[2F1-E-3] Agents: interaction and decision

[2F1-E-3-05] Multi-Agent Traffic Signal Control System Using Deep Q-Network

〇YOHEI KANZAKI¹, Keisuke Ohno¹, Eichi Takaya¹, Satoshi Kurihara¹ (1. Keio University)

Keywords:Multi-Agent, Deep Reinforcement Learning

In urban areas, temporal and economic losses due to traffic congestion are getting worse. It has a great influence on our lives. As a cause of traffic congestion, on ordinary road, inappropriate signal switching may be cited. Parameter manipulation in the general signal control is set based on experiences by human hands, and it is never optimal. Therefore, controlling traffic lights to improve traffic behavior is possible way to solve traffic congestion. In this study, we combine multi agent system with Deep Q-Network method. We used an intersection as an agent and conducted experiment in road environment with multiple intersections. As a result, it was shown that agents can perform appropriate parameter manipulation by mutual exchange of information among agents.