ABSTRACT Urban traffic systems exhibit complex and variable spatiotemporal characteristics, whereas traditional model‐based traffic signal control algorithms often lack the necessary flexibility. In response to this issue, researchers have developed traffic signal control algorithms on the basis of deep reinforcement learning (DRL). However, these DRL signal control algorithms are primarily trained and tested in traffic simulation environments due to safety considerations, and there is a significant lack of in‐depth discussion on how to effectively implement them in real‐world scenarios. This article first analyses the discrepancies between traffic simulation environments and real‐world settings, assessing how these differences impact the performance of DRL traffic signal control. Subsequently, it proposes a method based on pre‐trained DRL model selection and fine‐tuning, which constructs a library of pre‐trained DRL models, selects the most suitable pre‐trained model for specific intersection scenarios and performs targeted fine‐tuning. This approach enhances the applicability of DRL traffic signal control in real‐world environments while simplifying model training and online parameter update processes.
Li et al. (Thu,) studied this question.