PRGLight

Summary

PRGLight is a traffic light control algorithm. In PRGLight, we combine traffic prediction and traffic light control to jointly control the traffic light phase and traffic light duration.

The paper has been accepted by IJCAI 2021 Reinforcement Learning for Intelligent Transportation Systems (RL4ITS) Workshop.

Algorithm

• For the traffic light phase control, we design a novel pressure index, Pressure with Remaining Capacity of the Outgoing Lane (PRCOL), defined as PRCOL = N{in} * ( 1-N{out}/N{max}), where N{in} and N{out} are the number of vehicles on the incoming and outgoing lane respectively, N{max} is the capacity of the outgoing lane. We further design a Reinforcement Learning agent with PRCOL as reward. We adopt the structure of CoLight.
• For the traffic light duration control, we adopt a GNN module STGCN (paper, code). The duration is then decided by the light phase from the RL agent and the predicted traffic volume from the GNN. We design a flexible function to calculate the light duration from the real-time and future traffic condition.