BP(back propagation)算法是一种常用的神经网络学习算法,而基于Hadoop集群MapReduce编程模型的BP(MapReduce back propagation,MRBP)算法在处理大数据问题时,表现出良好的性能,因而得到了广泛应用.但是,由于该算法缺乏神经节点之间细...BP(back propagation)算法是一种常用的神经网络学习算法,而基于Hadoop集群MapReduce编程模型的BP(MapReduce back propagation,MRBP)算法在处理大数据问题时,表现出良好的性能,因而得到了广泛应用.但是,由于该算法缺乏神经节点之间细粒度结构并行的能力,当遇到数据维度较高、网络节点较多时,性能还显不足.另一方面,Hadoop集群计算节点通信不能由用户直接控制,现有基于集群系统的结构并行策略不能直接用于MRBP算法.为此,提出一种适合于Hadoop集群的结构并行MRBP(structure parallelism based MapReduce back propagation,SP-MRBP)算法,该算法将神经网络各层划分为多个结构,通过逐层并行-逐层集成(layer-wise parallelism,layer-wise ensemble,LPLE)的方式,实现了MRBP算法的结构并行.同时,推导出了SP-MRBP算法和MRBP算法计算时间解析表达式,以此分析了2种算法时间差和SP-MRBP算法最优并行规模.据了解,这是首次将结构并行策略引入MRBP算法中.实验表明,当神经网络规模较大时,SP-MRBP较之原算法,具有较好的性能.展开更多
Various force disturbances influence the thrust force of linear motors when a linear motor （LM） is running. Among all of force disturbances, the force ripple is the dominant while a linear motor runs in low speed. I...Various force disturbances influence the thrust force of linear motors when a linear motor （LM） is running. Among all of force disturbances, the force ripple is the dominant while a linear motor runs in low speed. In order to suppress the force ripple, back propagation（BP） neural network is proposed to learn the function of the force ripple of linear motors, and the acquisition method of training samples is proposed based on a disturbance observer. An off-line BP neural network is used mainly because of its high running efficiency and the real-time requirement of the servo control system of a linear motor. By using the function, the force ripple is on-line compensated according to the position of the LM. The experimental results show that the force ripple is effectively suppressed by the compensation of the BP neural network.展开更多
文摘BP(back propagation)算法是一种常用的神经网络学习算法,而基于Hadoop集群MapReduce编程模型的BP(MapReduce back propagation,MRBP)算法在处理大数据问题时,表现出良好的性能,因而得到了广泛应用.但是,由于该算法缺乏神经节点之间细粒度结构并行的能力,当遇到数据维度较高、网络节点较多时,性能还显不足.另一方面,Hadoop集群计算节点通信不能由用户直接控制,现有基于集群系统的结构并行策略不能直接用于MRBP算法.为此,提出一种适合于Hadoop集群的结构并行MRBP(structure parallelism based MapReduce back propagation,SP-MRBP)算法,该算法将神经网络各层划分为多个结构,通过逐层并行-逐层集成(layer-wise parallelism,layer-wise ensemble,LPLE)的方式,实现了MRBP算法的结构并行.同时,推导出了SP-MRBP算法和MRBP算法计算时间解析表达式,以此分析了2种算法时间差和SP-MRBP算法最优并行规模.据了解,这是首次将结构并行策略引入MRBP算法中.实验表明,当神经网络规模较大时,SP-MRBP较之原算法,具有较好的性能.
基金National Natural Science Foundation of China(No. 60474021)
文摘Various force disturbances influence the thrust force of linear motors when a linear motor （LM） is running. Among all of force disturbances, the force ripple is the dominant while a linear motor runs in low speed. In order to suppress the force ripple, back propagation（BP） neural network is proposed to learn the function of the force ripple of linear motors, and the acquisition method of training samples is proposed based on a disturbance observer. An off-line BP neural network is used mainly because of its high running efficiency and the real-time requirement of the servo control system of a linear motor. By using the function, the force ripple is on-line compensated according to the position of the LM. The experimental results show that the force ripple is effectively suppressed by the compensation of the BP neural network.
