基于k-最近邻分类增强学习的除冰机器人抓线控制  

Line-grasping control of de-icing robot based on k-nearest neighbor reinforcement learning

作　　者：魏书宁[1] 王耀南[1] 印峰[1] 杨易旻[1] 

机构地区：[1]湖南大学电气与信息工程学院,湖南长沙410082

出　　处：《控制理论与应用》2012年第4期470-476,共7页Control Theory & Applications

基　　金：国家科技支撑计划资助项目(2008BAF36B01)

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(k-nearest neighbor,KNN)分类算法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个最邻近状态并且对之加权.根据加权结果决定当前最优动作.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适应环境的能力,该方法能够克服机器人模型误差和姿态误差,以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.The flexible mechanical characteristic of power lines induces difficulties for line-grasping control for de-icing robots.To deal with this difficulty,we propose for de-icing robots a line-grasping control approach which combines the k-nearest neighbor（KNN） algorithm and the reinforcement-learning（RL）.In the learning iteration,the state-perception mechanism of the KNN algorithm selects k-nearest states and weights;from k-weighted states,an optimal action is determined.By expressing a continuous state by k-nearest discrete states in this way,this approach effectively ensures the convergence for the computation and avoids the curse of dimensionality occurred in traditional continuous state-space generalization methods.Abilities of RL in perception and adaptation to the environment make the line-grasping control to tolerate possible errors in robot model,errors of robot arm attitudes and interferences from the environment.The design procedures are presented in details.Simulation results of line-grasping control based on this approach are given.

关 键 词：除冰机器人 k–最近邻分类算法 增强学习 维数灾难 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]