Traffic Signal Timing via Parallel Reinforcement Learning

Nowadays, reinforcement learning is widely used to design intelligent control algorithms, which has gradually become one of the popular methods of signal control. We propose a new traffic signal control method, which applies parallel reinforcement learning methods to build a traffic signal control agent in the traffic micro-simulator. This method uses covariance adaptive matrix evolution strategy (CMA-ES) algorithm to train our system on computer cluster, with over 300 iterations and 500 populations in each iteration. In this paper, we provide preliminary results on how the parallel reinforcement learning methods perform in traffic signal control system.