TRAFFIC CONTROL BASED ON DEEP Q-NETWORK ALGORITHM WITH ADAPTIVE REWARD MECHANISM IN INTERSECTION NETWORK

• Main Author: Okto Fernandez, Eric
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/67152
• Institution: Institut Teknologi Bandung

Transportation demand has increased in the last few decades as human activities increase. One of the most negative impact is the increasing level of traffic congestion. A possible short-term solution for this problem is to utilize an adaptive traffic control algorithm. Most of the traffic control systems in Indonesia still utilize classic control algorithm with a predetermined green phase sequence. In this study, an adaptive traffic controller is proposed using a Reinforcement Learning algorithm. Reinforcement Learning algorithm will be applied into SUMO traffic simulation software. In designing control algorithm, a Reinforcement Learning algorithm called Deep- Q Network (DQN) is used. The action taken by DQN is to determine the traffic phase with various rewards, ranging from pressure to providing adaptive loads on pressure and queue length. DQN-based control algorithm with adaptive reward mechanism provide the best performance via vehicle throughput as much as 56,384 vehicles, followed by classic and conventional control method such as Webster (50,366), Max-Pressure (50,541) and Uniform (46,241). The increment of vehicle throughput in a region will increase the region productivity.