强化学习
计算机科学
学习迁移
传输(计算)
人工智能
并行计算
作者
Bo Wei,Jianxin Zhao,Yinuo Zhao,Feng Tian
标识
DOI:10.1109/bigcom61073.2023.00045
摘要
Autonomous Driving presents a promising solution to the issue of road accidents, which are mostly caused by human errors. The use of artificial intelligence technologies in this field has resulted in significant advancements in tasks such as object detection, path planning, and obstacle avoidance, leading to safer and more efficient transportation. Reinforcement learning (RL) is a powerful machine learning algorithm that has demonstrated effectiveness in various autonomous driving applications. However, the vanilla single RL policy is inadequate when faced with more complex transportation scenarios involving heavy and dynamic traffic. In this paper, we propose a novel OPtion-based multi-skill policy Transfer method with deep RL for autonomous driving, called "Opt-RL", to learn a more complex target policy by integrating basic skills from multiple source policies. An adaptive option learning module is designed to efficiently use learned skills in higher-level target domains, determining when and where to distil policies from different sources. We conduct experiments on challenging tasks in the Mujoco Maze2D benchmark and a simulated highway environment. Experimental results demonstrate that Opt-RL can achieve knowledge transfer among different levels of policies and successfully train a complex high-level decision-making policy by reasonably integrating multiple basic skills; it also achieves a longer safe driving distance 16% higher than the baseline DQN.
科研通智能强力驱动
Strongly Powered by AbleSci AI