基于k最近邻分类增强学习的除冰机器人抓线控制-控制理论与应用.PDF

第xx 卷第x 期 控制 理 论 与 应 用 Vol. xx No. x xxxx 年x 月 Control Theory Applications Xxx. xxxx 文文文章章章编编编号号号:2010 基基基于于于k最最最近近近邻邻邻分分分类类类增增增强强强学学学习习习的的的除除除冰冰冰机机机器器器人人人抓抓抓线线线控控控制制制 魏书宁, 王耀南, 印峰, 杨易旻 (湖南大学电气与信息工程学院,湖南长沙410082) 摘要: 输电线柔性结构特性给除冰机器人越障抓线控制带来极大困难. 本文提出了一种结合k-最近邻分类算法 （KNN ）和增强学习算法的抓线控制方法. 利用基于KNN算法的状态感知机制选择机器人当前状态k个最邻近状 态，并根据加权结果选择当前最优动作. 该方法可以得到机器人连续状态的离散表达形式，从而有效解决传统连 续状态泛化方法带来的计算收敛性和维数灾难问题. 借助增强学习算法探测和学习环境的能力，该方法能够克服 机器人模型和姿态误差及环境干扰等因素对抓线控制的影响. 文中给出了算法具体实现步骤，并给出了应用此方 法控制除冰机器人抓线的仿真实验. 关键词:除冰机器人; k-最近邻分类算法; 增强学习; 维数灾难 中图分类号: TP273 文献标识码: A Line- grasping control of de-icing robot based on k-nearest neighbor reinforcement learning Wei Shu-ning, Wang Yao-nan, Yin Feng, Yang Yi-min (College of Electrical and Information Engineering, Hunan University, Changsha Hunan 410082,China) Abstract: Flexible mechanical characteristic of power lines induces difficulties for line-grasping control of de-icing robot. A line-grasping control approach for de-icing robot combining with k-nearest neighbor (KNN) algorithm and Rein- forcement learning (RL) is proposed in this paper. In each learning iteration this approach makes use of KNN algorithm based state perception mechanism to select k-nearest states, and then the weighted k states are used to choose an optimal action. Expressing a continuous state by k-nearest discrete states, this approach effectively overcomes the computational convergence and the curse of dimension problems in traditional continuous state space generalization methods. By means of RL algorithm’s ability of perception and learning in environment, the approach can tolerate poss