Multi-agent Traffic Light Controls with SUMO

This study focuses on the optimization of Traffic Light System (TLS) control through the use of adaptive agents. The performance of adaptive cycle TLS was compared with fixed cycle TLS. Two different adaptive cycle TLS agents were investigated, reactive and Deep Q-Network (DQN) agents. The reactive agent adjusted its control signals based on traffic measures such as queue length, while the DQN agent employed a deep reinforcement learning algorithm to determine its control signals. Two sets of simulations were conducted to evaluate the performance of both approaches. The results showed that the adaptive cycle TLS was more effective in reducing waiting times and increasing traffic throughput than the fixed cycle TLS. Among the adaptive agents, the reactive agent outperformed the DQN agent, due to the difficulty in learning an optimal policy in the traffic control domain, which has a non-stationary and complex nature. This preliminary study showed the potential benefits of using adaptive agents for traffic light control, and further studies in various areas such as employing more advanced TLS control methods, expanding the scale of the study, and applying real-demand in the simulation, could be carried out in future work.