Liang Xiong∗ and Yu Du∗
Reinforcement learning, autonomous driving, longitudinal control, generating experience replay, adversarial reinforcement learning
Currently, autonomous driving longitudinal control algorithms based on twin delayed deep deterministic policy gradient have been extensively researched and have achieved relatively stable results. However, these algorithms also face a series of issues, such as low training efficiency and insufficient robustness in complex and challenging environments. To address these problems, an improved algorithm named enhanced twin delayed deep deterministic policy gradient has been proposed. This algorithm introduces feedforward neural networks to implement generative experience replay, enabling it to utilise experiences in the experience pool to generate additional virtual experiences. Experiences enhance the sampling efficiency of valuable experiences and alleviate the issue of low training efficiency caused by insufficient experiences. Additionally, the concept of adversarial reinforcement learning is introduced, where intelligent agents are alternately updated to refine their policies and value functions. This approach empowers the primary agent to better handle challenging scenarios, improve its generalisation capabilities, and enhance the algorithm’s robustness. A series of simulation experiments were conducted on the CARLA simulation platform, and the results showed that the algorithm exhibited significant improvements in training efficiency and robustness compared to traditional algorithms, soft actor–critic, and deep deterministic policy gradient algorithms in the context of longitudinal control tasks for autonomous vehicles. This indicates that the algorithm can effectively address issues related to low training efficiency and limited robustness caused by differences between simulation and real-world environments, ultimately enhancing the performance of longitudinal control in the field of autonomous driving. ∗ School of Robotics, Beijing Union University, Beijing Key Laboratory of Information Services Engineering, Beijing, 100101, China; e-mail: [email protected]; [email protected] Corresponding author: Yu Du
Important Links:
Go Back
