Evaluating the impact of data granularity on deep Q-network based smart traffic signal control

Deep Q-Networks (DQN) are a promising technology for AI-driven traffic signal control (TSC), but their training requires complex input data. Modeling can be conducted at either microscopic or macroscopic levels. While microscopic modeling captures detailed traffic dynamics, it requires extensive parameter calibration. In contrast, macroscopic modeling offers faster setup and reduced computational cost with less precision. To evaluate the trade-offs, the study compare models trained on both data types under two DQN configurations: one with fixed decision intervals, and another allowing decisions every second with enforced pauses after phase changes. All traffic data used in this study is synthetically generated using the SUMO traffic simulator, ensuring full control over experimental conditions and flow scenarios. Results show that macroscopic data enables faster convergence and comparable, if not better, performance. Although the microscopic model offers finer control, it suffers from instability when combined with coarse decision intervals. These findings highlight that high-fidelity data is not strictly necessary to train effective traffic signal control policies, which is particularly advantageous for large-scale urban simulations and city-scale digital twins.