基于遗传规划实现泛函网络神经元函数类型优化

泛函网络是最近提出的一种对神经网络的有效推广。与神经网络不同，它处理的是一般的泛函模型，其神经元函数不固定，而是可学习的，且在各个处理单元之间没有权值。同神经网络一样，至今还没有系统设计方法能够对给定问题设计出近似最优的结构。鉴于此，将整个泛函网络的设计分解为单个神经元的逐个设计；然后，在此框架下提出了基于遗传规划的单个神经元的设计方法，该方法可实现对神经元函数类型的优化。仿真实验表明，本方法是有效可行的，能用较小的网络规模获得更满意的泛化特性。     

遗传前馈神经网络在函数逼近中的应用

人工神经网络具有高计算能力、泛化能力和非线性映射等特点,被成功应用于众多领域,但缺乏用于确定其网络拓扑结构、激活函数和训练方法的规则。该文提出利用遗传算法优化前馈神经网络的方法,将网络结构、激活函数和训练方法等编码作为个体,发现最优或次优解,针对特定问题设计较理想的前馈神经网络。介绍遗传算法的具体步骤,对非线性函数逼近进行实验,结果表明优化后前馈神经网络的性能优于由经验确定的前馈神经网络,验证了本文方法的有效性。     

基于PSO训练的NN PID控制器设计及其FPGA实现*

提出了一种基于PSO学习、VHDL描述和FPGA实现的NNPID控制器设计方法。首先借助MATLAB系统仿真工具,在闭环控制系统中通过PSO优化算法训练前馈网络,得到优化的NNPID控制器参数;然后在FPGA集成开发环境下进行控制器的VHDL层次化设计,重点研究单个神经元和前馈网络的结构以及实现方式;最后对该控制器进行了闭环时序测试,并在一个具体的FPGA器件上实现。研究结果表明,PSO用于NNPID控制器训练速度快,VHDL描述和FPGA实现该控制器时序验证方便,而且控制器具有较好的鲁棒性。     

爽层感知器的灵敏度分析

灵敏度分析对于神经网络结构设计具有指导意义,已有的灵敏度分析方法往往针对特定的激活函数,并且对网络输入和权值扰动有严格的限制,该文采用另一种以1-ed^-λx/1_e^-λx型激活函数的倾斜度作为参数的函数形式逼近激活函数,得到一类神经元的统一的灵敏度解析表达式和网络灵敏度计算算法,该方法取消了对输入和权值扰动的限制,可以研究激活函数倾斜度对网络灵敏度的影响。计算机模拟试验证明了此方法的正确性,并且提出了网络结构设计的几条准则。     

多项式函数型回归神经网络模型及应用

文中利用回归神经网络既有前馈通路又有反馈通路的特点，将网络隐层中神经元的激活函数设置为可调多项式函数序列，提出了多项式函数型回归神经网络新模型，它不但具有传统回归神经网络的特点，而且具有较强的函数逼近能力，针对递归计算问题，提出了多项式函数型回归神经网络学习算法，并将该网络模型应用于多元多项式近似因式分解，其学习算法在多元多项式近似分解中体现了较强的优越性，通过算例分析表明，该算法十分有效，收敛速度快，计算精度高，可适用于递归计算问题领域，该文所提出的多项式函数型回归神经网络模型及学习算法对于代数符号近似计算有重要的指导意义。     

基于多层前馈神经网络的图像压缩的仿真研究

神经网络研究的再度兴起及其在图像编码中的应用，开辟了图像压缩的新途径。该文论述了多层前馈网络用于图像压缩的网络模型、原理、算法及关键技术，并通过大量的仿真实验说明了在BP网络图像压缩中，算法、激活函数和压缩率等参数的选择是至关重要的，它们与收敛时间以及重建图像的压缩性能息息相关。通过对实验结果的详细分析得知，BP网络图像压缩必须综合考虑压缩率、失真率和训练时间等因素，在学习规则、激活函数、隐层神经元数和压缩性能之间进行权衡，以满足实际应用。     

基于多项式基函数的泛函网络构造方法与逼近理论

用构造性方法证明:对于给定的r阶多项式函数,可以具体地构造出一个三层泛函网络,以任意精度逼近该多项式,所构造的网络的中问神经元个数仅与多项式基函数的阶数r有关,并能用r表达.该文所得结果对于基于多项式基函数的泛函网络逼近任意函数类的网络具体构造和逼近具有理论指导意义.     

具有多类型激活函数的RBF网络设计

提出多类型激活函数的径向基网络结构设计方式,网络隐层由薄板样条函数节点和高斯函数节点构成,分别采用改进的粒子群算法和前向局部优化算法对两种类型节点的参数进行学习.改进的粒子群算法综合运用粒子运动过程的启发式信息和样本的梯度信息进行种群迭代,减小了陷入局部极值的概率.前向局部优化算法是一种高斯节点的序贯学习算法,算法采用DFP方法对新增高斯节点的参数进行局部优化,提高了网络的逼近性能.通过解析和工程算例验证了多类型激活函数RBF网络的高拟合能力及其学习算法的有效性.     

机器学习中的核覆盖算法

基于统计学习理论的支持向量机（SVM）方法在样本空间或特征空间构造最优分类超平面解决了分类器的构造问题,但其本质是二分类的,且核函数中的参数难以确定,计算复杂性高．构造性学习算法根据训练样本构造性地设计分类网络,运行效率高,便于处理多分类问题,但存在所得的分界面零乱、测试计算量大的缺点．该文将SVM中的核函数法与构造性学习的覆盖算法相融合,给出一种新的核覆盖算法．新算法克服了以上两种模型的缺点,具有运算速度快、精度高、鲁棒性强的优点．其次．文中给出风险误差上界与覆盖个数的关系．最后给出实验模     

激活函数可调的RBF神经网络模型

RBF神经网络是一种新颖的前馈神经网络,相比传统的BP神经网络具有更强的函数逼近、分类能力,以及更快的学习速度.在实际应用和科学研究中,不同径向基函数的选择对RBF神经网络的性能有很大的影响,一般是根据经验或者实验效果来选择适当的径向基函数.本文提出一种激活函数可调的RBF神经网络模型,将具有不同表达能力的径向基函数通过线性组合的形式形成新的神经元激活函数,进而提高神经网络的逼近精度和泛化能力,对于有监督学习,给出了相应的各参数的训练算法和应用举例.     

A new optimization algorithm for single hidden layer feedforward neural networks

Feedforward neural networks are the most commonly used function approximation techniques in neural networks. By the universal approximation theorem, it is clear that a single-hidden layer feedforward neural network (FNN) is sufficient to approximate the corresponding desired outputs arbitrarily close. Some researchers use genetic algorithms (GAs) to explore the global optimal solution of the FNN structure. However, it is rather time consuming to use GA for the training of FNN. In this paper, we propose a new optimization algorithm for a single-hidden layer FNN. The method is based on the convex combination algorithm for massaging information in the hidden layer. In fact, this technique explores a continuum idea which combines the classic mutation and crossover strategies in GA together. The proposed method has the advantage over GA which requires a lot of preprocessing works in breaking down the data into a sequence of binary codes before learning or mutation can apply. Also, we set up a new error function to measure the performance of the FNN and obtain the optimal choice of the connection weights and thus the nonlinear optimization problem can be solved directly. Several computational experiments are used to illustrate the proposed algorithm, which has good exploration and exploitation capabilities in search of the optimal weight for single hidden layer FNNs.     

Intelligent forecasting system based on integration of electromagnetism-like mechanism and fuzzy neural network

Fuzzy neural network (FNN) architectures, in which fuzzy logic and artificial neural networks are integrated, have been proposed by many researchers. In addition to developing the architecture for the FNN models, evolution of the learning algorithms for the connection weights is also a very important. Researchers have proposed gradient descent methods such as the back propagation algorithm and evolution methods such as genetic algorithms (GA) for training FNN connection weights. In this paper, we integrate a new meta-heuristic algorithm, the electromagnetism-like mechanism (EM), into the FNN training process. The EM algorithm utilizes an attraction–repulsion mechanism to move the sample points towards the optimum. However, due to the characteristics of the repulsion mechanism, the EM algorithm does not settle easily into the local optimum. We use EM to develop an EM-based FNN (the EM-initialized FNN) model with fuzzy connection weights. Further, the EM-initialized FNN model is used to train fuzzy if–then rules for learning expert knowledge. The results of comparisons done of the performance of our EM-initialized FNN model to conventional FNN models and GA-initialized FNN models proposed by other researchers indicate that the performance of our EM-initialized FNN model is better than that of the other FNN models. In addition, our use of a fuzzy ranking method to eliminate redundant fuzzy connection weights in our FNN architecture results in improved performance over other FNN models.     

Adaptive fuzzy neuro-observer applied to low cost INS/GPS

A combined MEMS Inertial Navigation System (INS) with GPS is used to provide position and velocity data of land vehicles. Data fusion of INS and GPS measurements are commonly achieved through a conventional Extended Kalman filter (EKF). Considering the required accurate model of system together with perfect knowledge of predefined error models, the performance of the EKF is decreased due to unmodeled nonlinearities and unknown bias uncertainties of MEMS inertial sensors. Universal knowledge based approximators comprising of neural networks and fuzzy logic methods are capable of approximating the nonlinearities and the uncertainties of practical systems. First, in this paper, a new fuzzy neural network (FNN) function approximator is used to model unknown nonlinear systems. Second, the process of design and real-time implementation of an adaptive fuzzy neuro-observer (AFNO) in integrated low-cost INS/GPS positioning systems is proposed. To assess the long time performance of the proposed AFNO method, wide range tests of a real INS/GPS with a car vehicle have been performed. The unbiased estimation results of the AFNO show the superiority of the proposed method compared with the classic EKF and the adaptive neuro-observer (ANO) including a pure artificial neural network (ANN) function approximator.     

一种神经网络辨识的混合学习算法

文章提出了一种神经网络辨识的混合学习算法。采用具有递阶结构的遗传算法来获得神经网络拓扑结构和连接权值的全局次优解,之后由BP算法来进一步调整神经网络的连接权值,从而实现神经网络的自动优化设计。仿真结果表明,所得的神经网络结构简单、精度高,并具有良好的泛化能力。     

人工神经网络设计算法的研究

人工神经网络在计算机领域中仍然是一个有待进一步研究的问题。文中介绍了目前使用的几种不同的人工神经网络设计算法：BP算法、FP算法、多层前向网络的交叉覆盖设计算法。探讨了这些算法的应用和特点，并从原理和设计思路上进行了分析。对BP算法和交叉覆盖算法给出具体的测试结果，解释并说明了各种算法的优缺点。重点研究并介绍了目前流行的多层前向网络的交叉覆盖算法的设计思路和应用方法，测试结果证明了该方法在智能识别领域的有效性。     

基于模糊神经网络的网络业务分类研究

该文利用神经网络的自学习能力和模糊逻辑的动态性和及时性等特点,将模糊逻辑和神经网络有机地结合起来,构造出了四层模糊神经网络,并用训练神经网络的相应学习算法训练网络,将该模型用于网络业务源特征提取与分类的研究中,并与单纯的神经网络算法相比较。计算机仿真结果表明,模糊神经网络方法比神经网络算法更优越,该文的研究结果为解决网络业务源特征提取与分类奠定了基础。     

Lagrange优化前馈神经网络盲多用户检测算法

提出了一种前馈神经网络盲多用户检测算法.该算法利用恒模准则构造了一个代价函数,给出了该算法的约束条件.利用增广Lagrange函数方法对带约束的代价函数进行优化求解,获得前馈神经网络网络权值和参数的迭代公式,从而实现了盲多用户检测.利用计算机对增广Lagrange函数优化的前馈神经网络盲多用户检测算法进行仿真.仿真结果表明,新算法具有较好的误码率及收敛特性等性能.     

Image restoration techniques based on fuzzy neural networks

By establishing some suitable partitions of input and output spaces, a novel fuzzy neural network (FNN) which is called selection type FNN is developed. Such a system is a multilayer feedforward neural network, which can be a universal approximator with maximum norm. Based on a family of fuzzy inference rules that are of real senses, a simple and useful inference type FNN is constructed. As a result, the fusion of selection type FNN and inference type FNN results in a novel filter-FNN filter. It is simple in structure. And also it is convenient to design the learning algorithm for structural parameters. Further, FNN filter can efficiently suppress impulse noise superimposed on image and preserve fine image structure, simultaneously. Some examples are simulated to confirm the advantages of FNN filter over other filters, such as median filter and adaptive weighted fuzzy mean (AWFM) filter and so on, in suppression of noises and preservation of image structure.     

Forecasting of Software Reliability Using Neighborhood Fuzzy Particle Swarm Optimization Based Novel Neural Network

This paper proposes an artificial neural network (ANN) based software reliability model trained by novel particle swarm optimization (PSO) algorithm for enhanced forecasting of the reliability of software. The proposed ANN is developed considering the fault generation phenomenon during software testing with the fault complexity of different levels. We demonstrate the proposed model considering three types of faults residing in the software. We propose a neighborhood based fuzzy PSO algorithm for competent learning of the proposed ANN using software failure data. Fitting and prediction performances of the neighborhood fuzzy PSO based proposed neural network model are compared with the standard PSO based proposed neural network model and existing ANN based software reliability models in the literature through three real software failure data sets. We also compare the performance of the proposed PSO algorithm with the standard PSO algorithm through learning of the proposed ANN. Statistical analysis shows that the neighborhood fuzzy PSO based proposed neural network model has comparatively better fitting and predictive ability than the standard PSO based proposed neural network model and other ANN based software reliability models. Faster release of software is achievable by applying the proposed PSO based neural network model during the testing period.