbia.
METHODOLOGY
The QASCS algorithm for real-time safety optimization was trained using a microsimulation model of a real-world intersection that estimated interactions between cycle-level dynamic traffic parameters (i.e., traffic volume, shock wave area, and platoon ratio) as covariates. The model ran simulations for 500 iterations with randomized traffic volumes, using two reward functions: Return Level of Cycle (RLC) and Risk of Collision (ROC). The ROC function identified cycles with crash-prone traffic conditions, while RLC characterized the safety levels of even safe cycles with zero ROC. RLC values greater than or equal to zero indicate a high risk of crashes, while values less than zero imply no predicted crash risk. Additionally, a comparison between the existing real-world fully Actuated Signal Control (ASC) and the trained QASCS algorithm was conducted.
FINDINGS
- The first intersection saw a reduction in cycles with ROC from 103 to 18 (82 percent reduction) and 13 (87 percent reduction) when using ROC and RLC as reward functions, respectively.
- At the second intersection, the number of cycles with ROC decreased from 139 to 22 (84 percent reduction) and 29 (79 percent reduction) with ROC and RLC as reward functions, respectively.
- QASCS, with ROC as the objective function, reduced the total travel time per vehicle by an average of 16 percent at the two intersections. Using RLC as a reward led to a 7 percent reduction. The maximum queue length, the 95th percentile of queue length, and the number of stops decreased by 14, 39, and 32 percent using ROC as a reward, and by 16, 40, and 10 percent using RLC as a reward, respectively.
- Compared to an existing fully actuated signal controller, the developed algorithm decreased real-time crash risk by 43 to 45 percent at intersection approaches with CV market penetration rates (MPR) of 50 percent. It is important to note that MPR values below 20 percent may not provide significant safety benefits due to the lack of real-time information on vehicle positions and speeds.
