A Deep Reinforcement Learning Decision-Making Approach for Adaptive Cruise Control in Autonomous Vehicles

Abstract

In the evolving automobile industry, Adaptive Cruise Control (ACC) is key for aiding autonomous traffic navigation. Ideal ACC systems can decelerate to low speeds in stop-and-go traffic, maintain a safe following distance, minimize rear-end collision risks, and lessen the driver's need to continually adjust vehicle's speed to match traffic flow. In this paper, we offer a Deep Reinforcement Learning-based adaptive cruise control (DRL-ACC) system that creates safe, flexible, and responsive car-following policies agents. Instead of using discrete incremental and decremental values or a continuous action space, we suggest constructing a discrete high-level action space to accelerate, decelerate, and hold the current speed. We also provide a comprehensive, easyto-interpret multi-objective reward function that reflects safe, responsive, and rational traffic behavior. This strategy, trained on a single steady-state flow car-following scenario, promotes steadiness, responsiveness, and shows better generalization to diverse car-following scenarios. Results are also compared to the conventional Intelligent Driver Model (IDM). We further explore the model's potential to avoid rear-end collisions and facilitate future integration of lane-change maneuvers, which will increase its effectiveness in emergency situations.