Mätning av Distributionsförskjutning i Trafiksignalkontroll : Mått för att Bedöma Generalisering och Prestandaförsämring

Effective traffic signal control is crucial for managing urban mobility, reducing congestion, and minimizing environmental impact. Despite advancements, traditional and Reinforcement Learning (RL) Traffic Signal Control methods frequently experience performance degradation under unexpected or changing traffic conditions, known as distribution shift. This thesis represents a first step in addressing an underexplored research area: quantifying distribution shift between traffic scenarios and measuring the intrinsic difficulty of these scenarios. We propose metrics to quantify scenario-to-scenario distances, capturing an important portion of performance degradation due to distribution shifts. Among these metrics, those based on the GEH statistic—widely used in traffic modeling— demonstrate meaningful correlations with changes in travel time and throughput. Additionally, we introduce an intrinsic difficulty measure that successfully explains a significant proportion of performance variations, especially for traditional actuated control systems. Our experimental evaluation highlights these metrics as valuable tools for understanding performance impacts under distribution shifts. Although they do not fully account for all observed variations, particularly in RL-based systems, these metrics serve as a foundational step toward predictive quantification of scenario similarity and difficulty. Future research should further refine these approaches, extend them to multi-intersection networks, and validate them with empirical data.