Multilayer feedforward neural networks with single powers-of-twoweights

A new algorithm for designing multilayer feedforward neural networks with single powers-of-two weights is presented. By applying this algorithm, the digital hardware implementation of such networks becomes easier as a result of the elimination of multipliers. This proposed algorithm consists of two stages. First, the network is trained by using the standard backpropagation algorithm. Weights are then quantized to single powers-of-two values, and weights and slopes of activation functions are adjusted adaptively to reduce the sum of squared output errors to a specified level. Simulation results indicate that the multilayer feedforward neural networks with single powers-of-two weights obtained using the proposed algorithm have generalization performance similar to that of the original networks with continuous weights     

Neural network approach to land cover mapping

A pattern classification method is proposed for remote sensing data using neural networks. First, the authors apply the error backpropagation (BP) algorithm to classify the remote sensing data. In this case, the classification performance depends on a training data set. In order to get stable and precise classification results, the training data set is selected based on geographical information and Kohonen's self-organizing feature map. Using the training data set and the error backpropagation algorithm, a layered neural network is trained such that the training patterns are classified with a specified accuracy. After training the neural network, some pixels are deleted from the original training data set if they are incorrectly classified and a new training data set is built up. Once training is complete, a testing data set is classified by using the trained neural network. The classification results of LANDSAT TM data show that this approach produces excellent results which are more realistic and noiseless compared with a conventional Bayesian method     

Constrained least squares detector for OFDM/SDMA-based wireless networks

The two major obstacles toward high-capacity indoor wireless networks are distortion due to the indoor channel and the limited bandwidth which necessitates a high spectral efficiency. A combined orthogonal frequency division multiplexing (OFDM)/spatial division multiple access (SDMA) approach can efficiently tackle both obstacles and paves the way for cheap, high-capacity wireless indoor networks. The channel distortion due to multipath propagation is efficiently mitigated with OFDM while the bandwidth efficiency can be increased with the use of SDMA. However, to keep the cost of an indoor wireless network comparable to its wired counterpart's cost, low-complexity SDMA processors with good performance are of special interest. In this paper, we propose a new multiuser SDMA detector which is designed for constant modulus signals. This constrained least squares (CLS) receiver, which deterministically exploits the constant modulus nature of the subcarrier modulation to achieve better separation, is compared in terms of performance and complexity with the zero forcing (ZF) and the minimum mean square error (MMSE) receiver. Additionally, since the CLS detector relies on reliable channel knowledge at the receiver, we propose a strategy for estimating the multiple input multiple output (MIMO) channels. Simulations for a Hiperlan II-based case-study show that the CLS detector significantly outperforms the ZF detector and comes close to the performance of the MMSE detector for QPSK. For higher order M-PSK, the CLS detector outperforms the MMSF detector. Furthermore, the estimation complexity for the CLS detector is substantially lower than that for the MMSE detector which additionally requires estimation of the noise power.     

Blocky artefact reduction using an adaptive constrained leastsquares method

A blocky artefact reduction algorithm using the constrained least squares (CLS) approach is described. The authors use a new objective function which effectively constrains the relationship between not only the block boundary pixels but also the inner pixels. By gracefully reducing the visible discontinuities along the block boundaries, the proposed algorithm shows excellent noise reduction performance     

Linear and non‐linear space diversity combining algorithms over fading channels in TDMA systems

This paper investigates the performances of various adaptive algorithms for space diversity combining in time division multiple access (TDMA) digital cellular mobile radio systems. Two linear adaptive algorithms are investigated, the least mean square (LMS) and the square root Kalman (SRK) algorithm. These algorithms are based on the minimization of the mean‐square error. However, the optimal performance can only be obtained using algorithms satisfying the minimum bit error rate (BER) criterion. This criterion can be satisfied using non‐linear signal processing techniques such as artificial neural networks. An artificial neural network combiner model is developed, based on the recurrent neural network (RNN) structure, trained using the real‐time recurrent learning (RTRL) algorithm. It is shown that, for channels characterized by Rician fading, the artificial neural network combiners based on the RNN structure are able to provide significant improvements in the BER performance in comparison with the linear techniques. In particular, improvements are evident in time‐varying channels dominated by inter‐symbol interference. Copyright © 2001 John Wiley & Sons, Ltd.     

Pattern characteristics of linear arrays using the constrained least squares distribution

The constrained least squares (CLS) distribution is a method for obtaining distribution functions that yield low sidelobe patterns with specified constraints on the aperture efficiency, and are especially useful for the transmit patterns of active array antennas. The widely used Taylor distribution optimizes only pattern performance while the CLS distribution optimizes pattern performance while taking into account the constraints on both the peak element amplitude and the total effective radiated voltage (ERV) of the aperture distribution. The paper compares the pattern characteristics of linear arrays with CLS and Taylor distributions. The results help to establish guidelines on when a CLS distribution would be preferable over a Taylor distribution when a specified aperture efficiency is important.     

用于光流计算的一种神经网络算法

本文给出了一种利用神经网络计算光流场的新算法。整个计算过程分为三个阶段:神经网络模型参数的估计,轮廓边界垂直速度分量的动态测量以及光流场的计算。通过网络能量函数和运动的约束误差函数的比较对网络参数进行估计。用一个动态算法迭代运行非线性光流场计算方法以使神经网络能量函数达到最小,同时也对垂直速度分量进行动态估计。由模拟试验结果讨论了影响神经网络收敛性能的若干因素。     

一种快速多层神经网络训练算法及其在系统辩识中的应用

本文提出了一种新的多层前馈神经网络快速训练方法.该算法是基于指数加权局部最小二乘(EWLLS)目标函数及殴几里得方向集(EDS)方法的,在训练过程中,通过估计局部期望输出,多层神经网络可以被分解成若干个自适应线性神经元(Adaline),而Adaline是通过EDS方法进行训练的.该算法的性能是通过将其应用于系统辩识中加以说明的.     

Adaptive neuro-fuzzy inference system (ANFIS) digital predistorter for RF power amplifier linearization

This paper describes an adaptive digital predistorter (ADP) for RF power amplifier (PA) linearization using an adaptive neuro-fuzzy inference system (ANFIS). The ANFIS predistorter (PD) employs the advantage of real-time modeling of the PA's responses in determining the PD's functions. The amplitude and phase corrections for the PD are represented in an easy-to-understand fuzzy if-then rule, while the parameters involved in the fuzzy representation are trained using neural networks algorithms, namely gradient-descent and least squares estimate (LSE). Experimental results show that a 26.3-dB improvement in linearity for a two-tone signal is obtained, while a distorted WCDMA signal is suppressed by at least 12 dB. The adaptability of the ANFIS PD to instantaneous variation in PA responses through time is also demonstrated, and results show that the ANFIS PD is capable of adapting to simulated environmental changes, which is a topic often omitted by researchers in this area. Further testing demonstrated that the tuning parameters involved in the training could be reduced by more than half for a fairly nonlinear PA without significantly degrading the suppression capability.     

Signal modelling through adaptive Volterra representations

In adaptive Volterra signal modelling, the recursive implementations of the method of least squares commonly employ a weighting factor which has constraints. Here, the least squares estimation is reformulated as a constrained optimisation problem and solved using the Lagrange programming neural networks. The weighting factor is adapted and found to be optimal     

An approximation of the radiation integral for distorted reflectorantennas using surface-error decomposition

The physical optics/aperture integration (PO/AI) formulation is often used to analyze the radiation patterns of reflector antennas. In this study, the PO/AI radiation integrals for distorted reflector antennas are addressed. The surface error of the antennas is approximated by a series of surface expansion functions. The radiation integral is decomposed into a series of radiation-type integrals, each of which corresponds to one of the surface expansion functions. Each of these radiation-type integrals is then weighted by amplitude coefficients. The advantage of performing the decomposition is that each of the radiation-type integrals can be computed and the pattern data stored. The computation of the pattern for a distorted reflector antenna with a changing error profile is performed by recalling the pattern data for each perturbation term and weighting it with the amplitude coefficient. This facilitates rapid evaluation of the radiation integral in cases where the error profile is changing (for example, time-varying errors). The superposition of integrals presented in this paper was shown to be valid for surface-error profiles up to 0.1 λ rms amplitude     

A multilayer complex neural network training algorithm and its application in adaptive equalization

In this paper, the layer-by-layer optimizing algorithm for training multilayer neural network is extended for the case of a multilayer neural network whose inputs, weights, and activation functions are all complex. The updating of the weights of each layer in the network is based on the recursive least squares method. The performance of the proposed algorithm is demonstrated with application in adaptive complex communication channel equalization.     

随机petri网分析分组交换网中窗式流量控制机理

本文将随机Ｐｅｔｒｉ网与排队论相结合，对分组交换网中的窗式流量控制机理进行了描述与分析，建立了窗式流量控制机理的随机Ｐｅｔｒｉ网模型，并给出了随机Ｐｅｔｒｉ网模型的可达图及状态转移方程。为通信网的性能分析和评价提供了一种新的方法。     

Parametric signal restoration using artificial neural networks

The problem of parametric signal restoration given its blurred/nonlinearly distorted version contaminated by additive noise is discussed. It is postulated that feedforward artificial neural networks can be used to find a solution to this problem. The proposed estimator does not require iterative calculations that are normally performed using numerical methods for signal parameter estimation. Thus high speed is the main advantage of this approach. A two-stage neural network-based estimator architecture is considered in which the vector of measurements is projected on the signal subspace and the resulting features form the input to a feedforward neural network. The effect of noise on the estimator performance is analyzed and compared to the least-squares technique. It is shown, for low and moderate noise levels, that the two estimators are similar to each other in terms of their noise performance, provided the neural network approximates the inverse mapping from the measurement space to the parameter space with a negligible error. However, if the neural network is trained on noisy signal observations, the proposed technique is superior to the least-squares estimate (LSE) model fitting. Numerical examples are presented to support the analytical results. Problems for future research are addressed     

基于遗传算法的RBF神经网络在涡流传感器非线性补偿中应用

为了解决涡流传感器的非线性问题,应用遗传算法(GA)训练径向基函数(RBF)神经网络(NN)实现其非线性补偿.介绍非线性补偿的原理和网络训练方法.从实测数据出发,建立了涡流传感器的非线性补偿模型.该方法能同时优化网络结构和参数,具有全局寻优能力,补偿精度高、鲁棒性好、网络训练速度快、能实现在线软补偿.实验结果表明,所采用的涡流传感器非线性补偿方法是有效的和可行的.补偿后,最大非线性误差在0.5%范围内,具有良好的线性.     

Sliding mode controller by use of neural networks

This paper proposes a method to design a robust controller by use of a neural network. The trained neural network functions as a sliding mode controller which is robust against uncertainties. From the analysis of the neural network, it is proved that the switching surface is not the same as the sliding surface like conventional sliding mode control theory. The neural network shows that the switching surface should be a nonlinear surface because of a hard limitation on control inputs, even if the designed sliding surface is linear. From the result of estimating the robustness of neural networks, we propose that generalization of neural networks which are used as controllers should be measured by the robustness. Numerical simulations show that the controller is robust against uncertainties and robustness can be improved by the proposed method.     

Adaptive Principal component EXtraction (APEX) and applications

The authors describe a neural network model (APEX) for multiple principal component extraction. All the synaptic weights of the model are trained with the normalized Hebbian learning rule. The network structure features a hierarchical set of lateral connections among the output units which serve the purpose of weight orthogonalization. This structure also allows the size of the model to grow or shrink without need for retraining the old units. The exponential convergence of the network is formally proved while there is significant performance improvement over previous methods. By establishing an important connection with the recursive least squares algorithm they have been able to provide the optimal size for the learning step-size parameter which leads to a significant improvement in the convergence speed. This is in contrast with previous neural PCA models which lack such numerical advantages. The APEX algorithm is also parallelizable allowing the concurrent extraction of multiple principal components. Furthermore, APEX is shown to be applicable to the constrained PCA problem where the signal variance is maximized under external orthogonality constraints. They then study various principal component analysis (PCA) applications that might benefit from the adaptive solution offered by APEX. In particular they discuss applications in spectral estimation, signal detection and image compression and filtering, while other application domains are also briefly outlined     

基于HBF神经网络的自适应观测器

传统的RBF(Radial Basis Function)神经元基函数通常把高斯类型与单一宽度作为每个神经元的激活函数,这些特性限制了网络神经元的性能,特别是在处理复杂的非线性建模问题上.为了克服这个限制,本文应用了具有类似RBF网络,但激活函数不同-超基函数HBF(Hyper Basis Function)的网络.结合RBF网络,分析了HBF网络的结构、基函数形式及基函数对网络的影响,利用决策树算法计算了网络中心.在此基础上,提出了一种基于HBF神经网络的自适应观测器设计方法,并通过引入Lyapunov函数,证明了这种观测器设计方法的稳定性;最后通过仿真验证了这种HBF神经网络观测器能很好地观测系统的状态值.     

Linearly constrained block adaptive filtering algorithm withoptimum convergence factors

A new linearly constrained adaptive filtering algorithm, the linearly constrained optimum block adaptive (LCOBA) algorithm, is presented. The LCOBA algorithm processes data in blocks and uses variable convergence factors which are optimised in a least square sense. It is superior to Frost's linearly constrained least mean squares algorithm at achieving the conflicting goals of fast convergence with little steady-state error. In addition, its computational requirements generally tend to be smaller than that of the Frost algorithm, as the block length is increased     

Gradient-based learning applied to document recognition

Multilayer neural networks trained with the back-propagation algorithm constitute the best example of a successful gradient based learning technique. Given an appropriate network architecture, gradient-based learning algorithms can be used to synthesize a complex decision surface that can classify high-dimensional patterns, such as handwritten characters, with minimal preprocessing. This paper reviews various methods applied to handwritten character recognition and compares them on a standard handwritten digit recognition task. Convolutional neural networks, which are specifically designed to deal with the variability of 2D shapes, are shown to outperform all other techniques. Real-life document recognition systems are composed of multiple modules including field extraction, segmentation recognition, and language modeling. A new learning paradigm, called graph transformer networks (GTN), allows such multimodule systems to be trained globally using gradient-based methods so as to minimize an overall performance measure. Two systems for online handwriting recognition are described. Experiments demonstrate the advantage of global training, and the flexibility of graph transformer networks. A graph transformer network for reading a bank cheque is also described. It uses convolutional neural network character recognizers combined with global training techniques to provide record accuracy on business and personal cheques. It is deployed commercially and reads several million cheques per day