Artificial Intelligence and Soft Computing - ICAISC 2004

Information system design and optimum sizing is a very complex task. Theoretical research and practitioners often tackle the optimization problem by applying specific techniques for the optimization of individual design phases, usually leading to local optima. Conversely, this paper proposes the definition of a design methodology based on an evolutionary approach to the optimization of the client/server-farm distributed structure, which is typical of a distributed information technology (IT) architecture. The optimization problem consists of finding the minimum-cost physical systems that satisfy all architectural requirements given by the designer. The proposed methodology allows for the identification of the architectural solution that minimizes costs, against different information system requirements and multiple design alternatives, thorough a genetic-based exploration of the solution space. Experimental results show that costs can be significantly reduced with respect to conventional approaches adopted by IT designers and available in the professional literature.

This paper introduces the Clifford Support Vector Machines as a generalization of the real- and complex- valued Support Vector Machines. The major advantage of this approach is that one requires only one CSVM which can admit multiple multivector inputs and it can carry multi-class classification. In contrast one would need many real valued SVMs for a multi-class problem which is time consuming.

At the beginning, a short description of uncertain variables and a basic decision problem for a class of uncertain systems are presented. The main part of the paper is devoted to a review of new applications of the uncertain variables to: nonparametric decision problems, allocation and project management under uncertainty, systems with uncertain and random parameters, systems with distributed knowledge, pattern recognition and selected practical problems. The presentation is based on the author’s book Analysis and Decision Making in Uncertain Systems (Springer, 2004).

Blind source separation (BSS) and related methods, e.g., ICA are generally based on a wide class of unsupervised learning algorithms and they found potential applications in many areas from engineering to psychology and neuroscience. The recent trends in BSS is to consider problems in the framework of probabilistic generative and tree structured graphical models and exploit a priori knowledge about true nature and structure of latent (hidden) variables or sources such as statistical independence, spatio-temporal decorrelation, sparseness, smoothness or linear predictability. The goal of BSS can be considered as estimation of sources and parameters of a mixing system or more generally as finding a new reduced or compressed representation for the observed (sensor) data that can be interpreted as physically meaningful coding or blind source extraction. The key issue is to find a such transformation or coding (linear or nonlinear) which has true physical meaning and interpretation. In this paper, we briefly review some promising linear models and approaches to blind source separation and extraction using various criteria and assumptions.

Quality of neural network mappings may be evaluated by visual inspection of hidden and output node activities for the training dataset. This paper discusses how to visualize such multidimensional data, introducing a new projection on a lattice of hypercube nodes. It also discusses what type of information one may expect from visualization of the activity of hidden and output layers. Detailed analysis of the activity of RBF hidden nodes using this type of visualization is presented in the companion paper.

Scatterograms of images of training vectors in the hidden space help to evaluate the quality of neural network mappings and understand internal representations created by the hidden layers. Visualization of these representations leads to interesting conclusions about optimal architectures and training of such networks. The usefulness of visualization techniques is illustrated on parity problems solved with RBF networks.

In this paper incomplete data are assumed to be decision tables with missing attribute values. We discuss two main cases of missing attribute values: lost values (a value was recorded but it is unavailable) and “do not care” conditions (the original values were irrelevant). Through the entire paper the same calculus, based on computations of blocks of attribute-value pairs, is used. Complete data are characterized by the indiscernibility relations, a basic idea of rough set theory. Incomplete data are characterized by characteristic relations. Using characteristic relations, lower and upper approximations are generalized for incomplete data. Finally, from three definitions of such approximations certain and possible rule sets may be induced.

Some definitions and theorems concerning positive continuous-time and discrete-time linear systems are presented. The notion of a positive estimator maping a positive cone into a positive cone is introduced. A multi-layer perceptron and a radial neural network approximating the nonlinear estimator are proposed.A neural network modeling the dynamics of a positive nonlinear dynamical system is also proposed. The new neural networks are verified and illustrated by an example.

The paper deals with support of natural intelligence employing the concept of two factors utility which characterizes the sustainable development. The first factor represents the long-term expected profit of generalized capital investment.The second factor represents the worse case profit necessary for survival of crises. Using that concept it is possible to support the individual decisions connected with choosing the best alternative. It enables also to support the cooperation strategies connected with exploitation of innovations and new technologies.

A neural network model is developed for the top-down selective attention (TDSA), which estimates the most probable sensory input signal based on previous knowledge and filters out irrelevant sensory signals for high-confidence perception of noisy and confusing signals. The TDSA is modeled as an adaptation process to minimize the attention error, which is implemented by the error backpropagation algorithm for the multilayer Perceptron classifiers. Sequential recognition of superimposed patterns one by one is also possible. The developed TDSA model is applied to the recognition tasks of two-pattern images, superimposed handwritten characters, and noise-corrupted speeches.

The main goal of this paper is to present, through some of main ANN models and based techniques, their capability in real world industrial dilemmas solution. Several examples of real world applications and especially industrial ones have been presented and discussed.

ActiveMath is a web-based intelligent tutoring system for mathematics. This article presents the technical and pedagogical goals of ActiveMath, its principles of design and architecture, its knowledge representation, and its adaptive behavior. In particular, we concentrate on those features that rely on AI-techniques.

Basic rules of inference used in classical logic are Modus Ponens (MP) and Modus Tollens (MT). These two reasoning patterns start from some general knowledge about reality, expressed by true implication, ”if Φ then Ψ”. Then basing on true premise Φ we arrive at true conclusion Ψ (MP), or from negation of true conclusion Ψ we get negation of true premise Φ (MT).In reasoning from data (data mining) we also use rules ”if Φ then Ψ”, called decision rules, to express our knowledge about reality, but in this case the meaning of the expression is different. It does not express general knowledge but refers to partial facts. Therefore decision rules are not true or false but probable (possible) only. In this paper we compare inference rules and decision rules in the context of decision networks, proposed by the author as a new approach to analyze reasoning patterns in data.

To analyze adaptation capabilities of individuals and agents in constantly changing environments, we suggested using of connectionist methodology and the solution of sequences of different pattern recognition tasks. Each time after the task change, we start training from previous perceptron weight vector. We found that large values of components of the weight vector decrease the gradient and learning speed. A noise injected into the desired outputs of the perceptron is used as a “natural” method to control the weight growth and adaptation to new environment. To help artificial population to withstand lengthy sequences of strong catastrophes, populations with offspring and ”genetic” inheritance of the noise intensity parameter have to be created. It was found that the optimal interval for the noise intensity follows power of environmental changes. To improve the survivability of synthetic populations, we suggest “mother’s training”, and partial protection of offspring from artificially corrupted training signals. New simulation methodology could help explain known technical, biological, psychological and social phenomena and behaviors in quantitative way.

Given a data set describing a number of pairwise comparisons of reference objects made by a decision maker (DM), we wish to find a set of robust decision rules constituting a preference model of the DM. To accomplish this, we are constructing rough approximations of the comprehensive preference relation, called outranking, known from these pairwise comparisons. The rough approximations of the outranking relation are constructed using the Lorenz dominance relation on degrees of preference on particular criteria for pairs of reference objects being compared. The Lorenz dominance is used for its ability of drawing more robust conclusions from preference ordered data than the Pareto dominance. The rough approximations become a starting point for mining “if..., then...” decision rules constituting a logical preference model. Application of the set of decision rules to a new set of objects gives a fuzzy outranking graph. Positive and negative flows are calculated for each object in the graph, giving arguments about its strength and weakness. Aggregation of both arguments by the Net Flow Score procedure leads to a final ranking. The approach can be applied to support multicriteria choice and ranking of objects when the input information is a set of pairwise comparisons of some reference objects.

Paper presents absolutely new ideas about needs and possibilities of automatic understanding of the image semantic content. The idea under consideration can be found as next step on the way starting from capturing of the images in digital form as two–dimensional data structures, next going throw images processing as a tool for enhancement of the images visibility and readability, applying images analysis algorithms for extracting selected features of the images (or parts of images e.g. objects), and ending on the algorithms devoted to images classification and recognition. In the paper we try to explain, why all procedures mentioned above can not give us full satisfaction, when we do need understand image semantic sense, not only describe the image in terms of selected features and/or classes. The general idea of automatic images understanding is presented as well as some remarks about the successful applications of such ides for increasing potential possibilities and performance of computer vision systems dedicated to advanced medical images analysis.

The paper discusses the application of a multi layered high order neural network in modelling and fault diagnosis of dynamic processes. Dynamic properties can be obtained by adding a finite impulse response filter to the neuron. A combinatorial algorithm is used for selecting the network structure. If a linear activation function is used, the universal approximation capabilities of these networks are easy to prove. To show the applicability of such networks, the modelling and fault detection results for the two-tank-system are presented in the last part.

This paper presents a momentum modification of two RLS algoritms: momentum RLS and UD momentum RLS, each in classical and linear version. All methods are tested on two standart benchmarks. The results are discussed.

In this paper we present a parallel realization of QR learning algorithm. We use the Householder reflection method to transform matrices to QR form. Parallel structures and performance discusion are included.

Three neural network methods, feed forward back propagation (FFBP), radial basis function (RBF) and generalized regression neural network (GRNN) were employed for rainfall-runoff modelling of Turkish hydrometeorologic data. It was seen that all three different ANN algorithms compared well with conventional multi linear regression (MLR) technique. It was seen that only GRNN technique did not provide negative flow estimations for some observations. The rainfall-runoff correlogram was successfully used in determination of the input layer node number.

Population based methods, and among them, the population learning algorithm (PLA), can be used to train artificial neural networks. The paper studies the probability distribution of solution time to a sub-optimal target in the example implementation of the PLA-trained artificial neural network. The distribution is estimated by means of the computational experiment. Graphical analysis technique is used to compare the theoretical and empirical distributions and estimate parameters of the distributions. It has been observed that the solution time to a sub-optimal target value fits a two parameter exponential distribution.

In this paper, we have developed a parallel approach for minimizing the projection error in Sammon’s mapping applied in combination with the self-organizing map (SOM). In the so-called integrated mapping, Sammon’s algorithm takes into account the learning flow of the self-organizing neural network. As a final result in the integrated mapping, we need to visualize the neurons-winners of the SOM. The criterion of visualization quality is the projection error of Sammon’s mapping.

In this paper we present experiments comparing different training algorithms for Radial Basis Functions (RBF) neural networks. In particular we compare the classical training which consist of a unsupervised training of centers followed by a supervised training of the weights at the output, with the full supervised training by gradient descent proposed recently in same papers. We conclude that a fully supervised training performs generally better. We also compare Batch training with Online training and we conclude that Online training suppose a reduction in the number of iterations.

This paper deals with the problem of identification of continuous time dynamic neural networks when the measurements are given only at discrete time moments, not necessarily uniformly distributed. It is shown that the modified adjoint system, generating the gradient of the performance index, is a continuous-time system with jumps of state variables at moments corresponding to moments of measurements.

Building an ensemble of classifiers is an useful way to improve the performance. In the case of neural networks the bibliography has centered on the use of Multilayer Feedforward (MF). However, there are other interesting networks like Radial Basis Functions (RBF) that can be used as elements of the ensemble. Furthermore, as pointed out recently, the network RBF can also be trained by gradient descent, so all the methods of constructing the ensemble designed for MF are also applicable to RBF. In this paper we present the results of using eleven methods to construct a ensemble of RBF networks. The results show that the best method is in general the Simple Ensemble.

This paper deals with some theoretic and numerical issues in the learning algorithm for Procedure Neural Networks (PNNs). In PNNs the weights are time functions and can be expended by some basis functions. The properties of PNNs vary with the choice of weights functions. Orthodoxy basis functions have many advances in expending the weight functions and save training time in PNNs learning. In this paper several kinds of orthodoxy functions are proposed. Also the algorithm convergence of PNNs training is discussed.

This paper presents a new parameter and confidence estimation techniques for static GMDH neural networks. The main objective is to show how to employ the outer-bounding ellipsoid algorithm to solve such a challenging task that occurs in many practical situations. In particular, the proposed approach can be relatively easy applied in robust fault diagnosis schemes.

Visualization of MLP error surfaces helps to understand the influence of network structure and training data on neural learning dynamics. PCA is used to determine two orthogonal directions that capture almost all variance in the weight space. 3-dimensional plots show many aspects of the original error surfaces.

A neural network based knowledge discovery method for single fault detection in electronics circuits is presented. A functional equivalence of Radial Basis Function (RBF) neural network and Takagi-Sugeno (TS) fuzzy system is used in this process. A specially modified incremental RBF network training scheme suitable for rule discovery is used. Next, the RBF neural network is converted into the TS fuzzy system. A set of linguistic rules for detection of circuit catastrophic faults are obtained (100% detection accuracy was achieved for the tested electronic circuit).

We propose here to use a space-filling curve (SFC) as a tool to introduce a new metric in I _d defined as a distance along the space-filling curve. This metric is to be used inside radial functions instead of the Euclidean or the Mahalanobis distance. This approach is equivalent to using SFC to pre-process the input data before training the RBF net. All the network tuning operations are performed in one dimension. Furthermore, we introduce a new method of computing the weights of linear output neuron, which is based on connection between RBF net and Nadaraya-Watson kernel regression estimators.

Chaotic neural network with external inputs has been used as a mixed input pattern separator. In contrast to previous work on the subject, highly chaotic dynamical system (LLE ≈ 0.6) is applied here. The network is based on a “dynamical mapping” scheme as an effective framework for cortical mapping. This feature allows for a more effective pattern retrieval and separation by the network.

Training multilayer feedforward neural networks corresponds to the global minimization of the network error function. To address this problem we utilize the Snyman and Fatti [1] approach by considering a system of second order differential equations of the form, ẍ$=-\nabla E(x)$, where x is the vector of network weights and $\nabla E$ is the gradient of the network error function E. Equilibrium points of the above system of differential equations correspond to optimizers of the network error function. The proposed approach is described and experimental results are discussed.

Barycentric plotting, achieved by placing gaussian kernels in distant corners of the feature space and projecting multidimensional output of neural network on a plane, provides information about the process of training and certain features of the network. Additional visual guides added to the plot show tendencies and irregularities in the training process.

The problem of locating centers for radial basis functions in neural networks is discussed. The proposed approach allows us to apply the results from the theory of optimum experimental designs. In typical cases we are able to compose optimal centers’ locations from the known univariate experiment designs.

We consider fixed non–linear combining rules as a strategy for information fusion in neural network based machine learning problems and compare them with adaptive ones. In this strategy, essential part of the work is overloaded to examiner–operator who ought to split the data “reliably”. In small sample size situations, non–trainable combination rules allow creating committee with performance comparable or even lower with that obtained with more sophisticated information fusion methods. Fixed rule’s solutions are easier interpreted by end users.

Hamiltonian Neural Networks based orthogonal filters are universal signal processors. The structure of such processors rely on family of Hurwitz-Radon matrices. To illustrate, we propose in this paper a procedure of nonlinear mapping synthesis. Hence, we propose here system modeling and learning architectures which are suitable for very large scale implementations.

This paper deals with the design of the recurrent neural network structure approximating the quasi-inverses of the multi-dimensional Sierpiński space-filling curves. The network consists of two subsequent recurrent layers. The network structure shows that computing the quasi-inverse of the Sierpiński space-filling curve can be performed massively parallel.

Many of the reasoning methods are suitable neither for fuzzy control nor for fuzzy modeling. In the paper some possible reasoning methods are compared from this point of view. The author proposes new methods for fuzzy control, better than Mamdani, Larsen, Tsukamoto. Fuzzy systems are described by a set of rules using connectives ”and”, ”or”, ”also”. Different aggregation operators, as triangular norms and mean operations, are used for interpretation of these connectives. In the paper are discussed possible interpretations for if ... then rules, as different implications and other operations, in combination with defuzzification methods. Examples of the systems with PID fuzzy controllers are presented using different reasoning methods, aggregation operations, and linear and nonlinear plants. Some best methods are proposed.

In the paper we study flexible neuro-fuzzy systems based on a compromise fuzzy implication. The appropriate neuro-fuzzy structures are developed and the influence of a compromise parameter on their performance is investigated. The results are illustrated on typical benchmarks.

In this paper, a method for the reduction of noise in a speech signal is introduced. In the implementing of the method, firstly a high resolution frequency map of the signal is obtained. Each frequency band component of the signal is then segmented. Fuzzy inference system (FIS) is used for the determination of the noise contents of the segments. The output of the FIS is the suppression level of the segment. If the FIS decides that the segment contains only noise, then the segment is deleted or if the FIS decides that the segment is noise free, it is allowed to be passed without suppression. Since the signal to noise ratio (SNR) varies from case to case, the limits of the membership functions are tuned accordingly. This self tuning capability gives flexibility and robustness of the system.

In this paper general diagnostic model of cars’ current supply circuit is presented. The set of defects with appropriate symptom signals is described. Binary diagnostic matrix is built. Fuzzy-neural models bank is presented. Some results of computer analysis of experimental research are described.

Hierarchical fuzzy systems allow for reducing number of rules and for prioritization of rules. To retain fuzziness, intermediate signals should be fuzzy. Transferring fuzzy signal is computationally demanding. Special form of hierarchical fuzzy system is proposed to reduce computational burden.

Short term trends, particularly attractive for neural network analysis, can be used profitably in scenarios such as option trading, but only with significant risk. To predict stock trends, we exploit Emotional Learning Based Fuzzy Inference System (ELFIS). ELFIS has the advantage of low computational complexity in comparison with other multi-objective optimization methods. The performance of ELFIS in the prediction of stock prices will be compared with that of Adaptive Network Based Fuzzy Inference System (ANFIS). Simulations show better performance for ELFIS.

The paper presents a method of digital implementation of a fuzzy Petri net (FPN) based on asynchronous fuzzy RS flip-flop. The FPN can be viewed as a formal description of a control algorithm of a plant. The idea of FPN and its dynamics were characterized. A digital architecture of the asynchronous fuzzy RS flip-flop was given. A conception of the hardware implementation of the FPN was described.

Process of implementing operations’algorithms for ordered fuzzy numbers (OFN’s)are presented. First version of the program in the Delphi environment is created that uses algorithms dedicated to trapezoidal-type membership relations (functions). More useful implementation is a Fuzzy Calculator which allows counting with OFN’s of general type membership relations and is equipped with a graphical shell.

Ordered fuzzy number is an ordered pair of continuous real functions defined on the interval [0, 1]. Such numbers have been introduced by the author and his co-workers as an enlargement of classical fuzzy numbers by requiring a membership relation. It was done in order to define four algebraic operations between them, i.e. addition, subtraction, multiplication and division, in a way that renders them an algebra. Further, a normed topology is introduced which makes them a Banach space, and even more, a Banach algebra with unity. General form of linear functional on this space is presented which makes possible to define a large family of defuzzification methods of that class of numbers.

In this paper we present results of application of information cirterions to neuro-fuzzy systems (NFS) design. The criterions come from autoregression estimation theory and are employed to describe the level of NFS quality. Based on this method the preferred size of systems is determined. Various criterions are compared and discussed.

In this paper, we propose a generalization of the definition of an IF-set, an intuitionistic fuzzy set, and related hesitation degrees. We flexibilize the original method of computing these values by the use of triangular norms. Next, we present its application to group decision making problems.

A technique for Fuzzy Cognitive Maps learning, which is based on the minimization of a properly defined objective function using the Particle Swarm Optimization algorithm, is presented. The workings of the technique are illustrated on an industrial process control problem. The obtained results support the claim that swarm intelligence algorithms can be a valuable tool for Fuzzy Cognitive Maps learning, alleviating deficiencies of Fuzzy Cognitive Maps, and controlling the system’s convergence.

A system input-output response is modeled using a knowledge-based method of signal processing known as neuro-fuzzy logic. The paper presents a new method of the fuzzy model parameters tunning. Fuzzy model tuning procedures based on an evolutionary algorithm are also given. As an example, the analysis of the membership function kind is carried out for the fuzzy modeling of parameters, which are necessary to describe the state of a pressure vessel with water-steam mixture during accidental depressurizations.

Defining linguistic fuzzy evaluations of various quantities e.g. very small, average, approximately 5, a little more than 10, we usually assume that membership functions qualifying for these evaluations are of possibilistic character. The paper presents a comparison of a measurement realization by a technical instrument and of evaluation by man (uncertainty of the measurement realized by technical instrument is mainly of probabilistic character) to determine their common features. Next, the question is analyzed: what is the character of human evaluations. It is of great importance for the way in which operations of fuzzy arithmetic have to be realized and for Computing with Words.

The problem of determining a safe trajectory of a ship moving in a restricted area is presented. Goals and constraints are formulated for the optimal trajectory of ship movement in a fairway. The problem is presented as an optimization problem in a fuzzy environment. To solve it, the method of fuzzy linear programming is proposed. The results for a ship passing manoeuvre are given and conclusions are drawn.

In the paper the employment of fuzzy logic in feature models for software system families is presented. The fuzzy weights of some variable features are introduced. The approach is demonstrated on the example of the feature model describing car properties. The formulas resulting from the description of the feature tree and its constraints can be used as input for an expert system validating possible feature combinations.

In the paper, we present a new fuzzy relational system with multiple outputs for classification purposes. Rules in the system are more flexible than the rules in linguistic fuzzy systems because of the additional weights in rule consequents. The weights comes from an additional binary relation. Thanks to this, input and output fuzzy sets are related to each other with a certain degree. The size of the relations is determined by the number of input fuzzy sets and the number of output fuzzy sets for a given class. Simulation results confirmed the system ability to classify data.

In this study both classical type-1 and interval type-2 fuzzy systems have been compared with the perspective on overall output of both systems. Some analytical aspects have been examined for the case of two and three activated rules.

We propose a new similarity measure for intuitionistic fuzzy sets and show its usefulness in medical diagnostic reasoning. We point out advantages of this new concept over the most commonly used similarity measures being just the counterparts of distances. The measure we propose involves both similarity and dissimilarity.

In this paper, an evolutionary algorithm based on an immune system activity to handle constraints is discussed for three-criteria optimisation problem of finding a set of Pareto-suboptimal task assignments in parallel systems. This approach deals with a modified genetic algorithm cooperating with a main evolutionary algorithm. An immune system activity is emulated by a modified genetic algorithm to handle constraints. Some numerical results are submitted.

We have considered the problem of job scheduling on a single machine with deadlines. The objective is to find a feasible job sequence (satisfying the deadlines) to minimize the sum of weighted completion times. Since the problem is NP-hard, heuristics have to be used. Methods of artificial intelligence: simulated annealing, neural networks and genetic algorithms, are some of the recent approaches. We propose a very effective parallel genetic algorithm PGA and methods of determining lower and upper bounds of the objective function. Since there are difficulties with determining the initial population of PGA for this scheduling problem, therefore the algorithm also adds random generated unfeasible solutions to the population. We announce a method of elimination of these kind of solutions. The examined algorithms are implemented in Ada95 and MPI. Results of computational experiments are reported for a set of randomly generated test problems.

This paper deals with the two-level, partially stochastic optimization method named Two Level Adaptive Evolutionary Computation (TLAEC). New adaptation mechanism is embedded in the method. The aim of the paper is to present an algorithm based on TLAEC method, solving so-called development problem. A mathematical model of this problem assumes the form of mixed discrete-continuous programming. A concept of the algorithm is described in the paper and the proposed, new adaptation mechanism that is introduced in the algorithm is described in detail. The results of computation experiments as well as their analysis are also given.

The aim of the paper is to present a special type of adaptive evolutionary method, named here Two-Level Adaptive Evolutionary Computation (TLAEC). The method consists in combination of evolutionary computation with deterministic optimization algorithms in a hierarchy system. Novelty of the method consists also in a new type of adaptation mechanism. Post optimal analysis of the lower level optimization task is utilized in order to modify probability distribution for new genotype generations. The formal description of the method is presented in the paper. The application of this method to a mixed, discrete-continuous linear optimization task is given as an example.

An effect of the loss of diversity and the rapid concentration of an initially broadly spread population is analyzed for models of phenotypic evolution. The impact of the population size, different selection schemes, and dimensionality of the search space on population diversity is studied. Obtained results confirm common opinions that in large populations the diversity is greater and that the stronger selection pressure causes the faster concentration. High dimensional search spaces do not restrain the concentration of a population. In this case, populations cluster even faster than in one-dimensional search space.

There are many heuristic algorithms for clustering, from which the most important are the hierarchical methods of agglomeration, especially the Ward’s method. Among the iterative methods the most universally used is the C–means method and it’s generalizations. These methods have many advantages, but they are more or less dependent on the distribution of points in space and the shape of clusters. In this paper the problem of clustering is treated as a problem of optimization of a certain quality index. For that problem the author proposes two solutions: a hierarchical partitioning algorithm and an evolutionary algorithm.

In the paper, a new evolutionary approach to induction of oblique decision trees is described. In each non-terminal node, the specialized evolutionary algorithm is applied to search for a splitting hyper-plane. The feature selection is embedded into the algorithm, which allows to eliminate redundant and noisy features at each node. The experimental evaluation of the proposed approach is presented on both synthetic and real datasets.

The goal of this paper is to demonstrate the rate at which building blocks evolve in time during run of GA. We compare the building block propagation rate for the simple genetic algorithm (SGA) with the building block propagation rate for the migration (island) genetic algorithm (MPGA). The results are checked against the lower bound given by Holland’s schema theorem. Using genetic programming, we made symbolic regression on the number of individuals matching given schema versus generation number. As a result, a new expression describing the propagation rate was found, and its correctness was confirmed in all cases considered in the paper.

The success of artificial neural network evolution is determined by many factors. One of them is the fitness function used in genetic algorithm. Fitness function determines selection pressure and therefore influences the direction of evolution. It decides, whether received artificial neural network will be able to fulfill its tasks. Three fitness functions are proposed and examined in the paper, every one of them gives different selection pressure. Comparison and discussion of obtained results for every function is made.

The method of rule extraction from a neural network based on the genetic approach with Pareto optimization is presented in the paper. The idea of Pareto optimization is shortly described and the details of developed method such as fitness function, genetic operators and the structure of chromosome are shown. The method was tested with well known benchmark data sets. The results of these experiments are presented and discussed.

This paper deals with the problems of hierarchical representation of skeletal structures and their optimization by means of an evolutionary algorithm. We describe the main advantages of using hierarchical graph structures to represent the designed objects as well as specialized genetic operators able to work on these graphs A top-down algorithm allowing for optimization of structures at different levels of hierarchy is also introduced. We illustrate the proposed method with examples of its application to the problem of designing optimal transmission towers.

We describe new results in developing of a satisfiability checker for probabilistic logic based on the genetic algorithm approach combined with a local search procedure. Computational experiences show that problems with 200 propositional letters can be solved. They are, to the best of our knowledge, the largest PSAT-problems reported in the literature.

In this paper posibility of design and optimization of combinational digital circuits using modified evolutionary algorithm is presented. Modification of evolutionary algorithm depends on introduction of multilayer chromosomes and genetic operators operating on them. Design results for four combinational circuits obtained using this method are compared with described in literature methods: Karnaugh Maps, Quine-McCluskey and NGA and MGA genetic algorithms. Described evolutionary algorithm leads in many cases to better results.

In this paper modified version of roulette selection for evolution algorithms – the fan selection, is presented. This method depends on increase of survive probability of better individuals at the expense of worse individuals. Test functions chosen from literature are used for determination of quality of proposed method. Results obtained for fan selection are compared with results obtained using roulette selection and elitist selection.

Genetic algorithm is very useful method for global search of large search space and has been applied to various problems. It has two kinds of important search mechanisms, crossover and mutation. Especially many researchers have more interested in crossover operator than mutation operator because crossover operator has charge of the responsibility of local search. In this paper we introduce a new crossover operator avoiding the drawback of conventional crossovers. We compare it to several crossover operators for travelling salesman problem (TSP) for showing the performance of proposed crossover.

Biomarkers are molecular parameters associated with presence and severity of specific disease states. Search for biological markers of cancer in proteomic profiles is a relatively new but very active research area. This paper presents a novel approach to feature selection and thus biomarker identification. The proposed method is based on blind separation of sources and selection of features from a reduced set of components.

A set of violins submitted to a competition has been evaluated by the jury from the viewpoint of several criteria and then ranked from the best to the worst. The sound of the instruments played by violinists during the competition has been recorded digitally and then processed to obtain sound attributes. Given the jury’s ranking of violins according to sound quality criteria, we are inferring from the sound characteristics a preference model of the jury in the form of “if..., then...” decision rules. This preference model explains the given ranking and permits to build a ranking of a new set of violins according to this policy. The inference follows the scheme of an inductive supervised learning. For this, we are applying a special computational tool called Dominance-based Rough Set Approach (DRSA). The new set of attributes derived from the energy of consecutive halftones of the chromatic scales played on four strings has proved a good accuracy of the approximation.

This paper presents the variable precision fuzzy rough sets (VPFRS) model, which constitutes a generalisation of the extended variable precision rough set (VPRS) concept. The notion of the α-inclusion error based on the fuzzy implication operators will be introduced. Additionally to extending the basic definition of the fuzzy rough approximations, an idea of the weighted mean fuzzy rough approximations will be given. In an illustrating example the most popular residual implicators will be used.

The paper describes a method of creating a new information storage structure based on rough sets [1] and fuzzy sets [4]. The original element in the presented structure is a fuzzy indiscernibility relation between elements, which can be interpreted as two elements are indiscernible with degree r, where r is a number from the [1,4] interval. The authors show the implementation of the structure, and its application to intelligent searching of textual databases or Web resources, to improve and strengthen the existing searching mechanisms. The elements in the comparison process are words and sentences treated as sequences of letters and words, respectively. Given strings can be compared to other sequences of letters, Websites keywords, for instance. The important fact to be noted is that the described similarity does not concern the semantics of words or sentences it bases on letter subsequences only, but otherwise it can be used for finding semantic similarity also. The algorithm is used for partial structurizing the sets of synonyms for search engines. Finally, some remarks and comments on the further research are presented.

In this paper we presented a general solution to compose rough-neuro-fuzzy architectures. Monotonic properties of fuzzy implications were assumed to derive fuzzy systems in the case of missing features. The fuzzy implications satisfying Fodor’s lemma used in logical approach and t-norms used in Mamdani approach are discussed.

In this paper we presented a general solution to compose rough-neuro-fuzzy architectures. The fuzzy system in the case of missing features is derived without the assumption that used fuzzy implication is monotonic. The proposed solution is also suitable for the monotonic fuzzy implications satisfying Fodor’s lemma. The architecture based on the Zadeh and Willmott fuzzy implications is derived as the special case of the proposed general solution.

In this paper we introduce a new class of algebras, called extended residuated lattices. Basing on this structure we present an algebraic generalization of approximation operators and rough sets determined by abstract counterparts of fuzzy logical operations. We show formal properties of these structures taking into account several classes of fuzzy relations.

A fault diagnosis system using rough sets based classification techniques is developed for cage induction machines broken bar detection. The proposed algorithm uses the stator current and motor speed as input. Several features are extracted from the frequency spectrum of the current signal resulting from FFT. A Rough Sets based classifier is then developed and applied to distinguish between different motor conditions. A series of experiments using a three phase 3 hp cage induction machine performed in different load and fault conditions are used to provide data for training and then testing the classifier. Experimental results confirm the efficiency of the proposed algorithm for detecting the existence and severity of broken bar faults.

This paper presents a novel approach to financial time series analysis and prediction. It is mainly devoted to the problem of forecasting university facility and administrative cost recovery. However, it can also be used in other areas of a similar nature. The methodology incorporates a two-stage hybrid mechanism for selection of prediction-relevant features and for forecasting based on this selected sub-space of attributes. The first module of the methodology employs the theory of rough sets (RS) while the second part is based upon artificial neural networks (ANN).

We address a situation when more than one feature subset allows for linear separability of given data sets. Such situation can occur if a small number of cases is represented in a highly dimensional feature space.The method of the feature selection based on minimisation of a special criterion function is here analysed. This criterion function is convex and piecewise-linear (CPL). The proposed method allows to evaluate different feature subsets enabling linear separability and to choose the best one among them. A comparison of this method with the Support Vector Machines is also included.

The paper presents a method of signal analysis which is based on the parameter space consideration. The parameter space is created during the short-time analysis of the signal. The general schema of the approach consists of using a time window sliding in time along a signal. After choosing some particular parameters one observes their changes in a sliding window and analyzes the data in a multidimensional parameter space. For recognition and detection of different system states we propose to perform the clustering in the parameter space. The presented approach was used for analysis of EEG signals and some vibroacoustic signals taken form the combustion engine.

The paper presents two methods offering flexible solutions to cluster-analysis problems. The first one employs a genetic-algorithm-based Travelling-Salesman-Problem-solution, and the second one – self-organizing Kohonen networks. The operation of both techniques has been illustrated with the use of synthetic data set and then they have been tested by means of real-life, multidimensional Mushrooms Database (8124 records) available from the FTP server of the University of California at Irvine (ftp.ics.uci.edu).

The paper presents a modification of the self-organizing Kohonen networks for more efficient coping with complex, multidimensional cluster-analysis problems. The essence of modification consists in allowing the neuron chain – as the learning progresses – to disconnect and later to reconnect again. First, the operation of the modified approach has been illustrated by means of synthetic data set. Then, this technique has been tested with the use of a real-life, complex, multidimensional data set (Pen-Based Recognition of Handwritten Digits Database) available from the FTP server of the University of California at Irvine (ftp.ics.uci.edu).

The k Nearest Centroid Neighbor (k-NCN) is a relatively new powerful decision rule based on the concept of so-called surrounding neighborhood. Its main drawback is however slow classification, with complexity O(nk) per sample. In this work, we try to alleviate this disadvantage of k-NCN by limiting the set of the candidates for NCN neighbors for a given sample. It is based on an intuition that in most cases the NCN neighbors are located relatively close to the given sample. During the learning phase we estimate the fraction of the training set which should be examined only to approximate the “real” k-NCN rule. Similar modifications are applied also to ensemble of NCN classifiers, called voting k-NCN. Experimental results indicate that the accuracy of the original k-NCN and voting k-NCN may be preserved while the classification costs significantly reduced.

Decision tree algorithms deal with continuous variables by finding split points which provide best separation of objects belonging to different classes. Such criteria can also be used to augment methods which require or prefer symbolic data. A tool for continuous data discretization based on the SSV criterion (designed for decision trees) has been constructed. It significantly improves the performance of Naive Bayes Classifier. The combination of the two methods has been tested on 15 datasets from UCI repository and compared with similar approaches. The comparison confirms the robustness of the system.

This paper is an continuation of the accompanying paper with the same main title. The first paper reviewed instance selection algorithms, here results of empirical comparison and comments are presented. Several test were performed mostly on benchmark data sets from the machine learning repository at UCI. Instance selection algorithms were tested with neural networks and machine learning algorithms.

Partial Memory Learning (PML) is a machine learning paradigm in which only a subset of cases generated from an original training set is used for classification. This paper concerns a new method for partial memory learning. The SBL-PM-M method is a completely new model. We evaluate the performance of the new algorithm on several real-world datasets and compare it to a few other PML systems and to the base classifier.

The support vector machine (SVM) constitutes one of the most successful current learning algorithms with excellent classification accuracy in large real-life problems and strong theoretical background. However, a SVM solution is given by a not intuitive classification in terms of extreme values of the training set and the size of a SVM classifier scales with the number of training data. Generalized relevance learning vector quantization (GRLVQ) has recently been introduced as a simple though powerful expansion of basic LVQ. Unlike SVM, it provides a very intuitive classification in terms of prototypical vectors the number of which is independent of the size of the training set. Here, we discuss GRLVQ in comparison to the SVM and point out its beneficial theoretical properties which are similar to SVM whereby providing sparse and intuitive solutions. In addition, the competitive performance of GRLVQ is demonstrated in one experiment from computational biology.

Several methods were proposed to reduce the number of instances (vectors) in the learning set. Some of them extract only bad vectors while others try to remove as many instances as possible without significant degradation of the reduced dataset for learning. Several strategies to shrink training sets are compared here using different neural and machine learning classification algorithms. In part II (the accompanying paper) results on benchmarks databases have been presented.

The recent results of the research into construction of syntactic pattern recognition-based expert systems are presented. The model of syntactic pattern recognition has been defined with the use of GDPLL(k) grammars and parsers, and the model has been successfully applied as an efficient tool for inference support in several expert systems. Nevertheless, one of the main problems of practical application of GDPLL(k) grammars consists in difficulties in defining the grammar from the sample of a pattern language. In the paper we present the first achievement in the field of grammatical inferencing of GDPLL(k) grammars: an algorithm of automatic construction of a GDPLL(k) grammar from a so-called polynomial specification of the language.

In this paper there is a review of the selected issues on recognition and safety estimation of voice commands in natural language given by the operator of the technological device. A view is offered of the complexity of the recognition process of words and commands using neural networks made of a few layers of neurons. The paper presents some research results of speech recognition and automatic command recognition with artificial neural networks. The first part of the paper introduces a new conception of an intelligent layer of two-way voice communication of the technological device with the operator and discusses the general topics and issues. The second part is devoted to a discussion of more specific topics of the automatic command recognition and safety estimation that have led to interesting new approaches and techniques.

A neural network based method for classification of meteorological data is proposed in the paper. The method consists of two phases. First, a non-linear projection of the data space is performed by means of radial basis functions. The neural gas algorithm is used for determining locations of the basis functions. Second, a nonlinearly projected data is allocated to different classes by means of a competitive network layer. Nonlinear data transformation was necessary for obtaining linear separability of 6 classes of the meteorological data defined in 8 dimensions.

The strength of data integrity, message authentication and pseudonym generation mechanisms in the design of secure multimedia communication applications relies on the quality of the message digest algorithms used. In this paper, we propose Support Vector Machines based evaluation benchmarks to assess the message digest function quality since there is lack of practical tests to be applied to message digest algorithms in the emerging field of designing secure information and communication systems especially for the delivery of multimedia content, where the issues of copyright protection and security in transactions are outstanding.

A novel procedure based on genetic algorithms is presented for the evaluation and production of digital fingerprints in the design of secure communication systems. These digital fingerprints are computed using the methodology of un-keyed one-way functions (hash functions). The problem of evaluating the quality of such functions is formulated as a global optimization one in the space spanned by all possible messages and is approached from a practical viewpoint by involving genetic algorithms, contrary to the very few similar research efforts existing in the literature that are of only theoretical interest. Moreover, the problem of producing digital fingerprints of good quality for use in communication systems is formulated in terms of a hash function constructed by involving the genetic algorithm procedure, exclusively utilizing the crossover operator and the steady-state reproduction method and omitting its random components. The promising results herein obtained illustrate the importance of applying genetic algorithms in communication systems security design.

This paper deals with Chinese web page classification based on text contents. It includes setting up dictionary, departing words, extracting feature, designing classifiers, etc. First, feature compression problem is discussed in detail, which several methods are introduced reducing the feature dimension from more than 120,000 to 5,000. Second, several classifiers are selected to pattern training, such as linear neural perceptron, self-organization mapping networks, radial basis function neural networks, and support vector machine. In the last of this paper experiments on Chinese people daily web version are illustrated and classification results are compared.

This paper introduces a new fuzzy clustering method with constraints in time domain which may be used to signal analysis. Proposed method makes it possible to include a natural constraints for signal analysis using fuzzy clustering, that is, the neighbouring samples of signal belong to the same cluster. The well-known from the literature fuzzy c-regression models can be obtained as a special case of the method proposed in this paper.

The paper describes a proposed method of modification of self-organizing map and its application to special cluster analysis (delivering the information about selected essential feature, which values derive from an ordered set) and also to estimation of the selected (basic) feature for newly occurring patterns. The utilization of this technique in the issue of real estate appraisal has been described. The visualizations of a cluster map, selected estimation maps and numerical results for this problem have been presented.

The aim of the article is to present a possible way of joining the information coming from the modelling system with the unsupervised clustering method, fuzzy c–means method. The practical application of the proposed approach will be presented via problem of bankruptcy prediction.

This paper presents a hybrid method to use both the idea of projection and Toeplitz Matrix approaches to describe the feature vectors of an image and hence identifying it. The method applies two different tools. The main one is Toeplitz forms and the second is Neural Networks. The image model considered in this work are some selected Arabic scripts. The letter is first projected on 12 axes, then the lengths of these axes are measured and afterwards for the sake of classification and recognition these lengths are compared with the ones in the data base. The method has proved its high efficiency upon the other known approaches. Toeplitz model has shown its successful role in improving the description of the image feature vectors and hence increasing the rate of recognition. The overall algorithm has reached a very low rate of misclassification. Both machine and hand written cases have been studied. In this paper, examples of handwritten scripts are considered.

The option to reject an example in order to avoid the risk of a costly potential misclassification is well-explored in the pattern recognition literature. In this paper, we look at this issue from the perspective of statistical learning theory. Specifically, we look at ways of modeling the problem of learning with an embedded reject option, in terms of minimizing an appropriately defined risk functional, and discuss the applicability thereof of some fundamental principles of learning, such as minimizing empirical risk and structural risk. Finally, we present some directions for further theoretical work on this problem.

A new impulsive noise elimination filter, entitled Resilient Neural Network based impulsive noise removing filter (RF), which shows a high performance at the restoration of images corrupted by impulsive noise, is proposed in this paper. The RF uses Chi-square goodness-of-fit test in order to find corrupted pixels more accurately. The corrupted pixels are replaced by new values which were estimated by using the proposed RF. Extensive simulation results show that the proposed filter achieves a superior performance to the other filters mentioned in this paper in the cases of being effective in noise suppression and detail preservation, especially when the noise density is very high.

Synthetic Aperture Radar (SAR) raw data are characterized by a high entropy content. As a result, conventional SAR compression techniques (such as block adaptive quantization and its variants) do not provide fully satisfactory performances. In this paper, a novel methodology for SAR raw data compression is presented, based on discrete wavelet transform (DWT). The correlation between the DWT coefficients of a SAR image at different resolutions is exploited to predict each coefficient in a subband mainly from the (spatially) corresponding ones in the immediately lower resolution subbands. Prediction is carried out by classical multi-layer perceptron (MLP) neural networks, all of which share the same, quite simple topology. Experiments carried out show that the proposed approach provides noticeably better results than most state-of-the-art SAR compression techniques.

Current technology allows the acquisition, transmission, storing, and manipulation of large collections of images. Images are retrieved basing on similarity of features where features of the query specification are compared with features from the image database to determine which images match similarly with given features. Feature extraction is a crucial part for any of such retrieval systems.In this paper we propose effective method for image representation which utilizes fuzzy features such as color and fuzzy radial moments.

Biometrics identification methods proved to be very efficient, more natural and easy for users than traditional methods of human identification. The future of biometrics leads to passive physiological methods based on images of such parts of human body as face and ear. The article presents a novel geometrical method of feature extraction from human ear images in order to perform human identification.

We propose a technique for detecting significant changes in a scene automatically, based on images acquired at different times. Compared to conventional luminance difference methods, the proposed technique does not require an arbitrarily-determined threshold for deciding how much change in pixel values amounts to a significant change in the scene. The technique can be used to detect the changes that occured in the scene, even when the images are of different spectral domains.

In this paper, we propose a vision-based approach to detect a face direction from a single monocular view of a face by using a facial feature called facial triangle. Specifically, the proposed method introduces formulas to detect face rotation, horizontally and vertically, using the facial triangle. Our method makes no assumption about the structure of the face and produces an accurate estimate of face direction.

In this paper a combination of algorithms useful for image compression standard is discussed. The main algorithm, named predictive vector quantization (PVQ), is based on competitive neural networks quantizer and neural networks predictor. Additionally, the noiseless Huffman coding is used. The experimental results are presented and the performance of the algorithm is discussed.

Many correlation measures have been already proposed for image matching. The special group with quite different statistical properties constitute the nonparametric measures. Their virtue in the task of image matching lies mostly in the known distribution function and resistance against local image fluctuations and outliers. In this paper the fuzzy enhancement of the nonparametric measures is proposed. It allows for better representation of the local relations among image pixels. The presented concepts are underpinned by many experiments which results are also provided and discussed.

The fundamental matrix combines the mutual relation of the corresponding points in the two images of an observed scene. This relation, known also as an epipolar geometry, allows for a further depth reconstruction, image rectification or camera calibration. Thus, computation of the fundamental matrix has been one of the most important problems of computer vision. Many linear and non-linear methods were already proposed to solve this problem. However, due to the nature of image processing there is no unique solution and each method exhibits some exclusive properties. In this paper a neural approach to the computation of the fundamental matrix is proposed. For this purpose, the special configuration of the back-propagation neural network was developed. Both, linear and non-linear versions are also discussed.

We present a new method based on CMAC neural network, used as classifier in a frontal face detection system. The gray level and the position of the pixels of an input image are directly presented to the network. Due to the simple structure of CMAC, with only one trainable layer, the training phase is very fast. The proposed method has been tested on a data set containing 960 faces and 20000 non-faces, selected among difficult face and non-face patterns. The results of experimentations exhibit an error rate of 8.5%, which is a reasonable result considering the simple structure of system and the important number of difficult patterns in the test dataset.

We propose a new active stereo vision system using a human-like vergence control method. The proposed system uses a bottom-up saliency map model with a human-like selective attention function in order to select an interesting region in each camera. This system compares the landmarks as to whether the selective region in each camera finds the same region. If the left and right cameras successfully find the same landmark, the implemented vision system focuses on that landmark. Using motor encoder information, we can automatically obtain depth data even when occlusion problem occurs. Experimental results show that the proposed convergence method is very effective in implementing an active stereo system and it can also can be applied to a visual surveillance system for discriminating between a real human face and a photograph.

One of possible solutions in creating an automatic system of face recognition is application of auto-associative neural networks for remembering and recognise two-dimensional face images. Experiments with applying the Hopfield network and twolayer perceptron confirmed the possibility of remembering and reproducing face images, even if partially covered or disturbed. Limited technical possibilities enable using only low definition images. It is due to the fact that computations take a long time, and the number of remembered faces is relatively small. Proper working of the network is influenced in a significant way by light and facial expressions.

Traditionally, in face authentication/identification methods the presumption concerning face/head symmetry is used. For novel applications that concern creating techniques by means of which it is possible to reproduce the extraordinary complexity of skin, muscle, eye and hair movements, which convey emotion, gesture, psychological state or psycho-sociological traits, we begin to create new research direction called Biometrics of Asymmetrical Face.In the paper, it is presented a novel type of 2D precise normalized model of a face for modern and prospect authentication/identification techniques creation in live biometrics that called Czestochowa-face model. Result of creating the ophthalmogeometrical technique, which is based on Czestochowa-face model, is also given. Some attention has been drawn to interdisciplinary research context. Beside the first-hand usage, the new type of face model and the technique may be employed in the areas of human-computer interaction, identification of cognition-psyche type of personality for personnel management, and so on.

In this paper, we will introduce a method for wafer die position detection using corner detector. We present a hierarchical gray level corner detector (HGLCD) to detect die position accurately. HGLCD divides the corner region into many homocentric circles and get corner response and angle about each circle. The experiments have shown that the new corner detector is more accurate and more efficient in its performance than other two popular corner detectors.

One of the monolithic goals of Computer Vision (CV) is to automatically interpret general digital images of arbitrary scenes. Although this goal has produced a vast array of research, a solution to the general problem has not been found. The difficulty of this goal has caused the field to focus on smaller, more constrained problems related with the different tasks involved, such as: noise removal, smoothing, and sharpening of contrast -low-level-; segmentation of images to isolate objects and regions, and description and recognition of the segmented regions -intermediate-level-; and interpretation of the scene -high-level-.

In the paper a multi-layer neural network and its application to texture segmentation is presented. The generalized network is built using two types of elements: CU – clustering units and DCB – data completion blocks. Clustering units are composed of Kohonen networks. Each Kohonen network is a self-organizing map (SOM) trained to be able to distinguish, in an unsupervised way, certain clusters in the input data. Data completion blocks are placed between CU and their aim is to prepare data for the CU. This paper presents a sample application of a double-layer network to automatic texture segmentation. The method has been evaluated on both artificial and real images, and the results achieved are presented.

In the paper multi-robot environment with STRIPS representation is considered. Under some assumptions such problems can be modelled as a STRIPS system (for instance Block World environment) with one initial state and disjunction of goal states. If STRIPS planning problem is invertible then it is possible to apply machinery for planning in the presence of incomplete information to solve the inverted problem and then to find a solution for the original problem. To reduce computational complexity of this approach a transformation to Linear Programming problem is proposed. Simulations illustrate the reduced problem.

This paper proposes a sensor based navigation method with fuzzy combiner for navigation of mobile robot in uncertain environments. The proposed navigator consists of two main behaviors: a reaching the middle of a collision-free space behavior, and goal-seeking behavior. The fuzzy combiner can fuse low-level behaviors so that the mobile robot can go for the goal position without colliding with obstacles. The fuzzy combiner is a soft switch that chooses more then one low-level action to be active with different degrees through fuzzy combination at each time step. The output of the navigation level is fed into a fuzzy tracking controller that takes into account the dynamics of the mobile robot. Computer simulation have been conducted to illustrate the performance of the proposed fuzzy combiner of behaviors by a series of experiments on the emulator of wheeled mobile robot Pioneer-2DX.

In this paper, it is indicated that the decision unit of mobile robot should be performed in the form of a hardware which owns some features of intelligence. For this reason, the synthesis of the decision unit should be executed in such a way in order to achieve these features in its logical structure. In the paper, the process of robot moving, over the plane with obstacles, is treated as an extensive game with the nature. In this game the decision unit chooses its moving strategies in the same manner as a human being in the identical situation. The synthesis of a symbolic expression representing the game tree, performing by the computer, is introduced. In the symbolic expression some features of intelligence are taken into consideration. It is shown that this symbolic expression is transformed by the computer into another symbolic expressions which unequivocally indicates on the logical structure of the hardware playing the role of the decision unit.

This paper considers the problem of mobile robot localization. The localization is done using a particle filter built on a highly accurate probabilistic model of laser scan and a histogram based representation of sensor readings. A histogram matching exploits sensor data coming from the laser and data obtained from the existing map. Experimental results indicate feasibility of the proposed approach for navigation.

There is presented the design of the feedforward neural network for calculation of coefficients of the robot model. Proposed method distinguishes the degrees of freedom and improves the performance of the network using information about the control signals. A numerical example for calculation of the neural network model of Puma 560 robot is presented.

The paper addresses the problem of a real time collision-free movement coordination in a multi-robot environment. An architecture of the control system that is designed to control a movement of a team of mobile robots performing their navigational tasks is presented. The proposed approach to the problem of coordination utilize the normal form games. The cooperative concept of solution that provides a “fair” distribution of costs is used. An arbiter was introduced to provide unique solution when multiple ones exist. Results of simulation of the proposed method, made for two robots are presented.

In this study, the application of the single input single output (SISO) neural generalized predictive control (NGPC) of a three joint robotic manipulator with the comparison of the SISO generalized predictive control (GPC) is presented. Dynamics modeling of the robotic manipulator was made by using the Lagrange-Euler equations. The frictional effects, the random disturbance, the state of carrying and falling load were added to dynamics model. The sinusoidal trajectory principle is used for position reference and velocity reference trajectories. The results show that the NGPC-SISO algorithm performs better than GPC-SISO algorithm and the influence of the load changes and disturbances to the NGPC-SISO is less than that of the GPC-SISO with sinusoidal trajectory.

An agent-oriented architecture for personalized information system is considered in this paper. This architecture is composed of four types of agents and is based on FIPA standard. The emphasis is put on performance evaluation, which is based on system dynamics. In FIPA, the most extensive models of dynamics are statecharts. Performance statecharts are such statecharts that are extended by probability distributions of activity duration times and probability distribution for solving non-determinism. Performance analysis of the personalized information system using performance statecharts is carried over in the paper.

We present a neural approach to worm detection designed as a part of a multi-agent system intended to manage IP networks. The efficiency of virus recognition is about 95%.

In the paper a hybrid, agent-based system of evolving neural networks dedicated to time-series prediction is presented. First the idea of a multi-agent predicting system is introduced. Second some aspects of the system concerning management of collective intelligence and evolutionary design of a predicting neural network are discussed. Then a hybrid solution – an evolutionary multi-agent system (EMAS) is proposed. Finally selected results of the experiments are presented.

In this paper, I propose a Personalized Digital Library System (PDLS) based on a new mobile multi agent platform. This new platform is developed by improving the DECAF (Distributed Environment-Centered Agent Framework) which is one of the conventional distributed agent development toolkits. Also, a mobile ORB (Object Request Broker), Voyager, and a new multi agent negotiation algorithm are adopted to develop the new platform. The new mobile multi agent platform is for mobile multi agents as well as the distributed environment, whereas the DECAF is for the distributed and non-mobile environment. From the results of the simulation, the searched time of PDLS is lower, as the numbers of servers and agents are increased. And the user satisfaction is four times greater than the conventional client-server model. Therefore, the new platform has some optimality and higher performance in the distributed mobile environment.

This paper introduces FOOD, a new Dataflow model that goes beyond the limitations of the existing models, and targets the implementation of multi-agent systems. The central notion of FOOD, the Dataunit, which expresses all the dynamics that are required in multi-agent systems, is presented in details. Then it is shown how the FOOD model, despite its simplicity, has an expression power that goes beyond object-oriented languages in term of dynamics and mutability.

The paper presents an agent-based architecture facilitating implemetation of parallel evolutionary algorithms, utilising the novel concept of a flock. The model proposed is an extension to classical regional parallel evolutionary algorithm. Flocks introduce additional level of organisation of the system, allowing for separation of distribution and evolution issues, and thus opening possibility of dynamic reconfiguration of subpopulations adequately to the structure of the problem being solved. Selected experimental results illustrate the idea “at work”.

This paper describes a framework for a statistical anomaly prediction system using Quickprop neural network forecasting model, which predicts unauthorized invasions of user based on previous observations and takes further action before intrusion occurs. The experimental study is performed using real data provided by a major Corporate Bank. A comparative evaluation of the Quickprop neural network over the traditional neural network models was carried out using mean absolute percentage error on a prediction data set and a better prediction accuracy has been observed. Further, in order to make a legitimate comparison, the dataset was divided into two statistically equivalent subsets, viz. the training and the prediction sets, using genetic algorithm.

New elements of the parallel simulated annealing method are proposed to solve the permutation flow shop scheduling problem with the criterion of total completion time (F*||C _Sum ). This problem is more difficult to optimize than $F^{*}||C_{\rm max}$ (minimizing the makespan). Simulated annealing belongs to the artificial intelligence methods, which are commonly used to solve NP-hard combinatorial optimization problems. In the parallel algorithm, we propose a new acceptance probability function, multi-step technique, dynamic long-term memory and backtrack-jump. Computational experiments (for Taillard’s benchmarks ta001-ta050, [8]) are given and compared with results yielded by the best algorithms discussed in literature [10]. We also present new referential solutions for ta051-ta080 instances (published on our benchmark page [1]), which so far have no solutions.

An algorithm of simulated annealing for the job shop scheduling problem is presented. The proposed algorithm restarts with a new value every time the previous algorithm finishes. To begin the process of annealing, the starting point is a randomly generated schedule with the condition that the initial value of the makespan of the schedule does not surpass a previously established upper bound. The experimental results show the importance of using upper bounds in simulated annealing in order to more quickly approach good solutions.

The advent of the Internet has strongly influenced modern software systems. Existing intranet solutions are being gradually replaced with www services available everywhere and at any time. The availability of the wide area network has resulted in unprecedented opportunities of remote and distributed cooperation of large groups of people and organizations.Originally expert systems have been used for internal purposes in different organizations. The purpose of this work is to summarize the results of the project aiming to revise the functionality and architecture of traditional expert systems in terms of modern trends in web-based systems. Both knowledge representation, development of the knowledge base and interface solutions have been reconsidered. The proposed solutions are based on the pilot implementation of the internet expert system.

The paper introduces an approach to natural language processing based on perceiving natural language as a tool of human communication, tool describing subjects of this communication. This approach reveals parallel syntactic and semantic attempts to natural language processing. Both attempts are reflected in the paradigms of information granulation and granular computing. Duality of granular spaces derived from syntax and semantic of natural language statements is considered. The directions for future research are outlined

Model checking of temporal logic has already been proposed for automatic planning. In this paper, we introduce a simple adaptation of the ATL model checking algorithm that returns a strategy to achieve given goal. We point out that the algorithm generalizes minimaxing, and that ATL models generalize traditional game trees. The paper ends with suggestions about other game theory concepts that can be transfered to ATL-based planning.

In this paper we want to draw Reader’s attention to the issue of impact of separate measurement of features (attributes) from which we want to make inferences. It turns out, that the fact of separate measurements implies algorithmic simplifications for many forms of reasoning in DST. Basic theorems and algorithms exploiting this are given.

The paper deals with ideas of decision support with use of computer based systems. A negotiation problem is considered related to joint realization of a risky innovative project by two parties. The utility function approach and cooperative game concepts are utilized in the decision support. An experimental system has been constructed and selected computational results are presented.

This paper presents an application of ant colony optimisation and genetic algorithm to rescue operation planning. It considers the task as the multiple travelling salesmen problem and proposes suitable heuristics in order to improve the performance of the selected techniques. Then it applies the implemented solutions to a real data. The paper concludes with comparison of the implementations and discussion on the aspects of the utilisation of the proposed heuristics.

A tabular system takes the form of a table with columns described by attributes and rows specifying rules. Systems with non-atomic values are considered since they expressive power is much higher than the one of classical attributive decision tables. Such system can reduced to concise form through gluing of rows with similar values. Efficient reduction can lead to minimization, which can be performed with or without overlapping. The paper gives the idea of such reduction, provides two algorithms and presents some numerical results.

In this paper we examine the application of temporal difference methods in learning a linear state value function approximation in a game of give-away checkers. Empirical results show that the TD(λ) algorithm can be successfully used to improve playing policy quality in this domain. Training games with strong and random opponents were considered. Results show that learning only on negative game outcomes improved performance of the learning player against strong opponents.

This paper describes the results of applying artificial neural networks to the double dummy bridge problem. Several feedforward neural networks were trained using resilient backpropagation algorithm to estimate the number of tricks to take by players NS in fully revealed contract bridge deals. Training deals were the only data presented to the networks. The best networks were able to perfectly point the number of tricks in more than one third of deals and gained about 80% accuracy when one trick error was permitted. Only in less than 5% of deals the error exceeded 2 tricks.

Our discussion in this article centers on the application of ant algorithms to the non-bifurcated multicommodity flow problem. We propose a general framework of ant algorithm that can be applied to the design of static flows in connection-oriented computer networks. Next, through numerical simulation, we study the influence of algorithm’s parameters setting on the quality of solutions. We compare and discuss two variants of the algorithm: without and with feasible initial solution.

During modeling protein structure prediction, it is a fundamental operation and often as a preprocess of in specific tasks that a very large categorical data sets are partitioned into disjoint and homogeneous clusters. The classical k-modes algorithm is a partial solution to such problems. This work presents a parallel implementation of the k-modes algorithm based on the message passing model. The proposed algorithm exploits the inherent data-parallelism in the k-means style algorithm. Tested with the amino acid data sets on a maximum of 8 nodes the algorithm has demonstrated a very good relative speedup and scaleup in the size of the data set.

Algorithms of reinforcement learning usually employ consecutive agent’s actions to construct gradients estimators to adjust agent’s policy. The policy is then the result of some kind of stochastic approximation. Because of slowness of stochastic approximation, such algorithms are usually much too slow to be employed, e.g. in real-time adaptive control.In this paper we analyze replacing the stochastic approximation with the estimation based on the entire available history of an agent-environment interaction. We design an algorithm of reinforcement learning in continuous space/action domain that is of orders of magnitude faster then the classical methods.

Recently, we observe rapid growth of methodologies and their applications in creating 3D virtual object structures. The aim of this paper is to carry out the solid modeling ensures better flexibility of the software in creating the 3D structures of intelligent Micro-ElectroMechanical ( MEMS ). This paper deals with modelling of 3D structure of surface micromachined accelerometers. Keeping knowledge about predefined accelerometer structure we start optimisa-tion procedure exploiting deterministic optimisation technique.

This paper aims to inform about special diagnostic models designed for reasoning about causes of observed symptoms, e.g. a state of an object. The paper introduces the general idea of inverse models produced by numerical inversion of simulation results, where the simulation is being run by known cause-effect models transforming state features into diagnostic symptoms. Special attention has been drawn to a strict and interval modelling as well as to global and local models. Suggested methodology is focused on the identification of local models such as pattern based models, i.e. models spanned on particular sets of selected data. Although the presented approach is addressed to diagnostics it may be easily extended to other applications.

This paper presents an efficient neural network for solving constrained nonlinear optimization problems. More specifically, a two-stage neural network architecture is developed and its internal parameters are computed using the valid-subspace technique. The main advantage of the developed network is that it treats optimization and constraint terms in different stages with no interference with each other. Moreover, the proposed approach does not require specification of penalty or weighting parameters for its initialization.

In this paper, the design of Fuzzy Sliding Mode Control (FSMC) based on Genetic Algorithms (GAs) is studied. Using the proposed approach, a robust controller is designed for building structures under earthquake excitation. It is found that the building structure under the proposed control method could sustain in safety and stability when the system is subjected to external disturbances.

In the paper three neural networks state variables estimators of the induction motor are considered, which recreate rotor angular speed, rotor flux and stator current components in the rotor flux reference frame. Input variables for the neural estimators are the components of stator current and voltage to allow for sensor less control of induction motor drive. Performance of the estimators is compared for the networks trained using static, dynamic and mixed sets of data. Intention of the analysis is to find the best way the training data are obtained that assures possibly high accuracy of the estimators.

This paper considers the problem of designing optimal stabilizing controllers for the systems that can be modelled by the Takagi-Sugeno (TS) fuzzy model. Contrary to difficult optimal control problems dealing with fixed cost functions directly, we pursue the strategy in which the cost function is determined during the design process. This approach makes the problem easy to solve and yields stabilizing controllers which satisfy the inherent robustness of optimal controllers. The design procedure of this paper consists of solving LMIs (linear matrix inequalities). The applicability of the proposed method is illustrated via an example.

The main objective of this paper is to present a new method based on Multiobjective evolutionary algorithm for control of the multivariable and nonlinear systems. Problem design considers time domain specifications such as overshoot, rising time, settling time and stationary error as well as interaction effects. Genetic algorithms are employed to satisfy time domain design specifications, that are not considered in an explicit way in the standard nonlinear control theory. Adaptation, setpoint tracking and satisfaction of temporary response specifications are the advantages of this method that be shown by some simulations.

Heating, Ventilating and Air Conditioning (HVAC) plant, is a multivariable, nonlinear and non minimum phase system, that its control is very difficult. In this paper, we apply a robust sliding mode fuzzy controller to HVAC system. Our proposed method can achieve very robust and satisfactory performance and could be used to get the desired performance levels. The response time is also very fast despite the fact that the control strategy is based on bounded rationality. To evaluate the usefulness of the proposed method, we compare the response of this method with PID controller. The simulation results show that proposed method has the better control performance than PID controller.

In this paper, a problem of input-output complex systems identification is presented. The description of a complex system is given by a description of each system element and structure. Local and global identification problems are formulated, and a different approach to the identification problem is discussed. Based on the multi-criteria concept, the globally optimal model with respect to the quality of local models is presented.

Melanoma is a very dangerous skin cancer. In this paper we present results of experiments on three melanoma data sets. Two data mining tools were used, a new system called IRIM (Interesting Rule Induction Module) and well established LEM2 (Learning from Examples Module, version 2), both are components of the same data mining system LERS (Learning from Examples based on Rough Sets). Typically IRIM induces the strongest rules that are possible for a data set. IRIM does not need any preliminary discretization or preprocessing of missing attribute values. Though performance of IRIM and LEM2 is fully comparable, IRIM provides an additional opportunity to induce unexpected and strong rules supplying an important insight helpful for diagnosis of melanoma.

This paper presents an efficient technique for the detection of spiculated massesin the digitized mammogram to assist the attending radiologist in making his decisions. The presented technique consists of two stages, enhancement of spiculation masses followed by the segmentation process. Fuzzy Histogram Hyperbolization (FHH) algorithm is first used to improve the quality of the digitized mammogram images. The Fuzzy C-Mean (FCM) algorithm is then applied to the preprocessed image to initialize the segmentation. Four measures of quantifying enhancement have been developed in this work. Each measure is based on the statistical information obtained from the labelled region of interest and a border area surrounding it. The methodology is based on the assumption that target and background areas are accurately specified. We have tested the algorithms on digitized mammograms obtained from the Digital Databases for Mammographic Image Analysis Society (MIAS).

In the paper an artificial neural network designed for prediction of survival time is presented. The method aims at identifying areas in the feature space homogeneous from the point of view of survival experience. The proposed method is based on minimization of a piece-wise linear function. Appropriately designed dipolar criterion function is able to cope with censored data. Additional pruning phase prevents the network from over-fitting.

This paper presents decision algorithm based on fuzzy relation developed for the multistage pattern recognition. In this method – assuming that the learning set is given – first we find fuzzy relation in the product of feature and decision space as solution of an optimisation problem and next this relation is used in decision algorithm. The application of presented method to the computer-aided diagnosis of myocardial infraction is discussed and compared with algorithms based on statistical model.

In this article Sequential Minimal Optimization (SMO) approach as the solution of Support Vector Machines (SVMs) algorithm is applied to the ovarian cancer data classification problem. The comparison of different SVM models is presented in order to determine the chance of 60 months survival for a woman to be treated ovarian cancer. A cross-validation procedure is used for this purpose.

As fetal hypoxia may damage or kill the fetus, it is very important to monitor the infant so that any signs of fetal distress can be detected as soon as possible. In this paper, the performances of some artificial neural networks are evaluated, which eventually produce the suggested diagnosis of fetal hypoxia. Multilayer perceptron (MLP) structure with standard back propagation, MLP with fast back propagation (adaptive learning and momentum term added), Radial Basis Function (RBF) network structure trained by orthogonal least square algorithm, and Conic Section Function Neural Network (CSFNN) with adaptive learning were used for this purpose. Further more, Receiver Operating Characteristic (ROC) analysis is used to determine the accuracy of diagnostic test.

Computational diagnosis tools are becoming indispensable to support modern medical diagnosis. This research work introduces an hybrid soft computing scheme consisting of Fuzzy Cognitive Maps and the effective Active Hebbian Learning (AHL) algorithm for tumor characterization. The proposed method exploits human experts’ knowledge on histopathology expressed in descriptive terms and concepts and it is enhanced with Hebbian learning and then it classifies tumors based on the morphology of tissues. This method was validated in clinical data and the results enforce the effectiveness of the proposed approach.

In this paper, a multi-stage classification method is applied to a problem of larynx cancer diagnosis. The biochemical tumor markers, called CEA and SCC, as well as ferritin, and other factors, are used in order to produce the diagnosis. A neuro-fuzzy network is employed at every stage of the classification method. The networks reflect fuzzy IF-THEN rules, formulated based on the data containing measurements of the particular factors (attributes). The classification method is proposed to support a medical doctor decision, and provide additional useful information concerning the diagnosis.

Dealing with expert (human) knowledge consideration, the computer aided medical diagnosis dilemma is one of most interesting, but also one of the most difficult problems. Among difficulties contributing to the challenging nature of this problem, one can mention the need of fine classification. In this paper, we present a new classification approach founded on a tree like neural network based multiple-models structure, able to split a complex problem to a set of simpler sub-problems. This new concept has been used to design a Computer aided medical diagnostic tool that asserts auditory pathologies based on Brain-stem Evoked Response Auditory based biomedical test, which provides an ef-fective measure of the integrity of the auditory pathway.

In the paper the results of application of the neural network algorithm to identify the melanocytic spots are described. The algorithm was used to create, learning and making diagnoses on the ground of three bases containing the cases described in terms of fourteen attributes. The SNN (Statistica Neural Networks) was used to create the best neural networks and to find out of the hierarchy importance of descriptive attributes in computer assisted classification of melanocytic skin lessons.

The paper proposes a method of knowledge and certainty representation that is an alternative to fuzzy reasoning and classical probability based techniques. It makes possible to represent symptoms of different nature and is acceptable for physicians. The proposed solutions are based on the Dempster-Shafer theory of evidence, still, different definition for focal elements is proposed and specific method of basic probability assignment calculation for symptoms is suggested. It has been shown on examples of thyroid gland diseases diagnosis support that the method is efficient and numerically easy.

Myocardial ischemia is caused by a lack of oxygen and nutrients to the contractile cells and may lead to myocardial infarction with its severe consequence of heart failure and arrhythmia. An electrocardiogram (ECG) represents a recording of changes occurring in the electrical potentials between different sites on the skin as a result of the cardiac activity. Since the ECG is recorded easily and non–invasively, it becomes very important to provide means of reliable ischemia detection. Ischemic changes of the ECG frequently affect the entire repolarization wave shape. In this paper we propose a new classification methodology that draws from the disciplines of clustering and artificial neural networks, and apply it to the problem of myocardial ischemia detection. The results obtained are promising.

The aim of the paper is to present the application of the sequential and distributed evolutionary algorithms to selected structural optimization problems. The coupling of evolutionary algorithms with the finite element method and the boundary element method creates a computational intelligence technique that is very suitable in computer aided optimal design. Several numerical examples for shape, topology and material optimization are presented.

Two problems related to the concrete fatigue durability analysis, originated in papers [2,3] and based on the data banks taken from [1], are developed. The first problem deals with the neural simulation of the durability of concrete, defined as a number of cycles of compressive stresses N. The second problem corresponds to the identification of concrete strength associated with N. In the paper a neuro-fuzzy network FWNN (Fuzzy Weight NN), formulated in [4], was used to obtain better results than approximations by the standard (crisp) Back-Propagation NNs.

The paper is a continuation of [4], where a feed-forward neural network was used for generating samples in the Monte Carlo methods. The patterns for network training and testing were computed by an FEM program. A high numerical efficiency of neural generating MC samples does not correspond to the much more time consuming FEM computation of patterns. This question and an evaluation of the number of random inputs is discussed in the presented paper on an example of plane steel frame, called in [5] a calibrating frame.

The paper presents the application of Artificial Neural Networks (ANNs) for solution of an inverse problem [1]. Based on the dynamic characteristics of a plate, the neural identification of parameters of circular hole and additional mass have been performed. An emphasis was placed on the effective preparation of learning data, which were produced both by the finite element method and by experiment.

New methods of filtering of experimental data curves, based on the artificial neural networks and the wavelet analysis are presented in the paper. The thermomagnetic data curves were filtered using these methods. The obtained results were validated using the modified algorithm of the cubic spline approximation.

This paper presents an evolutionary negotiating agent to conduct negotiation tasks between power generating and consuming companies in electricity markets. The agent select the best negotiation strategy that meets the underlying company objectives and interests. It generates a sequence of improving strategy population as the outcome of a search modeled by the selection, crossover, and mutation genetic operators. Agents use a content specification language based on an extended object model to specify the requirements, constraints, and negotiation strategic rules, which are used by the negotiation server to conduct a negotiation. A design architecture for negotiation is presented with KQML communication primitives. Various software technologies have been used for implementation and tested in a C++ environment.

The article discusses a problem of time scheduling by evolutionary algorithm of combined heat and electric power production (CHP – Combined Heat and Power)in urban heat distribution network, aimed at securing maximum profit during a given time period. The developed time schedule takes into account the dynamics of the power station. The formulated model allows simulations of power plant’s co-operation with the heat distribution network and the heat collection tank.

Nonlinear state estimation is a useful approach to the monitoring of industrial (polymerization) processes. This paper investigates how this approach can be followed to the development of a soft sensor of the product quality (melt index). The bottleneck of the successful application of advanced state estimation algorithms is the identification of models that can accurately describe the process. This paper presents a semi-mechanistic modeling approach where neural networks describe the unknown phenomena of the system that cannot be formulated by prior knowledge based differential equations. Since in the presented semi-mechanistic model structure the neural network is a part of a nonlinear algebraic-differential equation set, there are no available direct input-output data to train the weights of the network. To handle this problem in this paper a simple, yet practically useful spline-smoothing based technique has been used. The results show that the developed semi-mechanistic model can be efficiently used for on-line state estimation.

The problem of time-frequency decomposition of signals by means of neural networks has been investigated. The paper contains formalization of the problem as an optimization task followed by a proposition of recurrent neural network that can be used to solve it. Depending on the applied base functions, the neural network can be used for calculation of several standard time-frequency signal representations including Gabor. However, it can be especially useful in research on new signal decompositions with non-orthogonal bases as well as a part of feature extraction blocks in neural classification systems. The theoretic considerations have been illustrated by an example of analysis of a signal with time-varying parameters.

The paper presents a new method for modelling of non-linear temperature sensor and correction of its dynamic errors by means of Artificial Neural Networks (ANNs). Feedforward multilayer ANNs with a moving window method and recurrent ANNs were applied. In the proposed correction technique an inverse dynamic model of the sensor is implemented by means of a neural corrector. ANN based modelling and correction technique has been evaluated experimentally for small platinum RTD immersed in oil. Recurrent ANN was used as a simulator for modelling sensor’s non-linear dynamics and to validate the correction technique.

A task of a stock index prediction is presented in this paper. Several issues are considered. The data is gathered at the concerned stock market (NIKKEI) and two other markets (NASDAQ and DAX). The data contains not only original numerical values from the markets but also indicators pre-processed in terms of technical analysis, i.e. the oscillators are calculated and the structures of a value chart are extracted. Selected data is input to a neural network that is functionally divided into separate modules. The prediction goal was next day opening value of Japanese stock market index NIKKEI with consideration of German and USA stock markets’ indexes. The average prediction error on the test set equals 43 points and the average percentage prediction error is equal to 0.27% while the average index volatility equals 0.96%.

The authors describe the structure of an Intelligent Prediction System (IPS) basing on neural networks simulated in the STATISTICA Neural Networks artificial intelligence package. The system has been adopted in the process of predicting regional heat use. The authors focus on the methods applied to select optimum neural networks in the prediction mode of the analysed phenomenon and present prediction results for the best networks.

In this paper we propose a new method for improving efficiency of database systems, using genetic algorithms. We show how table indexes can be coded in chromosomes, and a way to automatically find the best set of indexes for databases.

The paper presents a study of utility of two representation types: the log-polar representation and histograms of individual color components of grain surface images for recognition of wheat infected by the Fusarium genus fungi. The representations have been used as input data for a backpropagation type neural networks of various sizes of the hidden layer and – as a reference – for a Nearest Neighbor classifier. The best individual recognition rate has been obtained for the log-polar representation of Blue color component (98.8%). Quality assessment has been also done (in the sense of robustness of results obtained from neural networks of various structures).

Paper presents method of construction neural model of the sea bottom. Constructed model consisted of set of smaller models (local approximators). For model approximation was used set of RBF networks with various kernels, what enabled approximation of the entire model by networks with different structure. Experimental results show that in this way we can obtain better results than applying neural model based on local approximators with the same structure.

The paper presents fuzzy economic analysis of discrete transport systems in a function of system element reliability. Such approach can be used for a description of different problems related to transport logistics. No restriction on the system structure and on a kind of distribution is the main advantage of the proposed solution. Additionally the results of reliability and functional analysis can be used as a basis for economic aspects discussion related to discrete transport system. Fuzzy approach allows to reduce the problem of assumptions of reliability distributions.

Economic sanctions have been a prominent part of American statecraft since World War II, and increasingly so since the end of the Cold War. Sanction usually consists of a ban on the sale and shipment of products to a country and on the purchase of its exports. Our objective is to conduct a comprehensive empirical analysis of US economic sanctions on Iranain high tech industries by means of fuzzy logic approach. We will measure US sanction effects on Iranian high tech industries and rank high Iranian tech industries based on their vulnerability from US sanction. Our methodology is based on these steps: determining attributes of US sanction affected on Iranain high techs’, filtering selected attributes through fuzzy hypothesis test, and ranking of alternatives through Fuzzy Analytic Hierarchical Process (FAHP).

The advantages offered by the electronic component LED (Light Emitting Diode) have caused a quick and wide application of this device in replacement of incandescent lights. However, in its combined application, the relationship between the design variables and the desired effect or result is very complex and it becomes difficult to model by conventional techniques. This work consists of the development of a technique, through comparative analysis of neuro-fuzzy architectures, to make possible to obtain the luminous intensity values of brake lights using LEDs from design data.

Modeling of an absorption column performance using feed-forward type of neural network has been presented. The input and output data for training of the neural network are obtained from a rigorous model of the absorption column. The results obtained from the neural network models are then compared with the results obtained mainly from the simulation calculations. The results show that relatively simple neural network models can be used to model the steady state behavior of the column.

The article presents methods of computer ship’s position plotting by means of comparative methods. A new approach in comparative navigation is the use of artificial neural networks for plotting the ship’s position. Two main problems should be solved during ship’s positioning process: compressing (coding) image and recognition (interpolation) ship’s position.

In this research, forecasting models were built based on both univariate and multivariate analysis. Models built on multivariate fuzzy logic analysis were better in comparison to those built on other models. The performance of the models was tested by comparing one of the goodness-of-fit statistics, R2, and also by comparing actual sales with the forecasted sales of different types of garments. Five months sales data (August-December 2001) was used as back cast data in our models and a forecast was made for one month of the year 2002. The performance of the models was tested by comparing one of the goodness-of-fit statistics, R2, and also by comparing actual sales with the forecasted sales. An R2 of 0.93 was obtained for multivariate analysis (0.75 for univariate analysis), which is significantly higher than those of 0.90 and 0.75 found for Single Seasonal Exponential Smoothing and Winters’ three parameter model, respectively. Yet another model, based on artificial neural network approach, gave an R2 averaging 0.82 for multivariate analysis and 0.92 for univariate analysis.

In this paper we propose applying multidimensional decompositions for modeling improvement. Results generated by different models usually include both wanted and destructive components. Many of the components are common to all the models. Our aim is to find the basis variables with the positive and the negative influence on the modeling task. It will be perofrmed with multidimensional transforamtions such as ICA and PCA.