A Deep Reinforcement Learning Approach for Isolated Intersection Traffic Signal Control with Long-Short Term Memory Network

In this paper, a DRL algorithm based on long-short time memory (LSTM) network is proposed for the signal control problem of isolated intersection. The LSTM network is used to learn the sequence features of the state space, and a dueling network is used to separate the state and the action to a certain extent, so that the calculation of the state value function is no longer completely dependent on the action value. Setting the weighted sum of the average delay of the intersection and the average travel time of the vehicles as the reward function, we define the adjustment of the green time of current phase as the action. With the goal of maximizing the cumulative reward obtained by taking actions, the duration of signal phases is adjusted. Finally, an isolated intersection is taken as an example to test the effectiveness of the proposed method in Simulation of Urban Mobility.