A Comprehensive Survey on Value-Based and Policy-Based Reinforcement Learning for Traffic Optimization

Traffic congestion is an important issue in urban traffic, causing longer travel time, more fuel consumption, and more pollution. Conventional traffic signal control techniques such as fixed-timing or predefined rules are unable to respond to dynamic traffic, and therefore, requires innovative approach for optimize the traffic signal control. Reinforcement learning (RL)-based methods proved to be a promising solution, which allowed adaptive, self-improving traffic signal control for the optimization of traffic signal control through real-time data. This survey classifies the development of RL-based traffic signal control with emphasis on value-based and policy-based learning methods. These techniques have been exhaustively tested on real and simulation data to analyse their potential for optimizing traffic flow. The Finding suggest that policy-based methods perform better than value-based methods in dynamic and multi-agent traffic environments. This article highlights the main challenges and provides suggestions for further research focused on enhancing adaptive traffic control systems.