Neural networks, fuzzy systems, and evolutionary computation - the technologies featured in this conference - are often referred to collectively as soft computing. Like fuzzy logic this has an unfortunate overtone, but we are stuck with the term. I argue that the intuition that governed this grouping of technologies is that they are all human mimetic. Then I offer an overview of how to construct an artifact that, like you, uses those approaches to exercise judgement and innovate out of unforeseen problems.

We recently introduced a class of highly nonlinear associative memories called morphological associative memories (MAMs). Notable features of autoassociative morphological memories (AMMs) include optimal absolute storage capacity and one-step convergence. The fixed points can be characterized exactly in terms of the original patterns. Unfortunately, AMM fixed points include a large number of spurious memories. In this paper, we use a combination of a basic AMM model and the kernel method in order to eliminate most of the spurious memories while leaving other AMM properties intact. Furthermore, our new AMM model is more tolerant to noise than a basic AMM model and less dependent on kernel selection than the original kernel method.

A Pseudo-Outer Product based Fuzzy Neural Network using the Yager Rule of Inference called the POP-Yager FNN is proposed in this paper. The proposed POP-Yager FNN training consists of two phases: the fuzzy membership derivation phase using the Modified Learning Vector Quantization (MLVQ) method; and the rule identification phase using the novel one-pass LazyPOP learning algorithm. The proposed two-phase learning process effectively constructs the membership functions and identifies the fuzzy rules. Extensive experimental results based on the classification performance of the POP-Yager FNN using the Anderson's Iris data are presented for discussion. Results show that the POP-Yager FNN possesses excellent recall and generalization abilities.

A nonlinear dynamical brain theory with holographic undertones is presented. It leads to corticonic networks that can imitate important functional properties of the cortex. These include the handling of dynamic stimuli, autonomous learning and memory formation with no apparent cross-talk, and possibly vast storage-capacity.

Image analysis is an important requirement of many artificial intelligence systems. Though great effort has been devoted to inventing efficient algorithms for image analysis, there is still much work to be done. It is natural to turn to mammalian vision systems for guidance because they are the best known performers of visual tasks. The pulse- coupled neural network (PCNN) model of the cat visual cortex has proven to have interesting properties for image processing. This article describes the PCNN application to the processing of images of heterogeneous materials; specifically PCNN is applied to image denoising and image segmentation. Our results show that PCNNs do well at segmentation if we perform image smoothing prior to segmentation. We use PCNN for obth smoothing and segmentation. Combining smoothing and segmentation enable us to eliminate PCNN sensitivity to the setting of the various PCNN parameters whose optimal selection can be difficult and can vary even for the same problem. This approach makes image processing based on PCNN more automatic in our application and also results in better segmentation.

Support Vector Machines (SVMs) have become popular due to their accuracy in classifying sparse data sets. Their computational time can be virtually independent of the size of the feature vector. SVMs have been shown to out perform other learning machines on many data sets. In this paper, we use SVMs to detect a car in a lane of traffic. Digital pictures of various driving situations are used. The results from the SVM algorithm are compared to results from a standard neural network approach.

We have been experimenting with evolutionary approaches to create artifical neural networks that can play checkers at a level that is competitive with human experts. 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 coevolutionary manner, with 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 the piece differential. Initial results indicated that the best-evolved neural network earned a rating of 1750, placing it as a Class B player. This level of performance is competitive with many humans. More recent results have generated networks with ratings in the 1900s, in Class A, one level below expert as accepted by the American Checkers Foundation.

We consider the problem of pixel-by-pixel classification of a multi- spectral image using supervised learning. Conventional spuervised classification techniques such as maximum likelihood classification and less conventional ones s uch as neural networks, typically base such classifications solely on the spectral components of each pixel. It is easy to see why: the color of a pixel provides a nice, bounded, fixed dimensional space in which these classifiers work well. It is often the case however, that spectral information alone is not sufficient to correctly classify a pixel. Maybe spatial neighborhood information is required as well. Or maybe the raw spectral components do not themselves make for easy classification, but some arithmetic combination of them would. In either of these cases we have the problem of selecting suitable spatial, spectral or spatio-spectral features that allow the classifier to do its job well. The number of all possible such features is extremely large. How can we select a suitable subset? We have developed GENIE, a hybrid learning system that combines a genetic algorithm that searches a space of image processing operations for a set that can produce suitable feature planes, and a more conventional classifier which uses those feature planes to output a final classification. In this paper we show that the use of a hybrid GA provides significant advantages over using either a GA alone or more conventional classification methods alone. We present results using high-resolution IKONOS data, looking for regions of burned forest and for roads.

Decision tress have long been popular in classification as they use simple and easy-to-understand tests at each node. Most variants of decision trees test a single attribute at a node, leading to axis- parallel trees, where the test results in a hyperplane which is parallel to one of the dimensions in the attribute space. These trees can be rather large and inaccurate in cases where the concept to be learned is best approximated by oblique hyperplanes. In such cases, it may be more appropriate to use an oblique decision tree, where the decision at each node is a linear combination of the attributes. Oblique decision trees have not gained wide popularity in part due to the complexity of constructing good oblique splits and the tendency of existing splitting algorithms to get stuck in local minima. Several alternatives have been proposed to handle these problems including randomization in conjunction wiht deterministic hill-climbing and the use of simulated annealing. In this paper, we use evolutionary algorithms (EAs) to determine the split. EAs are well suited for this problem because of their global search properties, their tolerance to noisy fitness evaluations, and their scalability to large dimensional search spaces. We demonstrate our technique on a synthetic data set, and then we apply it to a practical problem from astronomy, namely, the classification of galaxies with a bent-double morphology. In addition, we describe our experiences with several split evaluation criteria. Our results suggest that, in some cases, the evolutionary approach is faster and more accurate than existing oblique decision tree algorithms. However, for our astronomical data, the accuracy is not significantly different than the axis-parallel trees.

In this work, the results for beam splitters and antireflection coatings by using a simple genetic algorithm are presented. The results obtained show the robustness of this technique when applied to complex search spaces.

Recent innovations in real-time machine vision, distributed computing, software architectures, high-speed communication, and mobile robotic systems are expanding the available technology for intelligent system development. These technologies allow the realization of intelligent systems that provide the capabilities for a user to experience events from remote locations and to interact with that location using an array of robotic devices. In this paper we describe research being done in the UCSD CVRR that will lead to the realization of a powerful and integrated traffic-incident detection, monitoring, and recovery system. Sensor clusters utilizing both rectilinear and omni-directional cameras will automate information gathering about the incident and provide a real- time televiewing interface to emergency response crews. Ultimately, this system will have a direct impact on reducing incident related highway congestion by improving the quality of information to which emergency personnel have access.

An improved fuzzy mesh analysis technique is applied to enhance the calculation efficiency of solving the soliton propagation equation. Propagation error due to the splitting of nonlinear soliton propagation equation can be avoided by the improved technique. The adaptive mesh control is applied to adjust the mesh size with the slope of the soliton profile such that the allocation of sampling points can be more effective. Therefore, the calculation accuracy as well as the efficiency of the fuzzy mesh analysis technique can be enhanced significantly.

State of the art fuzzy-logic based control is mainly implemented using electronic hardware or computer software. This requires interpretation of fuzzy logic concepts such as membership functions and fuzzy based rules, all of which have been thoroughly studied. However, the 2-D light-speed abilities of optical processing enables direct implementation of dual-input fuzzy logic inference engines. The optical equivalent of the membership function is generated in a straightforward manner and the same applies to rule tables and combination rules. Diffractive optical elements allow these optical inference engines to be compact in size and high on efficiency. This is done by binary optics and phase-only elements. Using the 2-D work-plane of optics, the ability of simple control over the wavelength and the polarization of light and the properties of diffractive elements, such an engine can deal with higher order data and lead the way to fast and dynamic fuzzy inferencing.

Complex designs in electronics are often assembled from the components, which are sufficiently large design units by themselves. Sometimes designers are looking for already manufactured components; in other situations, they are looking for design solutions to be incorporated into larger designs. Designers need to be sure that the selected components have the expected functional and physical characteristics. This paper describes the simulation-based methodology for verification of properties of real and virtual components in the process of large system design.

In this paper, we discuss a novel method, for multispectral and hyperspectral imagery data reduction. This method is based on singularity representation and integrates a rotational invariant visual object extraction and understanding technique based on the application of differential mapping to image processing. This new compression method applies Arnold's Differential Mapping Singularities Theory in the context of three-dimensional (3D) terrain and objects projection onto the two-dimensional (2D) image plane. It takes advantage of the fact that terrain features (particularly edges and singular points) and can be interpreted in terms of mapping singularities, which can be described by simple polynomials. We discuss the relationship between traditional approaches, including spatial and spectral decorrelation, and differential mapping singularities theory, or Catastrophe Theory (CT), in the context of multispectral image understanding and data reduction. CT maps 3D surfaces with exact results to construct a multispectral image-compression algorithm based on a finite set of singularities. This approach permits the rigorous mathematical description of a full set of singularities that describes the edges and other specific points of objects. The edges and specific points (degenerate critical points) are the products of mapping smooth 3D surfaces, which can be described by a simple set of polynomials that is suitable for image compression and automatic target recognition. The spectral signature for each extracted object was refined, and its dynamic diapason was compared with traditional methods.

The progress in soft computing and soft communication (SC2) is reviewed in detail including video/imaging compression, communications, processing, sensing and networking, based on 8 B0PS distributed hardware, allowing for full video frame evaluation, in real time.

Computational intelligence techniques have been successfully applied for solving control problems in modern networking architectures such as ATM and the Internet. The introduction of active networks offers a high level of flexibility in customizing the network infrastructure and introducing new functionality. There is a clear need for revisiting both the applicability of computational intelligence techniques in this new networking environment, as well as the provisions of active networking technology that computational intelligence techniques can exploit for improved operation. We elaborate on the characteristics of these technologies, their synergy and report on our study with applying computational intelligence techniques for improved routing on a novel active network resource management architecture.

Given a set of training data and a feature extraction tool, fuzzy membership functions are created using regression analysis on the extracted features. These membership functions are then used to classify a signal into one of two basic classes (namely, threat or non-threat). Alternatively, the dat can be classified into M groups, as desired. For this paper, the training data form a set of modeled infrared intensities for subpixel objects, of the types expected for a prototypical ballistic missile defense engagement scenario. The feature extraction took used is a form of local discriminant bases, as described by Coifman and Saito⁴. The top N features (typically two to four) are then piped pairwise through a regression tool to determine if any statistically significant trends occur. If a trend is discovered, then a membership function is created for the relationship; otherwise, membership functions are created for each feature independently. An example of each is given. Results indicate great flexibility in managing misclassification of targets (Leakage) versus classifying a non-target as a target (False Alarms), depending on the choice of membership functions. Results for using seven extracted features on performance data show < 1% Leakage corresponding to 13% False Alarms.

We propose a new classification algorithm to detect and classify targets of interest. It is based on an advanced brand of analytic geometry of manifolds, called theory of catastrophes. Physical Optics Corporation's (POC) scalable 3D model representation provides automatic and real-time analysis of a discrete frame of a sensed 2D imagery of terrain, urban, and target features. It then transforms this frame of discrete different-perspective 2D views of a target into a 3D continuous model called a pictogram. The unique local stereopsis feature of this modeling is the surprising ability to locally obtain a 3D pictogram from a single monoscopic photograph. The proposed 3D modeling, combined with more standard change detection algorithms and 3D terrain feature models, will constitute a novel classification algorithm and a new type of digital battlefield imagery for Imaging Systems.

A considerable amount of research has addressed the methods and objectives of model combination. Very little attention has been given to the question of how to obtain a good collection of models for combination. Here a rationale for inducive inference of multiple models of time series is developed in terms of algorithmic information theory. A model-based Kolmogorov sufficient statistic is described and is utilized in a recursive scheme for ranking models in a population. Ranks are assigned in such a way that the n top-ranked models are considered to be the best subset of n models to use in combination. The ranking scheme is appropriate for use in the selection operation of an evolutionary computation. The treatment is primarily theoretical, but several practical issues in ranking and model combination are addressed.

We address the problem of classification of speakers based on measurements of features obtained from their speech. The process is an adaption of biometric methods used to identify people. The process for speech differs since speech is not stationary. We therefore propose the classification of speakers b y the statistical distributions of parameters which may be accurately estimated by modern signal processing techniques. The intent is to develop a speaker clustering algorithm which is dependent of transmission channel and insensitive to language variations, and which may be re-trained, with minimal data, to include a new speaker. We demonstrate effectiveness on the problem of identification of the speakers gender, and present evidence that the methods may be extended to the general problem of speaker identification.

CMAC (Cerebellar Model Arithmetic Computer) have attractive properties of learning convergence and speed and can be ideal in the use of speech processing and enhancement. Many studies have used this special type of neural networks that imitate the human cerebellum in learning control and demonstrated successful results. In this paper CMAC is used to model the speech and noise pick up from a microphone in a vehicular environment. For storage and retrieval of learned data, the input speech and noise signals are quantized using the traditional equal-size quantization region. Results of the modeling were compared to that using the variable size amplitude spectral estimator (ASE). In addition speech enhancement simulations were also presented using the adaptive LMS-CMAC and the ASE-CMAC algorithm and have shown potential for real-time application. The ASE-CMAC produce a far better result especially in areas where the signal to noise ration is very low.

In this paper, Intelligent Optical Systems, Inc. reports on our progress in using neural network signal processing algorithms for the enhancement of sensor signals from a multigas optical sensor under development for NASA. We found that a 4x8x3 neural network yielded superior results over the last squares (LS), partial least squares (PLS), and principal components regression (PCR) algorithms in estimating oxygen, water vapor, and temperature.

This work mainly deals with the application of Cellular Automata to the simulation of Satellite Remote Sensing Images which is a typical real- life image analysis problem. Cellular Automata is applied to solve a diffusion equation, to simulate clouds behavior and the simulation results are compared with the actual images. The aim of this work is to develop a simulator for images as they are gathered by the AVHRR (Advanced Very High Resolution Radiometer) sensor carried aboard the NOAA (National Oceans and Atmospheric Administration) Polar Orbing satellites. The images simulated by means of the Cellular Automata are compared with actual data from the NOAA-14 satellite.

This article presents a generic computational framework necessary for the solution of image understanding problem. The hierarchical networks, represented dually as discrete and continuous structures, are data and algorithms at the same time within that framework. Such structures are able to perform both graph and diagrammatic operations being the basis of intelligence. Dual representation provides natural transformation of continuous image information into primary discrete structures, making the image available for analysis. The computational methods create further higher-level derivative structures, which can run top-bottom algorithms and play role of the context or measurement device, giving the ability to analyze. Symbols naturally emerge in such structures and symbolic operations work there in the combination with the new proposed simplified methods of computational intelligence. That makes images and scenes self-describing, and provides flexible ways of resolving uncertainty. Classification of images truly invariant to any transformation could be done via matching not their primary, but their derivative structures. The ability of image applications to resolve ambiguity and uncertainty in the real images requires tight integration of low-level image processing with high-level knowledge-based reasoning, which is the solution of the image understanding the problem. The proposed architecture does not require supercomputers, opening new ways for image applications.

In this paper, a modified evolutionary programming-based sampling frequency-sensitive network is proposed for pattern clustering. Many researchers study the neural networks for pattern clustering recently. The Kohenen Feature Maps (KFM) network and BP neural network are examples. But there are some problems with these models. For example, the network has a complicated structure and large amount of neurons. The neural network usually gets in unexpected local optimal solution. The results of pattern classification often correlate with the initial conditions. The fixed neural network structure is the major disadvantage for pattern clustering where the optimal number of patterns is unknown. Willie Chang presented a sampling frequency-sensitive network in 1997. The model ahs the advantages of simple structure and simple learning rules. But it also has fixed the architecture. The algorithm is porposed in this paper which effectively uses the sampling frequency-sensitive network and the powerful parallel search optimization tool EP (evolutionary programming) which is presented by Fogel,D.B.. The modified Hubert index and cluster splitting and merging algorithm are used in network architecture evolution. The rule of minimum mean square error is used to get the optimal parameters. The proposed method has an advantage of that the optical solution of neural network architecture and parameters can be get simultaneously. So the classification network can get the optimal number of clusters and the optimal vector quantization. The results of the experiment are given to prove that the neural network architecture can be changed for real world problems and get the optimal results.

Our main interest is to formulate algebraic description of both Geometrical and Fourier-approximations of Optics. We use triangular- norms based approach to formulate algebraic descriptions for both geometrical and Fourier-approximations of optics. To take into consideration real nonlinearity of recording media the measure theory is used. Unlimited plane wave as an universal set is considered. For geometrical optics an algebraic model as designed. Logical operators, parameterized by recording media operators, are defined. To extend dynamical range of negative recording media from linear to over-exposure one, non-additive measure is defined. Algebraic properties of the model in dependence on the approximating function choosing are discussed. Theoretical conclusions are illustrated by experimental measuring and numerical simulation for two-layered optical system. For Fourier- approximation Fourier-duality is used to design semi-ring by DeMorgan's law using, 4-f Fourier-holography setup constructs sequence of model's elements, corresponding to Peano's axioms. Convolution is an abstract addition and correlation is an abstract subtraction in the model Fuzzy- valued measure is defined and fuzzy-value logic is designed. Theoretical conclusions are confirmed by experimental demonstration of logical inference Generalized Modus Ponens realization.

A system has been developed which enables the detection and recognition of various gases. Plasma emission spectroscopy has been used to record spectra from volatile species of acetone, vinegar, and coffee beans, along with air and nitrogen spectra. The spectra have been uniquely processed and fed into an artificial neural network program for training and recognition of unknown gases. The system as a whole can be grouped into the emerging and diverse area of artificial nose technology. The sy stem has shown to provide a solution to the recognition of simple gases and odours (air, nitrogen, acetone) and could also satisfactorily recognise more complex samples (vinegar and coffee beans). Recognition is performed in seconds; this being a positive aspect for many artificial nose applications.