Capacity Region Evaluation for Traffic Signal Control

Performance comparisons between traffic signal controllers are often conducted without
checking whether a controller can handle the imposed traffic demand. When arrival rate exceeds a
controller’s effective service capacity, queues grow without bound and corrupt performance statistics.
In this work, we study the capacity region of a signalized intersection under stochastic arrivals. Using
SUMO, we evaluate a Fixed-Time (FT) controller and a Deep Q-Network (DQN) reinforcement learning
(RL) controller, and empirically identify their capacity regions by detecting whether queue lengths remain
bounded or grow over time under fixed arrival rates. The results show that the RL controller operates
stably over a larger capacity region than traditional methods, and achieves lower average travel time
and shorter queue lengths within the shared feasible region, without exceeding the theoretical maximum
service capacity.