Adaptive car-following model based on deep reinforcement learning and multi-strategy particle swarm optimization

In this study, a novel self-driving car-following framework was proposed, which combines a multi-strategy particle swarm optimization (MSPSO) algorithm with deep reinforcement learning (DRL) to achieve adaptive control across diverse driving styles. An adaptive following speed control algorithm was designed using DRL, and a reward function was formulated by incorporating driving characteristics such as safety, efficiency, moderateness and emergency braking. MSPSO algorithm was introduced to dynamically adjust the weights within the reward function. The optimal driving strategy was identified by the reinforcement learning (RL) algorithm through trial-and-error learning in a simulated environment. Experimental results demonstrated that the proposed algorithm significantly improved car-following efficiency in conservative driving scenarios, maintained a safe distance in aggressive driving and outperformed other methods in terms of comfort.