Adaptive Traffic Signal Control : Exploring Reward Definition For Reinforcement Learning

Abstract: As mobility grow in urban cities, traffic congestion become more frequent and troublesome. traffic signal is one way to decrease traffic congestion in urban areas but needs to be adjusted in order to take into account the stochasticity of traffic. Reinforcement learning (RL) has been the object of investigation of many recent papers as a promising approach to control such a stochastic environment. The goal of this paper is to analyze the feasibility of RL, particularly the use of Q-learning algorithm for adaptive traffic signal control in different traffic dynamics. A RL control was developed for an isolated multi-phase intersection using a microscopic traffic simulator known as Paramics. The novelty of this work consists of its methodology which uses a new generalized state space with different known reward definitions. The results of this study demonstrate the advantage of using RL over fixed signal plan, and yet exhibit different outcomes depending on the reward definitions and different traffic dynamics being considered.