[4N2-IS-3b] Agents (2/2)

[4N2-IS-3b-04] Improving Generalization Capability of Traffic Light Control by Introducing Time Series Information

〇Yuta Kinouchi¹, Sachiyo Arai¹ (1. Chiba University)

キーワード：reinforcement learning, traffic signal control

In recent years, the application of reinforcement learning to traffic light control(TLC) has been studied in order to alleviate traffic congestion. In the TLC problem, the state transition probability changes as the characteristics of the traffic flow change, so the traffic handling capability is not guaranteed for traffic flows other than the one used in the training. In addition, when traffic flows with various characteristics are used as training data, the performance is degraded because the same policy is used for traffic flows with different characteristics. In this study, we propose a control method that can perceive the global characteristics of traffic flows by incorporating time series information into the state input, and acquire different policy for each characteristic of the traffic flow. As a result, we obtained a shorter average travel time than the case where only the current state is used as the state input.