A Study on Possibility of Predictive Deep Reinforcement Learners for Isolated Intersection Signal Control

Abstract

Reinforcement Learning (RL) methods have been introduced to traffic control application for several decades. Traditional RL-based signal controls are mostly model-free that ignore the complex traffic states variation. Such treatment is not realistic due to external uncertainty of traffic. To fill this gap, an independent prediction module could be introduced to formulate a model-based RL. This study introduces queuing estimation models into deep-Qnetwork(DQN)-based signal control. The queuing situation could be reproduced and predicted by both input-output model and shock wave model. Through the empirical experiment, we confirm the necessity of prediction in RL-based signal control for isolated intersection.