Management of Traffic Signals using Deep Reinforcement Learning in Bidirectional Recurrent Neural Network in ITS

Abstract

The traffic flow management is primarily done through traffic signals, whose inefficient control causes numerous problems, such as long waiting time and huge waste of energy. To improve traffic flow efficiency, obtaining real-time traffic information as an input and dynamically adjusting the traffic signal duration accordingly is essential. Among the existing methods, Deep Reinforcement Learning (DRL) has shown to be the most effective solution. In this paper, a dynamic mechanism to control the traffic signal of a large scale road network is proposed using policy gradient method. A single agent is trained with spatio–temporal data of the multiple intersections of the network to alleviate congestion. The proposed system is implemented in two different types of deep bidirectional Recurrent Neural Network (RNN) - Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU). The simulation experiments demonstrate that the proposed system could reduce traffic congestion in terms of different simulation metrics during high density traffic flows.