A Deep Reinforcement Learning Framework with Memory Network to Coordinate Traffic Signal Control
IEEE Conference on Intelligent Transportation Systems, Proceedings, ITSC
Traditional traffic signal control(TSC) methods are difficult to adapt to dynamic traffic conditions. Nowadays, many people apply deep reinforcement learning (DRL) to TSC. However, the current widely used original DQN network in TSC does not consider the history states and action information that indeed have a certain impact on the state and action value function's prediction. Against this, we propose a framework named Memory Network Light (MNLight), which takes the history information into consideration through adding LSTM dual branches in DQN structure. Through comprehensive experimental evaluation, MNLight has been proven to be superior to the existed well-known traffic signal control methods.
Open Access Status
This publication is not available as open access