Rear-End Collision Avoidance Using Multi-Agent Deep Reinforcement Learning

This research paper offers a novel approach to detect rear-end collisions and increase the acceleration of the victim car to possibly avoid or minimize the impact of the collision. A Deep Q-Learning model is utilized to predict the collision, the optimum acceleration of the victim car and the optimum braking of the committer car. The study uses a multi-agent reinforcement learning framework in a simulated environment, focusing on synchronized responses between lead and following vehicles. A comprehensive reward function encourages safe driving behaviors, optimized braking, and successful obstacle detection and avoidance. The continuous action space allows for subtle vehicle control, enhancing the system's adaptability to various collision scenarios. Experimental results demonstrate an overall improvement in collision avoidance and impact reduction over 5000 training episodes. This approach addresses a critical gap in existing safety systems by specifically targeting rear-end collisions, offering the potential for a significant reduction in accident rates and severity. Future work includes real-world testing and expansion to more complex traffic scenarios.