PATRAS: Pattern-Aware Generalized Traffic Signal Control Based on Parallel Learning

The data-driven adaptive traffic signal control for real-world scenarios faces two significant challenges: adapting to diverse traffic patterns and addressing the generalization issue from the optimization environment to the deployment environment. To tackle these challenges, this paper proposes a novel approach for pattern-aware generalized traffic signal control (PATRAS) based on parallel learning. It establishes an optimization process comprising descriptive learning, predictive learning, and prescriptive learning, enabling robust and adaptable control for different traffic patterns. First, a lightweight and effective method is employed to characterize various traffic patterns using traffic flow trends. Next, for each specific traffic pattern, an offline reinforcement learning algorithm is applied to optimize the traffic signal control policy directly using collected traffic data instead of simulations. A decision support library is constructed to store and reuse these optimized policies. Lastly, a pattern-matching-based approach is used to recommend appropriate control policies from the library for the traffic pattern, guiding efficient operation of the actual transportation system. Experimental results demonstrate that the offline optimization method employed in PATRAS outperforms real-world demonstrations based on actuated control policies and other baselines in terms of generalization performance. Additionally, the recommendation method effectively improves traffic efficiency by directly applying pre-optimized policies to different traffic patterns.