The main thrust of this paper is to encourage the use of neural networks to process raw data for subsequent classification. This article addresses neural network techniques for processing raw pixel information. For this paper the definition of neural networks includes the conventional artificial neural networks such as the multilayer perceptrons and also biologically inspired processing techniques. Previously, we have successfully used the biologically inspired Gabor transform to process raw pixel information and segment images. In this paper we extend those ideas to both segment and track objects in multiframe sequences. It is also desirable for the neural network processing data to learn features for subsequent recognition. A common first step for processing raw data is to transform the data and use the transform coefficients as features for recognition. For example, handwritten English characters become linearly separable in the feature space of the low frequency Fourier coefficients. Much of human visual perception can be modelled by assuming low frequency Fourier as the feature space used by the human visual system. The optimum linear transform, with respect to reconstruction, is the Karhunen-Loeve transform (KLT). It has been shown that some neural network architectures can compute approximations to the KLT. The KLT coefficients can be used for recognition as well as for compression. We tested the use of the KLT on the problem of interfacing a nonverbal patient to a computer. The KLT uses an optimal basis set for object reconstruction. For object recognition, the KLT may not be optimal.

Two of our analog neural net chips have been integrated into board systems and are being used now in a variety of image recognition applications. One of the two circuits, the NET32K chip, has connections with a low resolution of between one and four bits. With this chip one can scan up to 32 kernels of a size of 16 X 16 pixels over an image. It is used mainly for extracting geometrical features from images, for such applications as image segmentation. The second of the chips, named ANNA, operates with a higher resolution of 6 bits in the weights and 3 bits in the states. It has been designed for implementing nets to recognize characters. The computation rates obtained with these circuits are 10 to 100 times faster than those of standard processors. With the NET32K chip we achieve between two and ten billion connections per second. With the ANNA chip we read over 150 characters per second, a tenfold increase compared with a digital signal processor.

We consider the classification of multiple objects in a scene with distortion and clutter present. Our opinions on the role for neural nets (NNs) in this application and the different properties that NNs must have to address this problem are advanced. A hierarchical/inference approach is suggested using correlation NNs for low-level operations and new classifier NNs with higher-order decision surfaces for the final decision NNs. Our concern is NN capacity and performance (in noise). Our capacity guidelines advanced concern the number of neurons, use of analog neurons, Ho-Kashyap (HK) NNs, and two new NNs with higher-order decision surfaces. Our noise performance guidelines advanced concern the number of neuron layers, hidden-layer neuron encoding, and robust HK NNs.

This paper describes the application of a visual pattern recognition neural network in a hybrid model based automatic target recognition (ATR) system. This neural network forms the feature extraction front end of the ATR and is derived from the Neocognitron network first proposed by K. Fukushima. For complex target recognition, modifications to the basic Neocognitron network paradigm were required to enhance robustness against image distortions due to undersampling (aliasing) and poor feature selection during training. The focus of the paper is on the enhancements, their rationale, and on the use of the network as a self- organizing feature extraction element of an ATR. Results of experiments with the overall ATR system against target imagery are shown and discussed.

An image segmentation system based on known cortical interactions and topography is presented. First, a space variant retinal-like representation of the image is obtained not only to provide with spatial focusing but also to reduce the computational load. Next, a layer of receptive fields for feature extraction was evolved from a set of Gabor functions with multiple orientational characteristics. Shunting competitive interactions among different features eliminate local ambiguities. Long-range interactions among `winning' cells, sharing similar orientation preferences, but with possible different spatial scales, are used to form a congruent description of the visual scene taking into account spatial context. The output of this layer is used to partition the image into emergent segments. Encouraging results of MRI images processing are presented.

Hierarchical artificial neural networks are designed to enhance edge measurement. The neural network comprises four subnets: the Edge Contour Detection subnet, the Maximum Detection subnet, the Gradient Adjustment subnet, and the Orientation Determination subnet. The interconnections between these subnets are fashioned in a hierarchical manner. In order for the neural network system to perform correctly and accurately, each of the neural subnets must be given suitable weights by learning. The learning is very difficult for the hierarchical neural networks because of the complicated hierarchical structure. In our learning algorithm the modularity is introduced for fast learning and good generalization, based on the analysis of the local concept and the distributed concept represented by the module. The amount of information which the nets need to learn is drastically reduced. Therefore, only a small number of training patterns are required to train the nets and still derive suitable weights for the nets to perform accurately and efficiently. The neural network is simulated on a MIPS M120-S machine running UNIX. For the test images degraded by random noise up to 20%, the true edges are detected and enhanced, the false edges are suppressed, the noise is eliminated, the weak edges are reinforced, and the missing edge elements are interpolated.

Texture is an important surface characteristic. Many industrial materials such as wood, textile, or paper are best characterized by their texture. Detection of defaults occurring on such materials or classification for quality control anD matching can be carried out through careful texture analysis. A system for the classification of pieces of wood used in the furniture industry is proposed. This paper is concerned with a neural network implementation of the features extraction and classification components of the proposed system. Texture appears differently depending at which spatial scale it is observed. A complete description of a texture thus implies an analysis at several spatial scales. We propose a compact pyramidal representation of the input image for multiresolution analysis. The feature extraction system is implemented on a multilayer artificial neural network. Each level of the pyramid, which is a representation of the input image at a given spatial resolution scale, is mapped into a layer of the neural network. A full resolution texture image is input at the base of the pyramid and a representation of the texture image at multiple resolutions is generated by the feedforward pyramid structure of the neural network. The receptive field of each neuron at a given pyramid level is preprogrammed as a discrete Gaussian low-pass filter. Meaningful characteristics of the textured image must be extracted if a good resolving power of the classifier must be achieved. Local dominant orientation is the principal feature which is extracted from the textured image. Local edge orientation is computed with a Sobel mask at four orientation angles (multiple of (pi) /4). The resulting intrinsic image, that is, the local dominant orientation image, is fed to the texture classification neural network. The classification network is a three-layer feedforward back-propagation neural network.

ANNIE is a neural network that removes noise and sharpens edges in digital images. For noise removal, ANNIE makes a weighted average of the values of the pixels over a certain neighborhood. For edge sharpening, ANNIE detects edges and applies a correction around them. Although averaging is a simple operation and needs only a two-layer neural network, detecting edges is more complex and demands several hidden layers. Based on Marr's theory of natural vision, the edge detection method uses zero-crossings in the image filtered by the ∇²G operator (where ∇² is the Laplacian operator and G stands for a two- dimensional Gaussian distribution), and uses two channels with different spatial frequencies. Edge detectors are tuned for vertical and horizontal orientations. Lateral inhibition implemented through one-step recursion achieves both edge relaxation and correlation of the two channels. Training by means of the quickprop algorithm determines the shapes of the weighted averaging filter and the edge correction filters, and the rules for edge relaxation and channel interaction. ANNIE uses pairs of pictures as training patterns: one picture is a reference for the output of the network and the same picture deteriorated by noise and/or blur is the input of the network.

This paper deals with the problem of using neural networks in the field of texture analysis in color images. Different texture parameters are computed on each color for a given texture. Those parameters are then given as inputs to a neural network which learns to recognize textures from a sample. Recognition performance and weighting matrix are analyzed.

This paper describes several different types of constraints that can be placed on multilayered feedforward neural networks which are used for automatic target recognition (ATR). We show how unconstrained networks are likely to give poor generalization on the ATR problem. We also show how the ATR problem requires a special type of classifier called a one-class classifier. The network constraints come in two forms: architectural constraints and learning constraints. Some of the constraints are used to improve generalization, while others are incorporated so that the network will be forced to perform one-class classification.

In this paper, a robust, and fast system for recognition as well as pose estimation of a 3-D object from a single 2-D perspective of it taken from an arbitrary viewpoint is developed. The approach is invariant to location, orientation, and scale of the object in the perspective. The silhouette of the object in the 2-D perspective is first normalized with respect to location and scale. A set of rotation invariant features derived from complex and orthogonal pseudo- Zernike moments of the image are then extracted. The next stage includes a bank of multilayer feed-forward neural networks (NN) each of which classifies the extracted features. The training set for these nets consists of perspective views of each object taken from several different viewing angles. The NNs in the bank differ in the size of their hidden layer nodes as well as their initial conditions but receive the same input. The classification decisions of all the nets are combined through a majority voting scheme. It is shown that this collective decision making yields better results compared to a single NN operating alone. After the object is classified, two of its pose parameters, namely elevation and aspect angles, are estimated by another module of NNs in a two-stage process. The first stage identifies the likely region of the space that the object is being viewed from. In the second stage, an NN estimator for the identified region is used to compute the pose angles. Extensive experimental studies involving clean and noisy images of seven military ground vehicles are carried out. The performance is compared to two other traditional methods, namely a nearest neighbor rule and a binary decision tree classifier and it is shown that our approach has major advantages over them.

A computer vision system is developed for 3-D object recognition using artificial neural networks and a knowledge-based top-down feedback analysis system. This computer vision system can adequately analyze an incomplete edge map provided by a low-level processor for 3-D representation and recognition using key features. The key features are selected using a priority assignment and then used in an artificial neural network for matching with model key features. The result of such matching is utilized in generating the model-driven top-down feedback analysis. From the incomplete edge map we try to pick a candidate pattern utilizing the key feature priority assignment. The highest priority is given for the most connected node and associated features. The features are space invariant structures and sets of orientation for edge primitives. These features are now mapped into real numbers. A Hopfield network is then applied with two levels of matching to reduce the search time. The first match is to choose the class of possible model, the second match is then to find the model closest to the data patterns. This model is then rotated in 3-D to find the best match with the incomplete edge patterns and to provide the additional features in 3-D. In the case of multiple objects, a dynamically interconnected search strategy is designed to recognize objects using one pattern at a time. This strategy is also useful in recognizing occluded objects. The experimental results presented show the capability and effectiveness of this system.

This model introduces rotation invariance into the Neocognitron model by way of rotation layers. The most likely orientation of the target is detected and used for further pattern classification using an algorithm similar to that for the Neocognitron model. The visual target recognition capability of this model is demonstrated.

The market for a product that can read handwritten forms, such as insurance applications, re- order forms, or checks, is enormous. Companies could save millions of dollars each year if they had an effective and efficient way to read handwritten forms into a computer without human intervention. Urged on by the potential gold mine that an adequate solution would yield, a number of companies and researchers have developed, and are developing, neural network-based solutions to this long-standing problem. This paper briefly outlines the current state-of-the-art in neural network-based handwriting recognition research and products. The first section of the paper examines the potential market for this technology. The next section outlines the steps in the recognition process, followed by a number of the basic issues that need to be dealt with to solve the recognition problem in a real-world setting. Next, an overview of current commercial solutions and research projects shows the different ways that neural networks are applied to the problem. This is followed by a breakdown of the current commercial market and the future outlook for neural network-based handwriting recognition technology.

The optimality of the Karhunen Loeve (KL) transform is well known. Since its basis is the eigenvector set of the covariance matrix, a statistical, not functional, representation of the variance in pattern ensembles is generated. By using the KL transform coefficients as a natural feature representation of a character image, the eigenvector set can be regarded as an unsupervised biological feature extractor for a (neural) classifier. The covariance matrix and its eigenvectors are obtained from 76,753 handwritten digits. This operation is a unique expense; once the basis set is calculated it forms a linear first layer of a three weight layer feed forward network. The subsequent nonlinear perceptron layers are trained using a scaled conjugate gradient algorithm that typically affords an order of magnitude reduction in computation over the ubiquitous back-propagation algorithm. In conjunction with a massively parallel computer, training is expedited such that tens of initially different random weight sets are trained and evaluated. Increase in training set size (up to 76,755 patterns) gives less accurate learning but improved generalization on the fixed disjoint test set. A neural classifier is realized that recognizes 96.1% of 15,000 handwritten digits from 944 different writers. This recognition is attributed to the energy compaction optimality of the KL transform.

This paper summarizes a research effort to explore the use of neural networks in processing hyperspectral imagery for the purpose of target detection and feature selection in an automatic target detection (ATR) scenario. Images containing 32 spectral bands in the 2.0 to 2.5 micrometers infrared range and with co-registered pixels were used to train and test a backpropagation neural network for detection of ground targets. The dimensionality of the original feature set was reduced using two methods, Karhunen-Loeve and Ruck's saliency technique. The results for the two feature selection techniques are compared using classifier performance as a metric. Finally, a neural network chip (ETANN) was used to test the feasibility of hardware implementation of the fusion processing.

The newest airborne and spaceborne imaging sensors, unlike their predecessors, generate data at tens to hundreds of wavelengths simultaneously. The advent of these imaging spectrometers greatly increases the detail of the information available to the analyst, but the concomitant disadvantages of vastly increased data volume and complexity have stimulated the search for more efficient and accurate hyperspectral data exploitation techniques. This paper describes two neural network models used in conjunction to analyze hyper-spectral image data. The first neural network is based on an unsupervised backward error propagation model. This network generates without supervision low dimensional discriminant feature vectors from the original high dimensional signals. These features are analyzed by the second neural network, which is a layered feed-forward neural network based on statistical principles, namely Parzen window density estimation and Bayes decision rule. This network estimates class-conditional probability density functions using training samples and classifies data points so that the expected risk is minimized. Results of applying these two neural networks to 210-band AVIRIS data are presented.

We describe a neural network based recognition scheme for 2-D objects. The Fourier Descriptors of the object boundary are taken as the features and they form the input to the neural network. A multilayered perceptron architecture is used for the classification, and a stochastic algorithm called Alopex is used for the network learning. The scheme is invariant to translation, rotation, and scale changes to the object. Taking isolated handwritten digits as the input data set, we show that the presented scheme gives very high recognition accuracy. The recognition scheme, learning algorithm, and simulation results are discussed in detail.

Classification of object signatures is not a trivial task. An object's signature only represents one particular aspect of the object's physical characteristics such as fluid density or voltage resistance over time. Fortunately, most objects produce a unique signature based on a number of features embedded within the signature. The objective of this presentation is to demonstrate and compare the usefulness of a variety of Neural Networks on object signature discrimination as well as to present the limitations and common problems associated with this task. The networks used are back propagation, recurrent back propagation, Kohonen, and the Padaline, each of which are well known networks and are easily obtained. Issues such as dynamic and cluttered environments are discussed with possible solutions, including pre/post filtering, fuzzy logic, and multiple networks. In support of the discussion, the preparation and results of applying each type of network to the radiometric signature of simple objects such as cubes, spheres, and cones are presented, thus demonstrating a variety of input, output, and pre/post- processing techniques.

An application of neural networks is the recognition of objects under translation, rotation, and scale change. Most existing networks for invariant object recognition require a huge number of connections and/or processing units. In this paper, we propose a new connectionist model for invariant object recognition for binary images with a reasonable network size. The network consists of five stages. The first stage shifts the object so that the centroid of the object coincides with the center of the image plane. The second stage is a variation of the polar- coordinate transformation used to obtain two N-dimensional representations of the input object. In this stage, the 0 axis is represented by the positions of the output units; therefore, any rotation of the original object becomes a cyclic shift of the output values of this stage. The third stage is a variation of the rapid transform, which provides invariant representations of cyclic-shift inputs. The next stage normalizes the outputs of the rapid transform to obtain scale invariance. The final stage is a nearest neighbor classifier. We tested the performance of the network for character recognition and good results were obtained with only one pattern per class in training.

Perceptron learning for the Bayesian classification problem has been analyzed in the framework of Markov diffusion under the influence of competing stochastic forces. The analytic solution for the one-layer architecture yields an immediate relationship between the statistics of the input signal and the weight configuration built by the Perceptron. The computer simulation of the Perceptron learning classification of image-like patterns governed by the Poisson distribution demonstrates a dependence of the learning dynamics on the number of layers and the size of the training data set.

Object recognition is a complex task involving simultaneous problems in grouping, segmentation, and matching. Previous work involved an objective function formulation of the problem, resulting in a uniform method of addressing problems in object recognition that have heretofore been approached by heterogenous complex vision systems. The complexity of our objective functions resulted in numerous optimization failures, not unexpectedly. Here we propose to prime the system with estimates of the objects parameters at a coarse, more abstract, scale. We discuss how this might be done. These initial values are expected to bring the state of the system closer to good minima.

Back propagation neural networks have been widely used as classifiers in many complex classification tasks. However, early experimental results show that as the number of classes involved in a classification task increases, the classification accuracy of these networks decreases, especially in the presence of noisy inputs. In addition, larger size networks are needed to be utilized in such cases, a fact that may not always be possible. In order to overcome both of these undesirable effects a new approach is proposed in this paper which utilizes multiple, relatively small size networks to perform the classification task. This approach has been applied on a machine printed character recognition experiment and it has demonstrated better classification accuracy than the one exhibited by the single, larger size, network approach.

This paper compares the performance of an artificial neural network technique to that of two conventional techniques in fusing (classifying) multispectral imagery. The true classification error rate is estimated by use of the k-fold cross-validation technique for a Bayesian classifier, a binary tree classifier, and a backpropagation neural network. The cascade correlation neural network is also described and its theory of operation is compared to that of the backpropagation neural network.

This work describes a neural network-based approach to multimodal real-world mapping and navigation for autonomous mobile robots in unknown environments. The system is built on top of a vector associative map to combine range data from stereo vision and ultrasonic rangefinders. Visual output from a boundary contour system is used to extract range data from a pair of 2-D images. In addition, range data from ultrasonic lasers is used to eliminate uncertainties, noise, and intrinsic errors introduced by the measurements. A recurrent competitive field used to model multimodal working memory excites a trajectory formation network which transforms desired temporal patterns (i.e., a trajectory formation pattern) into spatial patterns. The output of this network is processed by direction-sensitive cells which in turn activates the motor system that guides a mobile robot in unstructured environments. The model is capable of unsupervised, real-time, fast error-based learning of an unstructured environment.

We have applied a shift-invariant back propagation neural network to the cells' boundary detection of human corneal endothelium photomicrographs. The interconnection patterns of the neural network are constrained to be spatially invariant. Our experiments demonstrated that the generalizing ability of the neural network is dependent on its connectivity or complexity, i.e., the number of connections. To get a better generalization, we have proposed a modified back propagation learning rule, which reduces the complexity of the neural network during the learning procedure. The measure of the complexity of the neural network was defined as a formal entropy of the connectivity pattern. Simulations showed that the neural network trained by this learning rule has better generalizing ability but more computational complexity. In this paper a simplified function is investigated as the measure of the complexity of the neural network. This simplified measure is a first order approximation of the formal entropy measure. Simulations show that the simplified measure is effective to get better generalization and has less computational complexity than the original formal entropy measure.

Accurate identification of cloud type is an important aspect of weather forecasting. One of the primary applications of the remotely sensed cloud cover data is to provide synoptic cloud cover information over extensive data-sparse regions; particularly the oceans and deserts. In southeast Asia, information on cloud cover data is obtained from the infrared and visible channels by Geostationary Meteorological Satellite. These imageries contain data of clouds. By extracting the textural features embedded in the images, information on cloud types can be derived and mapped spatially. An artificial neural network is used as a classifier to identify different cloud types through comprehensive training cycles. The architecture of the network used in the present study is multilayered with feedforward and backpropagation. The study makes use of a classification scheme based on the SYNOP code of the World Meteorological Organization (WMO). The average cloud classification accuracy obtained in this study is 40%.

Famous artists' paintings, in general, allow for a large number of forgeries. In the work of a great Brazilian painter, Candido Portinari, we try to detect fake works through their image. To reach classifying results we must extract from digitalized images features that distinguish Portinari's original paintings from the false ones. So, it has been noted that the degree of variation of gray tones in a brush stroke reflects each painter's particular style on the painting. It is a feature that can be easily detected by the power spectrum of the detail image. However, as power spectra have large amounts of data, we use just some significant values of them for the classification. These selected pixels of the spectrum image are in a line which is perpendicular to the direction the brush passed, i.e., they indicate the variation of gray tones in the brush stroke. The data described previously are used as input for further classification by a backpropagation neural network. This neural net was exhaustively trained and has topology and parameters appropriate to the problem. Two output units indicate the major result: original or false.

A system is presented that applies a neural network to a video surveillance system. It consists of a pre-processing unit that extract high level information from images and introduces it in the neural network. This system can learn in operational conditions while under the supervision of an unskilled operator. It uses the error backpropagation learning algorithm in a multilayer perceptron structure. The results obtained show that the system performs well, and with a high degree of efficiency.

In this paper we present a neural network approach to shape recognition. The primary focus is the development of an effective representation method to increase the degree of robustness in recognition of shapes which may be blurred by noise. A space-perturbation neural network (SPNN) is described which is characterized by two important properties. (1) The network can be trained using error back-propagation with only noise-free data, which avoids the convergence problem due to possibly large variance in each shape class. (2) The space- perturbation arrangement enables the network to discover class features independent of random variations in shape and, hence, not blurred by random variations in the input. As a recognition task, the classification of four closed planar shapes was chosen. For comparison, the neural network classifier based upon extended training technique was implemented to perform the same task. Performance evaluation over 4000 testing shapes from various blur conditions showed that (1) the SPNN outperforms the extended training approach, especially when blur is present -- a median 25% reduction in the number of misclassification is achieved for the noisy case; (2) the SPNN performs about the same as the extended training approach for the noise-free case; and (3) the SPNN can be trained with a much reduced time consumption in comparison to the extended training scheme.

A neural network model, namely, Kohonen's Feature Map, together with the optimal feedforward network is used for variable font machine printed character recognition with tolerance to rotation, shift in position, and size errors. The determination of object orientation is found using the many rotated versions of individual symbols. Orientations are detected from printed text, but no knowledge of the context is used. The optimal Bayesian detector is derived, and it is shown that the optimal detector has the form of a feedforward network. This network together with the learning vector quantization (LVQ) approach is able to implement an inspection system which determines the orientation of the fonts. After the size normalization, rotation, and component finding process as a preprocessing step, the text becomes the input for the feature map. The feature map is trained first in an unsupervised manner. The algorithm is then adapted for supervised learning using improved LVQ technique. Rectangular and minimal spanning tree (MST) neighborhood topologies are experimented with. The results are encouraging, 87% of the characters of various fonts are correctly recognized even though the pattern is distorted in shape and transformed in a shift, size, and rotation invariant manner. Experimental results and comparisons are described.

The solution of the data association problem in a multi-target tracking scenario requires the computation of probabilities, (beta) _i^j, of assigning the i-th measurement to the j-th target. Previously, we have suggested a parallel structure based on a layered, asynchronous (sequential) Boltzmann machine for the estimation of the association probabilities. An efficient analog realization of this structure with stochastic neurons is presented here. The dynamics of this network are described by a vector Langevin equation, and as a result, the network approximates a purely synchronous Boltzmann machine, with potentially rapid convergence. Asymptotically, the probability (beta) _i^j equals the activation frequency of the quantized neuron output (upsilon) _ij, in a layered two-dimensional network. Design criteria for approximating the true association probabilities are described. The transient and steady state behaviors of each stochastic neuron, shown to be a diffusion process in a bounded region, are analyzed. The performance of the layered diffusion network is compared with a theoretical bound and also with the performance of an asynchronous Boltzmann machine.

The electronic architecture and silicon implementation of an artificial neuron which can be used to process and classify dynamic signals is described. The electrical circuit architecture is modeled after complex neurons in the vertebrate brain which have spatially extensive dendritic tree structures that support large numbers of synapses. The circuit is primarily analog and, as in the biological model system, is virtually immune to process variations and other factors which often plague more conventional circuits. The nonlinear circuit is sensitive to both temporal and spatial signal characteristics but does not make use of the conventional neural network concept of weights, and as such does not use multipliers, adders, look-up-tables, microprocessors, or other complex computational devices. We show that artificial neural networks with passive dendritic tree structures can be trained, using a specialized genetic algorithm, to produce control signals useful for target tracking and other dynamic signal processing applications.

This paper describes a discrete-time analog neural network solution to the data association, or data correlation problem. This work, which is an extension of previous investigations, was originally motivated by the earlier results of Sengupta and Iltis (1989). In this paper, we exploit the fact that the associated optimization problem is loosely described, and map the data association problem onto an analog discrete-time neural network connected in an on-center, off-surround configuration. This reduces the number of parameters required in the system design, thereby also reducing the system sensitivity to parameter variations, and leading to greater robustness. Results are presented for simulations performed on a typical workstation. Simulations were also performed with reduced precision numbers. The performance in both cases were not identical, and parameter adjustments in a specific direction are needed in the finite-precision case for acceptable performance.

Frequently, multipoint target tracking is achieved using a Kalman Filter or other means. Numerous papers have been published over the past decades on tracking of dynamic systems such as ships, planes, artillery shells, and control processes with Kalman Filters, particularly, when the mathematical equations of motion describing the dynamic system are available. Then, target tracking is a fairly straight forward procedure. In this paper, a back propagation neural network is successfully `trained' for tracking an artillery shell. It is a predictive neural network because its outputs are the future positions of the artillery shell.

In this paper, we propose a new scheme for image compression using neural networks. Image data compression deals with minimization of the amount of data required to represent an image while maintaining an acceptable quality. Several image compression techniques have been developed in recent years. We note that the coding performance of these techniques may be improved by employing adaptivity. Over the last few years neural network has emerged as an effective tool for solving a wide range of problems involving adaptivity and learning. A multilayer feed-forward neural network trained using the backward error propagation algorithm is used in many applications. However, this model is not suitable for image compression because of its poor coding performance. Recently, a self-organizing feature map (SOFM) algorithm has been proposed which yields a good coding performance. However, this algorithm requires a long training time because the network starts with random initial weights. In this paper we have used the backward error propagation algorithm (BEP) to quickly obtain the initial weights which are then used to speedup the training time required by the SOFM algorithm. The proposed approach (BEP-SOFM) combines the advantages of the two techniques and, hence, achieves a good coding performance in a shorter training time. Our simulation results demonstrate the potential gains using the proposed technique.

An error-less image compression of complex images has been achieved using a massively parallel computer. The algorithm involves utilization of a multi-level hierarchical structure of Kohonen type self-organizing learning vector quantization. The compression ratio increases greatly if a small amount of error is tolerated by limiting the number of templates employed. Utilization of DAP 610 enables the processing of compression and reconstruction in very short time. A few cases of error-less compression of several images, as well as some examples which achieved higher compression ratios by allowing a reasonable amount of error, are shown and compared.

The back-propagation neural network algorithm is an iterative technique for learning the relationship between an input and output. This algorithm has been successfully used in many real-world applications; however, it suffers from slow convergence problems and can get struck in local minima of the weight-error surface. A generalization of previously proposed activation functions has been developed using a free parameter for improved convergence. A modified algorithm based on these function is suggested by bounding the input data to the derivative evaluation in the backwards pass. The modified algorithm has demonstrated superior performance on the standard parity and encoder problems. Finally, a histogram normalization technique is presented for image data compression for improved convergence.

Vector quantization is a powerful technique for image coding. Generalized Lloyd Algorithm is a commonly acknowledged non-neural network benchmark algorithm for vector quantizer design. The main drawbacks of this algorithm are its computational complexity and large memory requirement. On the other hand, competitive learning algorithms in neural networks appear to be well-suited for use in vector quantization applications. In this paper, we first discuss three existing competitive learning algorithms for training a vector quantizer, Basic Competitive Learning Algorithm, Kohonen Self-Organizing Feature Map, and Frequence- Sensitive Competitive Learning Algorithm. Then, we introduce a new Near-Optimal Learning Algorithm (NOLA) which combines the advantages of both KSFM and K-means algorithms. It `on-line' generates codebooks with the least disturbance to previously learned results. Finally, we compare NOLA with the Generalized Lloyd Algorithm. The experimental results show that NOLA achieves a near-optimal performance with the entire execution time slightly more than one iteration of Generalized Lloyd Algorithm. This offers the potential for achieving real-time vector quantization. In addition, NOLA does not need to store the entire training set but only the previously trained results.

The basic idea of the system proposed in this paper lies in the fact that an image usually includes areas of different significance, so we have to code them in a different way to reach an accurate reproduction. Our system divides an image into areas of various importance which we code using wavelet transformations and neural networks for knowledge-based recognition. In this paper, we explain how the functional relationship between intensity and spatial frequency at the limits of human perception in vision (Contrast Sensitivity Threshold (CST) Curve) can guide one to choose the norm of the error metrics, the compression level in the wavelet hierarchy, and the coefficient quantization strategies to minimize the human perception of error. The CST curve is learned by a backpropagation neural network.

A vector quantizer based on artificial neural networks is developed for use in digital video data compression applications. A series of experiments investigating the edge performance of various distortion measures and experiments exploring various vector sizes are presented. The paper then describes a differential vector quantizer which preserves edge features and an adaptive algorithm, Frequency-Sensitive Competitive Learning, which is used to develop equiprobable vector quantizer codebooks. By using codebooks comprised of equiprobable codevectors, variable length coding is unnecessary which results in robust performance in the presence of channel bit errors. The resulting coder is efficient, robust, and permits real-time hardware realizations. The DVQ coder currently under construction is also described.

A practical neural networks paradigm is described for material handling. The learning algorithm is a modification of the cerebella model articulation controller (CMAC) developed by Albus. A table look-up approach detects faults by monitoring the output patterns from sensors and actuators. By analyzing the timing sequence, abnormal conditions can be detected. CMAC offers an alternative to conventional back-propagated, multilayered networks, with the advantage of rapid convergence. The approach appears to be more efficient for the on-line and real-time applications required in automated systems.

The advent of low cost, reliable sensor technology and ongoing dramatic improvements in computer price/performance have transformed the average manufacturing facility into a data- rich environment with millions of bytes of production information stored daily. This production data contains valuable information about the process that can be used by a neural network to model, control, and optimize the plant dynamics. This paper presents a perspective on the use of neural networks and fuzzy logic technology and outlines the methods that have been used to improve process performance in several applications in chemical/petrochemical production facilities.

The speed and accuracy of convergence of iterative optimization algorithms often depend critically upon the choice of a starting point. With a near optimum starting point, both speed and accuracy can be improved. A two step approach to optimization has been developed which utilizes the feedforward predictive capability of a neural network in conjunction with the feedback capability of an iterative optimization algorithm. This approach is taken in order to improve the speed of the iterative optimization algorithm, and also enhance the iterative algorithm's ability to locate a global optimum. This technique has been applied to the problem of system identification for continuous time transfer function models. The neural network is used to select an initial set of process parameters for a given model structure using unit step response data. We present results on the accuracy of the predictive capability of the neural network, and results showing the improved performance of the iterative nonlinear system identification algorithm.

To build advanced controllers one needs some arbitrary simplified models of the processes to be controlled. To overcome this situation, we are investigating the use of neural network techniques. The first problem is how to include dynamics in a neural net. We have considered recurrent neural networks and dynamical neurons. We have implemented an extended backpropagation to take into account the dynamics in the training phase. Then, a general method has been developed to build neural advanced controllers. This approach is based on the minimization of a closed-loop control criterion and on an architecture in which some control knowledge can be integrated. As a test example, we have chosen to control the pH of a mixture of a strong acid and a strong base; the strong non-linearity of this process results from the static gain being highly variable when the pH is around 7. The control performance of the neural controller is then compared with a geometric non-linear control. Finally, we show, in our example, that some neural network parameters can be tuned as can parameters in classical controllers.

Simplified ART1 performs unsupervised classification of binary input patterns, forming prototypes of object classes by grouping similar objects together. The model is based on ART1, but uses a new similarity measure and activation function which eliminate the need for sequential search through previously learned prototypes. The new similarity measure minimizes recoding of inputs into different categories, making stabilization faster. Simplified ART1 requires only half the weights of ART1, which makes it easier to implement. In summary, we introduce modifications to ART1 which produce a faster, more efficient and simpler model. Results obtained in software simulations are used to compare the performance of the ART1 and Simplified ART1 models.

We believe that an essential feature in machine learning is the real time satisfaction of multiple objectives such as identification, tracking, etc. The machine learning problem may be viewed as a nonlinear adaptive control problem where the environment plays the role of the `plant,' while the learner is the controller. Multiobjective optimization (MOO) in the control problem typically deals with simultaneous optimization of more than one objective, where each objective is described via a cost functional. In such a situation there often exists a region of tradeoff wherein one cost may be improved at the expense of others. Such a region is called the Pareto optimal (PO) set. A parameterization of this set simplifies the attainment of the existing tradeoff. Working within the Pareto set guaranties optimum tradeoff. As an example this algorithm is applied to the control of a dc motor.

In this study an intelligent control architecture is proposed for recognition and diagnosis of digital images captured by the vision system to control the robot manipulator in automated precision assembly. The architecture proposed provides enough flexibility to the assembly cell in processing a wide variety of products that require precision assembly. The system uses ART paradigm for part recognition and backpropagation for calibration. System architecture is tested using an IBM 7547 robot with a CCD camera for precision assembly of a printed circuit board (PCB) and satisfactory results are obtained.

Nuclear power plants are very complex systems. The diagnoses of transients or accident conditions is very difficult because a large amount of information, which is often noisy, or intermittent, or even incomplete, needs to be processed in real time. To demonstrate their potential application to nuclear power plants, neural networks are used to monitor the accident scenarios simulated by the training simulator of TVA's Watts Bar Nuclear Power Plant. A self-organization network is used to compress original data to reduce the total number of training patterns. Different accident scenarios are closely related to different key parameters which distinguish one accident scenario from another. Therefore, the accident scenarios can be monitored by a set of small size neural networks, called modular networks, each one of which monitors only one assigned accident scenario, to obtain fast training and recall. Sensitivity analysis is applied to select proper input variables for modular networks.

The design of an autonomous undersea vehicle (AUV) control system is a significant challenge in light of the highly uncertain nature of the ocean environment together with partially known nonlinear vehicle dynamics. This paper describes a neural network architecture called Cerebellar Model Arithmetic Computer (CMAC). CMAC is used to control a model of an autonomous underwater vehicle. The AUV model consists of two input parameters, the rudder and stern plane deflections, controlling six output parameters; forward velocity, vertical velocity, pitch angle, side velocity, roll angle, and yaw angle. Properties of CMAC and results of computer simulations for identification and control of the AUV model are presented.

Writing large-scale neural network software models for real-world applications is challenging. Computer-assisted model creation facilitates the development process by ensuring the accuracy of the software implementation and by relieving a programmer of a portion of the tedium of writing programs. Furthermore, a formal declarative representation for model specification allows for the application of tools for the synthesis, analysis, validation, and presentation of neural network software models.

There are many instances where a sequence of occurrences becomes highly significant when placed against a timeline. Examples of these occur in control environments such as manufacturing, military I&W, environmental control, and a myriad of other process control mechanisms. Classification of these time-based patterns using connectionist models is a flexible means of capturing and assessing such information. This investigation presents an application of connectionist models to classify time-based indications and warning (I&W) patterns. Three threat categories, with increasing degrees of activity overlap within each category, were selected as a problem set. A simulation was developed which generated realistic I&W messages for activities within each category. A recurrent network was then trained to estimate, based upon these simulated I&W messages, which of three possible states an aggressor nation might be in. Each network has an input unit for each of its activity indicators, some number of hidden units, and three outputs (one for each of the three possible states). Training was conducted using epochwise backpropagation through time.

We consider an nth-order SISO continuous nonlinear system of the form x equals f(x) + g(x).u with output y equals h(x) where x and u represent the states and the input, the objective is to transform the system in the state-apace using nonlinear feedback and a diffeomorphism (nonlinear transformation) to a linear system from input to output. This is the so-called Input- Output Feedback Linearization (IOFL) problem. In general, we assume the equations describing the system are not in normal (canonical) form. Assuming that the states x are measurable and the relative degree of the system is equal to the system order (n), a diffeomorphism z equals (phi) (x) can be found such that the system can be linearized in new coordinates z using only a nonlinear state feedback. Based on the traditional approaches in feedback linearization theory, the computation of the IOFL control law requires the exact knowledge of f, g, and h. If these are not known or partially known, an adaptive network can be used to learn such a control law. We employ high-order three layer networks to learn the IOFL control law based on a new method inspired by direct adaptive control architecture, details of which have been published elsewhere. The weights of the network (the controller) are adjusted using generalized delta rule algorithm. A linear reference model of order n is used as a teacher such that the closed-loop response of the resulting system follows the linear system response when both are excited with the same input. The diffeomorphism is implemented using a state variable filter (Paynter filter). Here we intend to report the experimental results obtained for a single link manipulator with flexible joint that is a feedback linearizable system. The transformation of nonlinear systems to linear ones is of interest since linear systems have been widely studied and there is a rich theoretical background as well as sophisticated design tools for controller synthesis. This new approach enables the discovery of IOFL control law without the exact knowledge of the system equations.

In this paper, we propose a learning control scheme for direct trajectory control of robotic manipulators. The main features are that we use a priori structure knowledge of robot dynamics in the design and the neural networks are not used to learn inverse dynamic models. The neural network controller is utilized to compensate the deviation due to the approximate models of robotic manipulators. In addition, true teaching signals of the neural network compensators are employed in the learning phase. Simulations are conducted to show the feasibility of the proposed method.

Atlanta will be the home of several special events during the next five years ranging from the 1996 Olympics to the 1994 Super Bowl. When combined with the existing special events (Braves, Falcons, and Hawks games, concerts, festivals, etc.), the need to effectively manage traffic flow from surface streets to interstate highways is apparent. This paper describes a system for traffic event response and management for intelligent navigation utilizing signals (TERMINUS) developed at Georgia Tech for adaptively managing special event traffic flows in the Atlanta, Georgia area. TERMINUS (the original name given Atlanta, Georgia based upon its role as a rail line terminating center) is an intelligent surface street signal control system designed to manage traffic flow in Metro Atlanta. The system consists of three components. The first is a traffic simulation of the downtown Atlanta area around Fulton County Stadium that models the flow of traffic when a stadium event lets out. Parameters for the surrounding area include modeling for events during various times of day (such as rush hour). The second component is a computer graphics interface with the simulation that shows the traffic flows achieved based upon intelligent control system execution. The final component is the intelligent control system that manages surface street light signals based upon feedback from control sensors that dynamically adapt the intelligent controller's decision making process. The intelligent controller is a neural network model that allows TERMINUS to control the configuration of surface street signals to optimize the flow of traffic away from special events.

A tool wear model based on Abductive Networks, which consists of a network of `polynomial' nodes, is described. The model relates the cutting parameters, components of the cutting force, and machining time to flank wear. Thus real time measurements of the cutting force can be used to monitor the machining process. The model is obtained by a training process in which the connectivity between the network's nodes and the polynomial coefficients of each node are determined by optimizing a performance criteria. Actual wear measurements of coated and uncoated carbide inserts were used for training and evaluating the established model.

A methodology is proposed for the clustering and classification of vibration signatures in the frequency domain. The technique is based on the technologies of neural networks and fuzzy clustering and it is especially suited for the problem of vibration analysis because it permits the incorporation of specific knowledge about the domain in a very simple manner, and because the system learns from actual process data. The system uses the backpropagation algorithm for classification of compressed signatures, where compression is used as a mechanism for noise removal and automatic feature extraction. The clustering system uses the Fuzzy C algorithm with a matrix of weights for the calculation of distances between patterns and centroids. The matrix is used to assign factors of importance to frequencies in the spectrum which are known to be related to particular defects. The two aspects of the analysis (clustering and classification) are complementary because in many cases the exact operating state of a machine cannot be assessed, and clustering may unveil classes of operating states that would not be discovered otherwise. Accurate results were obtained from testing the system on rolling element bearing data.

A major objective in this paper is the application of connectionist architectures for fast and robust on-line learning of dynamic relations used in robot control at the executive hierarchical level. The proposed new connectionist robot controllers as a new feature use decomposition of robot dynamics. In this way, this method enables the training of neural networks on the simpler input/output relations with significant reduction of learning time. The proposed controller structure comprises a form of intelligent feedforward control in the frame of decentralized control algorithm with feedback-error learning method. The other important features of these new algorithms are fast and robust convergence properties, because the problem of adjusting the weights of internal hidden units is considered as a problem of estimating parameters by recursive least squares method. From simulation examples of robot trajectory tracking it is shown that when a sufficiently trained network is desired the learning speed of the proposed algorithms is faster than that of the traditional back propagation algorithm.

This paper proposes an artificial neural-net based method for voltage control in power networks. Power system voltages are necessary to minimize a cost function of the voltage deviation from the target in order to keep sound power system conditions. Centralized control for the voltages is difficult to handle as the power system size becomes larger. Although on- line voltage control is required in power system control, the conventional nonlinear programming based approaches are quite time-consuming. In this paper, a decentralized scheme for voltage control is proposed to minimize the voltage deviations with an artificial neural network. It is expected that the voltage control problem is easier to handle in a decentralized way because of the problem characteristics. A three-layer perceptrons is utilized as an artificial neural network to estimate the optimal control variables. The proposed method allows us to alleviate the curse of dimensionality in power systems and make the computation much more efficient.

Intelligent control is an important field of AI application, which is closely related to machine learning, and the neurocontrol is a kind of intelligent control that controls actions of a physical system or a plant. Linear associative memory model is a good analytic tool for artificial neural networks. In this paper, we present a novel self-learning neurocontrol on the basis of the linear associative memory model to support intelligent control. Using our self-learning neurocontrol model, the learning process is viewed as an extension of one of J. Piaget's developmental stages. After a particular linear associative model developed by us is presented, a brief introduction to J. Piaget's cognitive theory is described as the basis of our self-learning style control. It follows that the neurocontrol model is presented, which usually includes two learning stages, viz. primary learning and high-level learning. As a demonstration of our neurocontrol model, an example is also presented with simulation techniques, called that `bird' catches an aim. The tentative experimental results show that the learning and controlling performance of this approach is surprisingly good. In conclusion, future research is pointed out to improve our self-learning neurocontrol model and explore other areas of application.

In a holographic optical learning network, the decay of multiply exposed holographic interconnections can adversely affect the training of the network. A new dynamic photorefractive holographic memory is described that allows an arbitrarily long sequence of adaptations by rejuvenating decayed holograms with a simple all-optical feedback loop.

Noise in an optoelectronic neural network alters the performance of the network. The shortest distance from a pattern to a decision plane limits the amount of system noise the network can tolerate. In this paper the noise model of a neural network implemented in optoelectronics is given, as well as an upper limit on the noise the network can tolerate and still classify patterns.

A description is first given of the electronic circuitry sufficient and necessary to imitate the action and optical functions of a phase conjugate hologram in the two dimensions of the Euclidean plane. An explanation is then given of its derivation from the underlying principles of UK patent No. GB 2 199 976 (automatic pattern recognition), and of its spatiotemporal (hypercube) applications in neural network form using the phenomena of superconductivity, the genetic code, and a simple geometrical solution of the traveling salesman problem as examples. The presentation includes a discussion of the degree of phase coherence which would be necessary to construct an artificial brain based on a dynamic holographic structure.

As reported previously, one can very efficiently convert an analog voltage to a clean M-valued digital number by using an M-zero neural network. Therefore, using these neural networks in conjunction with some analog arithmetic processors allows us to do the digital multiplication in a much more efficient way than the conventional binary multiplication. That is, we may have the advantages of the speed of the analog system and the accuracy of the M-valued digital system combined together in the new system. In this paper, we report the hardware design and the experimental result of the theoretical work that we previously published along this line.

Several potential applications for analog VLSI vector-matrix multiplier circuits are described. These applications include a bank of multi-dimensional convolution filters, a matrix-matrix multiplier, a discrete Fourier and Hartley transform processor, and a high speed two- dimensional template matcher.

Implementation of an artificial neural network to solve a character recognition problem on highly parallel, single-instruction-multiple-data (SIMD) machines is described. A SIMD simulator workbench developed to evaluate performance of various 1, 4, 8, and 16-bit processor architectures is used for experimentation. A feed-forward neural network with 25, 25, and 3 nodes at the input, hidden, and output layers is trained by error back-propagation to classify characters represented in a pixel grid in the presence of randomly introduced pixel-flip noise. Each node in the neural network is assigned to a processor, with weights distributed across the array. Outputs of each layer are broadcast to the inputs of the next layer via the array's interprocessor communication network. The available number of processors for the network is limited to an array of size 25 X 25. Experimental results from each of the architectures supported in the simulator workbench are presented and analyzed.

Neural networks are investigated for real-time matched filter selection in an optical correlator system. The input to the neural network is a sample space of the optical Fourier transform and the output is a pointer to the correct matched filter. Smaller feedforward network models were ported to analog neural hardware. Simulation and hardware results are discussed. Some architecture recommendations are suggested, specifically an associative memory for filter synthesis.

Optical implementation of Hopfield neural network with a programmable interconnect matrix is reported. Weighted N² interconnections between neurons are realized with a system of a LCLV and CRT controlled by a microcomputer.

A model of 2-dimensional reflexive optical neural networks using a hybrid optoelectronic system made up of a programmable liquid crystal television (LCTV) spatial light modulator with associative capability is presented in this paper.

Holographic photothermoplastic disk with two-dimensional Fourier-holograms registration under infrared laser heating of the recording layer is realized. It supplies high information recording density (10⁵ bit/mm²), and high quality of Fourier-holograms registration ((eta) equals 5 - 7%, contrast 70:1), in addition to the possibility of information rerecording and invariance to the shift of thin phase holograms. Holographic disk based scheme of the optical neural network with outer product implementation is suggested. Results of such neural network modeling are described.

The NNSIM development environment for neural networks aims to cover the path from abstract design to silicon realizations. In the paper, some of the underlying concepts are outlined. From a discussion on the relation between design and prototyping, two digital neural network implementation styles are introduced. Subsequently, the impact of area/time trade- offs on the architecture is analyzed. Finally, the required extensions to NNSIM are listed and a typical application is briefly reviewed.

One of the most important specific problems faced in the analog implementation of neural networks is the storage of synaptic weights. They must have a long storage lifetime, be compact in size, and compatible with a multiplier. They also must be adjustable to provide an on-chip learning capability. The purpose of the circuit presented in this paper was to test the implementation of dynamically refreshed analog memories using conventional CMOS technology with special care in achieving high resolution with small silicon area, and small read-write access time. The circuit prototype presented in this paper implements adjustable analog synaptic weights, compatible with an analog multiplier. The synaptic weights are dynamically refreshed with an on-chip system based on A/D and D/A conversion. The circuit has been fully tested, and a resolution of 7 bits per memory cell was measured. These are arranged in an array of 14 X 15 cells. Dimensions of the circuit, including refreshment system, are 1.7 mm X 1.7 mm in a conservative 2.4 micrometers CMOS process. Translated to a more efficient process (<EQ 1 (mu) CMOS process), this corresponds to a density of +/- 500 memory cells per square millimeter.

This presentation describes the real-time associative recognizing system based on neural network technology for distortion-invariant pattern recognition. The system consists of two cascaded parts, i.e., an improved MADALINE invariant net with local connecting structure, and a feed-forward layered net of probabilistic logic neuron (PLN). The simple learning algorithm is proposed for the system. General-purpose RAMs are major parts of the system so that real-time processing is available.

CMOS VLSI circuits implementing an analog neural network with on-chip contrastive Hebbian learning and capacitive synaptic weight storage have been designed and fabricated. Weights are refreshed by periodic repetition of the training data. To evaluate circuit performance in a medium-sized system, these circuits were used to build a 132 synapse neural network. An adaptive neural system, such as the one described in this paper, can compensate for imperfections in the components from which it is constructed, and thus it is possible to build this type of system using simple, silicon area-efficient analog circuits. Because these analog VLSI circuits are far more compact than their digital counterparts, analog VLSI neural network implementations are potentially more efficient than digital ones.

In this paper, two back propagation neural networks were trained to recognize four hollow cylinders. These cylinders are of the same size and thickness, but made of different materials. Two neural networks were used in this experiment. In case one a single hidden layer network was used, while in case two a hidden layer neural network was used. The acoustic data from these experiments was taken from references 3 and 7. The data are acoustic echoes of hollow aluminum, bronze, glass, and steel cylinders. The results seem to indicate that characteristics of the cylinders are more observable in the modulations of the frequency spectrum than in the time and spectral signals. The results show that a two hidden layer neural network can improve identification rate of a target.

This paper investigated a simple speaker identification system based on a backpropagation neural network. The network was trained to distinguish between a small base of speakers using parameters extracted from human speech which are typically speaker dependent. Different network configurations were tested and observations were made regarding training time, testing errors, and the ability to generalize. The results showed that an experimentally determined optimum network configuration produced no errors during testing and that the parameters used to represent each speaker are sufficient for generalization.

A prototype programmable analog neural computer has been assembled from over 100 custom VLSI modules containing neurons, synapses, routing switches, and programmable synaptic time constants. The modules are directly interconnected and arbitrary network configurations can be programmed. Connection symmetry and modular construction allow expansion of the network to any size. The network runs in real time analog mode, but connection architecture as well as neuron and synapse parameters are controlled by a digital host. Network performance is monitored by the host through an A/D interface and used in the implementation of learning algorithms. The machine is intended for real time, real world computations. In its current configuration maximal speed is equivalent to that of a digital machine capable of 10¹¹ FLOPS. The programmable synaptic time constants permit the real time computation of temporal patterns as they occur in speech and other acoustic signals. Several applications involving the dynamic decomposition and recognition of acoustical patterns including speech signals (phonemes) are described. The decomposition network is loosely based on the primary auditory system of higher vertebrates. It extracts and represents by the activity in different neuron arrays the following pattern primitives: frequency, bandwidth, amplitude, amplitude modulation, amplitude modulation frequency, frequency modulation, frequency modulation frequency, duration, sequence. The frequency tuned units are the first stage and form the input space for subsequent stages that extract the other primitives, e.g., bandwidth, amplitude modulation, etc., for different frequency bands. Acoustic input generates highly specific, relatively sparse distributed activity in this feature space, which is decoded and recognized by units trained by specific input patterns such as phonemes or diphones or active sonar patterns. Through simple feedback connections in conjunction with synaptic time constants the neurons can be transformed into spiking units resembling biological neurons and networks of such units can be used in simulations of small biological neural assemblies. A larger machine, with much higher component count, speed and density as well as higher resolution of synaptic weights and time constants is currently under development. Some specific design issues for the construction of larger machines including selection of optimal component parameters, high density interconnect methods, and control software are discussed.

A variety of artificial neural networks are evaluated for their classification abilities under noisy inputs. These networks include feedforward networks, localized basis function networks, and exemplar classifiers. The performance of radial basis function classifiers deteriorate rapidly in the presence of noise, but elliptical basis variants are able to adapt to extraneous input components quite robustly. For feedforward networks, selective pruning of weights based on an `optimal brain damage' approach helps in noise-tolerant classification. Results from a radar classification problem are presented.

We have used neural networks to separately retrieve atmospheric temperature and moisture vertical profiles from simulated microwave sounder data. Backpropagation networks, implemented on a PC or Sun SPARCstation 2, were trained on data that simulated the multichannel output of either the SSM/T-1 temperature- or SSM/T-2 moisture-sensing radiometers on board current Air Force DMSP weather satellites. Ground-truth information was obtained from the Phillips data, a collection of approximately 1600 validated radio- and rocketsonde measurements. Radiometer outputs were simulated from the ground-truth data using the Air Force program RADTRAN. For each temperature or moisture profile in the Phillips data, RADTRAN generated a corresponding simulated sounder output. These simulated radiometer outputs were combined with the ground truth to produce training and testing sets for the neural networks. For both temperature and moisture, the neural method produces atmospheric profiles from unfamiliar data that are comparable to or better than those obtained with current operational methods. Networks, however, were able to retrieve profiles from a much broader range of geographic areas and seasons than standard methods, and to do this using less externally supplied geographic or season information.

This paper compares three different approaches when used against a problem of present and practical interest: the classification of radar return data from two classes of aircraft. The three approaches are: the typical feature extraction approach used for target classification when dealing with radar data; a multi layer perceptron neural network approach and; a branched multi layer perceptron neural network approach. The comparison was performed under equal conditions and at restricted sensor parameter conditions to demonstrate the anticipated advantage of the neural network approach in that it will be able to classify targets when the sensor parameters are not suitable for the feature extraction approach to work. The classification rate was used as the measure of effectiveness. Up to date the feature extraction approach has provided a classification rate of 84.1%. The multi layer perceptron consists of a one hidden layer network and the best classification rate it has provided is 86.8%. The branched multi layer perceptron consists of two separate multi layer perceptron neural networks trained to recognize only one class of targets and the best classification rate it has provided is 54.9%. The discrepancy in performance between the two neural network approaches is perhaps due to the more general structure with improved discrimination power of the branched network.

One of the basic problems in pattern recognition is the detection of a pattern in noise. This problem becomes particularly difficult if the spectral content of the signal and noise overlap. In this case, high signal-to-noise ratios make detection of the signal very difficult. Noise cancellation using adaptive filters has been successful when the noise source is known. Another problem in pattern recognition involves recognizing the same pattern in different spatial positions. Some special high order neural networks have been shown to exhibit positional invariance, but these systems do not work well in noisy environments. The combined problem of identifying a target that varies in position and is embedded in noise can be approached by cascading systems that attempt to remove the noise and then detect the target with positionally invariant systems.

Improvements in the design of small unmanned underwater vehicles and man's ever increasing exploitation of the oceans, insures that autonomous pattern recognition systems will play an increasingly significant role in undersea systems. Unmanned underwater missions are an attractive alternative to dangerous and costly manned missions. Of national interest are exploration and mapping, environmental monitoring and cleanup, military applications as well as commercial exploitation. High performance classifiers are the greatest unsolved technical challenge to fielding autonomous undersea systems. Classification of man-made underwater objects is particularly difficult because the harsh sensing environment produces low resolution, low signal to noise data. Traditional approaches use single networks. These networks have performance limitations which are due in part to the restricted feature set size they can accommodate. Single networks are therefore not likely to perform optimally under varied environmental conditions. For similar reasons they have a problem taking advantage of multi- sensor data. In this paper, hierarchical neural network architectures, which address the limitations of single network classifiers, are developed. One of these architectures, the sequential classifier, is applied against a difficult side scan sonar data set containing a highly cluttered and variable environment. Performance of the sequential classifiers is compared to the traditional single network classifiers, whose feature sets have been optimized, and show significant improvement. Sequential classifiers have achieved a detection rate of .98 and a .002 false alarm rate. A subset of the networks have been demonstrated on special purpose hardware to run in real time.

Most neural networks have been used for problems of classification. We have undertaken a study using neural networks to predict continuous valued functions which are aperiodic or chaotic. In addition, we are considering a relatively new class of neural networks, ontogenic neural networks. Ontogenic neural networks are networks which generate their own topology during training. Cascade Correlation is one such network. In this study we used the Cascade Correlation neural network to answer two questions regarding prediction. First, how do the number of inputs affect prediction accuracy. Second, how do the number of training exemplars affect prediction accuracy. For these experiments, the Mackey-Glass equation was used with a Tau value of 17 which yields a correlation dimension of 2.1. Takens' theorem for this data set states that the number of inputs to obtain a smooth mapping should be 3 to 5. We were experimentally able to verify this. Experiments were run varying the number of training exemplars from 50 to 450. The results showed that there is an overall trend towards lower predictive rms error with a greater number of exemplars. However, there are good results obtained with only 50 exemplars which we are unable to explain at this time. In addition to these results, we discovered that the way in which predictive accuracy is generally represented, a graph of Mackey-Glass with the network output superimposed, can lead to erroneous conclusions

We present a new methodology of time series analysis with connectionist networks. Our approach is based on a careful analysis of the appropriateness of `neural' concepts within the connectionist methodology. This methodology shows one serious drawback: lack of robustness. We conjecture that it is possible to overcome this difficulty by introducing a scheme that allows monitoring and interpretation of characteristic net values during the process of parameter estimation. We present the concept of time series analysis with connectionist networks and describe our new methodology in detail. In order to demonstrate the usefulness of this approach we present results for an artificial time series sampled form the `van der Pol equation.'

The problem of estimating an unknown function from a finite number of noisy data points (examples) is an ill-posed problem of fundamental importance for many applications, such as machine vision, pattern recognition, and process control. Recently, several new computational techniques for non-parametric regression have been proposed by statisticians, and by researchers in artificial neural networks. However, there is little interaction between the two research communities. The goal of this paper is twofold. First, we present a critical survey of statistical and neural network techniques for non-parametric regression. Second, we present comparisons between a representative neural network technique called Constrained Topological Mapping, and several statistical methods, for low-dimensional regression problems.

In order to offer operating information to the monitoring system of a chip refiner, a neural network classifier is proposed for identifying the acoustic emission signals of wood species. In addition to classifying single wood species, the system is required to be able to recognize mixed species. The classification task is accomplished by a multilayer feedforward neural network in which both supervised and unsupervised learning are included. The simulations are run on the testing data, mixing two single species to represent mutually mixed wood species of five categories. If a signal is identified as a mixture, the network will indicate the corresponding component species according to a lookup table. Some expected classification accuracy is obtained on both single and mixed species identification and performance of classification is discussed based on the simulation results.

We used back propagation and Kohonen neural networks to classify IR spectra. In three experiments the performance of back propagation in classifying IR spectra into a number of predefined categories was evaluated. In another experiment how well a Kohonen network was able to find the same categories of IR spectra humans use was tested.

In this paper, we present an expert system architecture based on a new artificial neural network. Unlike other connectionist approaches, the proposed network paradigm is able to synthesize explicit production rules from the processing elements and the weighted connections of a trained network. The generated rules can be incorporated into an expert system to perform classification tasks such as engineering troubleshooting. The design of the neural network is based on the concepts of probabilistic inference. The network can identify the relevant attributes for classification using a statistical technique called residual analysis. Using the information theoretic weight of evidence measure, the weighted connections are established between the processing elements representing the important attribute values and classes. The proposed network is non-iterative and is therefore very efficient computationally. Since the topology of the network is deterministic, the heuristic functions of each element can be precisely understood and the internal associations directly analyzed to synthesize explicit and intuitive classification rules. This network has been shown previously to outperform the back propagation networks and ID3 in terms of computational efficiency and classification accuracy in certain types of supervised learning applications. Using a typical fault diagnosis task, we show in this paper that the proposed neural network can be used effectively to acquire knowledge for rule-based expert systems. Compared to other AI-based knowledge acquisition approaches using AQ and CN2 algorithms, our proposed approach has the fastest training time while producing the most effective classification rules.

The problem of identifying key landmarks in 3-dimensional surface data is of considerable interest in solving a number of difficult real-world tasks, including object recognition and image processing. The specific problem that we address in this research is to identify the specific landmarks (anatomical) in human surface data. This is a complex task, currently performed visually by an expert human operator. In order to replace these human operators and increase reliability of the data acquisition, we need to develop a computer algorithm which will utilize the interrelations between the 3-dimensional data to identify the landmarks of interest. The current presentation describes a method for designing, implementing, training, and testing a custom architecture neural network which will perform the landmark identification task. We discuss the performance of the net in relationship to human performance on the same task and how this net has been integrated with other AI and traditional programming methods to produce a powerful analysis tool for computer anthropometry.

Neural network techniques and those used in conventional artificial intelligence systems show promise for solving complex real world problems. The strengths and weaknesses of the separate techniques are well known, for example the former is capable of reasoning on noisy or incomplete data but has poor facilities for eliciting input/output correlations. On the other hand, the latter has poor domain knowledge elicitation but is much better at explaining the correlation between its input data and its output response. In certain valuable aspects the two techniques are complementary and this paper reports on the characteristics exhibited by a hybrid system formed from a set of neural networks and a classical expert system. The paper describes how a hybrid system based upon a number of artificial neural network subsystems (ANNS), each implementing a knowledge source, is attached to a rule based system. By fusing these neural networks into the knowledge based system in a transparent way a new architecture is formed which performs markedly better than a conventional rule based system while retaining the explanation facilities of the rule based approach. The paper discusses in detail a prototype system which operates in the domain of optical character recognition. The paper highlights the advantages and disadvantages of such a technique in terms of concepts that are applicable to many other real world problem domains.

Signal processing techniques are currently being developed by engineers and analysts to exploit subtle radar phenomena. Research efforts at the Georgia Tech Research Institute, sponsored by the U.S. Air Force, have been directed toward the collection and analysis of data on some of the more subtle characteristics of signals from complex domestic radar systems. Analysis efforts have included both classical and emerging techniques. Applying classical statistical methods of analysis to Georgia Tech's data base on such characteristics, analysts have obtained encouraging results in many areas of radar signal processing. There remain, however, areas involving slowly changing characteristics where an adaptive capability could be critical, particularly in the dense electromagnetic environments that can be anticipated. Recent rapid advances in neural network techniques, when accompanied by advances in parallel processing hardware that will be needed to implement the networks for future field use, have resulted in the potential to provide this adaptability to changing parametric conditions. To assess that potential, the U.S. Air Force initiated two separate programs at the Georgia Tech Research Institute. Results from one program include an evaluation of a number of neural network paradigms for their appropriateness for processing radar waveforms to extract subtle information contained there. Test results using selected neural networks with various forms of collected data are presented. Results from another ongoing effort include a discussion of an effort to implement in hardware one of the neural networks tested.

Hopfield proposed two types of neural networks; Discrete Hopfield Network (DHN) and Continuous Hopfield Network (CHN). Those have been used for solving the well-known traveling salesman problem in a sense of optimization. DHN, a stochastic model is simple to implement and fast in computing. However, DHN uses binary value for states of neurons and results in an approximate solution. On the other hand, CHN gives a near-optimal solution, but it takes too much time to simulate a differential equation which represents a main characteristic of CHN. A matching problem using a graph matching technique can be cast into an optimization problem. In this paper, a new method for two-dimensional object recognition by using a Hopfield neural network is presented. A Hybrid Hopfield Network (HHN), which combines the merit of both the Continuous Hopfield Network and the Discrete Hopfield Network, is described and some of the advantages such as reliability and speed are shown in this paper. The main idea behind the new network is that stable states of neurons are analyzed and predicted based upon the theory of CHN after the convergence in DHN.

Neural networks provide an alternative way to solve complex optimization problems. Instead of performing a program of instructions sequentially as in a traditional computer, neural network model explores many competing hypotheses simultaneously using its massively parallel net. The paper shows how to use the neural network approach to perform vertical micro-code compaction for a micro-programmed control unit. The compaction procedure includes two basic steps. The first step determines the compatibility classes and the second step selects a minimal subset to cover the control signals. Since the selection process is an NP- complete problem, to find an optimal solution is impractical. In this study, we employ a customized neural network to obtain the minimal subset. We first formalize this problem, and then define an `energy function' and map it to a two-layer fully connected neural network. The modified network has two types of neurons and can always obtain a valid solution.

After a short discussion on the problems related to the higher order correlations treatment in Hopfield neural net, we propose a modified architecture able to rearrange dynamically its topology in function of the input representation. The relations of this problem with computability problems are briefly considered, particularly in view of avoiding exponential time in computation. Some experimental results are shown for the recognition of particle traces in high energy accelerators and in speaker independent speed recognition.

The problem of spectral parameter estimation using Hopfield neural networks (HNN) is studied in this paper. Linear prediction method for least square (LS) estimates and Pisarenko's harmonical retrieval method for constrained least-square estimates by HNN are introduced and an improved HNN algorithm for spectral peak searching is derived. The key problem we attempt to solve is to map these actual optimization problems to proper energy functions.

In the computer interpretation of seismic data, the critical first step is to identify the general class of an unknown event. For example, the classification might be: teleseismic, regional, local, vehicular, or noise. We have studied self organizing neural networks (SONNs) for classifying such events. Both Kohonen and Adaptive Resonance Theory (ART) SONNs were developed and tested with a moderately large set of real seismic events. Given the detection of a seismic event and the corresponding signal, we compute the time-frequency distribution, its binary representation, and finally a shift-invariant representation, which is the magnitude of the two-dimensional Fourier transform (2-D FFT) of the binary time-frequency distribution. This preprocessed input is fed into the SONNs. The overall results based on 111 events (43 training and 68 test events) show that SONNs are able to group events that `look' similar. The Kohonen algorithm has the advantage of being able to develop a higher resolution set of clusters. We also find that the ART algorithm has the advantage of being able to develop a higher resolution set of clusters. We also find that the ART algorithm has an advantage in that the number of cluster groups do not need to be predefined. When a net type of event is detected, the ART network is able to handle the event rather gracefully. A modified Kohonen algorithm is developed to exploit each algorithm. The results from the SONNs together with an expert seismologist's classification are then used to derive fuzzy event classifications. The integration of the SONN into the Network Seismic Event Analyzer (NetSEA) expert system is described and a strategy to integrate the fuzzy classifications into the inference engine is also proposed.

The focus of this paper is on the classifier module for a real-time automatic target recognition system. It is a dynamic neural network that utilizes adaptive, on-line learning. The on-line learning component is required because of the changeability of the target detection scenario resulting in unpredictable feature space representation of targets and clutter. The system was successfully tested against IR imagery of target models.

One of the major problems in the development of computer-assisted systems for geologic mapping is how to individualize the system to meet user needs. Ideally, the system should be responsive to specifications of desired types of output structures. Also, the system should be able to incorporate the user's knowledge of regional characteristics into the feature extraction/selection and classification components. Automatic techniques for classification of remote sensing data typically require relatively large, labeled training sets which are well- organized with respect to the desired mapping between input and output patterns. The present paper focuses on the feature extraction/selection component of the system. Kohonen self- organizing feature maps in conjunction with image processing procedures for linear feature extraction are used for explorative data analysis, feature selection, and construction of exemplar patterns. The results of training Kohonen feature maps with different pattern sets and different feature combinations provide insight into the nature of pattern relationships which enables the user to develop sets of positive and negative training patterns for the classification component.

Fault detection and diagnosis are important issues in engineering. With proper fault detection and diagnosis schemes, factors such as safety, efficiency, and cost of system operations can be significantly improved. This paper presents the use of two popular artificial neural network paradigms, namely feedforward networks and Kohonen nets (trained by Learning Vector Quantization algorithm), to perform fault detection. For illustration purposes, the fault detection of single-phase squirrel-cage induction motors is discussed. Comparisons of the preliminary results obtained from the feedforward net and Kohonen net to perform single- phase induction motor fault detection, in terms of factors such as classification accuracy and training time, are presented and discussed.

In this work we explore the possibility of using neural network classifiers for the analysis of high-energy physics experimental data. The study deals with the search for the Minimal Standard Model Higgs boson using the data collected in 1990 by the DELPHI detector at LEP. A set of multilayer perceptrons was trained, using Monte Carlo simulated data, to discriminate between Higgs events and the most important background processes. As the signal-to- background ratio is very low it is very important to define suitable selection criteria to reject background events and to select the Higgs events with the best possible efficiency. The use of neural network classifiers allows us to obtain a very good detection efficiency for Higgs events and a complete rejection of all background events.

Neural networks are able to emulate economic subjects' behavior, without the utilization of a priori economic laws. This work states the possibility of explaining economic regularities about demand curve following connectionist models. Targets for training are generated by the model itself while it learns and acts with a cross-target method, i.e., the artificial neural network produces guesses both of the actions of the economic subject and of their effects. Actual effects are estimated by economic environment rules and results are used to train the effect guess mechanism; the evaluation of actions necessary to match guessed effects are, on the contrary, employed to train the decision mechanism (which guesses actions). Two kinds of learning -- short term and long term learning -- take place. The former only with local capability to react to the changes of the environment and the latter one with full capability to do it. For an observer the behavior produced by this kind of artificial agent (AA) matches the actions of a real agent as stated in economic handbooks, with goals and plans. Obviously, AA has no such symbolic entities, which are inventions of the observer. The assumption is that the observations of economists about real world agents' behavior suffer from the same bias. Economic regularities arising from present versions of the model encourage further developments, in particular toward the interaction of micro-models to build macroeconomic ones.

The determination of number of neurons (H) in hidden layers is very important as it affects the training time and generalization property of neural networks. A higher value of H may force the network to memorize (as opposed to generalize) the patterns which it has seen during training whereas a lower value of H would waste a great deal of training time in finding its optimal representation. It is thus important to devise some methods by which a proper selection of neurons in hidden layers can be made. In this paper, a procedure has been given which determines the number of separable regions (M) in binary error correcting codes (BECC). Thus it is possible to establish link between input training patterns (T), M, and H for such codes without running simulations. Theorems have been developed which provide justification of the use of implied minterm structure (IMS) to BECC. It is shown that BECC are nonlinearly separable (LS) and canonical. Investigations have also been conducted on systematic and nonsystematic codes to prove that systematic codes can be classified with a lesser value of H than the nonsystematic codes as systematic codes require less number of separable regions for their realization.

The importance of visualization of scientific data has increased over recent years. Much of this interest has arisen because of the growth in computing power capable of generating complex and interactive computer graphics. The use of scientific visualization has a wide range of applications and has proven its utility in many new and existing application areas. However, within the field of artificial networks, visualization techniques have been limited to comparatively simple models, such as Hinton boxes and phase diagrams. The techniques developed in other research areas have generally been ignored. This is especially surprising considering the strong geometric and physical analogies present within the field. This paper examines the potential for visualization techniques in the study and implementation of artificial neural networks. The paper is divided into three sections. The first section briefly reviews scientific visualization and acts as an introduction to the remainder of the paper. The second section illustrates the potential for the visualization of the static and temporal properties of artificial neural networks. The static characteristics can include the visualization of trained networks' weights while the temporal characteristics can include the visualization of the networks' adaptation process or their operation. The discussion is illustrated by examples drawn from a study of the Kohonen Feature Map model. The third section examines the utility of visualization as a practical design aid in the construction and development of real world embedded neural network applications. The potential problems and benefits experienced when applying visualization techniques are highlighted and examined.

Pattern recognition applications require algorithms be optimized to provide accurate and reproducible target identification. Approaches usually incorporate a combination preprocessing, feature extraction, and classification algorithms whose parameters have been adjusted for the best performance against a particular set of images. With the variety of neural network and statistical techniques available at each of these processing steps, choosing the correct algorithms for a particular application may be difficult. A Pattern Recognition Workstation (PRW) has been developed to assist in the selection of these algorithms. The workstation provides a variety of image degradation techniques to assist the user in assessing the performance of algorithms as a function of obscuration, noise levels, scale and rotation. Initial results are reported from preprocessors including the Contrast-Orientation-Ratio- Threshold-Maximum (CORT-X), Sobel and Laplacian, feature extractors including the Gabor Transform, Invariant Moments, and Fourier-Log-Polar Transform, and classifiers including Backpropagation and Bayes decision theory. The resulting class decision statistics are presented to assess robustness with respect to obscuration and noise levels.

This paper describes the NeuralGraphics software environment used to run interactive neural network training experiments. The NeuralGraphics environment is a collection of software tools, graphical displays, and demonstrations that allow users to easily adapt many of the current neural network paradigms to their particular classification problem. The paper discusses the paradigms implemented in the NeuralGraphics environment as well as the data files required to train and test the learning capability of selected neural networks.

Magnetic resonance imaging (MRI) of the liver has demonstrated to be quite sensitive in showing Hepatic Hemangioma as high intensity lesions in T2 weighted imaging sequence. Hepatic Hemangioma is a non-malignant tumor and has relative high occurrence rate among the general population. It is of importance to differentiate this benign abnormality from other high intensity malignant lesions, such as hepatoma, adenocarcinoma, or metastasis. The objective of our study was to investigate the feasibility of applying neural network to assist in the differentiation of the liver MRI lesions. Thirty-seven liver MRI studies were collected, this including twenty-three cases of hepatic hemangioma and fourteen cases of malignant tumors. all cases were clinically proven with the diagnosed pathological condition and verified by biopsy. Four quantitative features, adopted from published literatures and used clinically on a routine basis, were measured from MRI images. In this study, a multilayer and two layer backpropagation networks were used for performance comparison. By attempting various training methods, the accuracy of the two layer network had been improved from 74% to 83% by selecting the proper boundary set based on the euclidean distance for each data set in both classes when training the network.

The aim of the research introduced in this paper is to develop a unified neural network platform to model the behavior of cancerous cells. Neural networks are used to both extract features from the cell images and control the process for recognizing whether a sequence of cell images match the locomotive and social behavior of a cancerous cell and its probability to metastasize. The problem first tackled is the extraction of cell features from the images as, for example, the center of the cell. This paper gives an overview of the application and presents the results drawn from two neural network architectures, an `all connected' and a`locally connected' network, used for the extraction of cell centroid areas from images. Both networks are implemented on a distributed array of processors (DAP) and trained using the backpropagation learning algorithm.

The use of artificial neural networks and Bayesian decision theory is proposed to provide a diagnostic tool which is capable not simply of making a correct decision but of allowing expert knowledge to be incorporated leading to the least risky decision. A neural network is used to estimate a posteriori (class) probabilities conditioned on the input data and the expert knowledge is introduced in the form of subjectively assigned weightings on erroneous decisions. An example of a decision aid for the early diagnosis of heart attacks is presented.

In this paper, a neural network approach for the recognition of printed Chinese characters is developed. A multi-layer perceptron is trained as the classifier by using the back-propagation algorithm, and the Walsh functions are employed for feature extraction. Thirty similar Chinese characters (classes) are designed in the experimental domain. The network is initially trained with noisefree training samples, and is retrained gradually with misclassified noisy testing patterns to improve the robustness of the classifier. Through classifying a large set of 9000 unknown testing patterns of various noise degrees, a great augmentation in system robustness and an encouraging recognition performance are presented.

Regularization is a paradigm for performing image segmentation and edge detection, that can be implemented in a neural network type architecture. Various topics and problems pertaining to the use of regularization for image processing applications are discussed. Topics include data fusion, sensor blur, and the operation on partitioned images. A mathematical analysis of the different topics is presented, including a modification of the original regularization energy functional to perform data fusion.

A methodology employing artificial neural networks and fuzzy arithmetic in the diagnosis and control of complex systems such as pressurized water reactors is presented. Fuzzy numbers represent the linguistic values of plant-specific variables, e.g., performance or availability. The notion of a virtual instrument, i.e., a software-based measuring device calibrated to the idiosyncrasies of a specific system is used. Neural networks perform a mapping of physically measurable parameters to fuzzy numbers called Virtual Measurement Values (VMV). The methodology is tested with start-up data from an experimental nuclear reactor. The results demonstrate the very good capacity of such virtual instruments for failure-tolerance and suggest the possibility of developing alternative algorithms for diagnostics and control.