倪修华, 陈维山, 刘军考, 石胜君. 基于BP神经网络的被动步行稳定不动点的估算[J]. 机器人, 2010, 32(4): 478-483.
引用本文: 倪修华, 陈维山, 刘军考, 石胜君. 基于BP神经网络的被动步行稳定不动点的估算[J]. 机器人, 2010, 32(4): 478-483.
NI Xiuhua, CHEN Weishan, LIU Junkao, SHI Shengjun. Estimation of the Stable Fixed Point of Passive Dynamic Walking with BP Neural Network[J]. ROBOT, 2010, 32(4): 478-483.
Citation: NI Xiuhua, CHEN Weishan, LIU Junkao, SHI Shengjun. Estimation of the Stable Fixed Point of Passive Dynamic Walking with BP Neural Network[J]. ROBOT, 2010, 32(4): 478-483.

基于BP神经网络的被动步行稳定不动点的估算

Estimation of the Stable Fixed Point of Passive Dynamic Walking with BP Neural Network

  • 摘要: 建立了被动步行机器人的动力学模型.使用BP神经网络对被动步行稳定不动点进行估算,并将估算值作为Newton-Raphson迭代的初值来求解稳定不动点.该方法解决了以往利用非最简模型求解不动点时,由于初值选取不当所造成的搜索成功率低以及搜索时间长的问题.在获取训练样本时,参数变化较小时不动点的变化很小;利用这一特点,在计算相邻两个样本的不动点时,只使一个参数发生较小的变化,并将本次使用Newton-Raphson迭代搜索得到的稳定不动点作为搜索下一样本不动点时的初值.使用1000组随机产生的参数对所得神经网络进行测试,结果表明该方法可以大幅提高稳定不动点的搜索成功率,大幅缩短搜索时间.

     

    Abstract: The dynamic model of passive dynamic walking robot is given.BP(backpropagation) neural network(NN) is used to estimate the stable fixed point,and the estimated fixed point is taken as initial value for Newton-Raphson iteration to search the stable fixed point.This method solves the problems of low successful search rate and long search time caused by the improper initial value for the walking model which is not the simplest in previous study.Small variation of parameters leads to small variation of fixed points when acquiring training samples.When calculating fixed points of two successive samples,only one parameter has a small variation.The fixed point obtained with Newton-Raphson iteration is used as initial value for the next sample.1000 groups of parameters are generated randomly to test the NN performance.The results show that this method increases the successful search rate of stable fixed points significantly and shortens the search time greatly.

     

/

返回文章
返回