URBAN TRAFFIC NETWORK CONTROL SYSTEM USING ROBUST REINFORCEMENT LEARNING WITH RADIAL-DQN ALGORITHM

• Main Author: Shafly Hamzah, Muhammad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/84376
• Institution: Institut Teknologi Bandung

Urban traffic congestion is a problem that results in delays, pollution, and economic losses. Conventional traffic signal control methods are often ineffective in dealing with unpredictable conditions, especially in the presence of disturbances and inaccuracies in state observations. This research aims to address these limitations by developing a robust reinforcement learning (RL) agent using the RADIAL-DQN (Robust ADversarIAl Loss DQN) algorithm. The main objective of this study is to enhance the resilience of traffic signal control systems against environmental disturbances. The methodology involves the development of a RADIAL-DQN agent and comprehensive evaluation using the SUMO traffic simulation application. The evaluation was conducted in three main scenarios: a single intersection scenario, a seven-intersection scenario, and a Jakarta city traffic network scenario. Each scenario was tested under normal and adversarial conditions, with the application of Projected Gradient Descent (PGD) attacks to test the agent’s robustness. Analysis results demonstrate that the RADIAL-DQN agent consistently outperforms the standard DQN agent under adversarial conditions across all scenarios, showing up to a 37% improvement in queue length management. In the Jakarta scenario, Macroscopic Fundamental Diagram (MFD) analysis confirms the superiority of the robust agent in maintaining efficient traffic flow, with a 29% larger network capacity compared to the standard agent when facing adversarial attacks. This research shows that RADIAL-DQN offers a more reliable solution for urban traffic management in dynamic real-world conditions. In conclusion, RADIAL-DQN has significant potential for large-scale implementation to mitigate congestion and its negative impacts on the environment and economy. Keywords: Urban traffic management, adaptive traffic signal control, deep reinforcement learning, adversarial training, SUMO traffic simulation, RADIAL-DQN, Macroscopic Fundamental Diagram.