Evolutionary programming is a method for simulating evolution that has been investigated for almost 40 years. When originally introduced, the available computing equipment was quite slow and difficult to use as measured by current standards. This paper provides a series of experiments that follow the framework of the original approach from the early 1960s, brought up to date with current computing machinery. A brief review of evolutionary programming and its relationship to other methods of evolutionary computation, specifically genetic algorithms and evolution strategies, is also offered.

A new nonlinear feature extraction method called the maximum representation and discrimination feature (MRDF) method is presented for extraction of features from input image data. It implements transformations similar to the Sigma-Pi neural network. However, the weights of the MRDF are obtained in closed form, and offer advantages compared to nonlinear neural network implementations. The features extracted are useful for both object discrimination (classification) and object representation (pose estimation). We show its use in estimating the class and pose of images of real objects and rendered solid CAD models of machine parts from single views using a feature-space trajectory (FST) neural network classifier. We show more accurate classification and pose estimation results than are achieved by standard principal component analysis (PCA) and Fukunaga-Koontz (FK) feature extraction methods.

The 'mixture-of-experts (MOE)' is a popular architecture for function approximation. In the standard architecture, each expert is gated via a softmax function, and its domain of application is not very localized. This paper summarizes several recent results showing the advantages of using localized gating instead. These include a natural framework for model selection/adaptation by growing and shrinking the number of experts, modeling of non-stationary environments, improving the generalization performance and obtaining confidence intervals of network outputs. These results substantially increase the scope and power of MOE networks. Several simulation results are presented to support the theoretical arguments.

The Partial Least-Squares Regression (PLSR) approach to statistical calibration model development has been formulated using an inverse model. The inverse model PLSR algorithm is implemented using the Partial Least Squares neural NETwork (PLSNET) architecture. Generalized neural network learning rules derived from a statistical representation error criterion are presented. These learning rules will accommodate a quadratic optimization criterion, providing the linear solution. Optimization functions which grow less than quadratically can also be used to provide a robust solution when the empirical data contains impulsive and colored noise and outliers. The robust optimization criterion also accounts for the higher-order statistics associated with the input data. The inverse model PLSNET learning rules require fewer mathematical operations per weight update than the forward model robust PLSNET algorithms, resulting in faster convergence in many cases.

A novel genetic operator called cloning is introduced and tested in different applications of genetic algorithms. Essentially, the cloning monotonically increases the lengths of the chromosomes during the evolution. It is argued that, under these circumstances, the cloning operator can accommodate a multiresolution search strategy, where the search starts at coarser scales and is subsequently mapped to finer scales upon achieving some in-scale performance criteria. Although the practical implementation of cloning is application dependent, a few general requirements are stated. In the remainder of the paper, different implementations of the cloning operator are introduced and employed in distinct applications, namely, function optimization, object support reconstruction from the support of its autocorrelation and the shortest path problem in planar graphs. The first two cases present typical multiresolution approaches to search problems and their results show consistent improvements in convergence speed with respect to classical genetic algorithms. In the last problem, a cloning operator is incorporated in an evolutionary algorithm that builds a set of valid paths in a planar graph. It is demonstrated that cloning can enhance the ability of a genetic algorithm to explore the search space efficiently in some applications.

The objective of this research is to automate the classification of the temporal behavior of storm cloud systems based on measurements derived from consecutive satellite images. The motivation behind this study is to develop improved descriptions of cloud dynamics which can be used in general circulation models for prediction of global climate change. Analysis was applied to the International Satellite Cloud Climatology Project (ISCCP) low resolution cloud top pressure database for the first six days in April 1989. A total of 296 midlatitude storm cloud components were tracked between consecutive 3-hour time frames. For each pair of components, temporal correspondence events were classified as either (1) direct, (2) merge, (3) split, or (4) reject. The reject class, which was used primarily to categorize pairs of unrelated systems, included storm cloud system dissipation and creation as well. Statistical, neural network, and evolutionary techniques were developed for finding solutions to the storm cloud correspondence problem. Evolutionary techniques applied to the problem consisted of (1) a constraint-handling hybrid evolutionary technique and (2) a genetic local search algorithm. The results demonstrate the potential of evolutionary techniques to yield meteorologically feasible solutions, given appropriate constraints, to the two- frame storm tracking problem.

Recent developments in pulse coupled neural networks techniques provide an opportunity to extend the toolbox available for exploring new approaches to biosensor applications. This paper presents a demonstration of properties and limitations of new computational intelligence (CI) techniques as shown by and related to an application. New pulse coupled neural networks (PCNN) techniques are supplemented by combination with wavelet analysis and fine- tuned by radial basis functions. This toolbox is exercised to demonstrate its properties and limitations as related to the development of biosensor applications. The approach selected employs abstractions of biological models of peripheral vision and relates them to analysis of time series generated by biosensors such as chemosensors or motion detectors. Detection of targets (rare or interesting events) is facilitated by PCNN multi-scale image factorization. Interpretation of the resulting image set is aided by contrast enhancement and by segmentation using standard PCNNs. Wavelet coefficients provide supplemental discrimination and lead to characteristic sets of numbers useful in identifying image factors of interest. To complete the transition from acquisition of a complex, noisy image to recognition of targets of interest, radial basis function (RBF) analysis is appended. This five- step process (odor image generation, image factoring, PCNN analysis, wavelet analysis and RBF interpretation) was recently suggested, but is expanded and fully implemented here for the first time. This paper explores the properties and limitations of this approach for simulation of biosensors using small, incomplete sets of real-world data. The relationship between selection of appropriate design parameters and the need for supplementing the available data by simulation is investigated. Evolutionary computation is employed off line to explore and evaluate the possibilities and limitations. Sensor fault detection and RBF training vector generation are addressed. Results are analyzed to provide recommendations for further experimentation and collection of needed additional data without extraneous effort. This methodology is recommended for use in real-world applications where experimental data is difficult, expensive or time consuming to obtain.

In this paper, it is shown that Evolution Programs can be used to search the weight space for Bayesian training of a Neural Network. Bayesian Analysis is an integration problem (as opposed to an optimization problem) over weight space. The first application of the Bayesian method primarily focused on using a Gaussian approximation of the posterior distribution in an area of high probability in the weight space instead of using formal integration. More recently, training a neural network in a Bayesian fashion has been accomplished by searching weight space for areas of high probability density which obviates the need for the Gaussian assumption. In particular, a hybrid Monte-Carlo method was used to search weight space in a logical manner to obtain an arbitrarily close approximation of the integration involved in a Bayesian analysis. Genetic Algorithms have been used in the past to determine the weights in an ANN, and (with some slight modifications) are ideally suited for searching the weight space to approximate the Bayesian integration. In this respect, the Bayesian framework provides a simple and elegant way to apply Evolution Programs to the ANN training problem. While this paper concentrates on using ANNs as classifiers, the generalization to regression problems is straightforward.

Intelligent data fitting is often used in seismic data processing to reconstruct large three-dimensional data volumes based on relatively small amounts of 2D seismic and/or well- log data. This technique is useful for such applications because of the many different layers and lithologies in stratum. However, if such data fitting is performed over two different layers in the stratum, it often results in incorrect profiles. A more accurate way to perform such fitting is to first perform segmentation of the data into known classes, and then fit the data inside of each segmented class. In seismic data processing, interval velocity is one of the most important factors in rock type identification, i.e. lithology classification. To find a region with low velocity generally indicates a lithology of high porosity. Such a region is much more likely to contain oil and/or gas. Velocity data is more difficult to obtain than seismic waveform, which is obtained by 3D seismic prospecting. To obtain a vertical trace of interval velocity data requires extensive processing. Generally, such a velocity volume is obtained by means of a data fitting technique. This paper presents a general intelligent data fitting approach to reconstruct interval velocity volumes. This technique uses a process we term SEGFIT-segmentation followed by fitting. In SEGFIT, we use a fuzzy connected segmentation technique for the segmentation, which we term (lambda) -connected followed by a specialized fitting method which maintains the (lambda) -connectedness of the fitted data in each segmented class. In this paper we show that a real velocity volume can be obtained using this technique.

In this paper, an ultra-fast smart vision system-on-a-chip design is proposed to provide effective solutions for real time machine vision applications by taking advantages of recent advances in integrated sensing/processing designs, electronic neural networks, advanced microprocessors and sub- micron VLSI technology. The smart vision system mimics what is inherent in biological vision systems. It is programmable to perform vision processing in all levels such as image acquisition, image fusion, image analysis, and scene interpretation. A system-on-a-chip implementation of this smart vision system is shown to be feasible by integrating the whole system into a 3-cm by 3-cm chip design in a 0.18- micrometer CMOS technology. The system achieves one tea- operation-per-second computing power that is a two order-of- magnitude increase over the state-of-the-art microcomputer and DSP chips. Its high performance is due to massively parallel computing structures, high data throughput rates, fast learning capabilities, and advanced VLSI system-on-a-chip implementation. This highly integrated smart vision system can be used for various NASA scientific missions and other military, industrial or commercial vision applications.

Increase in the levels of carbon dioxide and other greenhouse gases over the next half-century may result in an increase in global mean temperature. The recent discoveries of possible advance of arctic tree line into the tundra and earlier greening of northern vegetation provide additional warnings that global warming may indeed be occurring. On the Earth surface, land cover and its changes affect the coupling between the biosphere and the atmosphere, and control many important Earth system processes. Satellite remote sensing provides long-term, repeated coverage over extended area and is the essential data source for monitoring climate changes. An Advanced Very-High Resolution Radiometer (AVHRR) Pathfinder dataset from 1987, in 1 degree latitude-longitude resolution, is used in this study. Two reflective channels, two thermal channels, and Normalized Difference Vegetation Index are the input parameters. In conjunction with a global vegetation ground truth, a multi-layer neural network is trained and used for global vegetation characterization. As the same type of vegetation may appear very differently over different parts of the Earth at any given time, global classification is more difficult than local classification. It is shown that a multitemporal approach, in which data from multiple dates are used, may improve the accuracy.

Sleep apnea is characterized by frequent prolonged interruptions of breathing during sleep. This syndrome causes severe sleep disorders and is often responsible for development of other diseases such as heart problems, high blood pressure and daytime fatigue, etc. After diagnosis, sleep apnea is often successfully treated by applying positive air pressure (CPAP) to the mouth and nose. Although effective, the (CPAP) equipment takes up a lot of space and the connected mask causes a lot of inconvenience for the patients. This raised interest in developing new techniques for treatment of sleep apnea syndrome. Several studies have indicated that electrical stimulation of the hypoglossal nerve and muscle in the tongue may be a useful method for treating patients with severe sleep apnea. In order to be able to successfully prevent the occurrence of apnea it is necessary to have some technique for early and fast on-line detection or prediction of the apnea events. This paper suggests using measurements of respiratory airflow (mouth temperature). The signal processing for this task includes the use of a short window FFT technique and uses an artificial back propagation neural net to model or predict the occurrence of apneas. The results show that early detection of respiratory interruption is possible and that the delay time for this is small.

A physician's decision support system consists of three components: (1) a comprehensive patient record and medical knowledge database, (2) information infrastructure for data storage, transfer, and (3) an analytical inference engine, accompanied by business operation database. Medical knowledge database provides the guideline for the selection of powerful clinical features or tests to be observed so that an accurate diagnosis as well as effective treatment can be quickly reached. With a tremendous amount of information stored in multiple data centers, it takes an effective information infrastructure to provide streamlined flow of information to the physician in a timely fashion. A real-time analytical inference engine mimics the physician's reasoning process. However due to incomplete, imperfect data and medical knowledge, a realistic output from this engine will be a list of options with associated confidence level, expected risk, so that the physician can make a well-informed final decision. Physicians are challenged to pursue the objective of ensuring an acceptable quality of care in an economically restrained environment. Therefore, business operation data have to be factored into the calculation of overall loss. Follow-up of diagnosis and treatment provides retrospective assessment of the accuracy and effectiveness of the existing inference engine.

Recently research on hybrid solutions for fault diagnosis problems. In this paper we propose a neural network implementation of diagnosis systems. Based on certainty factor modeling and Hebbian learning, the network features with local learning ability, quasi-causal representation of diagnosis process, and easy mechanism of knowledge manipulation.

This paper present the comparation and performance on no adaptive image segmentation techniques using illumination and adaptive image segmentation techniques. Results obtained on indoor plant reproduction by tissue culture, show the improve in time process, simplify the image segmentation process, experimental results are presented using common techniques in image processing and illumination, contrasted with adaptive image segmentation.

Most currently available vehicle identification systems use techniques such as R.F., microwave, or infrared to help identifying the vehicle. Transponders are usually installed in the vehicle in order to transmit the corresponding information to the sensory system. It is considered expensive to install a transponder in each vehicle and the malfunction of the transponder will result in the failure of the vehicle identification system. In this study, novel hybrid approach is proposed for automatic vehicle license plate recognition. A system prototype is built which can be used independently or cooperating with current vehicle identification system in identifying a vehicle. The prototype consists of four major modules including the module for license plate region identification, the module for character extraction from the license plate, the module for character recognition, and the module for the SimNet neuro-fuzzy system. To test the performance of the proposed system, three hundred and eighty vehicle image samples are taken by a digital camera. The license plate recognition success rate of the prototype is approximately 91% while the character recognition success rate of the prototype is approximately 97%.

Neural networks are a popular representation for inducing single-step predictors for chaotic times series. For complex time series it is often the case that a large number of hidden units must be used to reliably acquire appropriate predictors. This paper describes an evolutionary method that evolves a class of dynamic systems with a form similar to neural networks but requiring fewer computational units. Results for experiments on two popular chaotic times series are described and the current method's performance is shown to compare favorably with using larger neural networks.

Accurate electrical load forecasting is a necessary part of resource management for power generating companies. The better the hourly load forecast, the more closely the power generating assets of the company can be configured to minimize the cost. Automation of this process is a profitable goal and neural networks have shown promising results in achieving this goal. The most often used neural network to solve the electric load forecasting problem is the back-propagation neural network architecture. Although the performance of the back- propagation neural network architecture has been encouraging, it is worth noting that it suffers from the slow convergence problem and the difficulty of interpreting the answers that the architecture provides. A neural network architecture that does not suffer from the above mentioned drawbacks is the Fuzzy ARTMAP neural network, developed by Carpenter, Grossberg, and their colleagues at Boston University. In this work we applied the Fuzzy ARTMAP neural network to the electric load forecasting problem. We performed numerous experiments to forecast the electrical load. The experiments showed that the Fuzzy ARTMAP architecture yields as accurate electrical load forecasts as a back-propagation neural network with training time a small fraction of the training time required by the back-propagation neural network.

The 'Elton John' problem envisions candles blowing in the wind to motivate reasoning about prediction methods. The 'maximize smoothness but maintain trend' method has desirable parsimony when both data quantity and understanding are limited.

A neural network model has been designed to predict certain product properties which can be combined with a multivariate controller to improve the current operation of the crude fraction section of the refinery. The model used to predict the 95% naphtha cutoff point was trained using input vectors made up of 33 field inputs, which in turn were collected from actual refinery data. The model was successful in predicting the 95% cut off with a maximum error of 1.06 degree F in the training phase. In the operational phase the maximum error was 4.63 degree F. The paper also discusses issues related to the development of the specific neural network architecture and learning methodology used for this application.

This paper starts with an overview of a classical PID controller design. An account of how Neural Networks may be incorporated to provide control is such a setup. The example used in this paper is the problem of controlling a High Frequency Acoustics Platform (HFAP) in-flight. The HFAP is towed by a ship and flown in the water behind the ship to acquire acoustic data reflected from the sea floor. The stability of such a platform is of prime importance to the accuracy of data collected. Using fight data from previous runs of the platform, a Neural Network is trained. The trained network is then used to predict the behavior of the platform. These predictions may then be directly translated to control signals minimizing the platform's spatial deviations. In this paper results form the trained Neural Network on predicting the behavior of the platform are displayed. Network prediction results illustrating the ability of the network to operate with partial input are displayed. Displaying these results in contrast with conventional controller results given the same input parameters emphasizes the importance of such a feature. Finally the use of different network architectures and the cost of using these network, in terms of computing power is investigated.

This paper presents the design of a systematic failure detection and isolation system that uses the concept of failure sensitive variables (FSV) and artificial neural networks (ANN). The proposed approach was applied to tube leak detection in a utility boiler system. Results of the experimental testing are presented in the paper.

A system for control of two dimensional end milling operations is presented in this paper. In order to achieve high quality products and high production rates in 2-D end milling, it is important to minimize the tracking error and contouring error, and regulate the cutting force while maximizing the feedrate. Given the complexity of the cutting operations, it is difficult to determine what rules are needed to satisfactory accomplished the above. The Self-Organizing Fuzzy logic Controller is capable of accomplishing force regulation because it does not require prior knowledge of the system and is able to generate and modify the control rules. A hybrid controller is also presented in this paper, which is composed of an ordinary Fuzzy Logic Controller and a Self-Organizing Fuzzy Logic Controller. This new controller, called a Self- Paced Fuzzy Controller, will be used to minimize the contouring error. The benefit of this hybrid system is that initial rules can be provided to the system that can be updated as the controller learns about the system's performance. The control system is simulated and the results are presented.

The paper presents an intelligent architecture based a feedforward neural network with fuzzy Petri nets for modeling product quality in a CNC machining center. It discusses how the proposed architecture can be used for modeling, monitoring and control a product quality specification such as surface roughness. The surface roughness represents the output quality specification manufactured by a CNC machining center as a result of a milling process. The neural network approach employed the selected input parameters which defined by the machine operator via the CNC code. The fuzzy Petri nets approach utilized the exact input milling parameters, such as spindle speed, feed rate, tool diameter and coolant (off/on), which can be obtained via the machine or sensors system. An aim of the proposed architecture is to model the demanded quality of surface roughness as high, medium or low.

A simulated automobile factory paint shop is used as a testbed for exploring the emulation of human decision-making behavior. A discrete-events simulation of the paint shop as a collection of interacting Java actors is described. An evolutionary cognitive architecture is under development for building software actors to emulate humans in simulations of human- dominated complex systems. In this paper, the cognitive architecture is extended by implementing a persistent population of trial behaviors with an incremental fitness valuation update strategy, and by allowing a group of cognitive actors to share information. A proof-of-principle demonstration is presented.

Computer application technology is increasing in all sectors of industry and commerce. For this reason it is considered essential that today's students are conversant with the concepts and benefits of computer technology. This fuzzy logic simulation project is aimed at providing a student environment where these concepts can be easily and effectively taught, and to provide practical solutions to real industrial problems. This paper describes the development of such a system. A fuzzy logic system is modeled, analyzed and tested in an industrial environment. This involves the application of the software tool MATLAB to generate a mathematical model of a system and compare fuzzy logic algorithms with that of traditional control system design. Once the mathematical model was tested with a graphical simulation package, the system can then be implemented in a real-time environment. This allows any necessary changes to be made without interfering with the output of the system. The aim is to use such a system to broaden the available media for teaching and understanding fuzzy logic and as a consequence allow students to learn not only fuzzy logic, but control systems in a new and exciting manner, previously unavailable to them. Omron Electronics, Landis & Staefa, The Mathworks Inc., Auckland Institute of Technology and the University of Auckland are presently supporting this project.

We propose to design and to evaluate an on-board intelligent health assessment tool for rotorcraft machines, which is capable of detecting, identifying, and accommodating expected system degradations and unanticipated catastrophic failures in rotorcraft machines under an adverse operating environment. A fuzzy-based neural network paradigm with an on-line learning algorithm is developed to perform expert advising for the ground-based maintenance crew. A hierarchical fault diagnosis architecture is advocated to fulfill the time-critical and on- board needs in different levels of structural integrity over a global operating envelope. The research objective is to experimentally demonstrate the feasibility and flexibility of the proposed health monitoring procedure through numerical simulations of bearing faults in USAF MH-53J PAVE LOW helicopter transmissions. The proposed fault detection, identification and accommodation architecture is applicable to various generic rotorcraft machines. The proposed system will greatly reduce the operational and developmental costs and serve as an essential component in an autonomous control system.

This paper analyses the function of the mutation and recombination operators in evolution strategies (ES). By using three quadratic systems, this paper studies the suitable percentage of Gaussian mutation. Experimental results and Comparisons demonstrate that Gaussian mutation acts more important role in the evolutionary processing but the its percentage should not be too high. This paper tries to give a general idea to solve the practical implement of ES, especially for large sized problems.

A very large number of works concerning the area of Artificial Neural Networks (ANN) deal with implementation of these models, especially as digital or analogue CMOS integrated circuits. All of the presented implementations of A.N.N. have been supposed to be working in ideal conditions but real applications will be subject to global perturbations. Unfortunately, very few papers analyze the behavior of analogue implementation of neural network with such kind of perturbations. Since 1994, we have investigated the behavior modeling of electronic A.N.N. with global perturbation conditions. We have scrutinized the behavior analysis of a CMOS analogue implementation of synchronous Boltzmann Machine model with both ambient temperature and electrical perturbations (supply voltage) perturbation. In this paper we present, using our model, the analysis of these global perturbations effects on learning capability in a CMOS analogue implementation of synchronous Boltzmann Machine Simulation and experimental results have been exposed validating our concepts.

This paper presents a Kohonen-like mapping that eliminates or reduces four limitations of the Kohonen maps. The described network is invariant to scale, very resistant to 'automatic selection of feature dimensions,' results in strictly ordered clusters of ascending/descending magnitude, and may allow a greater amount of information to be gleaned from high dimensional data sets. The network treats each input component separately but each map is influenced via inter-map connections. Unfortunately, processing time increases combinatorially as the number of input components and number of neurons per component increases. As a demonstration, membership functions are constructed for a four variable data set with minimal parameter setting, the most crucial being the number of classes per input component.

With the Crew Assistant Military Aircraft (CAMA) a knowledge- based cockpit assistant system for future military transport aircraft is developed and tested to enhance situation awareness. Human-centered automation was the central principal for the development of CAMA, an approach to achieve advanced man-machine interaction, mainly by enhancing situation awareness. The CAMA-module Pilot Intent and Error Recognition (PIER) evaluates the pilot's activities and mission events in order to interpret and understand the pilot's actions in the context of the flight situation. Expected crew actions based on the flight plan are compared with the actual behavior shown by the crew. If discrepancies are detected the PIER module tries to figure out, whether the deviation was caused erroneously or by a sensible intent. By monitoring pilot actions as well as the mission context, the system is able to compare the pilot's action with a set of behavioral hypotheses. In case of an intentional deviation from the flight plan, the module checks, whether the behavior matches to the given set of behavior patterns of the pilot. Intent recognition can increase man-machine synergy by anticipating a need for assistance pertinent to the pilot's intent without having a pilot request. The interpretation of all possible situations with respect to intent recognition in terms of a reasoning process is based on a set of decision rules. To cope with the need of inferencing under uncertainty a fuzzy-logic approach is used. A weakness of the fuzzy-logic approach lies in the possibly ill-defined boundaries of the fuzzy sets. Self-Organizing Maps (SOM) as introduced and elaborated on by T. Kohonen are applied to improve the fuzzy set data and rule base complying with observed pilot behavior. Hierarchical cluster analysis is used to locate clusters of similar patterns in the maps. As introduced by Pedrycz, every feature is evaluated using fuzzy sets for each designated cluster. This approach allows to generate fuzzy sets and rules by use of a user-friendly and easily adjustable environment of development tools for data interpretation.

It is now possible to construct some sort of an intelligent knowledge base by using neural networks. But this is, so called, a static knowledge base which cannot evolve itself for finding optimized answers which suit an initial aim if the initial answers do not satisfy the aim; referring to (1), (2), (3), (4), (5). However, this is the system which can behave like our human beings within a limited condition. That is to say, this can behave in a good manner only if the knowledge bases which have been implemented beforehand have enough knowledge to treat a matter to be solved. But this is not the case of usual matter, because we commonly have new unknown things to overcome in order to satisfy an aim. Now we do not have any artificial knowledge or thoughts to treat them. Because this system lacks of treating mechanism of taking one data base after another autonomously. So, in order to overcome this problem, mechanism of evolving a new knowledge base by itself for treating them until a proper answer can be found has been studied.

The hidden layer neurons in a multi-layered feed-forward neural network serve a critical role. From one perspective, the hidden layer neurons establish (linear) decision boundaries in the feature space. These linear decision boundaries are then combined by succeeding layers leading to convex-open and thereafter arbitrarily shaped decision boundaries. In this paper we show that the use of unidirectional Gaussian lateral connections from a hidden layer neuron to an adjacent hidden layer leads to a much richer class of decision boundaries. In particular the proposed class of networks has the advantage of sigmoidal feed-forward networks (global characteristics) but with the added flexibility of being able to represent local structure. An algorithm to train the proposed network is presented and its training and validation performance shown using a simple classification problem.

In this paper the authors present the results of a research investigation on feature selection methods for neural network classifiers. As problems presented to computers for analysis become more complex and data dimensionality grows in size, traditional methods of feature extraction are being taxed beyond the limits of their usefulness. New methods of feature selection show promise in the laboratory, but need to be proven with real-world solutions. The purpose of this research is to compare the performance of newly proposed methods of selecting features on three challenging problems using non- artificial data. A feature saliency technique, and several variants of genetic algorithms, and random feature selection are compared and contrasted.

Designing a fuzzy system involves defining membership functions and constructing rules. Carrying out these two steps manually often results in a poorly performing system. Genetic Algorithms (GAs) has proved to be a useful tool for designing optimal fuzzy controller. In order to increase the efficiency and effectiveness of their application, parallel GAs (PAGs), evolving synchronously several populations with different balances between exploration and exploitation, have been implemented using a SIMD machine (APE100/Quadrics). The parameters to be identified are coded in such a way that the algorithm implicitly provides a compact fuzzy controller, by finding only necessary rules and removing useless inputs from them. Early results, working on a fuzzy controller implementing the wall-following task for a real vehicle as a test case, provided better fitness values in less generations with respect to previous experiments made using a sequential implementation of GAs.

The asymptotic expansions of the solutions within the framework of been found phase-space geometry for oscillatory neural model are constructed. The oscillatory neural networks consist of non-identical neurons are examined. The phenomenon of mutual neurons' synchronization has been analyzed.

Based on Amaris mathematical formulation of the self- organization of synaptic efficacies and neural response fields under the influence of external stimuli we show that if the map is a contraction, then the system has a unique equilibrium which is globally asymptotically stable; consequently the system acts as a stable encoder of external input stimuli. The system converges to a fixed point representing the steady- state of the neural activity which has as an upper bound the superposition of the spatial integrals of the weight function between neighboring neurons and the stimulus autocorrelation function.

This paper presents a comparison study of the rough set approach and Kohonen's vector quantization with the self- organizing algorithm. The main idea behind this research is the fact that neither the rough set nor Kohonen's neural network approaches require a prior knowledge of data distribution. The paradigms are compared in terms of their methods for the calculation of an accuracy of approximation and classification, reduction of non-significant attributes, minimal subset of attributes, and the uncertainty associated with the decision making process.

High connectivity of artificial neural network chip- embodiments combined with currently emerging 3-dimensionally stacked multichip modules for real-time applications of target classification require a scrutiny for low power technology insertion. Conventional CMOS high power consumption limits the allowable density of synapse/neuron elements. However Silicon-On-Insulator (SOI) technology has the potential for successful implementation of high density neural network because of the following unique features: (a) Operating voltage is reduced 3-fold from 5 to 1.5 volts, reducing power requirements by 9-fold; (b) Reduced substrate offers reduced capacitance and power and an increased speed; and, (c) Latch-up phenomenon is eliminated. Here we describe two practical winner/loser-take-all (W/LTA) circuits fabricated with 0.25 micrometer fully depleted SOI technology that are useful for neural networks and as compared to other such circuits offer considerable advantage of speed and performance. SPICE circuit simulations show that up to 9-bit resolution can be obtained between a winner and a loser input and with two cascaded circuits. Final characterization tests prove that constructing circuit elements from SOI technology would allow us to build large size neural networks for practical applications.

The application of artificial neural networks to the channel assignment problem for cellular code-division multiple access (CDMA) cellular networks has previously been investigated. CDMA takes advantage of voice activity and spatial isolation because its capacity is only interference limited, unlike time-division multiple access (TDMA) and frequency-division multiple access (FDMA) where capacities are bandwidth-limited. Any reduction in interference in CDMA translates linearly into increased capacity. To satisfy the high demands for new services and improved connectivity for mobile communications, microcellular and picocellular systems are being introduced. For these systems, there is a need to develop robust and efficient management procedures for the allocation of power and spectrum to maximize radio capacity. Topology-conserving mappings play an important role in the biological processing of sensory inputs. The same principles underlying Kohonen's self-organizing feature maps (SOFMs) are applied to the adaptive control of radio resources to minimize interference, hence, maximize capacity in direct-sequence (DS) CDMA networks. The approach based on SOFMs is applied to some published examples of both theoretical and empirical models of DS/CDMA microcellular networks in metropolitan areas. The results of the approach for these examples are informally compared to the performance of algorithms, based on Hopfield- Tank neural networks and on genetic algorithms, for the channel assignment problem.

Pulse Coupled Neural Networks (PCNN) are biologically inspired neural networks, mainly based on studies of the visual cortex of small mammals. The PCNN is very well suited as a pre- processor for image processing, particularly in connection with object isolation, edge detection and segmentation. Several implementations of PCNN on von Neumann computers, as well as on special parallel processing hardware devices (e.g. SIMD), exist. However, these implementations are not as flexible as required for many applications. Here we present an implementation in Field Programmable Gate Arrays (FPGA) together with a performance analysis. The FPGA hardware implementation may be considered a platform for further, extended implementations and easily expanded into various applications. The latter may include advanced on-line image analysis with close to real-time performance.

The study of function approximation is motivated by the human limitation and inability to register and manipulate with exact precision the behavior variations of the physical nature of a phenomenon. These variations are referred to as signals or signal functions. Many real world problem can be formulated as function approximation problems and from the viewpoint of artificial neural networks these can be seen as the problem of searching for a mapping that establishes a relationship from an input space to an output space through a process of network learning. Several paradigms of artificial neural networks (ANN) exist. Here we will be investigated a comparative of the ANN study of RBF with radial Polynomial Power of Sigmoids (PPS) in function approximation problems. Radial PPS are functions generated by linear combination of powers of sigmoids functions. The main objective of this paper is to show the advantages of the use of the radial PPS functions in relationship traditional RBF, through adaptive training and ridge regression techniques.

A Radial Basis Function (RBF) Iterative Construction Algorithm (RICA) is presented that autonomously determines the size of the network architecture needed to perform classification on a given data set. The algorithm uses a combination of a Gaussian goodness-of-fit measure and Mahalanobis distance clustering to calculate the number of hidden nodes needed and to estimate the parameters of the hidden node basis functions. An iterative minimum squared error reduction method is used to optimize the output layer weights. RICA is compared to several neural network algorithms, including a fixed architecture multilayer perceptron (MLP), a fixed architecture RBF, and an adaptive architecture MLP, using optical character recognition and infrared image data.

A novel approach to the problem of finding the weights of an adaptive array is presented. In cellular and satellite mobile communications systems, desired as well as interfering signals are mobile. Therefore, a fast tracking system is needed to constantly estimate the directions of those users and then adapt the radiation pattern of the antenna to direct multiple beams to desired users and nulls to sources of interference. In this paper, the computation of the optimum weights is approached as a mapping problem which can be modeled using a suitable artificial neural network trained with input output pairs. A study of a three-layer Radial Basis Function Neural Network (RBFNN) is conducted. RBFNN were used due to their ability for data interpolation in higher dimensions. The network weights are modified using the normalized cumulative delta rule. The performance of this network is compared to the Wiener solution. It was found that networks implementing these functions were successful in tracking mobile users as they move across the antenna's field of view.

We discuss possible new hardware and software techniques for handling very large databases such as image archives. In particular, we investigate how high capacity solid-state 'disks' could be used to speed the database processing by algorithms that require considerable memory space. One such algorithm, for example, called the RAM neural network, or weightless neural network, needs a number of large lookup tables to perform most efficiently. The solid state disks could provide fast storage both for the algorithm and the data. We also briefly discuss development of an algorithm to cluster images of similar objects. This algorithm could also benefit from a large cache of fast memory storage.

The paper describes the physio-associative temporal sensors integration algorithm which is motivated by the observed function of the thalamus and utilizes signals theory mathematics to model how a human efficiently perceives information in the environment. The algorithm is consistent with that of an aircraft pilot; namely, to estimate, filter, and predict sensed afferent signals and produce efferent controls under dynamic flight conditions. Dynamic sensor integration under uncertainty requires feature selection which can be formulated as an associative-learning problem in which sensed states are represented as current situational beliefs, and the information either excites or inhibits long-term memory associations. The objective of the learner/observer is to (1) abstract salient signals from the environment, (2) integrate the signal for real-time beliefs, and (3) compare beliefs to learned associations. Biologically, the paper models these processes from the biological systems of the eye, thalamus, and association-cortex; respectively. By selecting the optimal set of mutually non-exclusive sensors and comparing the integrated signal to learned associations, the physio-associative temporal algorithm maximizes the identification of targets in a simulated dynamic flight situation.

In this paper, we present a neural network based method that allows the optimal selection of a data fusion policy. We build dynamically the internal layer of a functional link network (FLN), we add to the classical FLN, a pruning algorithm, that allows to find the optimal architecture of the FLN and to define an optimal fusion policy. In order to use the FLN as a universal fusion operator, the functional expansion performed by its internal layer includes fusion operators. As the FLN minimize the mean square error (MSE) during the learning step, an optimal fusion policy is reached in the sense of the MSE. Some academic simulations validate our approach.

This paper is motivated by the work of Laden and Keefe, and addresses the topic of pitch class recognition. A neural net with one hidden layer is trained to recognize all thirty-six major, minor and diminished chords, which can be built over a chromatic scale that starts and ends in C. A harmonic complex representation is chosen for the chords. Each tone is represented by five partial harmonics. A three note chord consists of fifteen partials. Our net is trained with the Error Backpropagation algorithm. The effect of different learning rates and hidden layer sizes are studied. Experiments with a technique known as Bold Driver to speedup the learning are also conducted. Following the existing work, we examine the recognition of incomplete patterns, that is, chords with some harmonics missing. The recognition performance of the system could be significantly improved by adding noise in the training session, and using voting networks. Also the number of epochs needed to recognize all chords could be drastically reduced.

Events aboard the space shuttle such as crew movement, crew exercise, thruster firings, etc., disrupt the microgravity environment required for many on-board experiments. Automatic detection of these events would allow astronauts to minimize their impact on experiments. Hence, using Space Acceleration Measurement System (SAMS) data collected on the USMP-3 mission, a knowledge base is generated to aid in the detection of disruptive events aboard the USMP-4 mission. Input patterns containing power spectral density (PSD) information of SAMS data are used to train an Adaptive Resonance Theory 2-A (ART2- A) neural network. The ART2-A neural network has been chosen because it has the ability to automatically add clusters as new input patterns are presented. The weight vectors of the ART2-A are used as the knowledge base. Using characteristic frequencies and acceleration magnitudes determined by Principal Investigator Microgravity Services (PIMS), each weight vector is assigned a label or name representing a set of events. The labeled knowledge base is then tested by presenting input patterns created from data collected during an exercise event.

Although there is a large body of works on conventional target tracking techniques that are based primarily on Kalman filtering and probabilistic data association, there are very few practical techniques that can be shown to perform well under a high cluttered tracking environment. This is due to the difficulty of the combined target detection and measurement to track association problem. Furthermore, conventional techniques usually make some simplifying assumptions that are difficult to realized in practice, e.g. the clutter density is uniform, measurement noise is stationary, the target track is well defined, etc. Another weakness of the conventional techniques is that even if we have some special knowledge about target attributes, it is not easy to incorporate this knowledge into the tracking problem. This paper first presents an analysis of the target tracking problem using fuzzy logic theory. Subsequently, a number of fuzzy propositions that a fuzzy tracker can use to implement a data association algorithm are formulated. Finally, a fuzzy tracker is implemented based on the fuzzy association rules and Kalman filtering and its performance is compared against the performance of a standard PDA filter.

In this paper the use of a Genetic Algorithm to solve a constrained vehicle routing problem is explored. The problem is two-dimensional with obstacles represented as ellipses of uncertainty surrounding each obstacle point. A route is defined as a series of points through which the vehicle sequentially travels from the starting point to the ending point. The physical constraints of total route length and maximum turn angle are included and appear in the fitness function. In order to be valid, a route must go from start to finish without violating any constraint. The effects that different mutation rates and population sizes have on the algorithm's computation speed and ability to find a high quality route are also explored. Finally, possible applications of this algorithm to the problem of route planning for cruise missiles are discussed.

In this paper, we propose a neural network model with a faster learning speed and a good approximate capability in the function approximation for solving worst-case identification of nonlinear systems H_(infinity ) problems. Specifically, via the approximate transformable technique, we develop a Chebyshev Polynomials Based unified model neural network for solving the worst-case identification of nonlinear systems H_(infinity ) problems. Based on this approximate transformable technique, the relationship between the single-layered neural network and multi-layered perceptron neural network is derived. It is shown that the Chebyshev Polynomials Based unified model neural network can be represented as a functional link network that is based on Chebyshev polynomials. We also derive a new learning algorithm such that the infinity norm of the transfer function from the input to the output is under a prescribed level. It turns out that the Chebyshev Polynomials Based unified model neural network not only has the same capability of universal approximator, but also has a faster learning speed than multi-layered perceptron or the recurrent neural network in the deterministic worst-case identification of nonlinear systems H_(infinity ) problems.

This paper introduces a signal sorting method based on a self- organized probabilistic neural network (PNN) and mainly discusses its application for sorting signal in radar. A PNN sorting system for radar signal using many phase is described. The feasibility of this method is verified after lots of emulated experiments.

This paper first give a comprehensive discussion about the concept of artificial neural network its research methods and the relations with information processing. On the basis of such a discussion, we expound the mathematical similarity of artificial neural network and information processing. Then, the paper presents a new method of image recognition based on invariant features and neural network by using image Zernike transform. The method not only has the invariant properties for rotation, shift and scale of image object, but also has good fault tolerance and robustness. Meanwhile, it is also compared with statistical classifier and invariant moments recognition method.

In this paper, we propose complex Chebyshev Polynomial Based unified model neural network for the approximation of complex- valued function. Based on this approximate transformable technique, we have derived the relationship between the single-layered neural network and multi-layered perceptron neural network. It is shown that the complex Chebyshev Polynomial Based unified model neural network can be represented as a functional link network that are based on Chebyshev polynomial. We also derived a new learning algorithm for the proposed network. It turns out that the complex Chebyshev Polynomial Based unified model neural network not only has the same capability of universal approximator, but also has faster learning speed than conventional complex feedforward/recurrent neural network.

The U.S. Navy has been requested to provide insightful responses to questions regarding low and high resolution target discrimination and target classification capabilities for short and medium range ballistic missiles (SRBM/MRBM). Specific targets studied for this paper include the solid booster and the associated attitude control system (ACS) liquid divert thruster systems. Discriminants selected include booster and ACS separation debris, as well as fuel vent phenomena. Debris and vent cloud containment and elimination through Gaussian suppression techniques have been implemented for low resolution assessment for target detection and tracking. Target gradient edge intensities were extracted for aimpoint selection and will be added to the pattern referencing library database at NSWC. The results of this study indicate an increasing requirement for advanced image processing on the focal plane array of a generic LEAP (light exo-atmospheric projectile) type kill kinetic vehicle (KKV) in order to implement effective target and aimpoint detection/tracking correlation matching routines.

The resource allocating network (RAN) provides a simple and powerful method for on-line modeling with incremental growth in model complexity. However, the network growing algorithm is susceptible to outliers in the output domain. Pruning techniques subsequently proposed for RAN, though satisfactory for dealing with outliers in the input domain, are incapable of removing units grown in response to outliers in the output domain. The addition of a coarse scale unit in response to an output outlier results in a much larger network where units are wasted to negate the effect of the spurious unit. The resulting network generalizes poorly. In this paper, we discuss the problems associated with RAN in the presence of outliers, and provide a modified learning algorithm which recognizes and prunes units associated with spurious data. We also present a strategy to modify the remaining units, once a unit is pruned.

Auditory perception neurons also called inner hair cells (IHC) transform the mechanical movements of the basilar membrane into electrical impulses. The impulse coding of the neurons is the main information carrier in the auditory process and is the basis for improvements of cochlea implants as well as for low rate, high quality speech processing and compression. This paper shows how to compute the speech signal from the neural firing based on the analysis of the interspike interval histogram. This new approach solves problems which other standard analysis methods do no solve sufficiently well.

Search strategy is an important component of any system that uses autonomous agents to detect and neutralize mines. We describe a simple and efficient search strategy derived from research on the adaptive spatial behavior of animals. Electromagnetic sensor data are processed to obtain a discrete spatial target distribution. The target distribution is used as input for a dynamic diffusion process. The diffusion surface is used by the demining agent to optimize its spatial moves through a hill climbing technique. The agent chooses to move to the position with the highest diffusion surface value. If the same diffusion surface is available to all agents, the system can be scaled to guide an indefinite number of independent, non-interfering agents.

This paper describes a method for the evaluation of pavement condition through artificial neural networks using the MLP backpropagation technique. Two of the most used procedures for detecting the pavement conditions were applied: the 'overall severity index' and the 'irregularity index.' Tests with the model demonstrated that the simulation with the neural network gives better results than the procedures recommended by the highway officials. This network may also be applied for the construction of a graphic computer environment.

A neural network model that may give phase synchronization of activity patterns in ensembles with different types of complexity is described and simulated. This model is used to study the dynamic behavior of coupled phase oscillators with cosine interaction between them. Different schemes of the network architecture such as mean-field and nearest-neighbor interaction, symmetric and asymmetric types of coupling among elements are provided to deal with the processes of synchronization patterns formation. Dynamic pattern formation related to neural oscillations, patterns of phase an anti- phase synchronization of activities of elements within clusters remote from one another as well as clustering of different attractors are studied.

Rough set theory, introduced by Pawlak in the early 1980s, in a mathematical tool to deal with vagueness and uncertainty. In contract, for centuries, uncertainty was measured in terms of probability theory. In this paper a novel method based on the Bayesian approach is proposed to generate the rough set decisions. The results of this approach may be summarized as the following: (1) The classification accuracy of a concept can be calculated as a prior probability of the class. (2) The accuracy of approximation of each atomic event equals the posterior probability of the atomic event. The posterior probability can be calculated using the lower and the upper approximations of the event. These accuracy measures can then be used to derive the final decision. (3) Normalized class conditional probabilities can be used to determine the significance of attributes. In addition, a minimal (reduced) subset, which ensures a satisfactory quality of approximation, can be calculated as a product of the accuracy of approximation of each event and the frequency of the event in the original set. The reduced set, however, does not play any role in the decision making process if the proposed probabilistic approach is utilized.

In this paper, a new 3D reconstruction approach for 3D object recognition in neuro-vision system is presented. First, a phase based stereo matching using Hopfield neural network approach is presented. The stereo matching problems are treated in frequency domain by using local phase. Instead of matching feature or texture of images, the stereo matching process is performed by using local phases of left image and right image in stereo image pair. By using the windowed Fourier transform, the windowed Fourier phases can be calculated. Through the variable window Gabor filter, the local phases of image can also be obtained. The Hopfield neural network is adopted to implement the stereo matching process. A suitable architecture of neural network is established, so that the computation can be implemented efficiently in parallel. A suitable matching function is created by using the local phase property. The energy function for neural network is constructed with satisfying some necessary constraints. The stereo matching process then is carried to find the minimum energy corresponding to the solution of the problem. Second, a 3D object reconstruction neural network is constructed by using BP neural network. So the 3D configuration and shape can be reconstructed by this neural network. With multiple neural networks the 3D reconstruction processes can be performed in parallel. The examples for both synthetic and real images are shown in the experiment, and good results are obtained.

In this paper, a new three dimensional recognition method for intelligent assembly system is presented. In this method neural network technology is used to provide new methodologies for solving difficult computational problems in three dimensional recognition processes. The method can be divided into two parts. In the first part, phase based stereo matching techniques are used to find the correspondence between left and right image into stereo image pair. The Hopfield neural network is adopted to implement the stereo matching process. A suitable architecture of neural network is established, so that the computation can be implemented efficiently in parallel. A three dimensional object reconstruction neural network is constructed by using BP neural network. With the results of stereo matching, the 3D configuration and shape can be reconstructed. In the second part, the feature vector of 3D object is constructed by using 3D moment and its invariant. With the results obtained in first parts, ART2 neural network is adopted for neural network classifier. With the ART2 neural network classifier, the 3D objects can be recognized and classified. The method id tested with both synthetic and real parts in intelligent assembly system. Good results are obtained. It is proved through the experiments and actual applications that the method presented in this paper is correct and reliable. It is very suitable for intelligent assembly system.

In this paper, we propose to incorporate both spatial and frequency models of HVS into wavelet transform image coding. The process of wavelet transform decomposition, which splits the spatial frequency domain to several octave bands by dilation and translation of a single basic wavelet, is similar to that of frequency model HVS. Moreover, according to spatial model of HVS, some compact physical features like contours and regions are with highly visually significant to human vision system. Based on the spatial model, we apply fuzzy logic theory to detect visual significant edge points and based on these edge points to construct a Visual Perception Sensitive Map (VPSM) for wavelet coefficient thresholding scheme. Only the visual significant coefficients are retained and the rest are discard. This approach can achieve a high image compression ratio while minimizing the visual quality distortion of the reconstructed image. In addition, we develop an adaptive quantization scheme for the wavelet coefficients at each of the subbands. This quantization scheme is developed based on the HVS frequency model to minimize the visual errors caused by the quantization. In our image compression system, both the frequency and spatial aspects of HVS to the image have been taken into consideration. We preserve the highly visual perceptive wavelet coefficients and minimize the visual distortion of coefficients in each of the decomposed band. As a result, a high compression ratio with low visual distortion coder is obtained.

A common concern of neural network models has been the problem of relating the function of complex systems of neurons to what is known of individual neurons, their interconnections and offsets. In this paper, we propose a new model of neural networks that can control and produce the offset patterns of the input layer, the hidden layer, and the output layer neurons. It consists of the input layer for the signal patterns, the hidden layer for the offset patterns production, and memory part between the hidden layer and the output layer. The output of neurons is calculated by the offsets control parameter R_ofj. The input layer calculates the input patterns to be learned so that the proposed neural network can control and produce the offset patterns, and sends the results to the next layer. The hidden layer produces the offset patterns after receiving the pattern information from the input layer, and it sends the output information of the hidden layer to the memory part. The memory part stores the learned output patterns of the hidden layer after comparing it with the input pattern, and sends the stored information to the output layer after the entire learning. Simulation results show that the proposed neural network can produce the offset patterns and it can be efficiently applied in the logic circuit design and pattern classification.

This paper summarizes a new interpretation of the Hopfield- Tank model as it applies to the Planar Traveling Salesman Problem (TSP). We demonstrate that the network solves the TSP in a 'centroidal' manner, that is, it provides tours that are similar to those obtained by the traditional centroid algorithm. The traditional centroid algorithm computes the center of mass of the cities and then orders the cities by the corresponding central angles. This algorithm gives excellent results on near-circular city configurations, and abysmal results on near-linear city configurations. We demonstrate that for up to 30 randomly generated cities the centroid results are very competitive with well known heuristics, such as the nearest city and 2-opt, but after about 40 cities the centroid algorithm produces poor results in comparison. We claim that this effect is inherent to the Hopfield-Tank model and explains why such networks do not scale up.

In the analysis of molecular evolution, it is very frequent to consider M sequences at a time, where M greater than 2. The simultaneous study of the relationships among M sequences is a large and difficult problem. This paper presents a new approach to multiple protein sequence comparison based on Genetic Algorithms, (G.A.). In particular, it is described an algorithm for finding the alignment of three protein sequences; besides, it can be easily changed for finding the alignment of more than three sequences or for other types of sequences. The G.A. was originally developed for only two sequences comparison [Morato96].