Contents of Volume 13 (2003)

5/2003 4/2003 2/2003 2/2003 1/2003


  • [1] Dündar P. (Turkey): Tenacity of the thorny graphs of static interconnection networks, 591-598.

    When a network begins losing nodes or links there is, eventually, a loss in its effectiveness. Thus, a communication network must be constructed to be as stable as possible, not only with respect to the initial disruption, but also with respect to the possible reconstruction of the network. When any disruption happens in a communication network two questions are considered: How many vertices can still communicate? How difficult is it to reconnect the network? If a graph is considered as a modeling network, then the above questions can be answered by the graphs. Many graph parameters have been used to describe the stability of communication networks, including connectivity, integrity, and toughness and the binding number. The thorny graphs are special classes of graphs that represent some static interconnection networks. In this work, we have given the tenacity of thorny graphs of static interconnection networks.

  • [2] Grim J., Just P., Pudil P. (Czech Republic): Strictly modular probabilistic neural networks for pattern recognition, 599-615.

    Considering the statistical recognition of multidimensional binary observations we approximate the unknown class-conditional probability distributions by multivariate Bernoulli mixtures. We show that both the parameter optimization and the resulting Bayesian decision-making can be realized by a probabilistic neural network having strictly modular properties. In particular, the process of learning based on the EM algorithm can be performed by means of a sequential autonomous adaptation of neurons involving only the information from the input synapses and the interior of neurons. In this sense the probabilistic neural network can be designed automatically. The properties of the sequential strictly modular learning procedure are illustrated by numerical examples.

  • [3] Volná E. (Czech Republic): Emergence of modularity in evolved neural networks, 617-628.

    A novel method is described that allows us to study the emergence of the modular neural network structure through evolution. A preliminary design of modular neural networks is developed by evolutionary algorithm. The concept of emergence takes an important role in the study of the design of neural networks. The model presented in this paper might not only develop new functionality spontaneously but it could also grow and evolve its own structure autonomously. Network architecture emerges from an initial set of randomly connected networks.

  • [4] Nováková K., Kukal J. (Czech Republic): Database model of arithmetic network, 629-662.

    Arithmetic networks consist of neural, Boolean and fuzzy ones. Supposing the acyclic structure, decomposition of arithmetic network is possible. There are three results of our analysis: node unification, edge unification and network decomposition. We obtain only 14 node types and 4 edge types for realization of a wide class of traditional arithmetic networks from literature. The main result of our work is the splitting of the competitive neurons (nodes) to distance and soft extreme nodes. The side result of analysis is using the group of nodes instead of layer. It enables grouping the nodes of the same type but with the possibility of long interconnections. The main aim of our work was to realize the system of arithmetic networks in the SQL language on any SQL server. The database realization enables not only saving, watching and editing the network structures and parameters but also studying the response of archived networks. The learning process was not included because of being iterative in general and unrealizable without loops on database server at that time.

  • [5] Loo C. K., Rao M. V. C., Rajeswari M. (Malaysia): Growing Multi-Experts Network, 663-686.

    An endeavour is made in this paper to describe a constructive modular neural network called Growing Multi-Experts Network (GMN), which can approximate to us an unknown nonlinear function from observed input-output training data. In the GMN, the problem space is decomposed into overlapping regions by an expertise domain and the local expert models are graded according to their expertise level. The network output is computed by the smooth combination of local linear models. On the other hand, in order to avoid over-fitting problems, the GMN deploys a Redundant Experts Removal Algorithm to remove the redundant local experts from the network. In addition, a Growing Neural Gas algorithm is used to generate an induced Delaunay triangulation that is highly desired for optimal function approximation. The GMN is tested by four benchmark problems to compare its performance with other modeling approaches. The performance of the GMN compares favorably with the existing techniques. Thus, it seems to be extremely promising to determine an optimal structure of the network with a lot of potentials to be exploited.

  • [6] Book review, 687-689.
  • [7] Contens volume 13 (2003), 691-693.
  • [8] Author's index volume 13 (2003), 695-697.


  • [1] Editorial, 453.

    Invited papers

  • [2] Fermüller C. G. (Austria): Theories of vagueness versus fuzzy logic: can logicians learn from philosophers?, 455-466.

    Motivated by the fact that logicians and computer scientists working in fuzzy logic hardly seem to take notice of the prolific and broad discourse on vagueness in analytic philosophy, we provide an overview of the most important topics and trends in the `vagueness debate'. In particular we list a range of different phenomena of vagueness that should be addressed by any full-fledged theory of vagueness. Moreover we propose a classification of theories of vagueness and suggest various criteria for their evaluation.

  • [3] Gerla B. (Italy): Many-valued logic and semirings, 467-480.

    In this paper we describe the relationship between many-valued logics (in particular Basic logic and \luk\ logic) and semirings. We will also give definitions of automata on BL-algebras and MV-algebras.

  • [4] Jenei S., Montagna F. (Hungary, Italy): A proof of standard completeness for non-commutative monoidal t-norm logic, 481-489.

    In [JM] we proved that the monoidal t-norm logic MTL introduced by Esteva and Godo in [EG] is the logic of left-continuous t-norms and their residuals. Recently, the Rumenian school, P. Hájek and others investigated in deep non-commutative t-norms. Thus it is natural to look for the logic of left-continuous non-commutative t-norms. This is precisely what we do in this paper. The proof is a combination of the method used in [JM] and of results by J. Kühr in [K] and by P. Hájek in [H2].

  • [5] Jipsen P. (USA): An overview of generalized basic logic algebras, 491-500.

    Contributed papers

  • [6] Barták R., Müller T., Rudová H. (Czech Republic): Minimal perturbation problem - A formal view, 501-511.

    Formulation of many real-life problems evolves as the problem is being solved. These changes are typically initiated by a user intervention or by changes in the environment. In this paper, we propose a formal description of a so called minimal perturbation problem that allows an ``automated" modification of the (partial) solution when the problem formulation changes. Our model is defined for constraint satisfaction problems with emphasis put on finding a solution anytime even for over-constrained problems.

    Several systems supporting development and application of graphical Markov models are widely used; perhaps the most famous are HUGIN and NETICA, which are supporting Bayesian networks. The goal of this paper is to introduce system MUDIM, which is intended to support non-graphical multidimensional models, namely compositional models. The basic idea of these models resembles jig-saw puzzle, where a picture must be assembled from a great number of pieces, each bearing a small part of a picture. Analogously, compositional models of a multidimensional distribution are assembled (composed) of a great number of low-dimensional distributions.

    One of the advantages of this approach is that the same apparatus that is based on operators of composition, can be applied for description of both probabilistic and possibilistic models. This is also the goal for future MUDIM development, to extend it in the way that it will be able to process both probabilistic and possibilistic models.

  • [7] Jiroušek R. (Czech Republic): On experimental system for multidimensional model development MUDIM, 513-520.

    Several systems supporting development and application of graphical Markov models are widely used; perhaps the most famous are HUGIN and NETICA, which are supporting Bayesian networks. The goal of this paper is to introduce system MUDIM, which is intended to support non-graphical multidimensional models, namely compositional models. The basic idea of these models resembles jig-saw puzzle, where a picture must be assembled from a great number of pieces, each bearing a small part of a picture. Analogously, compositional models of a multidimensional distribution are assembled (composed) of a great number of low-dimensional distributions.

    One of the advantages of this approach is that the same apparatus that is based on operators of composition, can be applied for description of both probabilistic and possibilistic models. This is also the goal for future MUDIM development, to extend it in the way that it will be able to process both probabilistic and possibilistic models.

  • [8] Krajči S. (Czech Republic): Cluster based efficient generation of fuzzy concepts, 521-530.

    The cluster analysis and the formal concept analysis are both used to identify significant groups of similar objects. The Rice & Siff's algorithm joins these two methods for a two-valued object-attribute (O-A) model and often significantly reduces the amount of concepts and the complexity. We consider an O-A model with graded degrees of attributes. We define a new type of one-sided fuzzification of a conceptual lattice. We generalize the Rice & Siff's algorithm for this case wrt a fixed metric. We prove the basic properties of this new lattice, metric and algorithm and discuss it on a real example.

  • [9] Kramosil I. (Czech Republic): Possibilistic lattice-valued almost-measurability relation, 531-539.

    Possibilistic measures are usually defined as set functions ascribing to each subset of the universe of discourse a real number from the unit interval and obeying some well-known simple conditions. For the number of reasons, as a more realistic version of this model, let us consider partial possibilistic measures defined only for certain subsets and ascribing to them, instead of real numbers, elements from a more general structure. As a rule, a complete lattice will play this role, so let us pick up rather the qualitative and comparative than the quantitative features of particular degrees of possibility. Following the ideas of the standard measure theory, we define the inner and the outer measure induced by the partial lattice-valued possibilistic measure in question. A subset of the basic universe is defined as almost measurable, if the difference (or rather distance) between the values of the inner and the outer measure ascribed to this set does not exceed, in the sense of the partial ordering relation defined in the used complete lattice, some given threshold value (a "small" fixed element from this lattice). Properties of systems of almost measurable sets are investigated in greater detail and some assertions related to them are introduced.

  • [10] Kroupa T. (Czech Republic): On application of Choquet integral in possibilistic information theory, 541-548.

    The aim of this paper is to introduce the Choquet integral representation of some information quantities in the possibility theory. A possibilistic T-independence concept is further analyzed with respect to its information-theoretic properties. The main result is then the introduction of a so called general measure of T-dependence. It is further proven that the general measure of T-dependence exhibits significant properties from an information-theoretic point of view and can be conceived as an apt analogy of the well-known probabilistic mutual information.

  • [11] Metcalfe G., Olivetti N., Gabbay D. (United Kingdom, Italy): Proof theory for product logics, 549-558.

    In this paper we present analytic sequent and hypersequent calculi for Product logic Π, an important t-norm based fuzzy logic with conjunction interpreted as multiplication on the real unit interval [0,1], and Cancellative hoop logic CHL, a related logic with product conjunction interpreted on the real unit interval with 0 removed.

  • [12] Novák V. (Czech Republic): Descriptions in the full fuzzy type theory, 559-569.

    In this paper, we extend the fuzzy type theory (FTT) by the description operator \i_{\alpha(0\alpha)} whose interpretation is a function which assigns a fuzzy set from M_{0\alpha} and element from M_{\alpha} and is thus similar to the defuzzification operation introduced in the fuzzy set theory. The full fuzzy type theory is obtained when extending the FTT by the description operator together with a proper axiom. Some basic properties of the description operator have been proved as well as the completeness of the full FTT.

  • [13] Perfilieva I. (Czech Republic): Solvability of a system of fuzzy relation equations: easy-to-check conditions, 571-579.

    The problem of solvability of a system of fuzzy relation equations with respect to unknown fuzzy relation is considered. A number of new criteria of the so called Mamdani relation to be a solution to the system is suggested. At the same time those criteria are sufficient conditions of a solvability of the system in general. A new, easy-to-check criterion of a solvability of the system with special fuzzy parameters is found.

  • [14] Slobodová A. (Slovakia): Decision making under uncertainty and mixed models, 581-589.

    We use relations between undirected graphs and conditional independence to introduce a new class of graphical representations, expected utility networks with both discrete and continuous variables and discuss some of their structural properties. We want to show that in these networks node separation with respect to the probability and utility subgraphs implies conditional utility independence, and conditional independent decisions can be effectively decentralized. An application to decision making in expected utility networks in mixed models is introduced.


  • [1] Dunis C. L., Triantafyllidis J. A. (United Kingdom): Alternative forecasting techniques for predicting company insolvencies: The UK example (1980-2001), 326-360.

    The motivation for this paper lies with the prediction of company insolvencies in the UK using macroeconomic factors and adopting four alternative modelling approaches: ordinary least squares (OLS), ARMA, Logit and neural network regression (NNR). ARMA models are used as a benchmark for assessing the results of the other forecasting methods.

    Using relevant macroeconomic variables, the four models are first estimated in-sample with quarterly data for the period ranging from 1980Q1 until 1998Q1. Then, assessing traditional forecasting accuracy criteria and using the ARMA methodology as a benchmark, these models are tested out-of-sample over the period 1998Q2-2001Q1.

    The results clearly show that the NNR models achieve superior performance in predicting UK company insolvencies compared to the benchmark ARMA model and the remaining two models. This indicates that Neural Networks are a very promising tool in evaluating company insolvencies in terms of predictive accuracy and robustness.

  • [2] Kramosil I. (Czech Republic): Extensions of partial lattice-valued possibility measures, 361-384.

    For a number of reasons, in the real world the degrees of uncertainty, in particular, the degrees of possibilistic measures ascribed to various events, need not be definable quantitatively, by real numbers from the unit interval, but rather only qualitatively (greater than, smaller than,...). Moreover, the values of possibilistic measures need not be known or defined for every event from the field of events under consideration. To describe this situation mathematically, we investigate partial non-numerical (as a matter of fact, lattice-valued) possibilistic measures. Using the operation of pseudo-complement, definable in complete lattices, the partial lattice-valued necessity measures, induced by the given possibilistic ones, are introduced. Three extensions of given partial possibilistic and necessity measures to the whole system of events under investigation are defined, including those inspired by the idea of inner and outer measures common in the standard measure theory, and some assertions showing their various properties and mutual relations are presented and proved.

  • [3] Votruba Z., Novák M., Voráčová Š. (Czech Republic): Problem of dimensionality in predictive diagnostics, 385-392.

    Predictive diagnostics (PD) is well known as a significant and powerful approach how to increase reliability of a complex system. Nevertheless, its efficiency susceptibly reacts to the dimension of the task (namely the number of significant parameters or markers) and the level of uncertainty. The quantitative evaluation of this effect done on geometric considerations is presented. The obtained results demonstrate that due to the moderate impact of uncertainty on the system, the dimension of the task should be held at a value as low as possible, lower than 10. Otherwise, the efficient utilization of predictive diagnostics by contemporaneous analytical tools is very difficult and controversial.

  • [4] Kiong L. C., Rajeswari M., Rao M. V. C. (Malaysia): Nonlinear dynamic system identification and control via self-regulating modular neural network, 393-420.

    An endeavor is made in this paper to describe a self-regulating constructive multi-model neural network called Self-regulating Growing Multi-Experts Network (SGMN) that can approximate an unknown nonlinear function from observed input-output training data. The proposed network is devised to overcome the redundancy problems of Gaussian neural networks that use square mesh partition method. In the SGMN, the problem space is decomposed into overlapping regions by expertise domain and the local expert models are graded according to their expertise level. The network output is computed by a smooth combination of local polynomial models. In order to avoid an over-fitting problem, the SGMN deploys a Redundant Experts Removal Algorithm to remove the redundant local experts from the network. In addition, the Fully Self-Organized Simplified Adaptive Resonance Theory (FOSART) is modified and adopted to generate an induced Delaunay triangulation that is highly desired for optimal function approximation. Self-adaptive learning rates Gradient Descent learning rules are employed in a supervised learning phase. A parametric control at epoch terminations and performance based on local incremental experts insertions are incorporated. A variety of examples is solved from literature to establish the efficacy of SGMN. Discrete time nonlinear dynamic system modeling and water bath temperature control have been found to give excellent results via this novel neural network.

  • [5] Hogo M., Snorek M. (Czech Republic): Temporal web usage mining using the modified Kohonen SOM, 421-437.

    Temporal web usage mining involves application of data mining techniques on temporal web usage data to discover temporal patterns, which describes the temporal behavior of users on the Internet. Clustering is one of the most important functions in web usage mining. The clusters and associations in web usage mining do not necessarily have crisp boundaries. Researchers have studied the possibility of using fuzzy and rough sets in web mining clustering applications. Recent research presented in [1] introduced the adaptation of Kohonen SOM based on the properties of rough sets theory to find the interval set clusters. This paper introduces the temporal web usage mining of web users on one educational web site, using the adapted Kohonen SOM proposed in [1]. The paper also describes experiments design including data cleaning, data preparation, data segmentation, and interval set clustering process as well as interpretation and analysis of the obtained results.

  • [6] Book review, 439-444.
  • [7] Book review, 445-446.
  • [8] Book review, 447-448.
  • [9] Erratum, 449.
  • [10] Conference announcements, 451.


  • [1] Editorial, 221.
  • [2] Čerňanský M., Beňušková Ľ. (Slovakia): Simple recurrent network trained by RTRL and extended Kalman filter algorithms, 223-234.

    Recurrent neural networks (RNNs) have much larger potential than classical feed-forward neural networks. Their output responses depend also on the time position of a given input and they can be successfully used in spatio-temporal task processing. RNNs are often used in the cognitive science community to process symbol sequences that represent various natural language structures. Usually they are trained by common gradient-based algorithms such as real time recurrent learning (RTRL) or backpropagation through time (BPTT). This work compares the RTRL algorithm that represents gradient based approaches with extended Kalman filter (EKF) methodology adopted for training the Elman's simple recurrent network (SRN). We used data sets containing recursive structures inspired by studies of cognitive science community and trained SRN for the next symbol prediction task. The EKF approach, although computationally more expensive, shows higher robustness and the resulting next symbol prediction performance is higher.

  • [3] Farkaš I. (Slovakia): Lexical acquisition and developing semantic map, 235-245.

    In this paper, we describe a self-organizing neural network model that addresses the process of early lexical acquisition in young children. The growing lexicon is modeled by combined semantic word representations based on distributional statistics of words and on grounded semantic features of words. Changing semantic word representations are assumed to model the maturation of word meaning and serve as inputs to the growing semantic map. The model has been tested on a real child-directed parental language corpus and as a result, the map demonstrates the emergence and reorganization of various word categories, as quantified by two measures.

  • [4] Jakša R. (Slovakia): Automatic modularization of ANNs, 247-254.

    We propose automatic modularization method for artificial neural networks (ANNs). We treat modularization as an optimization task, therefore the optimization criteria are defined and the topology capable of continuous iterative modularization is introduced. The modularization process starts with unstructured plain network topology and iteratively builds up a modular structure. Automatic modularization approach not only learns to map inputs to outputs but it also tries to discover a structure of knowledge represented by training patterns.

  • [5] Kvasnička V., Pospíchal J. (Slovakia): Replicator theory of coevolution of genes and memes, 255-266.

    A simple replication theory of coevolution of genes and memes is proposed. A population composed of couples of genes and memes, the so-called m-genes, is subjected to Darwinian evolution. Three different types of operations over m-genes are introduced: Replication (an m-gene is replicated with mutations onto an offspring m-gene), interaction (a memetic transfer from a donor to an acceptor), and extinction (an m-gene is eliminated). Computer simulations of the present model allow us to identify different mechanisms of gene and meme coevolutions.

  • [6] Makula M., Beňušková Ľ. (Slovakia): Analysis of state space of RNNs trained on a chaotic symbolic sequence, 267-276.

    We investigate solutions provided by the finite-context predictive model called neural prediction machine (NPM) built on the recurrent layer of two types of recurrent neural networks (RNNs). One type is the first-order Elman's simple recurrent network (SRN) trained for the next symbol prediction by the technique of extended Kalman filter (EKF). The other type of RNN is an interesting unsupervised counterpart to the "classical'' SRN, that is a recurrent version of the Bienenstock, Cooper, Munro (BCM) network that performs a kind of time-conditional projection pursuit. As experimental data we chose a complex symbolic sequence with both long and short memory structures. We compared the solutions achieved by both types of the RNNs with Markov models to find out whether training can improve initial solutions reached by random network dynamics that can be interpreted as an iterated function system (IFS). The results of our simulations indicate that SRN trained by EKF achieves better next symbol prediction than its unsupervised counterpart. Recurrent BCM network can provide only the Markovian solution that is not able to cover long memory structures in sequence and thus beat SRN.

  • [7] Pospíchal J., Kvasnička V. (Slovakia): Simulated annealing construction of shortest paths on incomplete graphs, 277-290.

    Simulated annealing construction of shortest (spanning/nonspanning and closed/open) paths on general connected graphs is discussed. A brief graph-theoretical analysis of the problem is given. A theorem has been proved that for connected graphs the shortest paths are semielementary, that is each edge on the path is visited at most twice in opposite directions. This observation considerably reduces the search space. Tasks may be further specified depending on whether the initial and terminal vertices are given or not. Similarly, in construction of shortest open paths a subtask is considered when the path must visit a prescribed subset of graph vertices. Illustrative calculations demonstrate that the proposed method results for incomplete graphs in the paths that are closely related to optimal solutions.

  • [8] Rosipal R. (Slovakia): Kernel partial least squares for nonlinear regression and discrimination, 291-300.

    This paper summarizes recent results on applying the method of partial least squares (PLS) in a reproducing kernel Hilbert space (RKHS). A previously proposed nonlinear kernel-based PLS regression model has proven to be competitive with other regularized regression methods in RKHS. In this paper the use of kernel PLS for discrimination is discussed. A new methodology for classification is then proposed. This is based on kernel PLS dimensionality reduction of the original data space followed by a support vector classifier. Good results using this method on a two-class classification problem are reported here.

  • [9] Sinčák P., Hric M., Vaľo R., Horanský P., Karel P. (Slovakia): Using MF-ARTMAP neural network for financial data analysis, 301-310.

    The paper deals with application of MF-ARTMAP neural network on financial fraud data. The focus was on classification of data into 5 types of fraud based on expert knowledge with the aim to achieve the tool with highest classification accuracy. The fraud was characterized by 22 features and the verbal features were encoded into numerical values to be able to use them in the classification procedure. The results show that in the case of sufficient data (fraud) representation neural networks could be used with success; in case there are rather small examples, expert generated rules are preferred.

  • [10] Tiňo P., Polčicová G. (Slovakia): Topographic organization of user preference patterns in collaborative filtering, 311-324.

    We introduce topographic versions of two latent class models for collaborative filtering. Topographic organization of latent classes makes orientation in rating/preference patterns captured by the latent classes easier and more systematic. Furthermore, since we deal with probabilistic models of the data, we can readily use tools from probability and information theories to interpret and visualize information extracted by the model. We apply our models to a large collection of user ratings for films.


  • [1] Cornet B. Th. M., Rothkrantz L. J. M. (Netherlands): Recognition of Car License Plates using a neocognitron type of Artificial Neural Network, 115-132.

    In this paper, we describe the application of a combined neocognitron type of the neural network classifier in a generic Car License Plate Recognition (CLPR) system. The suggested system contains an image processor, a segment processor and five coupled neocognitron network classifiers that act as a character recognizer. The presented model of the system depends neither on the specific license plate image features nor on the license plates character style and size. Combining neocognitron classifiers were motivated by the fact that manually tuning a training set for a large neocognitron network is tedious. It is shown how the training set tuning for a large neocognitron network can be avoided. By connecting small neocognitrons specifically trained on ambiguous character classes, the performance of the recognizer in our CLPR was improved easily. The use of a neocognitron recognizer contributes significantly to the generality of a CLPR system. Besides, character recognition rates of 94% are realized using the proposed neocognitron.

  • [2] Popescu Th. D. (Romania): Change Detection in Nonstationary Time Series in Linear Regression Framework, 133-150.

    The problem of change detection in nonstationary time series using linear regression models is addressed. It is assumed that the data can by accurately described by a linear regression model with piece-wise constant parameters. Due to the limitations of some classical approaches, based upon the innovation of one autoregressive (AR) model, most algorithms for the change detection presented make use of two AR models: one is a reference model, and the other one is a current model updated via a sliding block. Changes are detected when a suitable "distance" between these two models is high. Three "distance" measures are considered in the paper: cepstral distance, log-likelihood ratio (justified by GLR) and a distance involving the cross-entropy of the two conditional probabilities laws (divergence test). Other methods based on the quadratic forms of Gaussian random variables are also discussed in the paper. Finally, a change detection algorithm using three models and the evolution of Akaike Information Criterion is presented. All the presented algorithms constituted the object of evaluation by multiple simulation and have been used to change detection in some nonstationary financial and economic time series.

  • [3] Kiong L. C., Rajeswari M., Rao M. V. C. (Malaysia): Extrapolation detection and novelty-based node insertion for sequential growing multi-experts network, 151-176.

    Artificial neural networks (ANNs) have been used to construct empirical nonlinear models of process data. Because networks are not based on the physical theory and contain nonlinearities, their predictions are suspect when extrapolating beyond the range of original training data. Standard networks give no indication of possible errors due to extrapolation. This paper describes a sequential supervised learning scheme for the recently formalized Growing Multi-experts Network (GMN). It is shown that the Certainty Factor can be generated by the GMN that can be taken as an extrapolation detector for the GMN. The On-line GMN identification algorithm is presented and its performance is evaluated. The capability of the GMN to extrapolate is also indicated. Four benchmark experiments are dealt with to demonstrate the effectiveness and utility of the GMN as a universal function approximator.

  • [4] Čespiva L. (Czech Republic): The PT core benchmark measurement of scaled number of processors, 177-186.

    The developed and tested part of the scalable benchmark code, the Pythagorean triples core, has been applied to the platform system of a scalable number of processors. The measurement has been performed on the system cluster consisting of 16 Pentium CPUs. The number of nodes of selected subclusters of an equivalent or a different performance of CPUs is scaled by the factor of 2. The core has been running in different conditions (homogeneous subcluster, heterogeneous subcluster, computationally free nodes and/or occupied nodes, etc.).

    A group of four measurements of the scalable number of processors has been selected and displayed in four characteristic blocks of the elapsed time windows comparable with those of the previous paper. The characteristic exponential curves fit well to the measured points under the normal conditions of task run. The maximum deviations of the two exponential parameters in all presented cases do not exceed 5 percent.

  • [5] Zapranis A., Sivridis S. (Greece, United Kingdom): Extending Vasicek with Neural Regression, 187-210.

    In this paper we contrast linear parametric estimation with non-parametric non-linear neural estimation of the reversion speed ? in the context of the Vasicek model, which is routinely being used for deriving the term structure of the short rate. The sampling parameters of the short-rate, even its realization, were varied widely. Neural regression was employed in an attempt to identify a possibly non-linear relationship, and from that to extract a measure of instantaneous reversion speed (a local equivalent of reversion speed). Neural network models outperformed consistently the linear estimator in terms of explained variability by more than 10%, indicating a degree of non-linearity in the underlying relationship.

  • [6] Book review, 211-214.
  • [7] Software package review, 215-219.


  • [1] Editorial, 1.
  • [2] Francisco Javier López Aligué, Miguel Macías Macías, Horacio Manuel González Velasco, Carlos Javier García Orellana, María Isabel Acevedo Sotoca (Spain): A Neural-Network-Based Classifier for Large Pattern Recognition, 3-14.

    In this paper a classification system, which consists of a neural network and a decision element, is presented, both parts processing information in series. For the neural network, we propose a training algorithm based on the direct equalization of weights and components of prototype vectors, and a neuronal function that detects similarities between its inputs and the weights. This systematics allows, in addition to a good performance in recognition, an easy, time-controlled reprogramming process of the network, even for large patterns. To test and validate the system, a real classifier is presented and studied, a classifier that is designed to recognize segmented handwritten characters corresponding to the NIST SD19 database and with which good results for digits and lower-case letters are obtained.

  • [3] Tobely T. E., Tsuruta N., Amamiya M. (Japan): The Competitive Algorithm of the Hypercolumn Neural Network Toward Real-time Image Recognition, 15-39.

    The Hypercolumn (HCM) neural network model is an unsupervised competitive network consisting of hierarchical layers of the Hierarchical Self-Orga\-nizing Map (HSOM) neural networks arranged by similar to the cell planes in the Neocognitron (NC) neural network. The HCM model combines the advantages of both the HSOM and the NC while rejecting their disadvantages, and alleviates many difficulties associated with image recognition applications. It can recognize images with variant objects size, position, orientation, and spatial resolution. However, due to the hierarchical structure of the HCM model, the network spends a long time in the recognition. In this paper, the HCM model is introduced with a new competitive algorithm that reduces the network recognition time into a real-time range. The proposed algorithm uses the subset from the most discriminate codebook of the network weights to find the winner of each HSOM in the first layer of the HCM model.

  • [4] Faber J., Novák M., Svoboda P., Tatarinov V. (Czech Republic): Electrical Brain Wave Analysis during Hypnagogium, 41-54.

    Impaired wakefulness in machine operators poses a danger not only to themselves but often to the public at large as well. While on duty, such persons are expected to be continuously, i.e., without interruption, on the alert. For that purpose, we designed and carried out an experimental model of continuous vigilance monitoring using electroencephalography (EEG) and reaction time measured as the latency of the proband's reaction to sound stimulus. If constructed, the set together with other logical elements and an alarm system can be used for an automatic detection of vigilance and, possibly, also of arousal stimuli in cases of micro-sleep. We found the following new facts and confirmed the validity of some of the earlier ones:

    1. Vigilance is marked by alpha activity in the EEG record (oscillation of 8-3 Hz) and reaction-time (RT) of 200-400 ms (milliseconds). Sleep is characterized by theta and delta activities (4 - 7 and 0.5 - 3.5 Hz respectively) with no reaction.
    2. Between wakefulness and sleep there are at least two stages: relaxation with prolonged RT of 400 to 800 ills and increased EEG alpha, sometimes also beta activities. Then there is the hypnagogic phase with disintegrating alpha and growing theta or even delta activities and an RT of 800 ms up to 1200 ms.
    3. Changes in the EEG and its spectrum1 and its actual localization on the cranial surface exhibit individual differences; hence, no sharp limits for the above stages can be established.
    4. As for changes in the vigilance in the relaxation and hypnagogic phases as well as in the processes of mentation, the alpha and delta are the most significant, less so the theta and beta bands.
    5. The most suitable sites for the detection of the changes on the skull surface are temporo-parieto-occipital (TPO) regions, i.e., those over the posterior parts of the skull with the least muscle and oculo-motor artifacts and with the most energy for alpha and delta activities.
    6. In somnolence, the cortex does not behave as a whole, which means that different areas show different spectra while getting off to sleep, a fact easy to express by means of the alpha/delta ratio (by alpha/delta ratio we understand the ratio of average amplitude of EEG signal components in the alpha band to the average amplitude of EEG signal components in the delta band) separately for each of the cranial areas.
    7. At sleep onset, the alpha/delta ratio undergoes changes; it is greater than 1 in wakefulness, less than 1 in sleep, and approaches to 1 as the person goes to sleep.
    8. In the course of sleep with zero reactivity, the cortex already behaves as a whole, i.e., all cranial areas have similar or the same spectrograms, with the alpha/delta coefficient being less than 1 all over the skull.
    9. Sometimes, the spectrogram taken during mentation (e.g., while undergoing psychological tests) resembles that of somnolence, with the alpha/delta coefficient being greater than 1. However, there are differences: in somnolence, the delta activity is increased all over its band, i.e., from 0,5 to 3,5 Hz, while during mentation it is increased solely in the slow delta activity band (0.5 to 2 Hz). In somnolence, theta is on the increase, but not so the mentation. In the hypnagogic phase, alpha becomes completely extinct-unlike in mentation.
    10. As follows from the above listed facts, not everyone applying for an automatic alarm detector of vigilance can be provided with one at random and expect it to go off at the first sign of slumber. Conversely, every applicant ought to be treated as a proband, i.e., tested with simultaneous EEG registration, EEG analysis, determination of the best suitable area on the cranial surface and EEG frequency, separately for vigilance, relaxation, the hypnagogic phase and mentation, and - in keeping with the above rules - have individual parameters of the alarm device adjusted accordingly.
  • [5] Tatarinov V. (Czech Republic): Alertness Detection of System Operator, 55-73.

    Attention decrease and an eventual micro-sleep of an artificial system operator is very dangerous and its early detection can prevent great losses. This work2 deals with an early detection of micro-sleep based on analysis of an electroencefalographic activity of the brain. There are classic spectral methods - the Discrete Fourier Transform and parametric methods - autoregressive models used for signal processing here. An influence of a band pass filter characteristic on classification is investigated. For the detection of the micro-sleep multi-layer perceptron, radial basis function (RBF) and the learning vector quantization (LVQ) neural networks are used. The k-nearest neighbor as a representative of non-parametric methods is examined. The last method used here is based on the Bayesian theory and its coefficients are found using the maximum likelihood estimation.

  • [6] Češpiva L. (Czech Republic): Pythagorean triples core of the scalable benchmark code, 75-88.

    The already announced paper presents continuation of the serial of articles concerning the simple parallel scalable benchmark code. The new Pythagorean triples core has been implemented, tested and used for measurement. Several preliminary tests have shown suitability of the core routines for benchmark practice. The Pythagorean triples core extends functional properties of the scalable benchmark code, i.e. primes core functionality.

    A group of six performance routines has been tested. Four Pythagorean triple generators/selectors have signifficant run time over the elapsed time window. The routines of significant run time have been taken into the final test of hardware performance. Performance routines are in general of two types. Four routines generate numerical values by a mathematical formula. Two routines have been implemented selecting Pythagorean triples by means of testing the difference of squares of triangle side lengths.

    Three hardware platforms have been tested: PC 486, an old type of Pentium, belongs to one of the first Pentium models of seriál production (called here shortly archaic Pentium) and one of the highest performance Pentium 900 still used in the computational world for graphical animation.

  • [7] Coufal D., Matucha P., Uhlířová H., Lomský B., Forczek S. T., Matucha M. (Czech Republic): Analysis of Coniferous Forest Damage: Effects of Trichloroacetic Acid, Sulphur, Fluorine and Chlorine on Needle Loss of Norway Spruce, 89-102.

    Trichloroacetic acid (TCA), secondary air pollutant (SAP) and a product of photooxidation of volatile chlorinated C2-hydrocarbons has phytotoxic properties and negatively influences on the state of forest health in general. The present knowledge shows the uptake of TCA by the spruce from the atmosphere by precipitation over soil, roots, and transpiration stream up into needles, where it affects the photosynthetic apparatus of the plant. To judge the role of TCA in forest ecosystems its effects along with other stressors must be followed. Those include, above all, emissions of sulphur oxides and fluoride from burning energetic coal. From monitoring the selected stressors - the content of total sulphur, fluorine and chlorine besides TCA - in spruce needles on selected stands, a positive correlation between the needle loss (as a measure of the spruce forest damage) and the content of S, F, and TCA was found. In this way the negative effect of TCA was confirmed.

  • [8] Book review, 103-105.
  • [9] Contens volume 12 (2002), 107-110.
  • [10] Author's index volume 12 (2002), 111-113.

1 Actually the term pseudo-spectrum should be used here; because of quasi-periodic and quasi-stationary time-series representing EEG signals, the spectrum in strong mathematical sense does not exist. Nevertheless, as is typical in contemporary literature, the term spectrum will be used here, too. As the often applied fast Fourier analysis is not fully correct here, leading to hardly compatible results, the analysis based on the use of Gabor filtration with special polynomial filtration function of the 50th order was used here.

2 Supported by the grants ME478 and VZ 210000024 of the Ministry of Education, Czech Republic.

Thanks to