A neural network approach to obtaining upper and lower bounds on the Bayes error rate for pattern recognition problems is presented. The approach is developed using the key concept of resubstitution and leave-one-out testing from conventional nonparametric error estimation techniques. The neural network approach is evaluated by applying it to several 8D, two-class `toy' problems, where the Bayes error rate is known. The neural network error estimate for a high-dimensional problem with an unknown Bayes error rate is also compared to error estimates obtained using conventional nonparametric estimation techniques. Using the neural network procedure, the upper bound of the Bayes error rate is reliably found for problems with complex decision boundary surfaces. Alternative testing approaches are suggested for reducing the difference between the bounds and the true Bayes rate.

This paper proposes a feedforward neural network structure to realize the mathematical morphology operations (MMO), namely: HMT (Hit and Miss Transformation), dilation, erosion, opening, closing, and the union and intersection (for multistructuring elements and multioperators) of them. Different kinds of operations can be implemented by assigning the weights, the threshold values and the architecture of the network according to the operation itself to be implemented. A general expression relating the weight value, threshold value to the configuration of the structuring element for different operations is derived, the assigning of the values becomes merely straight-forward training procedure of the proposed network. Also, it is proved that with a single hidden layer, all the MMO can be implemented by the ANN. The most interesting aspect of the method proposed is the reduction of on-line operation steps, which for conventional MMO algorithm consist of a series of operations processed consecutively. As a extension of the method, Boolean function implementations of the operations are also proposed, in which, the concept of collection of basic `And' structuring elements is presented. We prove that all MMO sequence can be implemented by a 2-layer logic gate array (or 3 layers in the sense of node levels).

When the dimension N of the input vector is much larger than the number M of different training patterns to be learned, a one-layered, hard-limited perceptron with N input nodes and P neurons (P > equals Log₂M) is generally sufficient to accomplish the learning- recognition task. The recognition should be very robust and very fast if an optimum noniterative learning scheme is applied to the perceptron learning process. This paper concentrates at the discussion of two special characteristics of this novel pattern recognition system: the automatic feature extraction and the automatic feature competition. An unedited video movie recorded on a series of learning-recognition experiments may demonstrate these characteristics of the novel system in real time.

Binary self-coded filters (BSCFs) are easily implemented optically. BSCFs are based on the optimization functions of the Hopfield model by nonsynchronous iterative neuron algorithms. Less filters are needed to carry out the same recognition task, compared with other methods. The error-tolerance ability is also very strong. All target objects can be correctively recognized when the characteristic codes are properly chosen. Starting from different initial states, we can obtain solution close to the overall optimum of the net.

Tactical communications are inherently short and exhibit a great deal of channel variability. A novel speaker recognition technique was developed which uses on-line training to circumvent the need for excessive speaker or channel modeling. The technique incorporates both feature set fusion and classifier fusion. Separate classifiers are trained for each feature set: liftered LPC cepstra, RASTA liftered cepstra concomitant with delta cepstra. For each classifier, the results of the individual (feature) classifiers are adjudicated to rank the speakers. A final step adjudicates the results of different classifiers to determine the correct speaker.

An attempt is made to find a new way for better diagnosis of hepatisis through application of artificial neural network theory. Learning from a given sample set, the neural network is used to establish a nonlinear mapping between various factors, such as symptoms, signs, and laboratorial experiments, and diagnosis of hepatisis. It is proved that the used network and values of weight after learning are available to the identification of equivalent class of a new pattern of hepatisis. In this paper, the knowledge learning and learning algorithms used in diagnosis are mainly discussed, an optimal generalization algorithm based on the error decrease algorithm and used to train multilayer feedforward is presented; meanwhile, the application results and their effectiveness are introduced.

In fault-tolerant real-time systems, software errors can be detected, recovered from, and reported using a recovery controller. A simulator has been developed to test the performance of an ANN (Artificial Neural Network) based recovery controller. The simulator for a highly reliable, fault-tolerant multiprocessor telecommunications exchange provides a real-world application to test the effectiveness of the ANN recovery controller. The ANN provides the software recovery controller with the adaptability to determine recovery actions for faults that were previously unseen or not anticipated. This reduces the number of times that human intervention is required to recover the system, and thus reduces the total down-time for the system. The ANN is trained by inserting known faults into the simulated real-time system. The system collects data on the characteristics observed when the fault is detected and uses the back propagation learning algorithm to classify the observed characteristics and recent history of recovery actions into an appropriate recovery action such as restarting or terminating a process, initializing the operating system for a specific processor, or reloading a data base for a processor of group of processors. Once the neural network has been trained, it is used to determine the appropriate responses for faults that occur during the operation of the system.

We proposed a neural network that can interpret multi-meaning Chinese words correctly by using context information. The self-organized network, designed for translating Chinese to English, builds a context according to key words of the processed text and utilizes it to interpret multi-meaning words correctly. The network is generated automatically basing on a Chinese-English dictionary and a knowledge-base of weights, and can adapt to the change of contexts. Simulation experiments have proved that the network worked as expected.

The performance of microwave digital radio systems is severely degraded by multipath fading and also by the nonlinear power amplifier at the transmitter, the combined effects of these result in a time varying nonlinear channel. In this paper, the multilayer perceptron (MLP) based equalizer is investigated for use on such channels. It is found to outperform the conventional approach, that is, the use of a transversal equalizer at the receiver operating with a signal predistorter placed before the non-linear power amplifier at the transmitter. An L- level non-linear function is used as a node activation for the MLP. Two update schemes are investigated, one being a complex version of the back propagation algorithm, the other a complex version of the delta-bar-delta algorithm.

This paper presents our continuing work in the area of nonlinear system analysis and simulation applying ANN methods. The system simulation for the payload of communications satellite is the process of predicting the overall system performance from the performance of each component in the system. The important part of this system simulation is the analysis and simulation of nonlinearity in the system, such as power transfer characteristics, and intermodulation. The model and simulation of nonlinearity for a communications satellite channel contained nonlinear components is described in this paper. Polynomial approximation by least squares and artificial neural networks simulation have been used to approximate the tested data curve of a single tone power transfer characteristics for each nonlinear element in the system. These methods have been applied to the Communications Satellite System Simulation Software Package that is developed by us in the Xi'an Institute of Space Radio Technology. Both of these methods have also been compared with the test data of real system of DFH-3 communications satellite. The results show that the approximation method using ANN concept presented in this article is more accurate and of theoretical importance and practical value in the analysis of nonlinear system problem.

We describe a neural network learning algorithm that implements differential learning in a generalized backpropagation framework. The algorithm regulates model complexity during the learning procedure, generating the best low-complexity approximation for the Bayes-optimal classifier allowed by the training sample. It learns to recognize handwritten digits of the AT&T DB1 database. Learning is done with little human intervention. The algorithm generates a simple neural network classifier from the benchmark partitioning of the database; the classifier has 650 total parameters and exhibits a test sample error rate of 1.3%.

A biologically—inspired modification to MAXNET is proposed. Unlike the original net where the weights are constant, the weights in the new net are dynamically changed. Consequently, the modified net achieves a drastic improvement in convergence rate. A simple hardware implementation for the modified net is presented.

This paper presents a systematic evaluation of the individual effects of training parameters: learning rate, momentum rate, number of hidden layer nodes, and processing element's transfer function, on the performance of back propagation networks in predicting quality characteristics of end milled parts. Multi-sensor signatures (acoustic emission, spindle vibration, and cutting force components) acquired during circular end-milling of 4140 steel and the corresponding measured quality characteristics (surface roughness and bore tolerance) were used to train the networks. The network is part of a proposed Intelligent Machining Monitoring and Diagnostic System for Quality Assurance of Machined Parts. The network performances were evaluated using four different criteria: maximum error, RMS error, mean error and number of training cycles. One of the results obtained shows that hyperbolic tangent transfer function gave a better performance than the sigmoid and sine functions respectively. Optimum combinations of training parameters have been observed. The effects of various combinations of training parameters are presented.

One of the major problems in supervised learning of neural networks is the inevitable local minima inherent in the cost function f(W,D). This often makes classic gradient-descent-based learning algorithms that calculate the weight updates for each iteration according to (Delta) W(t) equals -(eta) (DOT)$DEL_wf(W,D) powerless. In this paper we describe a new strategy to solve this problem, which, adaptively, changes the learning rate and manipulates the gradient estimator simultaneously. The idea is to implicitly convert the local- minima-laden cost function f((DOT)) into a sequence of its smoothed versions {f_{(beta t})}^T_tequals1, which, subject to the parameter (beta) _t, bears less details at time t equals 1 and gradually more later on, the learning is actually performed on this sequence of functionals. The corresponding smoothed global minima obtained in this way, {W_t}^T_tequals1, thus progressively approximate W--the desired global minimum. Experimental results on a nonconvex function minimization problem and a typical neural network learning task are given, analyses and discussions of some important issues are provided.

This paper describes a decision fusion method based on fuzzy logic and genetic algorithms. For the fusion process the generalized mean aggregation connective is used. The optimal parameters of the generalized mean are found by a genetic algorithm both with elitist and nonelitist strategy. The results of both strategies are compared. The decision fusion method proposed is tested on a vibration monitoring problem. The decisions from multiple sensors to be fused are obtained by neural networks. First vibration spectra are compressed by recirculation networks. Next classification of compressed signatures is performed for each sensor separately by backpropagation networks. The output of backpropagation networks is the input to the fuzzy fusion center performing the generalized mean operation.

Lexical combination presents a number of intriguing problems for cognitive science. By studying the empirical phenomena of combination we can derive constraints on models of the representation of individual lexical items. One particular phenomenon that symbolic models have been unable to accommodate is `semantic interaction'. Medin & Shoben (1988) have shown that properties associated with nouns by subjects vary with the choice of adjective. For example, wooden spoons are not just made of a different material: the phrase is interpreted as denoting a `larger' object. However, the adjective wooden is not generally held to carry implications as to size. We report experimental results showing similar effects across a range of properties for a single adjective in combination with different nouns from a single semantic field. It is this more radical dependence of interpretative features on lexical partners that we term `semantic interaction'. The phenomenon described by Medin and Shoben cannot be accounted for by the Selective Modification model, the most complete model hitherto. We show that a case-based reasoning system could account for earlier data because of the particular examples chosen, but that such a model could not handle semantic interaction. A neural network system is presented that does handle semantic interaction.

The proposed fuzzy expert network is an event-driven, acyclic neural network designed for knowledge learning on a fuzzy expert system. Initially, the network is constructed according to a primitive (rough) expert rules including the input and output linguistic variables and values of the system. For each inference rule, it corresponds to an inference network, which contains five types of nodes: Input, Membership-Function, AND, OR, and Defuzzification Nodes. We propose a two-phase learning procedure for the inference network. The first phase is the competitive backpropagation (CBP) training phase, and the second phase is the rule- pruning phase. The CBP learning algorithm in the training phase enables the network to learn the fuzzy rules as precisely as backpropagation-type learning algorithms and yet as quickly as competitive-type learning algorithms. After the CBP training, the rule-pruning process is performed to delete redundant weight connections for simple network structures and yet compatible retrieving performance.

This paper is a tutorial on the use of fuzzy logic system, providing a broad-scope overview, especially on fuzzy logic design, and also the results of research of fuzzy application in robotics. At first, the fuzzy sets and fuzzy rules are explained as well as fuzzy logic theory regarding our research. A neural network (connectionist) model for fuzzy logic control and decision system in their neural network structure and learning ability. The description of the system by using mathematical model, linguistic rules, and parameter distribution are discussed.

The Naval Air Warfare Center, China Lake has developed a real time, hardware and software system designed to implement and evaluate biologically inspired retinal and cortical models. The hardware is based on the Adaptive Solutions Inc. massively parallel CNAPS system COHO boards. Each COHO board is a standard size 6U VME card featuring 256 fixed point, RISC processors running at 20 MHz in a SIMD configuration. Each COHO board has a Companion board built to support a real time VSB interface to an imaging seeker, a NTSC camera and to other COHO boards. The system is designed to have multiple SIMD machines each performing different Corticomorphic functions. The system level software has been developed which allows a high level description of Corticomorphic structures to be translated into the native microcode of the CNAPS chips. Corticomorphic structures are those neural structures with a form similar to that of the retina, the lateral geniculate nucleus or the visual cortex. This real time hardware system is designed to be shrunk into a volume compatible with air launched tactical missiles. Initial versions of the software and hardware have been completed and are in the early stages of integration with a missile seeker.

Automated target recognition (ATR) software has been designed to perform image segmentation and scene analysis. Specifically, this software was developed as a package for the Army's Minefield and Reconnaissance and Detector (MIRADOR) program. MIRADOR is an on/off road, remote control, multisensor system designed to detect buried and surface- emplaced metallic and nonmetallic antitank mines. The basic requirements for this ATR software were the following: (1) an ability to separate target objects from the background in low signal-noise conditions; (2) an ability to handle a relatively high dynamic range in imaging light levels; (3) the ability to compensate for or remove light source effects such as shadows; and (4) the ability to identify target objects as mines. The image segmentation and target evaluation was performed using an integrated and parallel processing approach. Three basic techniques (texture analysis, edge enhancement, and contrast enhancement) were used collectively to extract all potential mine target shapes from the basic image. Target evaluation was then performed using a combination of size, geometrical, and fractal characteristics, which resulted in a calculated probability for each target shape. Overall results with this algorithm were quite good, though there is a tradeoff between detection confidence and the number of false alarms. This technology also has applications in the areas of hazardous waste site remediation, archaeology, and law enforcement.

This paper describes the design of an undersea mine detection system and compares the performance of various neural network models for classification of features extracted from side-scan sonar images. Techniques for region of interest and statistical feature extraction are described. Subsequent feature analysis verifies the need for neural network processing. Several different neural and conventional pattern classifiers are compared including: k-Nearest Neighbors, Backprop, Quickprop, and LVQ. Results using the Naval Image Database from Coastal Systems Station (Panama City, FL) indicate neural networks have consistently superior performance over conventional classifiers. Concepts for further performance improvements are also discussed including: alternative image preprocessing and classifier fusion.

In this paper, it is investigated how closely neural networks can approach the optimum classification of radar textures. To this end, a factorization technique is presented which aids convergence to the best possible solution obtainable from the training data. This factorization scheme is designed to be fully general. The specific performances of the factorized networks are studied, in this radar clutter classification problem, when applied to uncorrelated K distributed images. These results are then compared with the maximum likelihood performance and the performances of various intuitive and approximate classification schemes. Furthermore, preliminary network results are presented for the classification of correlated processes and these results are also compared to results obtained using classical techniques.

A serious degradation in detection probability of conventional Constant False Alarm Rate (CFAR) processors used in the automatic detection of radar targets results from a reduction in the number of available reference cells. Several factors such as any constraints on the radar system used (in terms of resolution and sampling time), presence of interfering targets and nonstationary clutter may contribute to the reduction in the number of reference cells. This paper presents a novel neural network-based CFAR detection scheme (referred to as NN- CFAR scheme) that offers robust performance in the face of loss of reference cells. This scheme employs a multilayer feedforward neural network trained by error backpropagation approach using the optimal detector as the teacher. The excellent pattern classification capabilities of trained neural networks are exploited in this application to effectively counter performance degradations due to reduced reference window sizes. In particular it is demonstrated that a neural network implementation of the CFAR detection scheme provides an efficient approach for accommodating more input parameters without increasing design complexity for countering the information loss due to reduced reference window size. Precise quantitative performance evaluation of the NN-CFAR scheme are conducted in a variety of situations that include both homogeneous and nonhomogeneous clutter backgrounds and the target detection performance is compared with that of the traditional CA-CFAR scheme to highlight the benefits.

We present here an algorithm which performs radar cross-section estimation by using techniques based on simulated annealing. Standard simulated annealing approaches to image restoration attempt to categorize each image element as belonging to one of a small number of predefined image states or values. This is restrictive for tasks such as radar cross-section estimation and we present here an algorithm which is capable of producing a real-valued output. This is achieved by introducing an edge detection stage into the simulated annealing process. The action of the annealing algorithm may be viewed as a filter which adapts to local image structure. We present results which demonstrate this behavior and in so doing allow us to estimate the residual noise levels we might expect.

This paper presents an edge detection algorithm using Hopfield neural network. This algorithm brings up a new concept which is different from those conventional differentiation operators, such as Sobel and Laplacian. In this algorithm, an image is considered a dynamic system which is completely depicted by an energy function. In other words, an image is described by a set of interconnected neurons. Every pixel in the image is represented by a neuron which is connected to all other neurons but not to itself. The weight of connection between two neurons is described as being a function of contrast of gray-level values and the distance between pixels. The initial state of each neuron represents the normalized gray-level value of the corresponding pixel in the original image. As a result of Hopfield network analysis, output of neurons is modified until the convergence. Even though the outputs are analog, they are close to zero in all regions except edges where the corresponding neurons have near 1.0 output values. A robust threshold on the output level of the converged network can be easily set up at 0.5 level to extract edges. The experimental results are presented to show the effectiveness and capability of this algorithm.

We describe hardware that has been built to compress video in real time using full-search vector quantization (VQ). This architecture implements a differential-vector-quantization (DVQ) algorithm which features entropy-biased codebooks designed using an artificial neural network. A special-purpose digital associative memory, the VAMPIRE chip, performs the VQ processing. We describe the DVQ algorithm, its adaptations for sampled NTSC composite- color video, and details of its hardware implementation. We conclude by presenting results drawn from real-time operation of the DVQ hardware.

Over the past several years we have been interested in the supervised classification of ultrasonic images of the liver based on quantitative texture features. Our most recent efforts are concerned with the inclusion of features computed from Markov random fields. After adding four such features to our existing model containing 17 features, we employed stepwise discriminant analysis to identify the features that could best discriminate among 184 previously classified normal and abnormal ultrasonic images. Three of the four features derived from Markov random field models were identified by stepwise discriminant analysis as being good discrimination along with 6 existing features. From these results we constructed a backpropagation neural network with an input layer consisting of 9 nodes. We found that this new model yielded slightly better results when compared to earlier models. Our most recent results yielded a sensitivity of 81%, a specificity of 77% and an overall accuracy of 79%.

A Hopfield type neural network was applied to optimize binary correlation synthetic discriminant functions (SDFs). Rotation invariance is achieved while the target object rotates in a certain angle range and a ratio for judgement which is defined as the ratio of the peak value of the average absolute value of a specific point set is given. The optimized binary SDFs (BSDFs) provide the control of the sidelobe levels and the expected shape of the output correlation functions as well as its peak intensity. The simulation results show that when the true target object is presented to the optimized filter, not only the correlation peak is higher than that of the false target objects, but also the order of the magnitude of the ratio for judgement is at least 1 greater than that of the false target objects. The filters perform quite well.

The goal of this research is to develop a vision system that is capable of recognizing objects based on past experience. This paper introduces the highest level of this system, which consists of a neural network that is capable of learning to recognize 3D objects. Knowledge about objects is acquired by learning their various views, guises, or aspects. Learning occurs on two levels. First, supervised competitive learning is employed to teach the network to differentiate between different objects. The competition causes the unique differences between objects to be emphasized in this stage. Second, unsupervised cooperative learning is employed to self-organize the various aspects of a given object. This stage works in a manner similar to the ART family of self-organizing networks. The cooperative learning causes similarities between different aspects of the same object to be emphasized. The object recognition system is intended for use in a manufacturing environment, including tasks such as component identification, classification of visible quality defects, and visual product grading and sorting.

A machine that can read unconstrained handwritten words remains an unsolved problem. For example, automatic entry of handwritten documents into a computer is yet to be accomplished. Most systems attempt to segment letters of a word and read words one character at a time. Segmenting a handwritten word is very difficult and often, the confidence of the results is low. Another method which avoids segmentation altogether is to treat each word as a whole. This research investigates the use of Fourier Transform coefficients, computed from the whole word, for the recognition of handwritten words. To test this concept, the particular pattern recognition problem studied consisted of classifying four handwritten words, `Buffalo', `Vegas', `Washington', and `City' from the SUNY post office database. Several feature subsets of the Fourier coefficients are examined. The best recognition performance of 76.2% was achieved when the Karhunen-Loeve transform was computed on the Fourier coefficients and those features were fed into a multilayer perceptron.

As many early vision tasks the computation of depth-from-motion is an ill-posed problem but very useful in computer vision and rotor craft navigation. The collective computation capabilities of highly parallel neural networks provides new powerful techniques for optimization problems in high dimensional spaces. This paper reports an investigation of computation of depth from motion. As this problem is formulated as minimizing a cost or energy function, a massively parallel neural network approach is used for solving this problem by regularization techniques. This approach presents some similarities with biological visual systems. The neural solution developed here is a direct method avoiding the explicit optical flow estimation. We perform an evaluation on both synthetic and real world image sequence.

The most difficult stage of automated target recognition is segmentation. Current segmentation problems include faces and tactical targets; previous efforts to segment these objects have used intensity and motion cues. This paper develops a color preprocessing scheme to be used with the other segmentation techniques. A neural network is trained to identify the color of a desired object, eliminating all but that color from the scene. Gabor correlations and 2D wavelet transformations will be performed on stationary images; and 3D wavelet transforms on multispectral data will incorporate color and motion detection into the machine visual system. The paper will demonstrate that color and motion cues can enhance a computer segmentation system. Results from segmenting faces both from the AFIT data base and from video taped television are presented; results from tactical targets such as tanks and airplanes are also given. Color preprocessing is shown to greatly improve the segmentation in most cases.

The Accurate Automation Corporation (AAC) neural network processor (NNP) module is a fully programmable multiple instruction multiple data (MIMD) parallel processor optimized for the implementation of neural networks. The AAC NNP design fully exploits the intrinsic sparseness of neural network topologies. Moreover, by using a MIMD parallel processing architecture one can update multiple neurons in parallel with efficiency approaching 100 percent as the size of the network increases. Each AAC NNP module has 8 K neurons and 32 K interconnections and is capable of 140,000,000 connections per second with an eight processor array capable of over one billion connections per second.

The implementation of artificial neural networks (ANN) as CMOS analog integrated circuits shows several attractive features. Stochastic models, and especially the Boltzmann Machine shows a number of many attractive features. Numerous papers show that small size analog networks operate correctly. However, recent studies on artificial models point out that classification is their most successful application field: so real pattern recognition tasks will require large networks. On the other hand, all of the presented implementations of ANN have been supposed to be working in ideal conditions but real applications will subject to perturbations. For a digital implementation of ANN perturbation effects could be neglected in a fifth-order approximation. But for the analog and mixed digital/analog implementation cases, the behavior analysis of the neural network with perturbation conditions is inevitable. Unfortunately, very few papers analyze the behavior of analog neural networks with perturbation or their limitations. In this paper we present the analysis of a CMOS analog implementation of synchronous Boltzmann Machine model's behavior with physical temperature perturbations. The relation between the T parameter of the Boltzmann Machine's model and the physical temperature of circuit has been presented. Simulation results have been given, temperature effects compensation have been discussed, and experimental results have been exposed.

Analog VLSI Implementation of artificial neural networks for vision applications is studied in this paper. A locally connected, regular structure for 2D convolution is proposed for high speed image processing. First, a mathematical formulation is given to map the convolution to analog computing domain. Then a system level design is performed together with the design and the testing of the basic building blocks. Experiments using real laboratory images are conducted.

This paper presents an analog circuit design and implementation for an adaptive resonance theory neural network architecture called the augmented ART1 neural network (AART1-NN). Practical monolithic operational amplifiers (Op-Amps) LM741 and LM318 are selected to implement the circuit, and a simple compensation scheme is developed to adjust the Op-Amp electrical characteristics to meet the design requirement. A 7-node prototype circuit has been designed and verified using the Pspice circuit simulator run on a Sun workstation. Results simulated from the AART1-NN circuit using the LM741, LM318, and ideal Op-Amps are presented and compared.

Structural Load Alleviation Control Systems (SLACS) are one of the principal components of Active Control Technology, which allows aerospace structures to operate more efficiently by making them lighter, yet still capable of sustaining the high operational loads experienced in flight. A SLACS alleviates the bending and torsional moments caused by the forces and moments acting upon the aircraft wings. A SLACS must be designed to achieve a well distributed alleviation across the entire span of the wing. SLACS designed using deterministic methods, such as linear optimal, H^(infinity ), or quantitative methods have had only limited success, owing to controller complexities, and the unsteady effects of flight through turbulence. The main objective of this paper is to provide an alternative solution for a high dimensional aircraft control problem using artificial neural networks. These networks have nonlinear modelling capabilities, and can potentially be used to adapt on-line to account for time-varying aircraft characteristics. To address problems of persistent input excitation and slow convergence commonly faced when synthesizing large neural controllers, this neural solution requires that the control architecture be decomposed into a hybrid combination of smaller sub-networks, and linear quadratic regulators, employed together in parallel. The modelling technique is based on the traditional backpropagating multi-layered perceptron and the B-Spline associative memory networks. The B-Spline network generally has faster parameter convergence with minimal learning interference, and is therefore potentially more robust in on-line implementations. The results of digital simulations are used to demonstrate the effectiveness of such neural SLACS controlling the wing structure of a large transport aircraft in flight. The performance is assessed for both clear air and turbulent conditions.

A method is developed to design simulations of neural-network based transfer functions, applicable to both linear and nonlinear structures. The algorithm used to implement the trainable neural mechanism is backpropagation. Using the trained structures as building blocks, a neural architecture is constructed in order to drive systems from expected inputs to satisfactory transient and steady-state output performance, in effect, the scope of control compensation; this method results in the design of neural-net control compensators. The algorithms are coded in a PC-based prolog, traditionally used for rule-based logic and Artificial Intelligence, rather than for Neural or Fuzzy models. Given a sequence representing the time-sample of a desired control input trajectory that will drive the plant to a desired output response, such a control input will be modelled as the desired output layer of an antecedent network driven by an error vector consistent with the closed-loop system's commanded behavior. This Controller network is trained to provide such an output profile for all expected inputs, in accordance with arbitrary specifications of rise-time, permitted overshoot, settling time, etc. The control vectors are generated as a by-product of this training. Additionally, a correlation is investigated between classical control parameters and the characteristics of the weight matrices, threshold vectors, and representation traits of the converged neural nets.

A general architecture for neuro-genetic adaptive control is described and contrasted with purely neural approaches to adaptive control. The system is demonstrated on the attitude control problem for a rigid body (satellite) equipped with thrusters about each principal axis. By simulating the dynamic system and applying standard neural network techniques a locally predictive network is first trained to the prevailing dynamics. The inputs for the network are a small history of system states up to the present and a set of current control inputs, the outputs are the next system state. It is assumed that a hardware implementation of this network is used to evaluate hypothetical control inputs very rapidly. A genetic algorithm with a simple goal function then searches the space of hypothetical control inputs, whose fitness is evaluated by the neural network, so as to find a satisfactory set of control inputs before the end of the predicted time interval--the whole process is then repeated. The results indicate that such an architecture is able to master the attitude control problem for arbitrary slew angles, with arbitrary unknown dynamics, large unknown deterministic perturbing forces (which left to themselves induced chaotic motion), and noise in the sensor system.

A control design technique known as the Model Reference Neural Network (MRNN) method has recently been developed. In this method, neural network controllers are trained so that the controlled system response mimics that of a desired reference model. Since the controller can be trained using experimental test data consisting of command and response state data, it is equally applicable to linear and nonlinear systems. The MRNN procedure was experimentally evaluated by applying it to several systems which demonstrated nonlinear behavior typically found in servosystems, including significant stick-slip friction, backlash, and positionally dependent gravitational torques. The performance of the MRNN was then compared to both PID and linear model reference controllers. Experimental results indicate that the accuracy of the MRNN controller typically equals or exceeds the linear model reference controllers.

We had previously demonstrated the ability of the wave expansion neural networks (WENN) to develop artificial potential fields (APFs) which are useful for evaluating paths for point robots. In this paper, we described how WENNs can be used to develop APFs for larger objects.

The design of control algorithms for large space structures, possessing nonlinear dynamics which are often time-varying and likely ill-modeled, presents great challenges for all current methodologies. These limitations have led to the pursuit of a robust and fault tolerant structural controller. In the present paper, we propose the use of adaptive time-delay radial basis function (ATDRBF) networks as a learning controller in system identification and dynamic control of flexible structures. The ability of such neural networks to approximate arbitrary continuous functions offers an efficient means of vibration suppression and trajectory maneuvering for precision pointing capability. The ATDRBF network, which incorporates adaptive time-delays and interconnection weights, provides a feasible and flexible modeling technique to effectively capture all of the spatiotemporal interactions among the structure members. In the spirit of model reference adaptive control, we utilize the ATDRBF network as a building block to allow the neural network to function as a closed-loop controller. The controller regulate the dynamics of the nonlinear plant to follow a prespecified reference model asymptotically. This paper addresses the theoretical foundation of the architecture and demonstrates its applicability via specific examples.

In this paper we investigate the application of an artificial neural network to the computation of the instantaneous observer's heading in a static environment. This parameter has an important role in vision-based collision avoidance systems and relies upon accurately locating the focus of expansion (FOE) associated with the radial optical flow pattern arising as a consequence of the translational component of motion. The approach proposed in this paper is based on a feed-forward neural network able to compute the image coordinates of the FOE. It is assumed that the input signals are supplied to the network by a sensorial module computing the optical flow associated with a sequence of time-varying images of the viewed scene. A number of experiments have been performed both for theoretical and for realistic optical flow fields. In this latter real-world experiments the input sensorial module has been simulated through an Hopfield network. Experimental results show that the proposed neural architecture is able to recover the FOE position of testing flow fields with a mean error of 0.1 pixels for exact theoretical motion fields. Moreover it seems resistant to noise and its performances appear appreciable also in real world contexts.

We have designed a neural network semiautonomous robotic arm controller. This controller performs end-effector path planning, inverse kinematics, and joint control to move the end- effector to a commanded position. We have tested the adaptive neural joint controller and inverse kinematics in simulation. The joint controller has been tested on two real arms. These real arms are the Extendable Stiff Arm Manipulator (ESAM) and the Proto-Flight Manipulator Arm (PFMA). Both of these arms are very different, yet the same unmodified joint controller software can control them both. The controller has also shown tremendous adaptability to large payload variations. It has been shown to adapt to a 35 pound end-effector payload on the ESAM from a zeroed initial state. This ability to handle different arms and payloads is due to the fact that the controller makes no assumptions as to the arm's dynamics or payload. The same tests performed on a decentralized PD controller showed that the neural network controller is superior.

A multispectral, multiaperture, nonimaging sensor was simulated and constructed to show that the relative location of a robot arm and a specified target can be determined through Neural Network processing when the arm and target produce different spectral signatures. Data acquired from both computer simulation and actual hardware implementation was used to train an artificial Neural Network to yield the relative position in two dimensions of a robot arm and a target. The arm and target contained optical sources of different spectral characteristics which allows the sensor to discriminate between them. Simulation of the sensor gave an error distribution with a mean of zero and a standard deviation of 0.3 inches in each dimension across a work area of 6 by 10 inches. The actual sensor produced a standard deviation of approximately 0.8 inches using a limited number of training and test sets. No significant differences were found in the system performance where 9 or 18 apertures were used, indicating a minimum number of apertures required is equal to or less than nine.

In order to find a very efficient technique to compress, store, and transmit to earth information from a satellite we developed a scheme of chaotic neural net using a new technique of extraction of unstable orbits within a chaotic attractor without applying classical embedding dimensions. We illustrate this technique both from the theoretical and the experimental standpoint. From the theoretical standpoint we show that by this extraction technique it is possible to perform a series expansion of a chaotic dynamics directly through all its composing cycles. Finally, we show how to apply these new possibilities deriving from our new technique of chaos detection, characterization, and stabilization to design a chaotic neural net. Because it is possible to profit by all the skeleton of unstable periodic orbits (i.e., all the inner frequencies) characterizing a chaotic attractor to store information, this net can in principle display an exponential increasing of memory capacity with respect to classical attractor nets.

A new operational system to interpret satellite images is represented. The described method is adaptive. It is trained by examples. In the reported application a combination of textural and spectral measures is used as a feature vector. The adaptation or learning of the extracted feature vectors occurs by a self-organizing process. As a result a topological feature map is generated. The map is identified by known samples, examples of clouds. The map is used later on as a code book for cloud classification. The obtained verification results are good. The represented method is general in the sense that by reselecting features it can be applied to new problems.

We want to present a further development of our technique of backpropagation with stochastic preprocessing to recognize particle tracks in a silicon calorimeter on a satellite to detect cosmic ray composition. In the first release we applied our technique to distinguish between two classes of discrete patterns. In the present release we developed the stochastic preprocessing to deal with continuous patterns such as the energy deposited by a cosmic particle. From the theoretical standpoint we demonstrate that by such a preprocessing technique the neural net is able to represent the complexity of learning set in a polynomial and not exponential time. This work is a part of `Skynnet' international project supported by INFN (National Institute for Nuclear Physics) and partially devoted to the application of neural techniques for recognition of high energy particle tracks in spatial environment.

This paper describes a modified feedback BP NN: Sigma-Pi-linked ((Sigma) -(Pi) ) model. In order to increase input information, a `higher-order' terms of input patterns by using functional-linked of the input patterns are introduced. Finally, the approximation property of the NN has a powerful approximation capability than that of conventional BP neural network.

A neural network approach to stock market index prediction is presented. Actual data of the Wall Street Journal's Dow Jones Industrial Index has been used for a benchmark in our experiments where Radial Basis Function based neural networks have been designed to model these indices over the period from January 1988 to Dec 1992. A notable success has been achieved with the proposed model producing over 90% prediction accuracies observed based on monthly Dow Jones Industrial Index predictions. The model has also captured both moderate and heavy index fluctuations. The experiments conducted in this study demonstrated that the Radial Basis Function neural network represents an excellent candidate to predict stock market index.

In this contribution the special interest is focused on the prediction error distributions using feedforward multilayer neural network predictors. The procedure of prediction can be evaluated not only with respect to the mean absolute (or mean square) error but the frequency of higher errors are also taken into account. Two new modifications of the learning algorithms are suggested. The first one is based on error controlled input sample selection for efficient training. The second one is based on the minimization of the special criterial functions, which more reflect especially the great deviations. The functions are expressed in the form of minimax criterial function or in the form of weighted sum of higher order deviations between predicted and measured values. The classical backpropagation is used in the first case and the stochastical method of the statistical gradient is used in the second case. The efficiency of our approaches is demonstrated on the electric load forecasting problem in the west bohemian region in the Czech Republic. A reduction in frequency of higher errors in the everyday morning peaks forecasting were achieved.

In the context of M. Minsky's and S. Papert's theorems on the impossibility of evaluating simple linear predicates by parallel architectures we want to show how these limitations can be avoided by introducing a generalized input-dependent preprocessing technique that does not suppose any a-priori knowledge of input like in classical input filtering procedures. This technique can be formalized in a very general way and can be also deduced by meta- mathematical arguments. A further development of the same technique can be applied at level of learning procedure to introduce in such a way the complete notion of `dynamic perception'. From the experimental standpoint, we show two applications of the dynamic perceptron in particle track recognition in high-energy accelerators. Firstly, we show the amazing improvement of performances that can be obtained in a perceptron architecture with classical learning by adding our dynamic preprocessing technique, already introduced last year in another paper presented at this Conference. Secondly, we show the first results of this technique extended also at the level of learning procedure always applied to the problem of particle track recognition.

This paper summarizes the results of a signal taxonomy study of gamma ray burst (GRB) data acquired with sensors on-board the Pioneer-Venus Orbiter (PVO) spacecraft. GRB events produce large fluxes of gamma rays with durations of seconds to minutes and have been observed since the early 1970's. The true nature of GRB's is still unknown, and several competing theories exist. A fundamental point of contention among such theories is whether or not different types of GRB exist. If different types of GRB's are discovered in the existing PVO data base, the differences may correlate with their position or source characteristics. Hence, the goal of this project was to use artificial neural networks to perform signal taxonomy on the GRB data base to determine if unique classes or types of GRB's exist. A total of 26 signal features were identified, some of which can be associated directly with some characteristic of the GRB, such as duration, peak count rate, and gamma ray spectrum hardness. Additional features that were selected included the number of zero crossings in the wavelet transform and the fractal dimension of each signal. A self organizing neural network was used with the signal features to search for correlations among the signals contained in the database. The results of this analysis revealed an intrinsic dimensionality of 2 or 3 in the database. That is, it appears as though 2 or 3 distinct types of GRB may exist. In particular, two of the classes contain roughly 90% of the signals in the database of GRB signals we had to work with. These two classes are similar in characteristics but are still sufficiently distinct from one another to form separate categories. The third class of GRB is definitely distinct from the first two.

In this study, neural networks have been used to retrieve thermal profiles from near polar, sun-synchronous meteorological satellite data obtained from the TIROS Operational Vertical Sounder (TOVS). Data were collected using the SeaSpace TeraScan satellite tracking system for thirteen sites ranging from the Southwestern United States to Canada. Earth-centered radiances, latitude, longitude, elevation, and angular information (satellite zenith angle, solar zenith angle, scatter phase angle, and sun reflection angle) were used as inputs to a backpropagation neural network. The network architecture consisted of one hidden layer of 30 neurons. The output layer provided temperature at the meteorological `mandatory' levels as well as the surface. Truth consisted of the thermal profiles obtained from a conventional algorithm, the TOVS Export Package. The results demonstrate that thermal profiles with Root Mean Square Errors of less than 4 C (typically < 3 C) can be obtained from the trained neural network. As expected, the accuracy of the thermal profiles is greatest at higher altitudes. These results are obtained without the computational overhead and complexity of conventional approaches.

This paper presents a study of the Fuzzy ARTMAP neural network in designing cascaded gratings and Frequency Selective Surfaces (FSS) in general. Conventionally, trial-and-error procedures are used until an FSS matches the design criteria. One way of avoiding this laborious and manual process is to use neural networks. A neural network can be trained to predict the dimensions of the metallic patches (or apertures), their distance of separation, their shape, and the number of layers required in a multilayer structure which gives the desired frequency response. In the past, to achieve this goal, the backpropagation (backprop) learning algorithm was used in conjunction with an inversion algorithm. Unfortunately, the backprop algorithm sometimes has problems with convergence. In this work the Fuzzy ARTMAP neural networks is utilized. The Fuzzy ARTMAP is faster to train than the backprop and it does not require an inversion algorithm to solve the FSS problem. Most importantly, its convergence is guaranteed. Several results (frequency responses) from cascaded gratings for various angles of wave incidence, layer separation, width strips, and interstrip separation are presented and discussed.

This paper summarizes a research effort in finding the locations and sizes of faces in color images (photographs, video stills, etc.) if, in fact, faces are presented. Scenarios for using such a system include serving as the means of localizing skin for automatic color balancing during photo processing or it could be used as a front-end in a customs port of energy context for a system which identified persona non grata given a database of known faces. The approach presented here is a hybrid system including: a neural pre-processor, some conventional image processing steps, and a neural classifier as the final face/non-face discriminator. Neither the training (containing 17,655 faces) nor the test (containing 1829 faces) imagery databases were constrained in their content or quality. The results for the pilot system are reported along with a discussion for improving the current system.

The principal components of a collection of randomly sampled photos from a police mugshot album were extracted using a modification of the neural net described by Sanger. Principal component analysis provides a basis that `spans' the face space, from which each face in a population can be reconstructed simply by taking a proper linear combination of the basis components. The coefficients of this linear combination can serve as a measure of similarity between faces. In previous studies, the authors tried a mugshot album search strategy based on subjective similarity judgments. A network of global, subjective similarity judgments was established between each photo in a small data base (3000 photos). The witness chose the photos most similar to the target from the set displayed on the monitor. The computer used the similarity network to rerank the remaining photos in the album that had not yet been displayed, to select the next set of photos with the best fit for presentation. This process was continued until the target photo was located. In this study, we used the objective similarity measure based on principal component coefficients in place of the subjective judgments. Each image in the experimental database was automatically coded using the first 100 principal components. The same experimental procedure as used with the manually coded data base was conducted. The results are better than those achieved with the subjective method and encourage the use of this coding scheme on larger police albums (100,000 photos).

Researchers have for many years tried to develop machine recognition systems using video images of the human face as the input, with limited success. This paper presents a technique for recognizing individuals based on facial features using a novel multi-layer neural network architecture called `PWRNET'. We envision a real-time version of this technique to be used for high security applications. Two systems are proposed. One involves taking a grayscale video image and using it directly, the other involves decomposing the grayscale image into a series of binary images using the isodensity regions of the image. Isodensity regions are the areas within an image where the intensity is within a certain range. The binary image is produced by setting the pixels inside this intensity range to one, and the rest of the pixels in the image to zero. Features based on moments are subsequently extracted from these grayscale images. These features are then used for classification of the image. The classification is accomplished using an artificial neural network called `PWRNET', which produces a polynomial expression of the trained network. There is one neural network for each individual to be identified, with an output value which is either positive or negative identification. A detailed development of the design is presented, and identification for small population of individuals is presented. It is shown that the system is effective for variations in both scale and translation, which are considered to be reasonable variations for this type of facial identification.

Two verification systems, face and speaker, are fused to form a single identity verification system. The Karhunen-Loeve Transform (KLT) is used for dimensional reduction, and a back- propagation neural net is used for classification. Verification involved training a net for each individual in the database for two classes of outputs, `Joe' or `not Joe.' The base speaker identification system used Cepstral analysis for feature extraction and a distortion measure for classification. Verification in this case involved performing the KLT on the Cepstral coefficients and then classifying using a two-class neural net for each individual. KLT feature reduction is compared to alternative linear methods, and the KLT is found to provide superior performance. The fusion of the two base verification systems is shown to provide superior performance over either system alone.