Two distinct bp training algorithms on multilayer perceptron type neural - network are developed to improve the trained network ' s quickness , robustness and generality . they are brought out from the combination of existing investigations - dynamic learning rate , momentum terms and quadratic function of weights , and amendment of desired error function respectively . the simulative results of rotating machinery fault pattern classification illustrated the effectiveness of the new algorithms 针对多层感知器神经网络的标准bp训练算法存在推广能力差和训练速度慢等缺陷，首先结合已有的研究成果，给出了一种动态学习率和动量项相结合，并在误差函数中引入网络权值的二次型函数项的改进训练算法，提高了网络的快速性和推广能力。
The load of the dc motor has a great influence on the tracking effect of speed control system , so a kind of load disturbance observer is constructed to compensate the load disturbance , and a new variable structure controller is designed by employing lyapunov speed error function 针对实际直流电机负载对速度控制系统跟踪效果有较大影响的情况，提出了一种重构轧制负载扰动的观测器，进而对负载扰动进行前馈补偿控制，设计了一种新的利用lyapunov速度误差函数构造的变结构鲁棒控制器。
In the design of corpus , we carefully analyze the syllable distribution of corpus th - coss , then classify the prosodic characters of this corpus and present out the distribution of every prosodic character . based on prosodic character vector , we construct an error function which is used to select original corpus for simulation system , and show the distribution of prosodic characters for the original corpus . greedy algorithm and corpus self - adaptive process are expatiated to set theoretical foundation for text material search 在语料库分析与设计方面,首先统计th - coss语料库中音节分布情况,给出th - coss语料库韵律特征分类,并对每一种韵律特征进行统计,然后构造了一个基于韵律特征向量的误差函数,并采用该误差函数提取语料组成模拟系统的初始语料库,分析该库的韵律特征分布,最后阐述了greedy算法与语料自适应过程,为文本语料的搜索打下理论基础。
In order to improve the generalization ability and to make the system more accordant to the practical requirement , the paper introduces the k - l information distance to the error function of the traditional bp algorithm . the simulation proves that the generalization ability of the system trained with the modified algorithm is much better than that of other algorithms 为了进一步改进系统的泛化能力，使其满足报价的实际需要，本文利用神经网络的概率描述，通过研究k - l信息距离和神经网络泛化能力的关系，构造了一个新的神经网络学习误差函数，并将此法与其它各种算法的泛化结果进行比较。
The tests of e - o applications by our flux ktp has been realized , the results showed : optical waveguides fabricated by using an ion - exchange process , which have an exchange - ion concentration depth profile and refractive - index profile , is close to a complementary error - function distribution , optical homogeneity and device thermal stability is much better . amplitude modulation switch formed by our flux ktp has the contrast ratio of 150 : 1 and insert loss is 2 . 5 % at 1064 nm . high quality optical pulse with 1 ns width was cut successfully by using an e - o modulator from a laser pulse with 50 ns width , this modulator had run for three years , and the crystal did n ' t blackened , it showed our low conductivity flux ktp can endure high modulation voltage for a very long time Ktp晶体的电光应用试验表明：用离子交换法制作的电光波导，其离子交换浓度、折射率变化符合余弦误差函数，光学均匀性以及器件的温度稳定性较好；制作的强度调制电光开关，消光比为150 ： 1 ，对1064nm激光的插入损耗为2 . 5 ；制作的电光调制器用于激光脉冲整形试验，从脉冲宽度50ns的激光脉冲削出脉宽1ns的高质量光脉冲，该电光开关经过长达三年多的使用，没有出现晶体变黑现象，说明本实验的低电导率ktp晶体能够耐受长时间的调制电压。
