Current measures of an artificial neural network (ANN) capability are based on the V-C dimension and its variations. These measures may be underestimating the actual ANNs capabilities and hence overestimating the required number of examples for learning. This is caused by relying on a single invariant description of the problem set, which, in this case is cardinality, and requiring worst case geometric arrangements and colorings. A capability measure allows aligning the measure with desired characteristics of the problem sets. The mathematical framework has been established in which to express other desired invariant descriptors of a capability measure. New invariants are defined on the problem space that yield new capability measures of ANNs that are based on topological properties. A specific example of an invariant is given which is based on topological properties of the problem set and yields a new measure of ANN architecture.

The quality of a feedforward neural network that allows it to associate data not used in training is called generalization. A common method of creating the desired network is for the user to select the network architecture and allowing a training algorithm to evolve the synaptic weights between the neurons. A popular belief is that the network with the fewest number of hidden neurons that correctly learns a sufficient training set is a network with better generalization. This paper will contradict that belief. The optimization of generalization requires that the network not assume information that does not exist in the training data. Unfortunately, a network with the minimum number of hidden neurons may require assumptions of information that does not exist. The network then skews the surface that maps the input space to the output space in order to accommodate the minimum architecture which then sacrifices generalization.

We have now many kinds of human assisted systems and facilities like computer, AI, robot, Database, new media, the internetworkings, etc. And we rely on these very much. But it is true to say that we cannot put our trust fully in them. That is to say, we still have to have thoughts whether they are alright or reliable or useful or not. So here proves that it can be made by neural networks of today by taking an idea from behavior/mechanisms that our brains have in order to think out.

Numerical simulation of physics-based models has been applied to computer graphics animation due to the high degree of realism and automation it offers. However, the high cost of computation with numerical simulation is a major disadvantage compared to the more efficient geometric- based approaches. This paper shows a different approach to creating realistic simulations by using neural networks to observe and learn the dynamics of physics-based models. It also facilitates a means to solve the control problem associated with physics-based models efficiently and generate goal-based simulations. In the implementation, a regularization network is selected with sigmoidal units to emulate the dynamics of a three-linked pendulum subjected to a gravitational field. It is demonstrated by computer simulation that a feed-forward neural network is able to animate the motion of a pendulum using a limited set of data.

In this research, we investigate partitioning schemes for reducing the computational complexity of an error diffusion neural network (EDN) for the application of digital halftoning. We show that by partitioning the original image into k subimages, the time required to perform the halftoning using an EDN is reduced by as much as a factor of k. Motivated by this potential speedup, we introduce three approaches to partitioning with varying degrees of overlap and communication between the partitions. We quantitatively demonstrate that the Constrained Framing approach produces halftoned images whose quality is as good as the quality of halftoned images produced by the EDN without partitioning.

Boosting is a mechanism that combines a large number of weak classifiers into a single stronger classifier by taking a weighted majority vote. We show that even very simple classifiers can be combined to form a reliable classifier through boosting, even when the individual classifiers themselves are rather poor. We demonstrate these results on several interesting problems including image analysis, chemical weapon detection, and an artificial olfactory device. Some practical aspects are discussed, particularly as related to very large data sets, high dimensional data, and error improvement versus increased computation time.

The analytical approach based on modeling Takens phase space temporal transformation for nonlinear delay discrete iterations is proposed. This model allows to obtain expressions for functional matrix of the vector map reconstructed phase trajectories from delay discrete iterations in multidimensional phase space. On a base of the suggested mode, analytical expressions for embedding parameters of investigated attractor have been calculated. Some computer experiments have been implemented modeling of the vector feedback in nonlinear processes under investigation.

It is quite well-known that one-hidden-layer feed-forward neural networks (FNNs) can approximate any continuous function to any desired accuracy as long as enough hidden units are included. Due to this fact many developments in constructive neural networks have been concentrated on only constructive or adaptive one-hidden-layer FNNs. However, this fact does not necessarily imply that one-hidden-layer networks are the most efficient and the best network structure feasible, as one has no explicit guideline to properly select the network structure. Consequently, in practice it has been observed that networks with more than one hidden layer may perform better than the one-hidden- layer networks in some applications. In this paper, we propose a novel strategy for constructing a multi-hidden- layer FNN with regular connections. The new algorithm incorporates in part the policy for adding hidden units from a one-hidden-layer constructive algorithm, and has in part its own new policy for additional layer creation. Extensive simulations are performed for nonlinear noisy regression problems, and it is found that the proposed algorithm converges quite fast and produces networks with one or as many hidden layers/units as required, which are dictated by the complexity of the underlying problem.

In this paper certain simple procedures are presented for stabilization of a class of NN trained by Back Propagation algorithm with minimum number of failures. To quicken the network response with minimum number of oscillations, a slope parameter is utilized in the bipolar sigmoidal activation function and is appropriately chosen with the help of Lyapunov's stability theorem. Further a new weight update scheme is proposed for the backpropagation algorithm. The above procedures are applied and tested with XOR problem, Iris data and image data for the choices of slope parameter, learning rate and momentum factor; its performance in terms of local minima, learning speed are evaluated and compared with the performance of traditional BP algorithm.

In cascade-correlation (CC) and constructive one-hidden- layer networks, structural level adaptation is achieved by incorporating new hidden units with identical activation functions one at a time into the active evolutionary net. Functional level adaptation has not received considerable attention, since selecting the activation functions will increase the search space considerably, and a systematic and a rigorous algorithm for accomplishing the search will be required as well. In this paper, we present a new strategy that is applicable to both the fixed structure as well as the constructive network trainings by using different activation functions having hierarchical degrees of nonlinearities, as the constructive learning of a one- hidden-layer feed-forward neural network (FNN) is progressing. Specifically, the orthonormal Hermite polynomials are used as the activation functions of the hidden units, which have certain interesting properties that are beneficial in network training. Simulation results for several noisy regression problems have revealed that our scheme can produce FNNs that generalize much better than one-hidden-layer constructive FNNs with identical sigmoidal activation functions, in particular as applied to rather complicated problems.

On the basis of the analysis of scientific ideas reflecting the law in the structure and functioning the biological structures of a brain, and analysis and synthesis of knowledge, developed by various directions in Computer Science, also there were developed the bases of the theory of a new class neural-like growing networks, not having the analogue in world practice. In a base of neural-like growing networks the synthesis of knowledge developed by classical theories - semantic and neural of networks is. The first of them enable to form sense, as objects and connections between them in accordance with construction of the network. With thus each sense gets a separate a component of a network as top, connected to other tops. In common it quite corresponds to structure reflected in a brain, where each obvious concept is presented by certain structure and has designating symbol. Secondly, this network gets increased semantic clearness at the expense owing to formation not only connections between neural by elements, but also themselves of elements as such, i.e. here has a place not simply construction of a network by accommodation sense structures in environment neural of elements, and purely creation of most this environment, as of an equivalent of environment of memory. Thus neural-like growing networks are represented by the convenient apparatus for modeling of mechanisms of teleological thinking, as a fulfillment of certain psychophysiological of functions.

Third-generation (3G) wireless networks, based on a hierarchical cellular structure, support tiered levels of multimedia services. These services can be categorized as real-time and delay-sensitive, or non-real-time and delay- insensitive. Each call carries demand for one or more services in parallel; each with a guaranteed quality of service (QoS). Roaming is handled by handoff procedures between base stations (BSs) and the mobile subscribers (MSs) within the network. Metrics such as the probabilities of handoff failure, dropped calls and blocked calls; handoff transition time; and handoff rate are used to evaluate the handoff schemes, which also directly affects QoS. Previous researchers have proposed a fuzzy logic system (FLS) with neural encoding of the rule base and probabilistic neural network to solve the handoff decision as a pattern recognition problem in the set of MS signal measurements and mobility amid fading path uncertainties. Both neural approaches evalute only voice traffic in a closed, single- layer network of uniform cells. This paper proposed a new topology-preserving, self-organizing neural network (SONN) for both handoff and admission control as part of an overall resource allocation (RA) problem to support QoS in a three- layer, wideband CDMA HCS with dynamic loading of multimedia services. MS profiles include simultaneous service requirements, which are mapped to a new set of variables, defined in terms of the network radio resources (RRs). Simulations of the new SONN-based algorithms under various operating scenarios of MS mobility, dynamic loading, active set size, and RR bounds, using published traffic models of 3G services, compare their performance with earlier approaches.

In this paper, we propose a new measure of novelty detection for target selection in visual scenes. Our approach to the definition of novelty is based on the use of local kernels and Fisher information metric in the context of support vector machine regression. We discuss the applications in the specific context of visual saccades as a mechanism of search and discuss naturel generations of the approach in other contexts. We also propose natural regularization approaches arising from consideration of the problem that can be applied to learning machines including the SVM.

Standard approaches for the classification of high dimensional data require the selection of features, the projection of the features to a lower dimensional space, and the construction of the classifier in the lower dimensional space. Two fundamental issues arise in determining an appropriate projection to a lower dimensional space: the target dimensionality for the projection must be determined, and a particular projection must be selected from a specified family. We present an algorithm which is designed specifically for classification task and addresses both these issues. The family of nonlinear projections considered is based on interpoint distances - in particular, we consider point-to-subset distances. Our algorithm selects both the number of subsets to use and the subsets themselves. The methodology is applied to an artificial nose odorant classification task.

A rule-generating algorithm, Incremental Reduced Error Pruning (IREP), has been proposed by Furnkranz and Widmer. A modified IREP algorithm (RIPPERk) may be applied to raw data representing a classification problem. Introduced by Cohen, 1995, RIPPERk generates a set of hypotheses in the form of if-then rules. The resulting solution maybe coarse or compete, covering all outlyers in the classification data set.

The basic idea of theory of Marr's image edge-detecting is firstly to smooth original image with Gaussian function, then obtain the zero-cross map of Laplacian's transform of smoothed image. However, the residual between the original image and smoothed image remain s some feature points that may not be detected. Therefore, this paper firstly proposed a new smoothing operator which has low-pass characteristics similar to a Butterworth filter and limited spatial extent similar to a Gaussian function, then we constructed a class of edge-detecting operator that can be controlled more easily using Laplace's transform. The new edge-detecting operator also has closed forms that facilitate implementation, and allows us flexibility control feature- detecting accuracy compared to Marr's operator. In addition, Marr's edge-detecting operator is a special formulation of a new operator. Practical numerical experimental results showed that hose edge-detecting operators have some practical effect and reference value.

The fuzzy c-means algorithm is a useful tool for clustering real s-dimensional data. Typically, each observation consists of numerical values for s feature such as height, length, etc. In some cases, data sets contain vectors that are missing one or more feature values. For example, a particular datum might have the form: (254.3, x, 36.2, 112.7, x), where the second and fifth feature values are missing. The (standard) fuzzy c-means algorithm cannot be applied in this case since the required computations reference numerical features values for all s features of every data point. Two adaptations of fuzzy c-means to the incomplete data case are presented here. One adaptation replaces unknown feature values with additional variables that are optimized to prove an extrapolated data set yielding the smallest possible value of the fuzzy c-means criterion. Another approach uses only the available feature values in distance calculations, and then adjusts for the missing feature values by an appropriately chosen scaling of the computed distances. Numerical convergence properties of the adaptations and computational costs are discussed. Artificial data sets are used to demonstrate the two new approaches.

Evolutionary algorithms are often applied to tasks where the challenges is to find a superior solution. The engineering challenge concerns how to best design such algorithms in terms of their representation, variation operators, and selection. The distribution of fitness scores that is obtained when applying variation operators to parents can provide useful information for setting the parameters that are associated with those operators. Experiments presented here indicate that fitness distributions can also reveal information about the landscapes that surrounds particular parents and suggest that typical methods of self-adaption may not be very well suited for exploring the state space of possible solutions in the presence of multiple minima.

In typical GA application, the fitness assigned to a chromosome-represented individual in the context of a specified environment takes a deterministic calculable value. In many problems of interest, the fitness of an individual is stochastic, and the environment changes in unpredictable ways. These two factors contribute to an uncertainty that can be associated with the estimated fitness of the individual.

The objective of this study is to compare geometric-based and evolutionary techniques for tracking storm systems from sequences of satellite images. Analysis was applied to the International Satellite Cloud Climatology Project low resolution D1 database for selected storm systems during the month of September, 1988. During this time period there were two exceptionally long tracks of major hurricane systems, Hurricanes Gilbert and Helene. Cloud top pressure and cloud optical thickness were used to identify storm systems. The ability of the geometric-based and evolutionary techniques to generate tracks through storm regions was assessed. Differences in final tracking results between the two techniques resulted not only from the differences in methodology but also form differences in the type of preprocessed input used by each of the techniques. Tracking results were compared to results disseminated by the Colorado State/Tropical Prediction Center and maintained by the National Hurricane Center in Miami, Florida. For the hurricanes investigated in this study, both techniques were able to generate tracks which followed either most or some of the portions of the hurricanes. The evolutionary algorithm was in general able to maintain good continuity along the tracks but, with no knowledge of overall region movement, was unable to discern which of two possible directions would be best to pursue in cases where there were tow or more equally close storm systems components. The geometric method was able to maintain a smooth track close to the course of the hurricane except for confusion primarily at the beginning and/or end of tracks.

The NASA Scatterometer was a satellite radar system launched in August 1996 on Japan's Advanced Earth Observing System ADEOS to remotely sense ocean surface wind vectors. This radar measures backscattered power for the ocean at three azimuths and uses a non-linear wind retrieval algorithm to infer surface wind vectors. This paper presents a new approach to the wind retrieval process to estimate ocean- surface wind vectors. This new approach employs a genetic algorithm (GA) to find the wind vector solutions that minimize the likelihood function. The likelihood function is generated by summing the errors of the theoretical backscattered versus measured sigma-0 divided by standard deviation of the measurements in every wind vector cell. Currently, the NSCAT Project algorithm implements a 'special brute force' approach to finding the wind vector solutions that minimize the likelihood function. The paper also present comparisons of the result of using the GA approach versus the current NSCAT algorithm. The GA approach is shown to be more robust and immune to reaching sub-optical solutions by avoiding local minima.

In this article we developed a method for optimizing the structure of a fuzzy artificial neural networks through genetic algorithms. This genetic algorithm is used by optimizing the number of weight connections in a neural network structure, by the evolution of those structures as individuals in a population. It is found that the optimization of the neural network provides higher confidence accuracy of the suggested solution in a case based diagnostic system. The computational cost of the optimized network also improved considerably high.

The independent component analysis (ICA) problem involves finding a set of statistically independent signals from a set of measurements consisting of unknown, perhaps convolutive mixtures of those signals. This problem arises in many applications such as speech processing, communications, and biomedical signal processing. We present a method to blindly separate instantaneous mixtures of non- Gaussian signals using a genetic algorithm (GA) and higher order statistics. The GA searches for a separating matrix such that the resulting output signal are both statistically independent and strongly non-Gaussian as measured by the kurtosis. The GA uses a binary representation together with a coarse-to-fine strategy to speed convergence and avoid such bits. Using data from a simulated narrow band communications scenario, we examine the algorithm's performance as signal length and sensor noise level are varied. We compare this performance with that obtained using the ACI algorithm developed by Comon. We show that the GA is able to achieve good separation of dense signal constellations, and achieves better separation with lower mean-square estimation error than the ACI, albeit with much higher algorithmic complexity. The improvement in performance may be dramatic when the signal length is short.

This paper presents an overview of current ongoing research and design efforts conducted by Intelligent Optical Systems, Inc. in the area of hardware-based color segmentation. We discuss the specifics of the design of a microchip that combines a hardwired hybrid neural network with on-chip imaging. Several preliminary tests show high approximation ability of our scheme. The single-chip implement has many advantages. The final product will consists of an RGB pixel array with infinite color depth and a neural network capable of high speed image segmentation.

In this paper we discuss variants of digital-analog CMOS implementation of artificial neuron taught to logical threshold functions. The implementation is based on earlier suggested (beta) -comparator and three output amplifiers. Such a circuit can be taught only to threshold functions with positive weights of variables, which belong to the class of isotonous Boolean functions. However, most problems solved by artificial neural networks either require inhibitory inputs. If the input type is known beforehand, the problem of inverting the weight sign is solved trivially by inverting the respective variable. Otherwise, the neural should have synapses capable of forming the weight and type of the input during the learning, using only increment and decrement signals. A neuron with such synapses can learn an arbitrary threshold function of a certain number of variables. Synapse circuits are suggested with two or one memory element for storing positive and negative input weights. The results of SPICe simulation prove that the problem of teaching non-isotonous threshold functions to a neuron has stable solutions.

Recently the ASL at Tennessee Technological University was donated a GEP50 welder. The welding is done via off line point-to-point teaching. A state of the art robot was needed for research but because money was not available to purchase such an expensive item. It was therefore decided to upgrade the GEP50 control system to make the welder a multitasking robot. The robot has five degrees of freedom can be sufficient to pursue some research in robotics control. The problem was that the control system of the welder is limited to point-to-point control, using off-line teaching. To make the GEP50 a multitasking robot that can be controlled using different control strategies, the existing control system of the welder had to be replaced. The upgrade turned to be a low cost operation. This robot is currently in sue to test different advanced control strategies in the ASL. This work discusses all the steps and tasks undertaken during the upgrade operation. The hardware and software required or the upgrade are provided in this paper. The newly developed control system has been implemented and tested successfully.

The surfaces of asteroids consist of mineral grains mixed on scales of 10-110's of micrometers where linear mixing rules for near-IR spectra do not apply. Further, the spectral properties of the mineral components are strong nonlinear functions of grain size and chemical composition. Detailed models of these nonlinear properties exist, but are not amenable to analytic inversion, requiring relatively inefficient iterative solutions to extract physical data from reflectance spectra. However, the NASA Near-Earth Asteroid Rendezvous mission soon to orbit the asteroid Eros requires near real time spectral analysis of near-IR spectral data for mission operations planning. The S- Asteroid Spectral Interpreter is a software system which includes a neural network which has been trained to invert the nonlinear physical model, and conventional gradient descent algorithm which refines the output of the neural network to arrive at a rapid analysis of input spectra.

We have evolved a Foveal Visual Pre-processor: the Retina model. It is based on an artificial neural network organized to simulate the radial variation of the visual acuity. The information is encoded through the implementation of analogic and impulse neurons. The main interest of this model inspired of this model inspired by the vertebrate retina is its response to stationary or moving stimuli: they can be distinguished according to both their shapes and velocities. This model is adaptive and its multi-resolution characteristics allow the detection of a wide range of velocities. From impulse output signals of Retina, we extract pertinent parameters that encode the motion and pattern information thanks to a time frequency analysis. We study the influence of the different retina areas in the velocity extraction. Our system realizes a very good generalization for classification of stimuli with different level of luminance and noise. The properties of our Retina model: adaptivity, multi-resolution allow us to consider its application on a real time sequence images. A control module combined with this sensor enabled us to reach interesting result in applications such as selective tracking of stimuli, tracking of solid or dotted white line on highway or road, image exploration.

Applications of neural networks have been reported on a lot in recent years, but the research on how to find reliable guidelines to design neural networks is still in its infancy. This work intends to provide some ideas on how to find useful predefined network structures for at least certain parts of the neural net. By breaking off to a certain extend the so-called black-box character of the neural net, the performance of the networks can be improved and the solutions of the nets get more transparent and understandable at the same time. Additionally, the ability of the neural nets to generalize from some training patterns to unlearned data regions is improved substantially. In this work two commonly used engineering principles such as the technique of dimensional analysis and the Laplace- transformation are used to identify suitable topologies for neural networks. The integration of the dimensional analysis in the context of feed-forward neural networks is presented. In the second part of this work the use of the Laplace- transformation in neural networks is demonstrated. Even though at the moment the application of this technique has been shown in a linear time-invariant process, a future use of this method in nonlinear system is considered.

Smart structures incorporating actor and sensor elements usually show an essentially nonlinear and often additionally time-variant behavior. Neural networks as a nonlinear and adaptive computational tool provide interesting possibilities in the field of control of smart structures. Different neural network control methodologies have been proposed in the literature, almost all neglecting a priori knowledge about the plant to be controlled. In this paper an indirect neural network control approach for smart structure system is shown using a neural network plant model designed according to the results of dimensional analysis.

Successful signal classification requires the selection of a problem-oriented signal representation and the extraction of a feature vector for final classification. The choice of the optimum representation and optimum features needs some a prior knowledge of the inherent structure and symmetries of the time series as well as the differences between the classes to be distinguished. One central question is: which time series are really similar. An important task for successful signal classification is therefor a clear concept of similarity. A very powerful tool for signal analysis are time-frequency distributions. They are 2D functions that indicate the time-varying frequency content of 1D signals. Each bilinear time-frequency distribution corresponds to a kernel function that controls the properties of the integral transform. However, there is no single time-frequency distribution which is optimal for all problems. It is still an unsolved problem to determine the optimum time-frequency distribution for a given signal and analysis task. In this work a data-driven adaptive time-frequency distribution is presented. The kernel of the time-frequency distribution is parameterized and adapts to the maximal classification rate. The subsequent feature extraction results from calculating the joint moments of the time-frequency distribution. Dimensional analysis is used for defining a similarity concept for time series analysis through generation of a set of dimensionless classification numbers.

In image reconnaissance the analyst of remotely sensed imagery is confronted with large amounts of data. Especially the integration of multi-sensor data calls for support of the observer by automatic image processing algorithms. For this purpose we recently developed model based structural image analysis algorithms which deliver successful results. However, varying scenarios, different applications and changing image material often require a tuning of the algorithms. Therefore, we suggest techniques to support and automate the adaptation of the image processing to changing requirements. Our approach uses techniques form data mining to discover relationships between image properties and optimal parameter vectors. This paper addresses two points: a supervised tuning approach and suggestions for unsupervised tuning. For the supervised tuning a representative image database was set up, and a corresponding ground truth was interactively defined. The results of the structural image analysis for a set of parameters can be compared to the ground truth. For the example images the parameters were optimized using an evolutionary optimization loop. For the unsupervised tuning the data form the supervised optimization is analyzed. We present promising results form manual clustering and propose a clustering approach based on decision trees, and hierarchical and evolutionary cluster algorithms with different distance measures.

In this paper, we present new findings in constructing and applications of artificial neural networks that use a biologically inspired spiking neuron model. The used model is a point neuron with the interaction between neurons described by postsynaptic potentials. The synaptic plasticity is achieved by using a temporal correlation learning rule, specified as a function of time difference between the firings of pre- and post-synaptic neurons. Using this rule we show how certain associations between neurons in a network of spiking neurons can be implemented. As an example we analyze the dynamic properties of networks of laterally connected spiking neurons and we show their capability to self-organize into topological maps in response to external stimulation. In another application we explore the capability networks of spiking neurons to solve graph algorithms by using temporal coding of distances in a given spatial configuration. The paper underlines the importance of temporal dimension in artificial neural network information processing.

In the last decades biologists have gained much knowledge about neural firing in the auditory system. It is a challenging problem to use this knowledge for the improvement of hearing aids and cochlear implants. This study first present the model of a human cochlea, which transforms acoustic signals into auditory nerve impulses. Then a method is proposed, which reconstructs the nerve impulses into acoustic signals. This method will then be used on the impulse-output of a widely used cochlear implant, in order to get an impression of what patients actually perceive with such a device. Suggestions for the improvement of coding strategies will be made, based on the findings of this study.

Laser-based ultrasonic (LBU) measurement shows great promise for on-line monitoring of weld quality in tailor-welded blanks. Tailor-welded blanks are steel blanks made from plates of differing thickness and/or properties butt-welded together; they are used in automobile manufacturing to produce body, frame, and closure panels. LBU uses a pulsed laser to generate the ultrasound and a continuous wave laser interferometer to detect the ultrasound at the point of interrogation to perform ultrasonic inspection. LBU enables in-process measurements since there is no sensor contact or near-contact with the workpiece. The authors are using laser-generated plate waves to propagate from one plate into the weld nugget as a means of detecting defects.

Mapping landcover type from airborne/spaceborne sensors is an important classification problem in remote sensing. Due to advances in sensor technology, it is now possible to acquire hyperspectral data simultaneously in more than 100 bands, each of which measures the integrated response of a target over a narrow window of the electromagnetic spectrum. The bands are ordered by their wavelengths and spectrally adjacent bands are generally statistically correlated.

In this paper, a kind of fuzzy algorithm for learning vector quantization is developed and used as pattern classifiers with a supervised learning paradigm in artificial odor discrimination system. In this type of FLVQ, the neuron activation is derived through fuzziness of the input data, so that the neural system could deal with the statistical of the measurement error directly. During learning,the similarity between the training vector and the reference vectors are calculated, and the winning reference vector is updated in two ways. Firstly, by shifting the central position of the fuzzy reference vector toward or away from the input vector, and secondly, by modifying its fuzziness. Two types of fuzziness modifications are used, i.e., a constant modification factor and a variable modification factor. This type of FLVQ is different in nature with FALVQ, and in this paper, the performance of FNLVQ network is compared with that of FALVQ in artificial odor recognition system. Experimental results show that both FALVQ and FNLVQ provided high recognition probability in determining various learn-category of odors, however, the FNLVQ neural system has the ability to recognize the unlearn-category of odor that could not recognized by FALVQ neural system.

The analysis of use cases describing construction of simulation configuration files in a data/information management system can lead to the acquisition of new information and knowledge. In this application, a user creates a use case with an eXtensible Markup Language (XML) description representing a configuration file for simulation of a physical system. INtelligent agents analyze separate versions of the XML descriptions of a user and additionally, make comparisons of the descriptions with examples form a library of use cases. The agents can then make recommendations to a user on how to proceed or if tutoring is necessary. In a proof-of-concept test, new information is acquired and a user learns from the agent-facilitated tutoring.

Neural networks are increasingly being used as a forecasting tool in many forecasting problems. This paper discusses the application of neural networks in predicting daily foreign exchange rates between the USD, GBP as well as DEM. We approach the problem from a time-series analysis framework - where future exchange rates are forecasted solely using past exchange rates. This relies on the belief that the past prices and future prices are very close related, and interdependent. We present the result of training a neural network with historical USD-GBP data. The methodology used in explained, as well as the training process. We discuss the selection of inputs to the network, and present a comparison of using the actual exchange rates and the exchange rate differences as inputs. Price and rate differences are the preferred way of training neural network in financial applications. Results of both approaches are present together for comparison. We show that the network is able to learn the trends in the exchange rate movements correctly, and present the results of the prediction over several periods of time.

In modern day production, tool condition monitoring systems are needed to get better quality of jobs and to ensure reduction in the downtime of machine tools due to catastrophic tool failures. Tool condition monitors alert the operator about excessive tool wear and stop the machine in case of an impending breakage or collision of tool. A tool condition monitoring system based on measurement of thrust has been developed for a CNC gantry-drilling machine. The system, though performing well, has limitations due to its total dependence on single sensor input. In view of this, investigations have been carried out to adopt a multi- sensor approach for this system. The inputs of axial thrust, spindle motor current and vibrations are used and the decision regarding the condition of tool is made using ANNs. Initially, a training algorithm is used to learn the complex association between sensor inputs and drill wear. Later on the trained network is employed to assess the condition of drill on new sensory information. An ANN based on Error Back Propagation algorithm is employed. The paper discusses various aspects considered in choosing the design parameters for the NN. The experimental results are presented in this paper.

Gas Insulated substations (GIS) up to 500kV class have been widely accepted over conventional air insulated substation due to several advantages. However, the presence of floating metal particles and protrusions within the GIS at various locations could seriously affect the performance. The paper describes the method of detection of partial discharges for various type of discharging sources e.g. floating particles, protrusions of high voltage conductor and particles sticking on the surface of insulator. In order to identify the discharge source, a Neural Network program is developed to classify each of the above source on the basis of its characteristic pattern.

An important element of monitoring compliance of the Comprehensive Nuclear-Test-Ban Treaty (CTBT) is an infrasound network. For reliable monitoring, it is important to distinguish between nuclear explosions and other sources of infrasound. This will require signal classification after a detection is made.

Arcing in high-energy systems can have a detrimental effect on the operational performance, energy efficiency, life cycle and operating and support costs of a facility. In can occur in motors, switching networks, and transformers and can pose a serious threat to humans who operate or work around the systems. To reduce this risk and increase operational efficiency, it is necessary to develop a capability to diagnose single and multiple arcing events in order to provide an effective measure of system performance. This calculated parameter can then be used to provide an effective measure of system health as it relates to arcing and its deleterious effects. This paper details the development of a model-based matched filter for an antenna that recognizes single and/or multiple arcing events in a direct current motor and calculates a functional measure of activity and a confidence factor based on an estimate of how well the data fit the matched filter model parameters. A principal component analysis is then performed on the descriptive statistics calculated from the model's input data stream to develop cluster centers for classifying non- arcing and arching events that are invariant to system operation set point. This approach also has a deployment benefit in that the PCA decreases the computational load on the classifier system by reducing the order of the system.

This paper investigates the use of a beta value derived from a receiver operator characteristic curve for target recognition. Using a physiologically-motivated sensor-fusion algorithm, lower-level data is filtered and fused using a pulse-coupled neural network (PCNN) to represent the feature processing of the parvocellular and magnetocellular pathways. High level decision making includes feature association from the PCNN filter, information fusion, and selection of a signal-detection beta value that optimizes performance. A beta value is represent bias based on a likelihood ratio of Gaussian distributions that can be used as a decision strategy to discriminate between targets. By employing a beta value as the output of the physiologic- motivated sensor fusion algorithm, targets are classified based on the fusion of feature data.

In this paper we describe a new nonlinear principal component analyzer and apply it in connection with a new compression scheme to lossy compression of digitized mammograms. We use a 'neural-gas' network for codebook design and several linear and nonlinear principal component method as a preprocessing technique. First, we analyze mathematically the nonlinear, single-layer neural network and show that the equilibrium points of this system are global asymptotically stale. Both a regular Hebbian rule and an anti-Hebbian rule are used for the adaptation of the connection weights between the constituent units. The, we investigate the performance of the compression scheme depending on the blocksize, codebook and number of chosen principal components. The nonlinear principal component method shows the best compression reslut in combination with the 'neural-gas' network.

Traditional CT reconstruction is done using the technique of Filtered Backprojection. While this technique is widely employed in industrial and medical applications, it is not generally understood that FB has a fundamental flaw. Gibbs phenomena states any Fourier reconstruction will produce errors in the vicinity of all discontinuities, and that the error will equal 28 percent of the discontinuity. A number of years back, one of the authors proposed a biological perception model whereby biological neural networks perceive 3D images from stereo vision. The perception model proports an internal hard-wired neural network which emulates the external physical process. A process is repeated whereby erroneous unknown internal values are used to generate an emulated signal with is compared to external sensed data, generating an error signal. Feedback from the error signal is then sued to update the erroneous internal values. The process is repeated until the error signal no longer decrease. It was soon realized that the same method could be used to obtain CT from x-rays without having to do Fourier transforms. Neural networks have the additional potential for handling non-linearities and missing data. The technique has been applied to some coral images, collected at the Los Alamos high-energy x-ray facility. The initial images show considerable promise, in some instances showing more detail than the FB images obtained from the same data. Although routine production using this new method would require a massively parallel computer, the method shows promise, especially where refined detail is required.

Map Gauss filter is a linear adaptive filter commonly used to reduce speckle noise present in synthetic aperture radar images of remote sensing satellites. In this study was incorporating some modifications that allow us to maximize the signal-to-noise ratio at the same time almost total features of the image are preserved. To evalute the performance of the new filter, both original and filtered images were classified by a unsupervised technique known as Parallel Self-Organizing Map and the results of this classification were compared. The P-SOM is an algorithm with its own-organization mapping that is specific for parallel computing environment. As examples of applications, are presented the results of the classification for preprocessed original RADARSAT images using the Map gauss, Frost and Gamma filters.

A novel personnel surveillance system has been developed for airport security to detect and identify threatening objects, which are concealed ont he human body. The main advantage of this system over conventional metal detectors is that non- metallic objects such as plastic explosives and plastic guns are detectable. This system is based on millimeter-wave array technology and a holographic imaging algorithm to provide surveillance images of objects hidden beneath clothing in near real-time. The privacy algorithm is based on image processing filters and artificial neural networks. The algorithm examines the millimeter-wave surveillance images to locate and segment the threats and place them on either a silhouette of the person or a wire-frame humanoid representation. In this way, all human features are removed from the final image and personal privacy is maintained. This system is ideally suited for mass transportation centers such as airport checkpoints that require high throughput rates. The system is currently under going evaluation. This paper reports on results from an earlier initial test of portions of the privacy algorithm that detect hidden plastic objects.

We propose a novel method for simultaneous speckle reduction and data compression based on wavelets. The main feature of the method is that of preserving the geometrical shapes of the figures present in the noisy images. A fast algorithm, the dynamic perceptron, is applied to detect the regular shapes present in the noisy image. Another fast algorithm is then used to find the best wavelet basis in the rate- distortion sense. Subsequently, a soft thresholding is applied in the wavelet domain to significantly suppress the speckles of the synthetic aperture radar images, while maintaining bright reflections for subsequent detection.

The current state of user interfaces for large information spaces imposes an unmanageable cognitive burden upon the user. Determining how to get the right information into the right form with the right tool at the right time has become a monumental task. Interface agents address the problem of increasing task load by serving as either an assistant or associate, extracting and analyzing relevant information, providing information abstractions of that information, and providing timely, beneficial assistance to suers. Interface agents communicate with the user through the existing user interface and also adapt to user needs and behaviors. User modeling, on the other hand, is concerned with how to represent users' knowledge and interaction within a system to adapt the system to the needs of users. The inclusion of a user model within the overall system architecture allows the system to adapt its response to the preferences, biases, expertise level, goals and needs.

A series of challenges to making computational intelligence viable in the real world is presented. These challenges include the applicability of artificial neural networks, fuzzy logic and evolutionary computation to limited data set problems. Various design and use perspectives will be presented to explain the challenges. A special panel of experts will address the challenges.