A survey of reinforcement and deep reinforcement learning for coordination in intelligent traffic light control

Intelligent traffic signal control is required for a transportation system to function properly. In contrast to existing traffic signals, where rules are typically developed manually, an intelligent traffic signal control system should dynamically adapt to real-time traffic. The use of reinforcement learning for intelligent traffic signal control is a growing trend, and recent studies have shown promising results. Reinforcement learning (RL) enables a single agent to learn and perform optimal actions independently, whereas multi-agent reinforcement learning (MARL) enables traffic light controllers to learn, exchange and optimize their actions. However, none of the current studies has tested actual traffic data yet. This paper presents the primary techniques and methods (RL, DL, DRL, MARL, MADRL). The analysis of each technique, the learning of its strengths and limitations, in order to evaluate at which levels, they satisfy the requirements of urban traffic. The paper also lines some of the simulators, which perform adaptive traffic. Finally, we discuss the advantages, strengths, and weaknesses of the latest transformer models and graph neural network models.