Traffic Signal Control using Reinforcement Learning

Abstract

This thesis presents an in-depth examination of traffic light control optimization using Reinforcement Learning (RL) techniques. The research focuses on two specific RL algorithms: Deep Q-Learning (DQN) and Double Deep Q-Learning (DDQN), investigating their ability to reduce wait times at a four-way traffic intersection. The RL agents’ learning process is driven by a reward function based on waiting times, designed to guide the agents towards minimizing these times. A transition phase is implemented in the model, allowing for flexibility and responsiveness to changing traffic conditions. Deep Neural Networks (DNNs) are used as function approximators, facilitating the understanding of the association between state-action pairs. The architecture comprises five fully connected hidden layers, providing an effective means of approximating the Qvalues for the state-action pairs. Training data for the DNN is stored in an Experience Replay Memory, which is effectively a history of state, action, reward, and subsequent state. The study concludes that both DQN and DDQN agents demonstrated an increasing proficiency over time, indicating the successful application of RL techniques in traffic light control systems. This research contributes to the ongoing efforts to employ advanced RL techniques in optimizing traffic flow, with potential applications in intelligent transportation systems, smart cities, and autonomous vehicle navigation