夏伟,李慧云.基于深度强化学习的自动驾驶策略学习方法[J].集成技术,2017,6(3):29-40
基于深度强化学习的自动驾驶策略学习方法
Training Method of Automatic Driving Strategy Based on DeepReinforcement Learning
  
DOI:
中文关键词:  深度强化学习;自动驾驶;聚类;神经网络
英文关键词:deep reinforcement learning; autonomous vehicle; cluster; neural network
基金项目:
作者单位
夏伟 中国科学院深圳先进技术研究院 深圳 518055;中国科学院大学深圳先进技术学院 深圳 518055 
李慧云 中国科学院深圳先进技术研究院 深圳 518055;中国科学院大学深圳先进技术学院 深圳 518055 
摘要点击次数: 191
全文下载次数: 290
中文摘要:
      自动驾驶是人工智能研究的重要应用领域,文章提出了一种基于深度强化学习的自动驾驶策略模型学习方法。首先采用在线交互式学习方法对深度网络模型进行训练,并基于专业司机的经验数据对 模型进行预训练,进而结合经验池回放技术提高模型训练收敛速度,通过对状态空间进行聚类再采样,提高其独立同分布特性以及策略模型的泛化能力。通过与神经网络拟和 Q-迭代算法的比较,所提方法的训练时间可缩短 90% 以上,稳定性能提高超过 30%。以复杂度略高于训练集的测试道路长度为基准,与经验过滤的 Q-学习算法相比,采用聚类再采样的方法可以使策略模型的平均行驶距离提高 70% 以上。
英文摘要:
      Automatic drive is an important application field of artificial intelligence. In this paper, a novel training strategy for self-driving vehicles was investigated based on the deep reinforcement learning model. The proposed method involves a Q-learning algorithm with filtered experience replay and pre-training with experiences from professional drivers, which accelerates the training process due to reduced exploration spaces. By resampling the input state after clustering, generalization ability of the strategy can be improved due to the individual and independent distribution of the samples. Experimental results show that, in comparison with conventional neural fitted Q-iteration algorithm, the training efficiency and controlling stability can be improved more than 90% and 30% respectively by the proposed approach. Experimental results with more complex testing tracks show that, average travel distance can be improved more than 70% in comparison with the Q-learning algorithm by the proposed method.
查看全文  查看/发表评论  下载PDF阅读器
关闭
微信关注二维码 用微信扫一扫

美女

美女图片

美女

美女图片