Boosting the Training of Deep Reinforcement Learning Traffic Control by Using the World Model

Deep reinforcement learning (DRL) traffic control attracts increasing interest since it is widely viewed as a powerful tool for alleviating congestion that significantly impacts people’s daily lives. However, traditional DRL models require a substantial amount of interactive data that are difficult to collect in both field and simulation tests. To overcome these challenges, this article adopts a world model for better training of DRL models for traffic control. The proposed world model can quickly learn traffic dynamics and the influence of traffic control based on limited field and simulation data. The DRL control model can then be established based on the training data generated by the world model instead. Experimental results demonstrate that DRL traffic control using world models can achieve a noticeably better signal control strategy, providing a promising solution to optimize urban traffic.