Performance of Deep Reinforcement Learning for Adaptive Left-Turn Phase Traffic Light Control

As traffic conditions become more complex and demanding, traditional methods of traffic signal control often fall short. The application of artificial intelligence and machine learning algorithms to traffic light timing has proven to be highly promising. This research uses reinforcement learning to manage traffic light phases automatically and efficiently, enhancing traffic flow and reducing intersection queue lengths. This paper examines the effectiveness of deep reinforcement learning techniques in optimizing the adaptive control of left-turn phases at urban intersections. The study introduces two deep reinforcement learning algorithms and compares the performance of the Double Dueling Deep Q-Network (3DQN) with the standard Deep Q-Network (DQN). These value-based methods in our proposed method use reinforcement learning optimization to determine the green duration for each phase and select either the protected or permitted left-turn phase for the next cycle. The adaptive control system adjusts traffic light timings in real-time without human intervention, ensuring smoother and more efficient traffic flow, significantly reducing queue lengths. The 3DQN algorithm uses a target network that updates target Q values at a slower rate to stabilize training and minimize errors. The dueling network splits the neural network into two parts: one to estimate the expected reward and the other to assess the relative importance of each action. Simulations were conducted with both uniform and variable car flow distributions, under light and heavy traffic volumes. They show that controllers using the 3DQN algorithm outperform the DQN algorithm. The results also reveal that the 3DQN algorithm can reduce cumulative vehicle queue lengths by at least 26% in all cases, and up to 67% in scenarios with heavy and uniform traffic flow. This research is crucial in developing intelligent traffic control systems and reducing traffic delays. The study highlights the potential of adaptive control systems using reinforcement learning to optimize traffic light timings and mitigate vehicle queue lengths, supporting the advancement of intelligent traffic control systems capable of adapting to dynamic urban conditions.