REINFORCEMENT LEARNING-BASED TRAFFIC LIGHT DURATION CONTROL SYSTEM DEVELOPMENT

• Main Author: Pandu Irsyadi, Naufal
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/87727
• Institution: Institut Teknologi Bandung

fuel consumption, exhaust emissions, as well as the mental and physical health of the public. Fixed-duration traffic light control systems are unable to adapt to dynamic traffic conditions. Therefore, an adaptive traffic light duration control system is needed, with the objective of reducing congestion. This research aims to develop a traffic light duration control system based on Reinforcement Learning (RL) that can adjust traffic light durations in real-time. The research methodology follows the CRISP-DM (Cross Industry Standard Process for Data Mining) framework, which consists of the following phases: Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation, and Deployment. Traffic data is obtained through the Google Maps API and further synthesized into vehicle route data using the SUMO simulator to construct a simulated traffic environment. The RL model is trained using eight different algorithms: SARSA, Q-Learning, Deep Q-Network (DQN), Double Deep Q-Network (DDQN), REINFORCE, Deep Deterministic Policy Gradient (DDPG), Proximal Policy Optimization (PPO), and Soft Actor-Critic (SAC). Hyperparameter optimization is performed using the grid search approach to determine the best parameters for each algorithm. In this study, the evaluation results show that the off-policy policy gradient-based algorithm, DDPG, achieves the best performance in reducing congestion compared to other models. On weekday traffic, this model is able to reduce travel time by 30%, decrease queue length by 23%, and reduce the maximum queue length at intersections by 79%. On weekend traffic, it reduces travel time by 16%, decreases queue length by 14%, and reduces the maximum queue length at intersections by 75%.