Generative Adversarial Networks-Based Reinforcement Learning for Traffic Signal Control

Abstract

Reinforcement learning (RL) agents often struggles with generalization, performing poorly when encountering unseen state that do not exist during training. This limitation stems from their reliance on trial-and-error learning, where decision-making is guided primarily by reward from exprerienced states. In safety-critical application like traffic signal control, failure to handle unseen traffic scenarios can lead to suboptimal and unsafe actions. This study addresses this challenge by generating and incorporating unseen states into training of RL-based traffic signal controllers. By exposing agents with diverse and atypical traffic scenarios, the approach enchances roboustness and adaptibaility, enabling more reliable performance under real-world uncertainties. This paper presents the review of models, algorithms, applications, and open issues of both reinforcement learning and generative adversarial networks in addressing generalization to unseen state for RL model deployment.