Comparison of Q-Learning based Traffic Light Control Methods and Objective Functions

摘要

Traffic control is a cardinal issue in the life of urban areas. The traditional fixed duration traffic signal methods do not provide the most optimal solution while the volume of vehicles is constantly growing. Reinforcement learning is a promising approach for adaptive signal control that observe, learn, and select the optimal traffic light control action. In this paper, we present a comparison of Deep Reinforcement Learning based traffic optimization methods. This alternative way let the controllers to learn by the dynamics of the traffic and adapt to it. Our aim was to investigate the performance of advanced DQN variants in a single intersection environment. We examined four Q-Learning approaches: DQN, Double DQN, Dueling DQN and Double Dueling DQN with six different objective functions in a freely adaptable signal cycle environment. Our results show that the Double Dueling DQN based agent outperform the other models on the basis of different aspects and we can conclude about the possibilities of integration for real-life traffic management.