Deep Reinforcement Learning for Eco-Driving Control of Connected and Automated Vehicles at Consecutive Signalized Intersections

Abstract

Signalized intersections force vehicles into frequent idling and repeated acceleration, leading to increased energy consumption. To address this, eco-driving strategies targeting stop-free passage through intersections are considered crucial. This study applies the soft actor–critic (SAC) reinforcement-learning algorithm to develop a control policy that enables vehicle convoys to pass through two consecutive signalized intersections without stopping. By leveraging the maximum-entropy framework, SAC balances exploration and exploitation, allowing for robust and adaptive decision-making under complex traffic dynamics. Simulation experiments compare SAC with natural driving models—(noisy) intelligent driver models (IDM, N-IDM)—and other reinforcement-learning methods—deep deterministic policy gradient (DDPG) and proximal policy optimization (PPO)—under varied traffic conditions. Results demonstrate SAC reduces energy consumption by 21.38% (versus IDM), 21.18% (versus N-IDM), 0.24% (versus DDPG), and 4.03% (versus PPO) while maintaining competitive travel time and speed. These results confirm the effectiveness and robustness of SAC, highlighting its potential for deployment in real-world mixed-traffic environments.