Transfer Learning Method in Reinforcement Learning-based Traffic Signal Control

Abstract

Traffic signal control is becoming more important in intelligent transport systems. Existing studies managed to increase the traffic efficiency on the assumption of a stable traffic environment where no emergencies occur. However, accidents and road closures happen from time to time, and the existing studies cannot guarantee efficiency when such temporary changes happen in the road conditions. Thus, we designed a transfer learning method for existing reinforcement learning-based traffic signal control systems. Our proposed method uses parameters from the previous training model to initialize the new model to increase its initial performance, thus speeding up the learning process and reducing the time needed to adapt to road condition changes.