Learning a large number of simple local concepts is both faster and easier than learning a single global concept. Inspired by this principle of divide and conquer, a number of modular learning approaches have been proposed by the computational intelligence community. In modular learning, the classification/regression/clustering problem is first decomposed into a number of simpler subproblems, a module is learned for each of these subproblems, and finally their results are integrated by a suitable combining method. Mixtures of experts and clustering are two of the techniques that are describable in this paradigm. In this paper we present a broad framework for Generalized Associative Modular Learning Systems (GAMLS). Modularity is introduced through soft association of each training pattern with every module. The coupled problems of learning the module parameters and learning associations are solved iteratively using deterministic annealing. Starting at a high temperature with only one module, GAMLS framework automatically evolves the required number of modules through a systematic growing and pruning technique. Each phase begins by splitting every module in the previous phase into two, updating these new modules and then pruning and merging any redundant modules. A phase transition is induced by temperature decay. A number of existing modular learning problems, both unsupervised (clustering, mixture model density, mixture of principal components) and supervised (mixture of experts, radial basis function networks), can be effectively tackled in GAMLS. Case studies for clustering and regression using mixture of experts are provided for a number of datasets showing the efficacy of the GAMLS framework in evolving the right number of modules, inducing interpretable localizations among modules and robustness of the solution obtained. More importantly, this framework provides a unifying view for understanding and characterizing modular learning methods.

Current measures of an artificial neural networks (ANN) capability are the V-C dimension and its variations. These measures may be underestimating capabilities (in the primal sense) and hence overestimating the required number of examples for learning (in the dual sense). This is a result of relying on a single invariant description of the problem set, which is cardinality, and requiring worst case geometries and colorings. Generalization of a capability measure allows aligning the measure with desired characteristics of the problem sets. We present a mathematical framework in which to express other desired invariant descriptors of a capability measure, and guarantee proper application of the measure to ANNs. We define a collection of invariants defined on the problem space that yield new capability measures of ANNs. A specific example of an invariant is given which is based on geometric complexity of the problem set and yields a new measure of ANNs called the Ox-Cart dimension.

By overcoming weaknesses of the linear fuzzy system, our new piecewise nonlinear constructive method can effectively construct a reasonable fuzzy system with the near-optimal number of fuzzy rules. In addition, the new approach is capable of generating a commonly used fuzzy rule base with both meaningful input and output membership functions. The normal-fuzzy-reasoning-based nonlinear constructive approach provides us with a powerful tool to model a normal fuzzy system piece by piece (i.e. interval by interval) for both given data and any required accuracy. Additionally, the piecewise construct approach is a useful tool to discover meaningful fuzzy knowledge from raw data.

Applying generalized inverse learning to a feedforward neural network has been shown to be an effective tool in pattern recognition. The difficult computational step is finding the pseudo-inverse of a matrix. In this paper, we develop an efficient method using differential equations to calculate the pseudo-inverse.

The Vapnik-Chervonenkis (V-C) dimension of a set of functions representing a feed-forward, multi-layered, single output artificial neural network (ANN) with hard-limited activation functions can be evaluated using the Poincare polynomial of the implied hyperplane arrangement. This ANN geometrically is a hyperplane arrangement configured to dichotomize a signed set (i.e., a two-class set). Since it is known that the cut- intersections of the hyperplane arrangement forms a semi- lattice, then the Poincare polynomial can be used to evaluate certain geometric invariants of this semi-lattice, in particular, the cardinality of the resultant chamber set of the arrangements, which is shown to be the V-C dimension. From this theory comes a stable formula to compute the V-C dimension values.

Aim of this work is to demonstrate theoretically and experimentally how straightforwardly simple neural structures can obtain satisfying results in financial forecasting that can be easily used by market operators. The simplicity of the structures can allow indeed very flexible and user friendly implementations also for real-time forecasting. Such structure simplicity however has to be rightly understood. In fact, it is the result of a wide experimental research and a consequent theoretical demonstration devoted to outline a mathematical theorem for the definition of the optimal minimal neural structure for particular and very diffused typologies of financial data. The discussion of these theoretical and experimental results will be developed in this paper according to the following scheme: Deep theoretical discussion of the precedent points in terms of the 'generalization-learning theorem' for classical neural architectures. Recalling of the main principles underlying our 'dynamic perceptron' architecture presented and discussed elsewhere, also in precedent Orlando's SPIE Conferences. Partial neural implementation of these ideas by modification in a 'dynamic' sense of a classical back-propagation architecture. Application of the theoretical results discussed above to the time series of monetary cross-rates.

Proposed in this paper is a network which uses basis functions based on products of the input space variables raised to a variable power. These basis functions are introduced in regions of confusion obtained through vector quantization of the input space based on patterns which are erroneously classified by a simple linear classifier. The overall effect is thus of directly generating relevant higher order combinations of the input data in regions of maximum confusion. We present the complete architecture of the network and derive a training algorithm. Results using two synthetic data sets are provided.

Kosko proposed Bidirectional Associative Memory (BAM), where pairs of patterns (A_i,B_i) are encoded. When one of a pair of patterns is presented, the other is expected to be recalled. Irrespective of the number of pattern-pairs encoded, if dimensions of A_i and B_i are n and m respectively, a correlation matrix with (mn) elements is required to encode them; and at least O(mn) computation-time is required for recalling a pattern. It is believed that for practical applications (mn) is a large number. Moreover, to guarantee correct recalling of every encoded pattern, the correlation matrix may need to be augmented, which will increase the size of the matrix further. To overcome these problems, we propose a Three Layer BAM (TLBAM) and two novel encoding methods that require smaller size correlation-matrices. To encode p-pair of patterns, only p(m + n) elements are necessary. Thus, recalling time is also reduced. For instance, to encode three pattern-pairs from a recent paper (with n equals 288, m equals 280, and p equals 3) a correlation matrix of (288 X 280 equals) 80,640 elements is required. This encoding does not recall all three pairs correctly. Using one augmentation method the modified correlation matrix will have 89,600 elements for correct recall of all three pairs. Another augmentation method requires modified correlation matrix of 81,208 elements. Our novel encodings proposed here require two correlation matrices with only (288 X 3 + 3 X 280 equals) 1,704 elements.

The conditions of topological stabilization for Takens attractor are investigated when enlarging phase space dimension by means of topological dependences analysis and asymptotical estimations calculation. It has been shown that ideal topological stabilization is equivalent to presence of linearized segmentation properties in time series under investigation, these properties being invariant concerning scale of partition. The exact ideal topological stabilization appears to be impossible for nonlinear time series, so we have to consider asymptotical estimations of convergence to exact stabilization. Some statistical characteristics of investigated attractor have been defined by relatively- difference investigation of obtained topological curves.

Field operators use their eyes, ears, and nose to detect process behavior and to trigger corrective control actions. For instance: in daily practice, the experienced operator in sulfuric acid treatment of phosphate rock may observe froth color or bubble character to control process material in-flow. Or, similarly, (s)he may use acoustic sound of cavitation or boiling/flashing to increase or decrease material flow rates in tank levels. By contrast, process control computers continue to be limited to taking action on P, T, F, and A signals. Yet, there is sufficient evidence from the fields that visual and acoustic information can be used for control and identification. Smart in-situ sensors have facilitated potential mechanism for factory automation with promising industry applicability. In respond to these critical needs, a generic, structured health monitoring approach is proposed. The system assumes a given sensor suite will act as an on-line health usage monitor and at best provide the real-time control autonomy. The sensor suite can incorporate various types of sensory devices, from vibration accelerometers, directional microphones, machine vision CCDs, pressure gauges to temperature indicators. The decision can be shown in a visual on-board display or fed to the control block to invoke controller reconfigurration.

This paper deals with tube leak detection in industrial boilers. A decentralized information processing approach is used to detect and isolate the location of boilers tube leaks. Tube leak sensitive variables (TLSV) are used as the information source for detection and isolation. Such variables are already collected by the system for the purpose of control and monitoring. Given the TLSV, artificial neural networks are used to detect the presence of a leak and its location in the boiler. The proposed approach was successfully applied to tube leak detection and isolation in five subsystems of a utility boiler.

The electronic noise is a natural match for physiologically motivated chemical data analysis. Both the olfactory system and the electronic nose consist of an array of chemical sensing elements and a pattern recognition system. Physiologically motivated approaches to automated chemical analysis with electronic noses are discussed in this paper. Also, applications of electronic noses to environmental sensing, food processing, and medicine are referenced. The quantity and complexity of the data collected by sensor arrays can make conventional chemical analysis of data in an automated fashion difficult. One approach to odor or volatile compound identification is to build an array of sensors, where each sensor in the array is designed to respond to a specific chemical. With this approach, the number of unique sensors must be at least as great as the number of chemicals being monitored. It is both expensive and difficult to build highly selective chemical sensors. An alternative approach is to use sensors that have a broader response and rely on advanced information processing to discriminate between different chemicals. This latter approach was inspired by biological olfactory systems and is the approach incorporated in electronic noses to reduce the requirements on both the number and the selectivity of the sensors.

In this paper, we propose a new nonlinear adaptive inverse control via a unified model neural network. In order to overcome nonsystematic design and long training time in nonlinear adaptive inverse control, we propose the approximate transformable technique to obtain a Chebyshev Polynomials Based Unified Model (CPBUM) neural network for the feedforward/recurrent neural networks. It turns out that the proposed method can use less training time to get an inverse model. Finally, we apply this proposed method to control magnetic bearing system. The experimental results show that the proposed nonlinear adaptive inverse control architecture provides a greater flexibility and better performance in controlling magnetic bearing systems.

It is known that some types of variable bit rate (VBR) video traffic exhibit strong long term correlations and non- stationary behavior. Estimation of an accurate amount of bandwidth to support this traffic has been a challenge task using conventional algorithmic approaches. In this paper, we show that a radial basis function neural networks (RBFNN) is capable of learning the non-linear multi-dimensional mapping between different video traffic patterns, quality of service (QoS) requirements and the required bandwidth to support each call. In addition, RBFNN model adopts to new traffic scenarios and still produces accurate results. This approach bypass the modeling approach which requires detailed knowledge about the traffic statistical patterns. Our method employs 'on-line' measurements of the traffic count process over a monitoring period which is determined such that the error in estimating the bandwidth is minimized to less than 3% of the actual value. In order to simplify the design of the RBFNN, the input traffic is preprocessed through a lowpass filter in order to smooth all high frequency fluctuations. A large set of training data, representing different traffic patterns with different QoS requirements, was used to ensure that RBFNN can generalize and produce accurate results when confronted with new data. The reported results prove that the neurocomputing approach is effective in achieving more accurate results than other traditional methods, based upon mathematical or simulation analysis. This is primarily due to the fact that the unique learning and adaptive capabilities of NN enable them to extract and memorize rules from previous experience. Evidently, such unique capabilities poise NN to solve many of the problems encountered in the design of ATM networks.

This paper presents a fuzzy approach to the traditional Hopfield-Tank TSP model. It is shown that a fuzzy interpretation of the rows of the activation array reveals interesting features of the dynamics. For example, the initial 'centroid' nature of the network becomes obvious, and even more interesting is the emergence of a 'monotonic' phase that paves the way for the final, nearest-city, phase. This fuzzy approach sheds new light on the Hopfield-Tank model and exposes many previously unnoticed aspects of the network.

International stocks trading, currency and derivative contracts play an increasingly important role for many investors. Neural network is playing a dominant role in predicting the trends in stock markets and in currency speculation. In most economic applications, the success rate using neural networks is limited to 70 - 80%. By means of the new approach of GMDH (Group Method of Data Handling) neural network predictions can be improved further by 10 - 15%. It was observed in our study, that using GMDH for short, noisy or inaccurate data sample resulted in the best-simplified model. In the GMDH model accuracy of prediction is higher and the structure is simpler than that of the usual full physical model. As an example, prediction of the activity on the stock exchange in New York was considered. On the basis of observations in the period of Jan '95 to July '98, several variables of the stock market (S&P 500, Small Cap, Dow Jones, etc.) were predicted. A model portfolio using various stocks (Amgen, Merck, Office Depot, etc.) was built and its performance was evaluated based on neural network forecasting of the closing prices. Comparison of results was made with various neural network models such as Multilayer Perceptrons with Back Propagation, and the GMDH neural network. Variations of GMDH were studied and analysis of their performance is reported in the paper.

While the Hausdorff-Voronoi network has demonstrated promising capabilities for three-dimensional object recognition and classification, serious investigation and refinement of this network required something more than the ad hoc programs currently available. To fill this need an object-oriented mixed-language MATLAB^d and C++ software toolkit has been developed, together with a user-friendly graphical user interface.

The improved version of digital-analog CMOS implementation of an artificial neuron is discussed. This neuron is learnable to logical threshold functions, being functionally powerful and highly noise-stable. It is built on the basis of a previously suggested circuit consisting of synapses, (beta) -comparator and output amplifier. Every learnable synapse contains 5 minimum transistors and a capacitor for storing the results of the learning. It has been shown that higher non-linearity of the (beta) -comparator in the threshold zone can sharply increase the threshold of the realized functions and noise- stability of the neuron. To increase the minimum leap of voltage at the (beta) -comparator output in the threshold zone which is attainable during the teaching, it is suggested to use an output amplifier with threshold hysteresis. For this aim, the neuron has three output amplifiers with different thresholds. The output of the amplifier with the middle value of threshold is the output of the neuron; the outputs of the other two amplifiers are used during the teaching. The way is suggested of refreshing the voltages (found during the teaching) on the capacitors during the evaluation process. The results of SPICE simulation prove that the neuron is learnable to most complicated threshold functions of 10 and more variables and that it is capable to maintain the learned state for a long time. In the simulation, transistor modes MOSIS BSIM3v3.1 0.8 micrometer were used.

The early vision principle of redundancy reduction of 10⁸ sensor excitations is understandable from computer vision viewpoint toward sparse edge maps. It is only recently derived using a truly unsupervised learning paradigm of artificial neural networks (ANN). In fact, the biological vision, Hubel- Wiesel edge maps, is reproduced seeking the underlying independent components analyses (ICA) among 10² image samples by maximizing the ANN output entropy (partial)H(V)/(partial)[W] equals (partial)[W]/(partial)t. When a pair of newborn eyes or ears meet the bustling and hustling world without supervision, they seek ICA by comparing 2 sensory measurements (x₁(t), x₂(t))^T equalsV X(t). Assuming a linear and instantaneous mixture model of the external world X(t) equals [A] S(t), where both the mixing matrix ([A] equalsV [a₁, a₂] of ICA vectors and the source percentages (s₁(t), s₂(t))^T equalsV S(t) are unknown, we seek the independent sources <S(t) S^T(t)> approximately equals [I] where the approximated sign indicates that higher order statistics (HOS) may not be trivial. Without a teacher, the ANN weight matrix [W] equalsV [w₁, w₂] adjusts the outputs V(t) equals tanh([W]X(t)) approximately equals [W]X(t) until no desired outputs except the (Gaussian) 'garbage' (neither YES '1' nor NO '-1' but at linear may-be range 'origin 0') defined by Gaussian covariance <V(t) V(t)^T>_G equals [I] equals [W][A] <S(t) S^T(t)greater than [A]^T[W]^T. Thus, ANN obtains [W][A] approximately equals [I] without an explicit teacher, and discovers the internal knowledge representation [W], as the inverse of the external world matrix [A]^-1. To unify IC, PCA, ANN & HOS theories since 1991 (advanced by Jutten & Herault, Comon, Oja, Bell-Sejnowski, Amari-Cichocki, Cardoso), the LYAPONOV function L(v₁,...,v_n, w₁,...w_n,) equals E(v₁,...,v_n) - H(w₁,...w_n) is constructed as the HELMHOTZ free energy to prove both convergences of supervised energy E and unsupervised entropy H learning. Consequently, rather using the faithful but dumb computer: 'GARBAGE-IN, GARBAGE-OUT,' the smarter neurocomputer will be equipped with an unsupervised learning that extracts 'RAW INFO-IN, (until) GARBAGE-OUT' for sensory knowledge acquisition in enhancing Machine IQ. We must go beyond the LMS error energy, and apply HOS To ANN. We begin with the Auto- Regression (AR) which extrapolates from the past X(t) to the future u_i(t+1) equals w_i^TX(t) by varying the weight vector in minimizing LMS error energy E equals <[x(t+1) - u_i(t+1)]²> at the fixed point (partial)E/(partial)w_i equals 0 resulted in an exact Toplitz matrix inversion for a stationary covariance assumption. We generalize AR by a nonlinear output v_i(t+1) equals tanh(w_i^TX(t)) within E equals <[x(t+1) - v_i(t+1)]²>, and the gradient descent (partial)E/(partial)w_i equals - (partial)w_i/(partial)t. Further generalization is possible because of specific image/speech having a specific histogram whose gray scale statistics departs from that of Gaussian random variable and can be measured by the fourth order cumulant, Kurtosis K(v_i) equals <v_i⁴> - 3 <v_i²>² (K greater than or equal to 0 super-G for speeches, K less than or equal to 0 sub-G for images). Thus, the stationary value at (partial)K/(partial)w_i equals plus or minus 4 PTLw_i/(partial)t can de-mix unknown mixtures of noisy images/speeches without a teacher. This stationary statistics may be parallel implemented using the 'factorized pdf code: (rho) (v₁, v₂) equals (rho) (v₁) (rho) (v₂)' occurred at a maximal entropy algorithm improved by the natural gradient of Amari. Real world applications are given in Part II, (Wavelet Appl-VI, SPIE Proc. Vol. 3723) such as remote sensing subpixel composition, speech segmentation by means of ICA de-hyphenation, and cable TV bandwidth enhancement by simultaneously mixing sport and movie entertainment events.

The Anticipatory System (AS) formalism developed by Robert Rosen provides some insight into the problem of embedding intelligent behavior in machines. AS emulates the anticipatory behavior of biological systems. AS bases its behavior on its expectations about the near future and those expectations are modified as the system gains experience. The expectation is based on an internal model that is drawn from an appeal to physical reality. To be adaptive, the model must be able to update itself. To be practical, the model must run faster than real-time. The need for a physical model and the requirement that the model execute at extreme speeds, has held back the application of AS to practical problems. Two recent advances make it possible to consider the use of AS for practical intelligent sensors. First, advances in transducer technology make it possible to obtain previously unavailable data from which a model can be derived. For example, acoustic emissions (AE) can be fed into a Bayesian system identifier that enables the separation of a weak characterizing signal, such as the signature of pump cavitation precursors, from a strong masking signal, such as a pump vibration feature. The second advance is the development of extremely fast, but inexpensive, digital signal processing hardware on which it is possible to run an adaptive Bayesian-derived model faster than real-time. This paper reports the investigation of an AS using a model of cavitation based on hydrodynamic principles and Bayesian analysis of data from high-performance AE sensors.

The notion of perspective when supported in knowledge representation can allow the representation of multiple and varying points of view, some of which may even be inconsistent with one another. In an object-based knowledge representation methodology created and used by the authors, a perspective is defined by consolidating a number of objects and a number of those objects' associated attributes and method into a view. This view can help partition a knowledge domain into separate portions. A separate portion represents an individual's view of the knowledge domain. Representation of multiple and varying perspectives can add detail and context to the knowledge in a knowledge domain. The ability to create new perspectives may add to the existing knowledge as well as reveal paths to additional knowledge. A simple example is presented where perspectives are used to represent game playing strategies and levels of expertise in those strategies. Players' perspectives are adapted and changed to provide additional knowledge and insight into further game playing strategies. Results show improvement in the playing of the games. Additionally, a more complex problem for applying these techniques is introduced.

Despite the extensive research and the advent of several new information technologies in the last three decades, machine labeling of ground categories using remotely sensed data has not become a routine process. Considerable amount of human intervention is needed to achieve a level of acceptable labeling accuracy. A number of fundamental reasons may explain why machine labeling has not become automatic. In addition, there may be shortcomings in the methodology for labeling ground categories. The spatial information of a pixel, whether textural or contextual, relates a pixel to its surroundings. This information should be utilized to improve the performance of machine labeling of ground categories. Landsat-4 Thematic Mapper (TM) data taken in July 1982 over an area in the vicinity of Washington, D.C. are used in this study. On-line texture extraction by neural networks may not be the most efficient way to incorporate textural information into the labeling process. Texture features are pre-computed from co- occurrence matrices and then combined with a pixel's spectral and contextual information as the input to a neural network. The improvement in labeling accuracy with spatial information included is significant. The prospect of automatic generation of metadata consisting of ground categories, textural and contextual information is discussed.

When a serial offender strikes, it usually means that the investigation is unprecedented for that police agency. The volume of incoming leads and pieces of information in the case(s) can be overwhelming as evidenced by the thousands of leads gathered in the Ted Bundy Murders, Atlanta Child Murders, and the Green River Murders. Serial cases can be long term investigations in which the suspect remains unknown and continues to perpetrate crimes. With state and local murder investigative systems beginning to crop up, it will become important to manage that information in a timely and efficient way by developing computer programs to assist in that task. One vital function will be to compare violent crime cases from different jurisdictions so investigators can approach the investigation knowing that similar cases exist. CATCH (Computer Aided Tracking and Characterization of Homicides) is being developed to assist crime investigations by assessing likely characteristics of unknown offenders, by relating a specific crime case to other cases, and by providing a tool for clustering similar cases that may be attributed to the same offenders. CATCH is a collection of tools that assist the crime analyst in the investigation process by providing advanced data mining and visualization capabilities.These tools include clustering maps, query tools, geographic maps, timelines, etc. Each tool is designed to give the crime analyst a different view of the case data. The clustering tools in CATCH are based on artificial neural networks (ANNs). The ANNs learn to cluster similar cases from approximately 5000 murders and 3000 sexual assaults residing in a database. The clustering algorithm is applied to parameters describing modus operandi (MO), signature characteristics of the offenders, and other parameters describing the victim and offender. The proximity of cases within a two-dimensional representation of the clusters allows the analyst to identify similar or serial murders and sexual assaults.

A hybrid neural network approach is presented to estimate radio propagation characteristics and multiuser interference and to evaluate their combined impact on throughput, latency and information loss in third-generation (3G) wireless networks. The latter three performance parameters influence the quality of service (QoS) for multimedia services under consideration for 3G networks. These networks, based on a hierarchical architecture of overlaying macrocells on top of micro- and picocells, are planned to operate in mobile urban and indoor environments with service demands emanating from circuit-switched, packet-switched and satellite-based traffic sources. Candidate radio interfaces for these networks employ a form of wideband CDMA in 5-MHz and wider-bandwidth channels, with possible asynchronous operation of the mobile subscribers. The proposed neural network (NN) architecture allocates network resources to optimize QoS metrics. Parameters of the radio propagation channel are estimated, followed by control of an adaptive antenna array at the base station to minimize interference, and then joint multiuser detection is performed at the base station receiver. These adaptive processing stages are implemented as a sequence of NN techniques that provide their estimates as inputs to a final- stage Kohonen self-organizing feature map (SOFM). The SOFM optimizes the allocation of available network resources to satisfy QoS requirements for variable-rate voice, data and video services. As the first stage of the sequence, a modified feed-forward multilayer perceptron NN is trained on the pilot signals of the mobile subscribers to estimate the parameters of shadowing, multipath fading and delays on the uplinks. A recurrent NN (RNN) forms the second stage to control base stations' adaptive antenna arrays to minimize intra-cell interference. The third stage is based on a Hopfield NN (HNN), modified to detect multiple users on the uplink radio channels to mitigate multiaccess interference, control carrier-sense multiple-access (CSMA) protocols, and refine call handoff procedures. In the final stage, the Kohonen SOFM, operating in a hybrid continuous and discrete space, adaptively allocates the resources of antenna-based cell sectorization, activity monitoring, variable-rate coding, power control, handoff and caller admission to meet user demands for various multimedia services at minimum QoS levels. The performance of the NN cascade is evaluated through simulation of a candidate 3G wireless network using W-CDMA parameters in a small-cell environment. The simulated network consists of a representative number of cells. Mobile users with typical movement patterns are assumed. QoS requirements for different classes of multimedia services are considered. The proposed method is shown to provide relatively low probability of new call blocking and handoff dropping, while maintaining efficient use of the network's radio resources.

The objective of this study is to compare statistical and unsupervised neural network techniques for determination of correspondences between storm system regions extracted from sequences of satellite images. Analysis was applied to the International Satellite Cloud Climatology Project (ISCCP) low resolution D1 database for selected storm systems during the period April 5 - 9, 1989. Cloud top pressure was used to delineate regions of interest and cloud optical thickness combined with spatial location was used to track regions throughout a given time sequence. The ability of the k-nearest neighbor classifier and of self-organizing maps to determine correspondences between storm regions was assessed. The two techniques generally yielded similar associations between regions of interest throughout the time sequence. Differences in final tracking results between the two techniques occurred primarily as a result of differences in the collections of points from a region in a time step t₂ that corresponded to a region in an earlier time step t₁. The tracking results were also compared to the results obtained at the NASA Goddard Institute for Space Studies using sea level pressure data from the National Meteorological Center (NMC). For the storm systems investigated in this study, the storm tracks exhibited the same general tracking behavior with expected variations between cloud system storm centers and low sea level pressure centers.

A study of Parallel Self-Organizing Map (Parallel-SOM) is proposed to modify Self-Organizing Map for parallel computing environments. In this model, the conventional repeated learning procedure is modified to learn just once. The once learning manner is more similar to human learning and memorizing activities. During training, every connection between neurons of input/output layers is considered as an independent processor. In this way, all elements of every matrix are calculated simultaneously. This synchronization feature improves the weight updating sequence significantly. In this paper, the detail sequence of Parallel-SOM is demonstrated through the classification of coin for deeply understanding the properties of the proposed model. In conventional computing environment (one processor), Parallel- SOM can be implemented without the once learning and parallel weight updating features. As an application, its implementation for the classification of the meteorological radar images is also shown.

Evolving a single individual does not provide a solution when a set of individuals is needed to address different aspects of the problem. Genetic algorithms, however, are typically used to evolve a single individual. We have generalized previous work on modeling the immune system and developed a fitness method that allows a single run of a genetic algorithm to evolve a suite of individuals that compete with each other for fitness and yet cooperate to solve a problem. The method is a form of tournament fitness where individuals compete for fitness with other members of the population. The competitions vary throughout the run, resulting in a changing fitness environment. Experimental results and a preliminary model are described that verify the validity of the approach. We also present explanations for unexpected results from our previous work.

Traditional techniques for designing evolutionary algorithms rely on schema processing, minimizing expected losses, and emphasize certain genetic operators such as crossover. Unfortunately, these have failed to provide robust optimization performance. Recently, fitness distribution analysis has been proposed as an alternative tool for designing efficient evolutionary computations. This analysis has concentrated on obtaining very accurate expected improvement (EI) and probability of improvement (PI) statistics for specific mutation operators (using as many as 5000 Monte Carlo trials) on noiseless object functions. In practice, such extensive analysis might be computationally prohibitive and the objective functions might also be noisy. Experiments were designed here to determine the amount of sampling required to obtain useful estimates of the EI and PI both in the presence and absence of noise. Simulations indicate that useful statistics can be obtained in as few as 10 trials in the absence of noise. On noisy functions, however, the required number of trials increased as the 'signal to noise ratio' decreased.

This paper describes a genetic algorithm for designing materials. The genetic algorithm uses knowledge of the contributions that certain chemical fragments make to desired molecular properties. The technique proposes collections of fragments that, when combined into a molecule, are likely to possess specific properties. The proposed molecules can be assessed in a laboratory in order to confirm the degree to which property goals are met. In addition to proposing reasonable candidate materials, the genetic algorithm is both robust and flexible; it can be modified to include additional properties, to favor specific properties, or to enforce differing assumptions about the molecule structure of the materials considered. The genetic algorithm also possesses the potential to deliver candidates in a timely fashion.

Modern technical tasks often need the use of complex system models. In many complex cases the model parameters can be gained using neural networks, but these systems allow only a one-way simulation from the input values to the learned output values. If evaluation in the other direction is needed, these model allow no direct evaluation. This task can be solved using evolutionary algorithms, which are part of the computational intelligence. The term computational intelligence covers three special fields of the artificial intelligence, fuzzy logic, artificial neural networks and evolutionary algorithms. We will focus only on the topic of evolutionary algorithms and fuzzy logic. Evolutionary algorithms covers the fields of genetic algorithms, evolution strategies and evolutionary programming. These methods can be used to optimize technical problems. Evolutionary algorithms have certain advantages, if these problems have no mathematical properties, like steadiness or the possibility to obtain the derivatives. Fuzzy logic systems normally lack these properties. The use of a combination of evolutionary algorithms and fuzzy logic allow an evaluation of the learned simulation models in the direction form output to the input values. An example can be given from the field of screw rotor design.

An integral part of the Comprehensive Nuclear Test Ban Treaty monitoring is an international infrasonic monitoring network that is capable of detecting and verifying nuclear explosions. Reliable detection of such events must be made from data that may contain other sources of infrasonic phenomena. Infrasonic waves can also result from volcanic eruptions, mountain associated waves, auroral waves, earthquakes, meteors, avalanches, severe weather, quarry blasting, high-speed aircraft, gravity waves, and microbaroms. This paper shows that a feedforward multi-layer neural network discriminator, trained by backpropagation, is capable of distinguishing between two unique infrasonic events recorded from single station recordings with a relatively high degree of accuracy. The two types of infrasonic events used in this study are volcanic eruptions and a set of mountain associated waves recorded at Windless Bight, Antarctica. An important element for the successful classification of infrasonic events is the preprocessing techniques used to form a set of feature vectors that can be used to train and test the neural network. The preprocessing steps used in our analysis for the infrasonic data are similar to those techniques used in speech processing, specifically speech recognition. From the raw time-domain infrasonic data, a set of mel-frequency cepstral coefficients and their associated derivatives for each signal are used to form, a set of feature vectors. These feature vectors contain the pertinent characteristics of the data that can be used to classify the events of interest as opposed to using the raw data. A linear analysis was first performed on the feature vector space to determine the best combination of mel-frequency cepstral coefficients and derivatives. Then several simulations were run to distinguish between two different volcanic events, and mountain associated waves versus volcanic events, using their infrasonic characteristics.

Modern authentication systems require high-accuracy low complexity methods. High accuracy ensures secure access to sensitive data. Low computational requirements produce high transaction rates for large authentication populations. We propose a polynomial-based classification system that combines high-accuracy and low complexity using discriminative techniques. Traditionally polynomial classifiers have been difficult to use for authentication because of either low accuracy or problems associated with large training sets. We detail a new training method that solves these problems. The new method achieves high accuracy by implementing discriminative classification between in-class and out-of- class feature sets. A separable approach to the problem enables the method to be applied to large data sets. Storage is reduced by eliminating redundant correlations in the in- class and out-of-class sets. We also show several new techniques that can be applied to balance prior probabilities and facilitate low complexity retraining. We illustrate the method by applying it to the problem of speaker authentication using voice. We demonstrate the technique on a multisession speaker verification database collected over a one month period. Using a third order polynomial-based scheme, the new system gives less than one percent average equal error rate using only one minute of training data and less than five seconds of testing data per speaker.

Displacement cardiography techniques such as the ballistocardiogram and seismocardiogram use accelerometers to measure body motion caused by the beating heart. The radarcardiogram (RCG) measures this motion using highly sensitive radar developed at the Georgia Tech Research Institute. Combining the portability and non-invasiveness of radar along with neural network processing techniques opens a host of potential new applications including unknown person identification, stress measurement, and medical diagnosis. Correlation between displacement cardiography and the RCG will be discussed along with preliminary research using RCG data and a neural network to identify unknown persons. It was found that a neural network could accurately identify the RCG of an unknown individual out of a small pool of training data. In addition, the system was able to correctly reject individuals not within the training set.

Blind equalization for a nonminimum phase channel problem arises in digital communication. Here we develop a new method to achieve blind equalization problem for linear finite impulse response (FIR) systems, whether the systems are minimum phase or not. This new approach divides the problem into two parts. Firstly, it employs the characteristic of the linear system and take the tricepstrum method to estimate the original channel. Thus, nonminimum phase channel can be reconstructed and additive Gaussian noise will be restrained. Secondly, it utilizes the nonlinear characteristic of the neural network to establish an equalizer for the original channel. This is done by using the estimated channel as a reference system to train the neural network. The neural network can reduce the degree of model uncertainty and resist additive noise. Taking the Advantage of both linear and nonlinear systems, this scheme works well for both stationary and nonstationary cases.

This paper shows the results obtained in a system vision applied to plant reproduction by tissue culture using adaptive image segmentation and pattern matching algorithms, this analysis improves the number of tissue obtained and minimize errors, the image features of tissue are considered join to statistical analysis to determine the best match and results. Tests make on potato plants are used to present comparative results with original images processed with adaptive segmentation algorithm and non adaptive algorithms and pattern matching.

We consider the problem of matching domain-specific statistical structure to neural-network (NN) architecture. In past work we have considered this problem in the function approximation context; here we consider the pattern classification context. General Systems Methodology tools for finding problem-domain structure suffer exponential scaling of computation with respect to the number of variables considered. Therefore we introduce the use of Extended Dependency Analysis (EDA), which scales only polynomially in the number of variables, for the desired analysis. Based on EDA, we demonstrate a number of NN pre-structuring techniques applicable for building neural classifiers. An example is provided in which EDA results in significant dimension reduction of the input space, as well as capability for direct design of an NN classifier.

In this paper we present an innovative lossy plus lossless residual encoding scheme consisting of the following steps: (A) Dynamic pre-processing applied either to the original image in order to separate homogeneous parts of it; or to the histogram of the pixel values in order to generate three images each with the same size of the original one that superposed reconstruct exactly the source image. (B) Use of an efficient lossy compression scheme to pre-processed data in order to generate low bit rate images. (C) Definition of residuals by computing the differences between the lossy reconstructions and the pre-processed images. (D) Encode the residuals using an appropriate lossless technique. We applied this double scheme, with the two different pre-processing techniques, to some HST FITS images, obtaining from 1:4 to 1:6.4 lossless compression ratios.

In this paper we describe some of the most important types of neural networks applied in biomedical image processing. The networks described are variations of well-known architectures but are including image-relevant features in their structure. Convolutional neural networks, modified Hopfield networks, regularization networks and nonlinear principal component analysis neural networks are successfully applied in biomedical image classification, restoration and compression.

With the AC₄ algorithm proposed by Mohr and Henderson in 1986, a framework was proposed to solve the problem of matching between data and a semantic graph. This matching is usually associated with the notion of understanding. However, the very high combinatorial aspect of this problem makes very difficult to solve it by a computer. Moreover, there are very few bijective relations in practice. The high combinatorial aspect can be reduced with a local checking of the constraint satisfaction. With the AC₄ algorithm, this strategy can be used only when the matching is bijective. With the notion of FDCSP_BC this strategy was extended to non bijective relations. This case is encountered when we want to label over-segmented images. However, this extension is only adequate for a matching corresponding to surjective functions. The question of new extensions of this strategy to non surjective functions and non functional relations can be considered. In medical image analysis, the second case is often encountered when an unexpected object like a tumor appears. In that case, the data can not be mapped to the semantic graph, with a classical approach. In this paper we propose an extension of the FDCSP_BC to solve the constraint satisfaction problem for non functional relations.

This paper is concerned with the problem of separation of data, by a Neural based computer recognition system. To this end certain types of data which are 'tricky' are studied in order to see if they can be separated (i.e. classified) by a neural network or by a Kohonen based classifier. It is shown that there exist data which cannot simply be separated by a nearest distance classifier and yet can be treated well by a neural network, these correspond to the symmetry problem in images.

Pulse-coupled neural networks (PCNNs) have recently become fashionable for image processing. This paper discusses some of the advantages and disadvantages of PCNNs for performing image segmentation in the realm of medical diagnostics. PCNNs were tested with magnetic resonance imagery (MRI) of the brian and abdominal region and nuclear scintigraphic imagery of the lungs (V/Q scans). Our preliminary results show that PCNNs do well at contrast enhancement. They also do well at image segmentation when each segment is approximately uniform in intensity. However, there are limits to what PCNNs can do. For example, when intensity significantly varies across a single segment, that segment does not properly separate from other objects. Another problem with the PCNN is properly setting the various parameters so that a uniform response is achieved over a set of imagery. Sometimes, a set of parameters that properly segment objects in one image fail on a similar image.

The Pulsed Coupled Neural Network (PCNN) algorithm, when modified for use as an image processor, provides a unique method of multiplicative image decomposition (PCNN factorization). Because the factorization is ordered by levels of scene contrast, the first few factors contain the strong contrasts generally associated with shadows. The PCNN factorization effectively and automatically finds scene shadows. This is further developed here as a computationally effective shadow compensation algorithm with illustrative examples given, and is shown to be significantly more effective than histogram equalization. The advantage and disadvantages are discussed.

Present study concerns the problem of learning, pattern recognition and computational abilities of a homogeneous network composed from coupled bistable units. An efficient learning algorithm is developed. New possibilities for pattern recognition may be realized due to the developed technique that permits a reconstruction of a dynamical system using the distributions of its attractors. In both cases the updating procedure for the coupling matrix uses the minimization of least-mean-square errors between the applied and desired patterns.

Prediction and modeling in the case of non linear systems (or processes), especially of complex industrial processes are known being a class of involved problems. In this paper, we deal with the production yield prediction dilemma in VLSI manufacturing. An RBF neural networks based approach and its hardware implementation on a ZISC neural board have been presented. Experimental results comparing our approach with an expert have been reported and discussed.

One of the most important problems, for a machine control process is the system identification. To identify varying parameters which are dependent from other system's parameters (speed, voltage and currents, etc.), one must have an adaptive control system. Synchronous machines conventional vector control's implementation using PID controllers have been recently proposed presenting the best actual solution. It supposes an appropriated model of the plant. But real plant's parameters vary and the P.I.D. controller is not suitable because of the parameters variation and non-linearity introduced by the machine's physical structure. In this paper, we present an on-line dynamic adaptive neural based vector control system identifying the motor's parameters of a synchronous machine. We present and discuss a DSP based real- time implementation of our adaptive neuro-controller. Simulation and experimental results validating our approach have been reported.

Associative memory networks such as Radial Basis Functions, Neurofuzzy and Fuzzy Logic used for modelling nonlinear processes suffer from the curse of dimensionality (COD), in that as the input dimension increases the parameterization, computation cost, training data requirements, etc. increase exponentially. Here a new algorithm is introduced for the construction of a Delaunay input space partitioned optimal piecewise locally linear models to overcome the COD as well as generate locally linear models directly amenable to linear control and estimation algorithms. The training of the model is configured as a new mixture of experts network with a new fast decision rule derived using convex set theory. A very fast simulated reannealing (VFSR) algorithm is utilized to search a global optimal solution of the Delaunay input space partition. A benchmark non-linear time series is used to demonstrate the new approach.

This paper shows the procedure to derive a qualitative description of a single-input-single-output system. From the system numerical input-output data a fuzzy model is constructed using the mountain method for structure identification and the back-propagation method for parameter identification. The procedure is used in the implementation of a fuzzy logic based digital thermometer using a TMS320C50 DSP board. This thermometer is able to predict a patient's body temperature more quickly than that of a conventional thermometer with very good accuracy.

The stable properties of equilibrium point are the most important properties of associative memory neural network, which include local stability, domain of absorb and convergent rate. Because associative memory neural network has a lot of equilibrium points, and different equilibrium point has different stable properties, so it is an interesting and important research problem to reveal the quantitative relation between equilibrium point and its stable properties. In the paper, the following three results are proved: (1) the equilibrium point X* is locally exponentially stable if and only if the real parts of all eigenvalues of derivative (matrix) of network at X* are less than zero; (2) the fastest convergent speed of trajectory of equilibrium point X* is equal to the maximum of real parts of all eigenvalues of derivative (matrix) of network at X*; (3) the domain of absorb of equilibrium point X* is determined by the change rate of output function in the local neighborhood of X*, and its estimate can be obtained by the computation of a local characteristic function of X* defined in the paper. From all these results, people can see that the stable properties of a given equilibrium point of associative memory neural network are uniquely determined by the equilibrium point itself. So as a matter of fact, equilibrium point can be thought as an information point containing the important information about its stability.

Some data sets contain outlying data values which can degrade the quality of the clustering results obtained using standard techniques such as the fuzzy c-means algorithm. This note gives an extended family of fuzzy c-means type models, and attempts to empirically identify those members of the family which are least influenced by the presence of outliers. The form of the extended family of clustering criteria suggests an alternating optimization approach, is feasible, and specific algorithms for implementing the optimization of the models are stated. The implemented approach is then tested using various artificial data sets.

Artificial Neural Networks play a pivotal role in the branch of Artificial Intelligence. They can be trained efficiently for a variety of tasks using different methods, of which the Back Propagation method is one among them. The paper studies the choosing of various design parameters of a neural network for the Back Propagation method. The study shows that when these parameters are properly assigned, the training task of the net is greatly simplified. The character recognition problem has been chosen as a test case for this study. A sample space of different handwritten characters of the English alphabet was gathered. A Neural net is finally designed taking many the design aspects into consideration and trained for different styles of writing. Experimental results are reported and discussed. It has been found that an appropriate choice of the design parameters of the neural net for the Back Propagation method reduces the training time and improves the performance of the net.

This paper details efforts made to evolve neural networks for playing checkers. In particular, multilayer perceptrons were used as evaluation functions to compare the worth of alternative boards. The weights of these neural networks were evolved in a co-evolutionary manner, which networks competing only against other extant networks in the population. No external 'expert system' was used for comparison or evaluation. Feedback to the networks was limited to an overall point score based on the outcome of 10 games at each generation. No attempt was made to give credit to moves in isolation or to prescribe useful features beyond the possible inclusion of piece differential. When played in 100 games against rated human opponents, the final rating for the best evolved network was 1750, placing it as a Class B player. This level of performance is competitive with many humans.

Almost 40% of the total horticultural produce exported from Mexico the USA is tomato, and quality is fundamental for maintaining the market. Many fruits packed at the green-mature stage do not mature towards a red color as they were harvested before achieving its physiological maturity. Tomato gassed for advancing maturation does not respond on those fruits, and repacking is necessary at terminal markets, causing losses to the producer. Tomato spectral signatures are different on each maturity stage and tomato size was poorly correlated against peak wavelengths. A back-propagation neural network was used to predict tomato maturity using reflectance ratios as inputs. Higher success rates were achieved on tomato maturity stage recognition with neural networks than with discriminant analysis.