Quantifying Distribution Shift in Traffic Signal Control with Histogram-Based GEH Distance

TL;DR

The paper addresses distribution shift in traffic signal control by introducing a histogram-based representation of intersection demand and a GEH-based distance to quantify differences between traffic scenarios. Each scenario is represented by per-movement histograms collected over time, and a bucket-level GEH-mismatch test aggregates into a single distance across all movements. Empirical validation across 20 simulated scenarios with both a NEMA actuated controller and a FRAP++ RL controller demonstrates that larger distances predict higher travel times and lower throughput, with stronger predictive power for learning-based controllers; the method also outperforms a KL-based hourly-volume baseline. The approach is interpretable, policy-agnostic, and readily usable for benchmarking, training regime design, and monitoring under evolving traffic conditions, with potential extensions to multi-intersection networks and real-world data validation.

Abstract

Traffic signal control algorithms are vulnerable to distribution shift, where performance degrades under traffic conditions that differ from those seen during design or training. This paper introduces a principled approach to quantify distribution shift by representing traffic scenarios as demand histograms and comparing them with a GEH-based distance function. The method is policy-independent, interpretable, and leverages a widely used traffic engineering statistic. We validate the approach on 20 simulated scenarios using both a NEMA actuated controller and a reinforcement learning controller (FRAP++). Results show that larger scenario distances consistently correspond to increased travel time and reduced throughput, with particularly strong explanatory power for learning-based control. Overall, this method can predict performance degradation under distribution shift better than previously published techniques. These findings highlight the utility of the proposed framework for benchmarking, training regime design, and monitoring in adaptive traffic signal control.

Figures (7)

• Figure 1: Representation of an approach of a traffic intersection. The approach has three allowed traffic movements, that vehicles choose depending on their intended turn direction. To each movement, we assign an histogram that represents the vehicle demand over time.
• Figure 2: The simulated intersection with its traffic movements, as viewed in the netedit tool of the SUMO software. Note that while SUMO represents each lane-to-lane collection as a separate movement, we group together all movements by their incoming and outgoing aproach. This means that in, this case, each approach has the three traffic movements shown in Figure \ref{['fig:traffic_movements']}
• Figure 3: Travel time as a function of GEH distance for the 20 FRAP++ models
• Figure 4: Travel time as a function of GEH distance for the 20 NEMA models
• Figure 5: $R^2$ against P value for regressions of Figure \ref{['fig:frap_multi_plot_travel_time']} and Figure \ref{['fig:nema_multi_plot_travel_time']} for GEH distance