Optimization and control of trafic signals using artificial intelligence :a deep reinforcement learning approach

Abstract

Traffic congestion is consider as a real issue especially in crowded cities and urban areas. Intel- ligent transportation systems ITS leveraged distinct advanced techniques in order to optimize the traffic flow. In particular, traffic light controllers TLCs play an essential role in ITS by controlling the traffic flow at intersections. Many researchers have been working on developing algorithms and techniques to optimize TLCs behaviour. However, most of existing techniques rely on hand-crafted designs and pre-made assumptions about the traffic conditions, restricting the controller from acting optimally. In this report, we introduce a new traffic signal controller which is based on two machine learning paradigms: reinforcement learning (RL) and deep learning (DL). We address some com- mon challenges found in the literature, mainly state design, reward formulation, and agent model architecture. We propose a novel approach to formulate a better state and reward definitions which improves the learning of the agent and subsequently pushes the performance of the traffic light con- troller further. We consider the double deep Q-Network DDQN along with prioritized experience replay with a modified reward process for the architecture of the agent. We conduct the simulations using SUMO simulator interfaced with Python framework and we compare the performance of our proposal to traditional and state-of-the-art learning-based techniques.