Parameterized-action based deep reinforcement learning for intelligent traffic signal control

Abstract

Traffic Signal Control (TSC) is a crucial component in Intelligent Transportation Systems (ITS) for optimizing traffic flow. Deep Reinforcement Learning (DRL) techniques have emerged as leading approaches for TSC due to their promising performance. Most existing DRL-based approaches typically use discrete action spaces to predict the next action phase, without specifying the signal duration. In contrast, some studies employ continuous action spaces to determine signal phase timing within a fixed light cycle. To address the limitations of both approaches, we propose a flexible framework that predicts both the appropriate traffic light phase along with its associated duration. Our approach utilizes a Parameterized-action based deep reinforcement learning architecture to handle the combination of discrete-continuous actions. We evaluate our method using the Simulation of Urban MObility (SUMO) environment, comparing its efficiency against state-of-the-art techniques. Results demonstrate that our approach significantly outperforms traditional and learning-based methods.