Adaptive Traffic Signal Control (ATSC) is a technique to reduce traffic congestion by controlling intersection traffic signal efficiently. Early ATSC controlled intersection signal with a centralized computer and it is applied on actual roads, mainly in North America and Europe. Distributed ATSC has been attracting attention recently because of the increasing demand for real-time and scalability.
In my thesis, ATSC using Reinforcement Learning (RL) is introduced. ATSC using RL is one of the distributed ATSC techniques and is expected to overcome the constraints of existing technology. To implement an algorithm applicable to more realistic environments, various vehicle parameters are considered. Deep Q-Network (DQN) is used as a function approximation technique. Normal-mode algorithm is designed for general purpose of ATSC and eco-mode algorithm is designed focusing on minimizing vehicle emissions. The results are verified for three scenarios compared with the optimal signal transition following delay model using Simulation of Urban MObility (SUMO).