<![CDATA[Traffic congestion has become a major challenge in Indonesian urban areas due to rapid vehicle growth and the limited adaptability of conventional traffic signal control systems. Most existing Deep Reinforcement Learning (DRL)-based traffic signal control studies adopt a free-phase selection approach, which assumes full agent freedom in determining signal phases — an assumption fundamentally incompatible with fixed phase-sequence regulations in Indonesian urban infrastructure — and rely on synthetic traffic data that fails to represent motorcycle-dominated traffic conditions. Furthermore, existing DQN-based approaches treat all traffic density conditions uniformly, without utilizing IoT-derived density categories for context-aware decision-making. To address these gaps, this study proposes a manual phase rotation mechanism with constrained actions (15, 30, and 60 seconds) compatible with existing fixed-phase infrastructure without hardware modifications, real-world IoT CCTV data from four intersections in Metro City processed using the YOLOv11 model to generate Low, Medium, and High traffic density categories as a representative training foundation for Indonesian urban traffic conditions, and a category-based action bias mechanism that adjusts DQN Q-value estimates according to IoT-derived traffic density, enabling context-aware signal duration selection. The DQN agent interacts with the SUMO simulation environment through the TraCI interface, receiving real-time traffic states comprising vehicle count, queue length, waiting time, average speed, density category, and delta queue, and selecting optimal green signal durations based on an epsilon-greedy exploration strategy and experience replay mechanism over 1,100 training episodes. Training yielded a 39.2% improvement in total reward and a 6.6% reduction in average waiting time. The best-performing model, obtained at episode 1050, achieved an 8.6% reduction in average waiting time and an 11.7% increase in traffic throughput compared to the fixed-time baseline. These results demonstrate that the proposed framework contributes three concrete advances for adaptive traffic signal control, a constrained-action DQN that is fully compatible with real-world fixed-phase infrastructure, a real-world IoT CCTV dataset as a representative data foundation for Indonesian traffic conditions, and a category-based bias mechanism for context-aware control — collectively offering a deployable, infrastructure-compatible, and replicable solution for traffic authorities and local governments advancing the smart city agenda in Indonesia.]]>
Copyrights © 2026
