Artificial Intelligence and Soft Computing – ICAISC 2008

In this paper a Kohonen self-organizing competitive algorithm is considered. A formal approach to classification problem, basing on equivalence relations, is proposed. The Kohonen neural networks are considered as classifying systems. The main topic of this paper is proposal of applying stereographic projection as an input signals normalization procedure. Both theoretical justification is discussed and results of experiments are presented. It turns out that the introduced normalization procedure is effective.

In this paper we present a parallel realisation of Real-Time Recurrent Network (RTRN) learning algorithm. We introduce the cuboid architecture to parallelise computation of learning algorithms. Parallel neural network structures are explicitly presented and the performance discussion is included.

Novel convergence properties of identification algorithm for complex input-output systems, which uses recurrent neural networks, are derived. By the term “complex system” we understand a system containing interconnected sub processes (elementary processes), which can operate separately. Each element of the complex system is modeled by a multi-input, multi-output neural network. A model of the whole system is obtained by composing all neural networks into one global network. Stable learning algorithm of such a neural network is proposed. We derived sufficient condition of stability using the second Lyapunov method and proved that algorithm is stable even if stability conditions for some individual neural networks are not satisfied.

In the paper the new approach to create artificial neural networks (ANNs) is proposed. ANN’s are inspired by natural neural networks (NNNs) that receive data in time still tuning themselves. In opposite to them ANNs usually work on the training data (TD) acquired in the past and are totally available at the beginning of the adaptation process. Because of this the adaptation methods of the ANNs can be sometimes more effective than the natural training process observed in the NNNs. This paper presents the ability of ANNs to adapt more effectively than NNNs do if only all TD are known before the beginning of the adaptation process. The design and adaptation process of the proposed ANNs is divided into two stages. First, analyze or examining the set of TD. Second, the construction of neural network topology and weights computation. In the paper, two kinds of ANNs which use the proposed construction strategy are presented. The first kind of network is used for classification tasks and the second kind for feature extraction.

The paper presents a clustering technique based on dynamic self-organizing neural networks and its application to a large-scale and highly multidimensional WWW-newsgroup-document clustering problem. The collection of 19 997 documents (e-mail messages of different

Usenet-News

newsgroups) available at WWW server of the School of Computer Science, Carnegie Mellon University (www.cs.cmu.edu/ TextLearning/datasets.html) has been the subject of clustering. A broad comparative analysis with nine alternative clustering techniques has also been carried out demonstrating the superiority of the proposed approach in the considered problem.

The paper presents the design of municipal creditworthiness parameters. Further, a model is designed based on Learning Vector Quantization neural networks for municipal creditworthiness classification. The model is composed of Kohonen’s Self-organizing Feature Maps (unsupervised learning) whose outputs represent the input of the Learning Vector Quantization neural networks (supervised learning).

The calculation method for weights of orthogonal Fourier series neural networks on the grounds of multidimensional discrete Fourier transform is presented. The method proposed represents high speed of operation and outlier robustness. It allows easy reduction of network structure following its training process. The paper presents also the ways of applying the method to modelling of dynamic controlled systems. It is very easy to prepare a program which would allow to use the procedure proposed.

Some accuracy improvement of the neural network (NN) estimator is proposed in the paper. The estimator approximates stator current components in the rotor flux reference frame. Two approaches are considered: data mining with GMDH algorithm and gradual training of the NN in the desired frequency range. In both cases the accuracy of the estimator is significantly improved. Provided tests confirmed this feature and encourage to implement such an estimator it in a sensorless vector controlled induction motor drive.

In the paper the ensemble of dipolar neural networks (EDNN) for analysis of survival data is proposed. The tool is build on the base of the learning sets, which contain the data from clinical studies following patients response for a given treatment. Such datasets may contain incomplete (censored) information on patients failure times. The proposed method is able to cope with censored observations and as the result returns the aggregated Kaplan-Meier survival function. The prediction ability of the received tool as well as the significance of individual features is verified by the Brier score,

$\tilde{D}_{S,x}$

and

$\hat{D}_x$

measures of predictive accuracy.

The problem of binary optimization of a quadratic functional is discussed. By analyzing the generalized Hopfield model we obtain expressions describing the relationship between the depth of a local minimum and the size of the basin of attraction. Based on this, we present the probability of finding a local minimum as a function of the depth of the minimum. Such a relation can be used in optimization applications: it allows one, basing on a series of already found minima, to estimate the probability of finding a deeper minimum, and to decide in favor of or against further running the program. The iterative algorithm that allows us to represent any symmetric

N

×

N

matrix as a weighted Hebbian series of bipolar vectors with a given accuracy is proposed. It so proves that all conclusions about neural networks and optimization algorithms that are based on Hebbian matrices are true for any other type of matrix. The theory is in a good agreement with experimental results.

We apply normalized RBF networks to the problem of learning nonlinear regression functions. The parameters of the networks are learned by empirical risk minimization and complexity regularization. We study convergence of the RBF networks for various radial kernels as the number of training samples increases. The rates of convergence are also examined.

This paper presents a predictive control scheme integrated with economic optimisation. Two neural models are used: a dynamic one (for the control subproblem) and a steady-state one (for the economic optimisation subproblem). The algorithm is computationally efficient because it needs solving on-line only one quadratic programming problem. Unlike the classical control system structure, the necessity of repeating two nonlinear optimisation problems at each sampling instant is avoided.

The increasing complexity of technological processes implemented in present industrial installations causes serious problems in the modern control system design and analysis. Chemical refineries, electrical furnaces, water treatments and other industrial plants are complex systems and in some cases cannot be precisely described by classical mathematical models. On the other hand, modern industrial systems are subject to faults in their components. Due to these facts, fault-tolerant control design using soft computing methods is gaining more and more attention in recent years. In this paper, the model-based approach to fault detection and isolation using locally recurrent neural networks is presented. The paper contains a numerical example that illustrates the performance of the proposed locally recurrent neural network with respect to other well-known neural structures.

In this paper the results in leaves classification with non-parametrized one nearest neighbor and multilayer perceptron classifiers are presented. The feature vectors are composed of Fourier descriptors that are calculated for leaves contours with fast adaptive Fourier transform algorithm. An application of fast adaptive algorithm results in new fast adaptive Fourier descriptors.

Experimental results prove that the fast adaptive Fourier transform algorithm significantly accelerates the process of descriptors calculation and enables almost eightfold reduction in the number of contour data with no effect on classification performance. Moreover the neural network classifier gives higher accuracies of classification in comparison to the minimum distance one nearest neighbor classifier.

The method of using a lateral histogram for evaluating the number of holes (e.g., defects) from images is known to be fast but rather inaccurate. Our aim is to propose a method of improving its performance by learning, but keeping the speed of the original method. This task is accomplished by considering a multiclass pattern recognition problem with linearly ordered labels and a loss function, which measures absolute deviations of decisions from true classes.

Training data containing outliers are often a problem for supervised neural networks learning methods that may not always come up with acceptable performance. In this paper a new, robust to outliers learning algorithm, employing the concept of initial data analysis by the MCD (minimum covariance determinant) estimator, is proposed. Results of implementation and simulation of nets trained with the new algorithm and the traditional backpropagation (BP) algorithm and robust Lmls are presented and compared. The better performance and robustness against outliers for the new method are demonstrated.

In the paper we present a new approach based on application of neural networks to detect SQL attacks. SQL attacks are those attacks that take the advantage of using SQL statements to be performed. The problem of detection of this class of attacks is transformed to time series prediction problem. SQL queries are used as a source of events in a protected environment. To differentiate between normal SQL queries and those sent by an attacker, we divide SQL statements into tokens and pass them to our detection system, which predicts the next token, taking into account previously seen tokens. In the learning phase tokens are passed to a recurrent neural network (RNN) trained by backpropagation through time (BPTT) algorithm. Then, two coefficients of the rule are evaluated. The rule is used to interpret RNN output. In the testing phase RNN with the rule is examined against attacks and legal data to find out how evaluated rule affects efficiency of detecting attacks. All experiments were conducted on Jordan network. Experimental results show the relationship between the rule and a length of SQL queries.

This paper describes an approach of combining Monte Carlo simulation and neural nets. The approach is applied to model transport systems, with the accurate but computationally expensive Monte Carlo simulation used to train a neural net. Once trained the neural net can efficiently provide functional analysis and reliability predictions. No restriction on the system structure and on any kind of distribution is the main advantage of the proposed approach. The paper presents exemplar decision problem solved by proposed approach.

The cross-validation method is commonly applied in the design of Artificial Neural Networks (ANNs). In the paper the design of ANN is related to searching for an optimal value of the regularization coefficient or the number of neurons in the hidden layer of network. Instead of the cross-validation procedure, the Maximum of Marginal Likelihood (MML) criterion, taken from Bayesian approach, can be used. The MML criterion, applied to searching for the optimal values of design parameters of neural networks, is illustrated on two examples. The obtained results enable us to formulate conclusions that the MML criterion can be used instead of the cross-validation method (especially for small data sets), since it permits the design of ANNs without formulation of a validation set of patterns.

This paper presents type-2 fuzzy decision trees (T2FDTs) that employ type-2 fuzzy sets as values of attributes. A modified fuzzy double clustering algorithm is proposed as a method for generating type-2 fuzzy sets. This method allows to create T2FDTs that are easy to interpret and understand. To illustrate performace of the proposed T2FDTs and in order to compare them with results obtained for type-1 fuzzy decision trees (T1FDTs), two benchmark data sets, available on the internet, have been used.

In the paper we develop a new method for designing and reduction of neuro-fuzzy systems. The method is based on the concept of the weighted triangular norms. In subsequent stages we reduce number of inputs, number of rules and number of antecedents. Simulation results are given.

Tracking of multiple objects belongs to one of the fundamental tasks of computer vision. In this paper an improvement to the continuously adaptive mean shift tracking method is proposed. It consists in substitution of the probabilistic density function for the especially formed membership function. This makes possible design of tracking systems in terms of fuzzy logic. Additionally, a special data structure was developed to allow tracking of multiple objects at a time. It stores information on image regions which are active for tracking. By this it provides initial conditions for tracking in subsequent frames which also speeds up computations. The method was used and verified in an application of the road signs tracking in real time of 30 frames/s.

The main problem of known methods for Multiple Criteria Decision Making in the Intuitionistic Fuzzy setting is that they are generally based on the intermediate type reduction. Such approaches lead inevitable to the loss of important information. Another problem is the choice of an appropriate method for the local criteria aggregation taking into account their ranks. The aim of this paper is to present a new method which makes it possible to solve the first problem and facilitates the solution of the second one. The method is based on the Dempster-Shafer Theory (DST). It allows to solve the Multiple Criteria Decision Making problem without intermediate type reduction for different approaches to aggregation of the local criteria. The usefulness of elaborated method is illustrated with known example of Multiple Criteria Decision Making problem.

We reformulate and extend our previous works (cf. Kacprzyk, Wilbik and Zadrożny [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17]), mainly towards a more complex and realistic evaluation of results on the linguistic summarization of time series which is meant as the derivation of an linguistic quantifier driven aggregation of partial trends with respect to the dynamics of change, duration and variability. We use Zadeh’s calculus of linguistically quantified propositions but, in addition to the basic criterion of a degree of truth (validity), we also use a degree of specificity to make it possible to account for a frequent case that though the degree of truth of a very general (not specific) summary is high, its usefulness may be low due to its low specificity. We show an application to the absolute performance type analysis of daily quotations of an investment fund.

The work presents some results concerning analytical modeling using the Takagi-Sugeno fuzzy rule-based system, which can be used for exact fuzzy modeling of some class of conventional systems. A special attention was paid to the so called P2-TS systems, which use the polynomial membership functions of the second degree. Theorems provide necessary and sufficient conditions for transformation of fuzzy rules into the crisp model of the system and vice-versa.

Neuro-fuzzy systems show very good performance and the knowledge comprised within their structure is easily interpretable. To further improve their accuracy they can be combined into ensembles. In the paper we combine specially modified Mamdani neuro-fuzzy systems into an AdaBoost ensemble. The proposed modification improves the interpretability of knowledge by allowing merging the subsystems rule bases into one knowledge base. Simulations on two benchmarks shows excellent performance of the modified neuro-fuzzy systems.

A fuzzy predictive set–point optimizer which uses a nonlinear, fuzzy dynamic process model is proposed in the paper. The algorithm of the optimizer is formulated in such a way that only a numerically efficient, quadratic optimization problem must be solved at each algorithm iteration. It is demonstrated, using an example of a control system of a nonlinear MIMO control plant, that application of the optimizer based on a fuzzy model instead of a linear one can result in substantial improvement of control system operation. The fuzzy control plant model, the optimizer is based on, consists of local models in the form of control plant step responses. Thus, the model is easy to obtain and the proposed optimizer easy to design.

The paper proposes new definitions of (im)precision measures for type-2 fuzzy sets representing linguistic terms and linguistically quantified statements. The proposed imprecision measures extend similar concepts for traditional (type-1) fuzzy sets, cf. [1,2]. Applications of those new concepts to linguistic summarization of data are proposed in the context of the problem statement of finding the best summaries.

The growing expectations for service-oriented applications tailored to the customers needs comprise a definition of the service contract documents, for both, technical and non-technical service elements. The usage of the fuzzy methods and techniques allows for the customized prearrangement of such application systems, which could be developed at the costs of the standard software. In the paper the concepts associated with a service-oriented design of non-technical service aspects especially appropriate for the fuzzy description of them as the requirements destined for Service Level Agreements SLAs, are introduced. The proposed approach is demonstrated on the example of a service accessibility requirement, an important component of a contracted SLA.

This paper describes a new look on adaptation of fuzzy rules in fuzzy control process. New idea is based on properties of the Ordered Fuzzy Numbers. The Ordered fuzzy nunbers (OFN) are a new model of fuzzy numbers, presented a few years ago [15]. Important property and advantage of the new model of fuzzy numbers is simple realization of arithmetical operations. Thanks to that we can get neutral element of adding and multiplication in the same way like in real numbers. Easy way of calculating on the Ordered Fuzzy Numbers makes possible to use them in a fuzzy control process. In the [21] new methods of processing information for a fuzzy control system were presented. These methods basing on arithmetic of the Ordered Fuzzy Numbers.

The goal of that paper is to present a way to use a good arithmetical properties of Ordered Fuzzy Numbers in the process of rules adaptation for the fuzzy control system.

In the paper relational neuro-fuzzy systems are described with additional fuzzy relation connecting input and output linguistic fuzzy terms. Thanks to this the fuzzy rules have more complicated structure and can be better suited the task. Fuzzy clustering and relational equations are used to obtain the initial set of fuzzy rules and systems are then learned by the backpropagation algorithm.Simulations shows excellent performance of the modified neuro-fuzzy systems.

Presented paper shows a new approach to creating a fuzzy system based on an exclusive use of clustering algorithms, which determine the value of necessary parameters. The applied multisegment fuzzy system functions as a classifier. Each segment makes an independent fuzzy system with a defined knowledge base and uses singleton fuzzification, as well as fuzzy inference with product operation as the Cartesian product and well-matched membership functions. Defuzzification method is not used. Only the rule-firing level must be analysed and its value suffices to determine the class. The use of clustering algorithms has allowed a qualification of the number of rules in the base of fuzzy rules for each independent segment, as well as a specification of the centers of fuzzy sets used in the given rules. The calculated parameters have proved precise, so that no additional methods have been applied to correct their values. This procedure greatly simplifies the creation of a fuzzy system. The constructed fuzzy system has been tested on medical data that come from the Internet. In the future, those systems may help doctors with their everyday work.

In the paper the defuzzification of interval type-2 fuzzy sets is studied. The standard K-M type-reduction method and the uncertainty bounds approximate type-reduction are compared with the classical defuzzification of averaged type-1 fuzzy sets.

The Dempster-Shafer theory is convenient for implementation in models of medical diagnosis as it neglects dependence of symptoms. Yet, combination of two basic probability assignments that is defined in the theory is often criticized. The paper shows opportunities of combining that are created when the Dempster-Shafer theory is extended for fuzzy focal elements. The proposed method can help to avoid several disadvantages of the classical combination.

We address some crucial problem associated with text categorization, a local feature selection. It seems that intuitionistic fuzzy sets can be an effective and efficient tool making it possible to assess each term (from a feature set for each category) from a point of view of both its indicative and non-indicative ability. It is important especially for high dimensional problems to improve text filtering via a confident rejection of non-relevant documents. Moreover, we indicate that intuitionistic fuzzy sets are a good tool for the classification of imbalanced and overlapping classes, a commonly encountered case in text categorization.

Evolutionary Algorithms (EAs) are common optimization techniques based on the concept of Darwinian evolution. During the search for the global optimum of a search space, a traditional EA will often become trapped in a local optimum. The Scouting-Inspired Evolutionary Algorithms (SEAs) are a recently–introduced family of EAs that use a cross–generational memory mechanism to overcome this problem and discover solutions of higher fitness. The merit of the SEAs has been established in previous work with a number of two and three-dimensional test cases and a variety of configurations. In this paper, we will present two approaches to using SEAs to solve high–dimensional problems. The first one involves the use of Locality Sensitive Hashing (LSH) for the repository of individuals, whereas the second approach entails the use of scouting–driven mutation at a certain rate, the Scouting Rate. We will show that an SEA significantly improves the equivalent simple EA configuration with higher–dimensional problems in an expeditious manner.

In this paper a Memetic Algorithm (MA) is proposed for solving the Vehicles Routing Problem with Time Windows (VRPTW) multi-objective, using a constraint satisfaction heuristic that allows pruning of the search space to direct a search towards good solutions. An evolutionary heuristic is applied in order to establish the crossover and mutation between sub-routes. The results of MA demonstrate that the use of Constraints Satisfaction Technique permits MA to work more efficiently in the VRPTW.

Co-evolutionary algorithms are a special type of evolutionary algorithms, in which the fitness of each individual depends on other individuals’ fitness. Such algorithms are applicable in the case of problems for which the formulation of explicit fitness function is difficult or impossible. Co-evolutionary algorithms also maintain population diversity better than “classical” evolutionary algorithms. In this paper the agent-based version of co-operative co-evolutionary algorithm is presented and applied to multi-objective test problems. The proposed technique is also compared to two “classical” multi-objective evolutionary algorithms.

In this paper we present the problem of designing modular systems combined with the Bagging Algorithm. As component classifiers the Mamdani-type neuro fuzzy-systems are applied and trained using evolutionary methods. Experimental investigations presented in this paper include the classification performed by the modular system built by means of classic Bagging algorithm and its modified version which assigns evolutionary chosen weights to base classifiers.

In this paper we present an evolutionary method to create an interpretable modular system. It consists of many neuro-fuzzy structures which are merged using a very popular algorithm called AdaBoost. As the alternative to the backpropagation method to train all models a special evolutionary algorithm has been used based on the evolutionary strategy (

μ

,

λ

).

Convergence properties of genetic algorithms are investigated. For them some measures are introduced. A classification procedure is proposed for genetic algorithms based on a conjecture: the entropy and the fractal dimension of trajectories produced by them are quantities that characterize the classes of the algorithms. The role of these quantities as invariants of the algorithm classes is presented. The present approach can form a new method in construction and adaptation of genetic algorithms and their optimization based on dynamical systems theory.

In most of data mining systems decision trees are induced in a top-down manner. This greedy method is fast but can fail for certain classification problems. As an alternative a global approach based on evolutionary algorithms (EAs) can be applied. We developed

Global Decision Tree

(GDT) system, which learns a tree structure and tests in one run of the EA. Specialized genetic operators are used, which allow the system to exchange parts of trees, generate new sub-trees, prune existing ones as well as change the node type and the tests. The system is able to induce univariate, oblique and mixed decision trees. In the paper, we investigate how the

GDT

system can profit from a parallelization on a compute cluster. Both parallel implementation and distributed version of the induction are considered and significant speedups are obtained. Preliminary experimental results show that at least for certain problems the distributed version of the

GDT

system is more accurate than its panmictic predecessor.

At present, the optimization problem resolution is a topic of great interest, which has fostered the development of several computer methods forsolving them.

Particle Swarm Optimization (PSO) is a metaheuristics which has successfully been used in the resolution of a wider range of optimization problems, including neural network training and function minimization. In its original definition, PSO makes use, during the overall adaptive process, of a population made up by a fixed number of solutions.

This paper presents a new extension of PSO, called VarPSO, incorporating the concepts of age and neighborhood to allow varying the size of the population. In this way, the quality of the solution to be obtained will not be affected by the used swarm’s size.

The method here proposed is applied to the resolution of some complex functions, finding better results than those typically achieved using a fixed size population.

The paper describes the model of virtual stock market which is evolved by a genetic algorithm. The model consists of cooperating Agents that imitate behaviour of real investors. They act on the virtual market buying or selling stocks. The aim of the model is to generate stocks prices on a virtual market that are similar to real ones for a short period of time. Each Agent is described by its unique characteristics which determine his performance. The details of the model are presented in the paper. The applied genetic algorithm is generic one. Its main components such as: an individual, genetic operators and fitness function are described here, as well. The results of experiments investigating the role of genetic algorithm parameters are presented in the paper. Agent’s ability to predict the quotations values are presented and analysed. Future plans referring to the further development of the system are presented at the end of the paper.

One of the most serious problem concerning global optimization methods is their correct configuration. Usually algorithms are described by some number of external parameters for which optimal values strongly depend on the objective function. If there is a lack of knowledge on the function under consideration the optimization algorithms can by adjusted using trail-and-error method. Naturally, this kind of approach gives rise to many computational problems. Moreover, it can be applied only when a lot of function evaluations is allowed. In order to avoid trial-and-error method it is reasonable to use an optimization algorithm which is characterized by the highest degree of robustness according to the variations in its control parameters. In this paper, the robustness issue of evolutionary strategy with isotropic stable mutations is discussed. The experimental simulations are conducted with the help of special search environment - the so-called general search space.

The objective of the research presented in this paper is to investigate the application of a metaheuristic algorithm called

Ant Colony Algorithm

to petroleum engineering problems. This algorithm usually used for discrete domains, but with some modifications could be applied to continuous optimization. In this Paper, two examples with continuous and discrete parameters also known solutions and varying degrees of complexities are presented as an illustration for solving a large class of process optimization problems in petroleum engineering. Results of case studies show ability of Ant Colony Algorithm to provide fast and accurate solutions.

In the paper an application of evolutionary algorithm with multi-layer chromosomes to the design and multi-objective optimization of combinational digital circuits is presented. The optimization criterions are minimizations of: number of gates, number of transistors in the circuit, and circuit propagation time. Four combinational circuits, chosen from literature, are designed, and optimized using proposed method. Results obtained using this method are compared with results obtained by other methods. The results obtained using this method are in many cases better than those obtained using other methods.

The simple genetic algorithm (SGA) and its convergence analysis are main subjects of the article. The SGA is defined on a finite multi-set of potential problem solutions (individuals) together with mutation and selection operators, and appearing with some prescribed probabilities. The selection operation acts on the basis of the fitness function defined on individuals, and is fundamental for the problem considered. Generation of new population is realized by iterative actions of those operators written in the form of a transition operator acting on probability vectors. The transition operator is a Markov one. Conditions for convergence and asymptotic stability of the transition operator are formulated.

Heuristic approaches already proved their efficiency for the cases where real-world problems dynamically change in time and there is no effective way of prediction of the changes. Among them a mixed multi-swarm optimization (mSO) is regarded as the most efficient. The approach is a hybrid solution and it is based on two types of particle swarm optimization (PSO): pure PSO and quantum swarm optimization (QSO). Both types are applied in a set of simultaneously working sub-swarms. In spite of the fact that there appeared a series of publications discussing properties of this approach the motion mechanism of quantum particles was just briefly studied, and there is still some research to do. This paper presents the results of our research on this subject. The novelty is based on a new type of distributions of particles in a quantum cloud. Obtained results allow to derive some guidelines of an effective tuning of the mechanism of distribution in the quantum cloud and show that further improvement of mSO is possible.

This paper describes a method for production of an ensemble of general classifiers using unsupervised learning. The method uses the ‘divide and conquer’ strategy. Using competitive learning the feature space is divided into subregions where the classifiers are constructed. The structure of the ensemble starts from a single member and new members are being added during the training. The growth of the ensemble is self determined until the ensemble reaches the desired accuracy. The overall response of the ensemble to an input pattern is represented by the output of a winning member for the particular pattern. The method is generic, i.e. it is not bound to a specific type of classifier and it is suitable for parallel implementation. It is possible to use the method for data mining. A basic wide margin linear classifier is used in the experiments here. Experimental results achieved on artificial and real world data are presented and compared to the results of Gaussian SVM. Parallel implementation of the method is described.

This paper focuses on classification tasks. The goal of the paper is to propose a framework for adaptive and integrated machine classification and to investigate the effect of different adaptation and integration schemes. After having introduced several integration and adaptation schemes a framework for adaptive and integrated classification in the form of the software shell is proposed. The shell allows for integrating data pre-processing with data mining stages using population-based and A-Team techniques. The approach was validated experimentally. Experiment results have shown that integrated and adaptive classification outperforms traditional approaches.

We introduce a novel decision rule induction algorithm for solving the regression problem. There are only few approaches in which decision rules are applied to this type of prediction problems. The algorithm uses a single decision rule as a base classifier in the ensemble. Forward stagewise additive modeling is used in order to obtain the ensemble of decision rules. We consider two types of loss functions, the squared- and absolute-error loss, that are commonly used in regression problems. The minimization of empirical risk based on these loss functions is performed by two optimization techniques, the gradient boosting and the least angle technique. The main advantage of decision rules is their simplicity and good interpretability. The prediction model in the form of an ensemble of decision rules is powerful, which is shown by results of the experiment presented in the paper.

We present a novel approach to meta-learning, which is not just a ranking of methods, not just a strategy for building model committees, but an algorithm performing a search similar to what human experts do when analyzing data, solving full scope of data mining problems. The search through the space of possible solutions is driven by special mechanisms of machine generators based on meta-schemes. The approach facilitates using human experts knowledge to restrict the search space and gaining meta-knowledge in an automated manner. The conclusions help in further search and may also be passed to other meta-learners. All the functionality is included in our new general architecture for data mining, especially eligible for meta-learning tasks.

The development of effective interestingness measures that help in interpretation and evaluation of the discovered knowledge is an active research area in data mining and machine learning. In this paper, we consider a new Bayesian confirmation measure for ”

if..., then...

” rules proposed in [4]. We analyze this measure, called

Z

, with respect to valuable property M of monotonic dependency on the number of objects in the dataset satisfying or not the premise or the conclusion of the rule. The obtained results unveil interesting relationship between

Z

measure and two other simple and commonly used measures of rule support and anti-support, which leads to efficiency gains while searching for the best rules.

Learning in cases that are almost linearly separable is easy, but for highly non-separable problems all standard machine learning methods fail. Many strategies to build adaptive systems are based on the “divide-and-conquer” principle. Constructive neural network architectures with novel training methods allow to overcome some drawbacks of standard backpropagation MLP networks. They are able to handle complex multidimensional problems in reasonable time, creating models with small number of neurons. In this paper a comparison of our new constructive

c3sep

algorithm based on

k

-separability idea with several sequential constructive learning methods is reported. Tests have been performed on parity function, 3 artificial Monks problems, and a few benchmark problems. Simple and accurate solutions have been discovered using

c3sep

algorithm even in highly non-separable cases.

In order to classify an unseen (query) vector

q

with the

k

-Nearest Neighbors method (

k

-NN) one computes a similarity function between

q

and training vectors in a database. In the basic variant of the

k

-NN algorithm the predicted class of

q

is estimated by taking the majority class of the

q

’s

k

-nearest neighbors. Various similarity functions may be applied leading to different classification results. In this paper a heterogeneous similarity function is constructed out of different 1-component metrics by minimization of the number of classification errors the system makes on a training set. The HSFL-NN system, which has been introduced in this paper, on five tested datasets has given better results on unseen samples than the plain

k

-NN method with the optimally selected

k

parameter and the optimal homogeneous similarity function.

This paper presents a method of music tunes recognition based on adopted Hough transform. One can also find here experimental results showing the effectiveness of the presented solution. Perspectives of further work and quality improvements are also stated as a base for subsequent research.

In this paper we propose and analyse a

γ

-margin generalisation of the perceptron learning algorithm of Rosenblatt. The difference between the original approach and the

γ

-margin approach is only in the update step. We consider the behaviour of such a modified algorithm in both separable and non-separable case and also when the

γ

-margin is negative. We give the convergence proof of such a modified algorithm, similar to the classical proof by Novikoff. Moreover we show how to change the margin of the update step in the progress of the algorithm to obtain the maximal possible margin of separation. In application part, we show the connection of the maximal margin of separation with SVM methods.

In this paper we analyze Support Vector Machine (SVM) algorithm to the problem of chemical compounds screening with a desired activity, definition of hits. The support vector machine transforms the input data in an (unknown) high dimensional feature space and the kernel technique is applied to calculate the inner-product of feature data.The problem of automatically tuning multiple parameters for pattern recognition SVMs using our new introduced kernel for chemical compounds is considered. This is done by simple eigen analysis method which is applied to the matrix of the same dimension as the kernel matrix to find the structure of feature data, and to find the kernel parameter accordingly. We characterize distribution of data by the principle component analysis method.

The paper deals with the extraction of features for statistical pattern recognition. Bayes probability of correct classification is adopted as the extraction criterion. The problem with complete probabilistic information is discussed and next the Bayes-optimal feature extraction procedure for the supervised classfication is presented in detail. As method of solution of optimal feature extraction a genetic algorithm is proposed. Several computer experiments for wide spectrum of cases were made and their results demonstrating capability of proposed approach to solve feature extraction problem are presented.

We show that the localization problem for multilevel wireless sensor networks (WSNs) can be solved as a pattern recognition with the use of the Support Vector Machines (SVM) method. In this paper, we propose a novel hierarchical classification method that generalizes the SVM learning and that is based on discriminant functions structured in such a way that it contains the class hierarchy. We study a version of this solution, which uses a hierarchical SVM classifier. We present experimental results the hierarchical SVM classifier for localization in multilevel WSNs.

Shannon entropy used in standard top-down decision trees does not guarantee the best generalization. Split criteria based on generalized entropies offer different compromise between purity of nodes and overall information gain. Modified C4.5 decision trees based on Tsallis and Renyi entropies have been tested on several high-dimensional microarray datasets with interesting results. This approach may be used in any decision tree and information selection algorithm.

This article presents an example of a handwritten numbers classifier based on the immune system. We study mutual relations between the system operation parameters, as well as new mechanisms introduced in order to make the system work faster. To achieve the goal the weights inspired by the information theory have been inserted to the immune system.

Bayes’ rule is the basis of probabilistic reasoning. It enables to surmount information gaps. However, it requires the knowledge of prior distributions of probabilistic variables. If this distribution is not known then, according to the principle of indifference, the uniform distribution has to be assumed. The uniform distribution is frequently and heavily criticized. The paper presents a safe distribution of probability density that can be often used instead of the uniform distribution to surmount information gaps. According to the authors’ knowledge the concept of the safe distribution is new and unknown in the literature.

We consider a multi-class pattern recognition problem with linearly ordered labels and a loss function, which measures absolute deviations of decisions from true classes. In the bayesian setting the optimal decision rule is shown to be the median of a posteriori class probabilities. Then, we propose three approaches to constructing an empirical decision rule, based on a learning sequence. Our starting point is the Parzen-Rosenblatt kernel density estimator. The second and the third approach are based on radial bases functions (RBF) nets estimators of class densities.

In this paper, we propose a flexible parallel ant colony algorithm for classification rule discovery in the large databases. We call this algorithm Parallel Ant-Miner2. This model relies on the extension of real behavior of ants and data mining concepts. The artificial ants are firstly generated and separated into several groups. Each group is assigned a class label which is the consequent parts of the rules it should discover. Ants try to discover rules in parallel and then communicate with each other to update the pheromones in different paths. The communication methods help ants not to gather irrelevant terms of the rule. The parallel executions of ants reduce the speed of convergence and consequently make it possible to extract more new high quality rules by exploring all search space. Our experimental results show that the proposed model is more accurate than the other versions of Ant-Miner.

It is well-known that maintenance and evolution represent important stages in the lifecycle of any software system (about 66% from the total cost of the software systems development). That is why in this paper we are focusing on the problem of automating an essential activity that appears in the maintenance and evolution of software systems: the problem of identifying refactorings that would improve the structure of the system.

Refactoring

is the process of improving the design of software systems, by improving their internal structure, without altering the external behavior of the code. The aim of this paper is to introduce a new clustering algorithm,

CASYR

(

Clustering Algorithm for Software Systems Restructuring

), that can be used for improving software systems design, by identifying the appropriate refactorings. The proposed approach can be useful for assisting software engineers in their daily work of refactoring software systems. We evaluate our approach on a real software system and we also provide a comparison with previous approaches.

Recently, a considerable growth of interest in using Nonnegative Matrix Factorization (NMF) for pattern classification and data clustering has been observed. For nonnegative data (observations, data items, feature vectors) many problems of partitional clustering can be modeled in terms of a matrix factorization into two groups of vectors: the nonnegative centroid vectors and the binary vectors of cluster indicators. Hence our data partitional clustering problem boils down to a semi-binary NMF problem. Usually, NMF problems are solved with an alternating minimization of a given cost function with multiplicative algorithms. Since our NMF problem has a particular characteristics, we apply a different algorithm for updating the estimated factors than commonly-used, i.e. a binary update with simulated annealing steering. As a result, our algorithm outperforms some well-known algorithms for partitional clustering.

In this paper, we propose a new Association Rule Mining algorithm for Classification (ARMC). Our algorithm extracts the set of rules, specific to each class, using a fuzzy approach to select the items and does not require the user to provide thresholds. ARMC is experimentaly evaluated and compared to state of the art classification algorithms, namely CBA, PART and RIPPER. Results of experiments on standard UCI benchmarks show that our algorithm outperforms the above mentionned approaches in terms of mean accuracy.

In this paper two level model of handwritten word recognition is considered. On the first level the consecutive letters are recognized by the same classifier using preprocessed data from the optical device, while on the second level we try to recognize the whole word. From the other point of view we can treat the first level as a feature reduction level. Then, in this paper two different methods of feature reduction for handwritten word recognition algorithm are described. On the lower level different well-known in the literature methods are taken into account (for example multi layer perceptron, k-NN algorithm). The results of classification from the first level serve as a feature for the second level and two different cases are considered. The first one consist in taking into account the crisp result of classification from the first level while in the second approach we take into account the support vector of decision on this level. On the second level, in order to improve the word recognition accuracy, for both methods, the Probabilistic Character Level Language Model was applied. In this model, the assumption of first-order Markov dependence in the sequence of characters was made. Moreover, we comment the possibility of using Markov model in forward and backward directions. For both methods of feature reduction the appropriate word recognition algorithms are presented. In order to find the best solution, the Viterbi algorithm is used. A number of experiments were carried out to test the properties of the proposed methods of feature reduction. The experiment results are presented and concluded in the end of the paper.

This paper applies dynamic neural fields model [1,23,7] to multimodal interaction of sensory cues obtained from a mobile robot, and shows the impact of different temporal aspects of the integration to the precision of movements. We speculate that temporally uncoordinated sensory integration might be a reason for the poor motor skills of patients with autism. Accordingly, we make a simulation of orientation behavior and suggest that the results can be generalized for grasping and other movements that are performed in three dimensional space. Our experiments show that impact of temporal aspects of sensory integration on the precision of movement are concordant with behavioral studies of sensory integration dysfunction and of autism. Our simulation and the robot experiment may suggest ideas for understanding and training the motor skills of patients with sensory integration dysfunction, and autistic patients in particular, and are aimed to help design of games for behavioral training of autistic children.

This paper presents a novel approach to the problem of image reconstruction from projections using recurrent neural network. The reconstruction process is performed during the minimizing of the energy function in this network. Our method is of a great practical use in reconstruction from discrete fan-beam projections. Experimental results show that the appropriately designed neural network is able to reconstruct an image with better quality than obtained from conventional algorithms.

In this work we propose a new method for ultrasound imaging segmentation in objective to help doctors and specialists to interpret anatomical structure. The proposed method is based on fuzzy fusion theory in objective to extract the venous thrombosis contour in ultrasound images acquired in vivo case. The first obtained results by optimization algorithm, adapted to our particular problem case are presented.

This work describes a velocity field controller for a mobile platform based on the estimation of its position using odometry. It is a known fact that this approach has an unbounded error attached. The approach here proposed uses a position estimation obtained by odometry, which uses the six components of the velocity calculated on the basis of the displacement vectors resulted from optical flow estimation. The control loop proposed takes the position estimation as reference and performs a velocity field control. Results obtained shows that, for closed trajectories, the tracking error is bounded.

Automatic speech recognition (ASR) is an area where the task is to assign the correct phoneme or word sequence to an utterance. The idea behind the ASR segment-based approach is to treat one phoneme as a whole unit in every respect, in contrast with the frame-based approach where it is divided into equal-sized, smaller chunks. Doing this has many advantages, but also gives rise to some new problems. One of these is the detection of potential bounds between phones, which has an effect on both the recognition accuracy and the speed of the speech recognition system. In this paper we present three ways of boundary detection: first two simple algorithms are tested, then we will concentrate on our novel method which incorporates a spiking neuron. On examining the test results we find that the latter algorithm indeed proves successful: we were able to speed up the recognition process by 35.72% while also slightly improving the recognition performance.

Image processing algorithms are being intensively researched in the last decades. One of the most influential filtering tendencies is based on partial differential equations (PDE). Different kinds of modifications of classical linear process were already proposed. Most of them are based on non-linear or anisotropic process taking into consideration local descriptor of image structure. Main goal is to remove noise and simultaneously to decrease level of blurring important features (like edges). In this paper a new approach is presented, which introduces, into non-linear diffusion process, extra knowledge about geometric structures existing on an image. Algorithm scheme is proposed and results of numerical experiments are presented. Moreover, possibilities of algorithm application within cellular neural networks paradigm will be analysed.

A multiagent system to support a mobile robot motion planning has been presented. Baptized

MARCoPlan

(MutiAgent Remote Control motion Planning), this system deals with optimizing robot path. Considered as an agent, the robot has to optimize its motion from a start position to a final goal in a dynamic and unknown environment, on the one hand by the introduction of sub-goals, and on the other hand by the cooperation of multiagents. In fact, we propose to agentify the proximity environment (zones) of the robot; cooperation between theses zones agents will allow the selection of the best sub-goal to be reached. Therefore, the task of the planner agent to guide the robot to its destination in an optimized way will be easier.

MARCoPlan

is simulated and tested using randomly and dynamically generated problem instances with different distributions of obstacles. The tests verify some robustness of

MARCoPlan

with regard to environment changes. Moreover, the results highlight that the agentification and the cooperation improve the choice of the best path to the sub goals, then to the final goal.

Speaker identification is the process of identifying an unknown speaker from a set of known speakers. In a speaker identification or verification, the prime interest is not in recognizing the words but determining who is speaking the words. In systems of speaker identification, a test of signal from an unknown speaker is compared to all known speaker signals in the set. The signal that has the maximum probability is identified as the unknown speaker. In security systems based on speaker identification, faultless identification has huge meaning for safety.

In aim of increasing safety, in this work it was proposed own approach to user identification, based on independent speech and facial asymmetry. Extraction of the audio features of person’s speech is done using mechanism of cepstral speech analysis.

The part of the work that deals with face recognition was based on the technique of automatic authentication of a person with assumption that the use of automatically extracted, structural characteristics of the face asymmetry (in particular within the eyes and mouth regions as the most informative parts of the face) leads to improvement of the biometrical authentication systems.

Finally, the paper will show results of user identification.

Few years old children lift and manipulate unfamiliar objects more dexterously than today’s robots. Therefore, it has arisen an interest at the artificial intelligence community to look for inspiration on neurophysiological studies to design better models for the robots. The estimation of the friction coefficient of the object’s material is a crucial information in a human dexterous manipulation. Humans estimate the friction coefficient based on the responses of their tactile mechanoreceptors. In this paper, finite element analysis was used to model a finger and an object. Simulated human afferent responses were then obtained for different friction coefficients. Multiple multilayer perceptrons that received as input simulated human afferent responses, and gave as output an estimation of the friction coefficient, were trained and tested. A performance analysis was carried out to verify the influence of the following factors: number of hidden neurons, compression ratio of the input pattern, partitions of the input pattern.

The digital image watermarking field addresses the problem of digital image authentication and integrity. In this paper we propose a novel color image watermarking scheme based on image self-embedding and self-recovery techniques. The main idea of this algorithm is to embed a reduced content of the original image to itself, in order to be able to partially recover the deleted features from the watermarked image. Separately, the red and blue color channels are embedded, respectively in the wavelet domain by a compressed version of the original image, and in the spatial domain by binary encoded sequences generated from the original image. This allowed us, in detection stage, to prove the ownership, detect the altered blocks, and recover them. The detection and recovery bits extraction is computed using an ICA algorithm. The experimental results were satisfactory and show a high robustness against most common attacks as well as a reassuring rate of image recovery.

Recently, research for the summarization of video data has been studied a lot due to the proliferation of user created contents. Besides, the use of multiple cameras for the collection of the video data has been increasing, but most of them have used the multi-camera system either to cover the wide area or to track moving objects. This paper focuses on getting diverse views for a single event using multi-camera system and deals with the problem of summarizing event sequences collected in the office environment based on this perspective. Summarization includes camera view selection and event sequence summarization. View selection makes a single event sequence from multiple event sequences as selecting optimal views in each time, for which domain ontology based on the elements in an office environment and rules from questionnaire surveys have been used. Summarization generates a summarized sequence from a whole sequence, and the fuzzy rule-based system is used to approximate human decision making. The degrees of interests input by users are used in both parts. Finally, we have confirmed that the proposed method yields acceptable results using experiments of summarization.

In this paper we propose a new method for image retrieval with relevance feedback based on eliciting preferences from the decision-maker acquiring visual information from an image database. The proposed extension of the common approach to image retrieval with relevance feedback allows it to be applied to objects with non-homogenous colour and texture. This has been accomplished by the algorithms, which model user queries by an RBF neural network. As an example of application of this approach, we have used a content-based search in an atlas of species. An experimental comparison with the commonly used content-based image retrieval approach is presented.

The paper deals with image compression performed using an adaptive fast transform-based method. The point of departure is a base scheme for fast computation of certain discrete transforms. The scheme can be interpreted in terms of the neural architecture whose parameters (neurons’ weights) can be adjusted during learning on set data, here images. The same basic network topology enables realization of diverse transformations. The results obtained for the task of image compression are presented and evaluated.

The following paper introduces a set of novel descriptors of emotional speech, which allows for a significant increase in emotion classification performance. The proposed characteristics - statistical properties of Poincare Maps, derived for voiced-speech segments of utterances - are used in recognition in combinations with a variety of both commonly used and some other, original descriptors of emotional speech. The introduced features proved to provide useful information into a classification process. Emotion recognition is performed using binary decision trees, which perform extraction of different emotions at consecutive decision levels. Classification rates for the considered six-category problem, which involved anger, boredom, joy, fear, neutral and sadness, are at the level up to 79% for both speaker-dependent and speaker-independent cases.

The ability of a robot to do gaze-drawing and to gesture have become recognized as essential elements in achieving joint attention of the real world during human-robot interaction. However, the ability of a robot using such non-verbal actions to interrupt human action is unclear. We have conducted an experiment in which a robot tries to interrupt human action to demonstrate the effectiveness of an interactive robot acting simultaneously with a person. Using the results of this experiment, we developed a system that a robot can use to predict human motion so that the robot can automatically perform simultaneous behavior.

Principal rules defining the adaptation of ECG interpretation software in a distributed surveillance network are presented in this paper. Thanks to the pervasive access to wireless digital communication services, the intelligent monitoring networks automatically solve difficult medical cases thanks to the auto-adaptation of data interpretation and transmission to the variable patient status and technical constrains. The foundation of this innovative approach is the use of selected diagnostic parameters in a loopback modifying the running interpretive software. The auto adaptive process maximizes the general estimate of patient description quality aggregating the divergence values of particular parameters modulated by the medical relevance factor dependent on the status of patient. Our approach is motivated by the outcomes from the research on human experts behavior, statistics of the procedures reliability and usage as well as tests in a prototype client-server application. The tests yielded very promising results: the convergence of the remotely computed diagnostic outcome was achieved in over 80% of software adaptation attempts. Comparing to the rigid reporting mode, avoiding unnecessary computation extends the autonomy time by 65% and the transmission channel occupation was reduced by 3,1 to 5,6 times.

In this paper we reported initial work at development of a human-machine interface system for people with severe disabilities based on measurement of electrooculography signal (EOG). We proposed a system for detecting and predicting eyes position using EOG. We applied a zero-order Takagi-Sugeno-Kang model with modified inference procedure for this task. To calculate values of parameters of fuzzy system we used results of EOG signal segmentation. Experimentation shows the usefulness of the presented method for improving functionality of interface systems that can assist people with limited upper body mobility.

It is well-known that people with autism or mental retardation experience crucial problems in thinking and communicating using linguistic structures. Thus, we foresee the emergence of text-to-image conversion systems to let such people establish a bridge between linguistic expressions and the concepts these expressions refer to via relevant images. S2S is such a system for converting Turkish sentences into representative 3D scenes via the mediation of an HPSG-based NLP module. A precursor to S2S, a non-3D version, has been tested with a group of students with autism and mental retardation in a special education center and has provided promising results motivating the work presented in this paper.

The analysis of small molecular substructures (like enzyme active sites) in the whole protein structure can be supported by using methods of similarity searching. These methods allow to search the 3D structural patterns in a database of protein structures. However, the well-known methods of fold similarity searching like VAST or DALI are not appropriate for this task. Methods that benefit from a dependency between a spatial conformation and potential energy of protein structure seem to be more supportive. In the paper, we present a new version of the EAST (Energy Alignment Search Tool) algorithm that uses energy signatures in the process of similarity searching. This makes the algorithm not only more sensitive, but also eliminates disadvantages of previous implementations of our EAST method.

Applying semantic pervasive services to Biomedical research is providing a new breed of intelligent applications which can tackle with the heterogeneity and intrinsic complexity of biomedical information integration. Using semantics leverages the potential of enabling cross-interoperability among a variety of storage and data formats widely distributed both across the Internet and within individual organizations. In this paper, we present SYMBIOS, a fully-fledged biomedical information integration solution based on semantic pervasive services that combine a Service Oriented Architecture (SOA) and semantically-empowered techniques to ascertain biomedical information intelligent integration. We discuss our approach with a proof-of-concept implementation where the breakthroughs and efficiency of integrating the biomedical publications database MEDLine, the Database of Interacting Proteins (DIP) and the Munich Information Center for Protein Sequences (MIPS) has been tested.

The Polymerase Chain Reaction process is a well-known technique for the

in vitro

amplification of a DNA sequence. The success of a PCR depends on several parameters particularly the primer sequences used. Since the design of a suitable pair of primer involves a reasonable number of variables, which can have a range of different values, computer programs are commonly used to assist this task. This paper approaches the design of a pair of primer sequences as a search process throughout the space defined by all possible primer sequence pairs, directed by an evaluation function that combines the many variables involved in a primer design; an experiment and its results are discussed.

In this paper we develop a neuro-fuzzy system for stroke diagnosis. A novel concept of weights describing importance of antecedents and rules will be incorporated into construction of such systems. Simulation results based on 298 real stroke data will be presented.

Novel method for spleen’s semiautomatic accurate quantitative morphometry adaptable in diagnosis of splenomegally is described. The method is based on multiscale wavelet image decomposition, Bayesian inference that reveals the most probable structure delineation in the image and spectral method to smooth and approximate the most probable representation of a real contour hidden in a noisy or fuzzy data.

This paper, as well as coupled paper [2], deal with various aspects of scheduling algorithms dedicated for processing in parallel computing environments. In this paper, for the exemplary problem, namely the flow-shop scheduling problem with makespan criterion, there are proposed original methods for parallel analysis of a solution as well as a group of concentrated and/or distributed solutions, recommended for the use in metaheuristic approaches with single-thread trajectory. Such methods examine in parallel consecutive local sub-areas of the solution space, or a set of distributed solutions called population, located along the

single

trajectory passed through the space. Supplementary multi-thread search techniques applied in metaheuristics have been discussed in complementary our paper [2].

This paper proposes a novel model, based on the artificial immune system, to solve the problem of short-term load forecasting. An artificial immune system is trained to recognize antigens which encode sequences of load time series. The created immune memory is a representation of these sequences. In the forecast procedure a new incomplete antigen, containing only the first part of the sequence, is presented to the model. The second forecasted part of the sequence is reconstructed from activated antibodies. The model was verified using several real data examples of the short-term load forecast.

This paper presents a new algorithm for hypertext graph crawling. Using an ant as an agent in a hypertext graph significantly limits amount of irrelevant hypertext documents which must be downloaded in order to download a given number of relevant documents. Moreover, during all time of the crawling, artificial ants do not need a queue to central control crawling process. The proposed algorithm, called the Focused Ant Crawling Algorithm, for hypertext graph crawling, is better than the Shark-Search crawling algorithm and the algorithm with best-first search strategy utilizing a queue for the central control of the crawling process.

Clinical knowledge building upon evidence-based medicine is typically represented in textual guidelines, thus providing a rather informal description from a logical point of view. Further, the context which provides utility of these guidelines is not specified in any detail with respect to workflow and underlying motivations for decision-making. In addition, the level of detail is mostly static in the sense that measurements and decision values are fixed and intended for specific user groups. There is thus a lack of flexibility which disables knowledge to be shifted coherently between user levels in the entire workflow and decision process. In this paper, we will discuss formalizations of the underlying logical structures of guidelines from the viewpoint of being represented appropriately at each user level. Further, to establish a formal correctness criterion, the shift from one level of representation to another is required to be morphic in the categorical sense. General logics [7] is the selected generalized, and categorical, framework for our approach to flexible guideline representation. Our medical scope is dementia differential diagnosis based on consensus guidelines [2], and we focus on types of cognitive disorders as a prerequisite for further diagnostic tasks.

Reinforcement learning tackles the problem of how to act optimally given observations of the current world state. Agents that learn from reinforcements execute actions in an environment and receive feedback (reward) that can be used to guide the learning process. The distinguishing feature of reinforcement learning is that the model of the environment (i.e., effects of actions or the reward function) are not known in advance. Model-based approaches represent a class of reinforcement learning algorithms which learn the model of dynamics. This model can be used by the learning agent to simulate interactions with the environment. DynaQ and its extended version with prioritised sweeping are the most popular examples of model-based approaches. This paper shows that, contrary to common belief, DynaQ with prioritised sweeping may perform worse than pure DynaQ in domains where the agent can be easily misled by a sub-optimal solution.

A combined routing-scheduling problem is considered in the paper. It consists in scheduling of tasks on moving executors. The case with non-preemptive and independent tasks, unrelated executors as well as interval execution times to minimize the makespan is investigated. The worst-case scheduling problem based on an absolute regret is formulated. Solution algorithms of polynomial computational complexity, which use simulated annealing as well as tabu search approaches, are presented. The results of numerical experiments are given.

People mainly organize their experience as a kind of narratives. Story generated from user’s information in mobile environment can help share his experience with other people and recall his meaningful memory. In this paper, we propose a method that generates a story with Petri net and user contexts. In order to verify the usefulness of the proposed method, we show an example of generating user’s experience to story with user context in mobile environment. Comparison of user’s report and generated story confirms the validity of the automatic story generation.

Belief propagation (BP) is effective for computing marginal probabilities of a high dimensional probability distribution. Loopy belief propagation (LBP) is known not to compute precise marginal probabilities and not to guarantee its convergence. The fixed points of LBP are known to accord with the extrema of Bethe free energy. Hence, the fixed points are analyzed by minimizing the Bethe free energy.

In this paper, we consider the Bethe free energy in Gaussian distributions and analytically clarify the extrema, equivalently, the fixed points of LBP for some particular cases. The analytical results tell us a necessary condition for LBP convergence and the quantities which determine the accuracy of LBP in Gaussian distributions. Based on the analytical results, we perform numerical experiments of LBP and compare the results with analytical solutions.

The problem of algorithm selection for solving NP problems arises with the appearance of a variety of heuristic algorithms. The first works claimed the supremacy of some algorithm for a given problem. Subsequent works revealed the supremacy of algorithms only applied to a subset of instances. However, it was not explained why an algorithm solved better a subset of instances. In this respect, this work approaches the problem of explaining through causal model the interrelations between instances characteristics and the inner workings of algorithms. For validating the results of the proposed approach, a set of experiments was carried out in a study case of the Threshold Accepting algorithm to solve the Bin Packing problem. Finally, the proposed approach can be useful for redesigning the logic of heuristic algorithms and for justifying the use of an algorithm to solve an instance subset. This information could contribute to algorithm selection for NP problems.

A graphical method for the modeling of compromise in multiple criteria problems solution is proposed. The method is based on the analysis of the strategic games characteristics and takes into account both the players cooperation for the compromise solution searching and the influence of rejection of the better solutions in favor of the compromise ones. The visualization of the considered problem is based on triangle type representation of the local criteria and can be realized both in deterministic and interval or fuzzy versions. The methodology for building the models of compromise based on the comparative analysis of possible solutions is proposed. In comparison with the approaches based the on the polygon method in games theory [32], our proposition is evidently less algorithmic complex and seems as more suitable for the comparative analysis. Its usefulness becomes especially apparent is the case of small number of local criteria.

The paper presents neural networks as models for prediction of the water intake. For construction of prediction models three types of neural networks were used: linear network, multi-layer network with error backpropagation and Radial Basis Function network (RBF).

The prediction models were compared for obtaining optima quality prognosis. Prediction models were done for working days, Saturdays and Sundays. The research was done for selected nodes of water supply system: detached house node and nodes for 4 hydrophore stations from different pressure areas of water supply system. Models for Sundays were presented in detail. Further research concerning the creation of prognosis models should be directed towards constructing models not only for particular days, but also for the complete week, four seasons of the year: spring, summer, autumn and winter, and finally the entire year.

An application of neuro-fuzzy systems to supporting trading decisions is presented. The system has the ability to use expert knowledge and to be fitted to the learning data by various machine learning techniques. The proposed approach uses the backpropagation algorithm to determine system parameters on the basis of several indices. Experiments were made on past data showing relatively good performance of the proposed approach.

This paper demonstrates that AI methods can be applied to the development of intelligent IT systems. They also facilitate an in-depth analysis of the meaning presented in cognitive categorization information systems - in particular UBIAS systems (

Understanding Based Image Analysis Systems

). This paper also presents the IT mechanisms of object meaning description on selected examples of long bone fractures image analysis. The procedures for such semantic reasoning are based on the model of cognitive resonance and cognitive analysis. These have been applied to the task of interpreting the meaning of selected diagnostic images from the long bone fractures system as an intelligent analysis module in IT systems. The application presented in this paper is of a research character and it serves the preparation of efficient lesion detection methods applied to a dataset originating from of the long bone fractures structures.

The subject of this paper is artificial intelligence (AI) of non-player characters in computer games, i.e. bots. We develop an idea of game AI based on predictive control. Bot’s activity is defined by a currently realized plan. This plan results from an optimization process in which random plans are continuously generated and reselected.

We apply our idea to implement a bot for the game Half-Life. Our bot, Randomly Planning Fighter (RPF), defeats the bot earlier designed for Half-Life with the use of behavior-based techniques. The experiments prove that on-line planning can be feasible in rapidly changing environment of modern computer games.

Time reduction of steel scraps meltdown during the electic arc process is really a challenging problem. Typically the EAF process is stochastic without any determinism and only simple and naive rules are currently used to manage such processes. The goal of the paper is to present the way, which have been considered, to build an accurate model concerning different feature selection methods that would be helpful in predicting the end of the meltdown and maximum energy needed by the furnace.

The basic reinforcement learning algorithm, as Q-learning, is characterized by short time-consuming single learning step, however, the number of epochs necessary to achieve the optimal policy is not satisfactory. There are many methods that reduce the number of necessary epochs, like TD(

λ

> 0), Dyna or prioritized sweeping, but their learning time is considerable. This paper proposes a combination of Q-learning algorithm performed in incremental mode with executed in epoch mode method of acceleration based on environment model and distance to terminal state. This approach ensures the maintenance of short time of a single learning step and high efficiency comparable with Dyna or prioritized sweeping. Proposed algorithm is compared with Q(

λ

)-learning, Dyna-Q and prioritized sweeping in the experiments on three maze tasks. The time-consuming learning process and number of epochs necessary to reach the terminal state is used to evaluate the efficiency of compared algorithms.

Currently, we are developing an agent-based travel support system, in which ontologically demarcated data is used to facilitate personalized information provisioning. Recently we have shown how Open Travel Alliance golf-related messages can be reverse-engineered to create an

OTA ontology of golf

. The aim of this paper is to illustrate how these ontologies are going to be used in the system. In addition to the general scenario, details concerning implementation of needed translators will be discussed.

This paper presents a model designed under the formalism of Bayesian Networks to deal with the problem of collaborative recommendation. It has been designed to perform efficient and effective recommendations. We also consider the fact that the user can usually use vague ratings for the products, which might be represented as fuzzy labels. The complete proposal is evaluated with MovieLens.

Imitation learning is a promising route to instruct robotic multi-agent systems. However, imitating agents should be able to decide autonomously what behavior, observed in others, is interesting to copy. Here we investigate whether a simple recurrent network (Elman Net) can be used to extract meaningful chunks from a continuous sequence of observed actions. Results suggest that, even in spite of the high level of task specific noise, Elman nets can be used for isolating re-occurring action patterns in robots. Limitations and future directions are discussed.

In our work we are designing agent based support for workers in a virtual organization. In the system under development, all resources are to be ontologically demarcated and utilized through semantically- driven information processing techniques. This paper contains preliminary considerations concerning resource profiling and ontological matching methods which we intend to utilize within the system. Two applications that we are currently developing are used to illustrate the proposed approach.

Intelligent Agents and Semantic Web Services are two technologies with great potential in the future. Striking new applications can be developed by using them. However, in order for these technologies to be fully applicable to real settings, several problems remain unsolved. In particular, Semantic Web Services need for an upper software entity able to deal with them, while Intelligent Agents have historically suffered from a number of drawbacks not yet resolved. In this paper, we claim that integrating these technologies into a joint environment can overcome part of their respective problems while strengthening their advantages. The necessity for integrating these technologies and its benefits are analyzed in this work. Based on this study, we present SEMMAS, a framework for seamlessly integrating Intelligent Agents and Semantic Web Services. The fundamentals of the framework and its architecture are explained and a proof-of-concept implementation is described.

Graph transformations are a powerful notation formally describing different aspects of the software systems. Multiagent systems are one of the most promising ways to introduction of the parallel computation. The reason of difficulties in joining these approaches was centralized way of the graph representation. The GRADIS framework offers the possibility of splitting the graph describing the problem onto a few partial graphs, that can be maintained in different places. Moreover, the GRADIS distributed environment makes the application of old rules possible, in order to modify these set of partial graphs. Basing on this framework, we show how to introduce the multiagent system and we show an example estimation of improvement time complexity of the multiagent system in comparison to the centralized one.

We introduce a multi-agent logic

${ {\cal M}{\cal A}_{{\cal L}{\cal T}{\cal L}}}$

– a variant of the linear temporal logic LTL with embedded multi-agent knowledge with interacting agents. The logic is motivated by semantics based on potentially infinite runs with time points represented by clusters of states with distributed knowledge of the agents. We address properties of local and global knowledge modeled in this framework, consider modeling of interaction between agents by possibility to puss information from one agent to others via possible transitions within time clusters of states. Main question we are focused on is the satisfiability problem and decidability of the logic

${ {\cal M}{\cal A}_{{\cal L}{\cal T}{\cal L}}}$

. Key result is proposed algorithm which recognizes theorems of

${ {\cal M}{\cal A}_{{\cal L}{\cal T}{\cal L}}}$

(so we show that

${ {\cal M}{\cal A}_{{\cal L}{\cal T}{\cal L}}}$

is decidable). It is based on verification of validity for special normal reduced forms of rules in models with at most triple exponential size in the testing rules. In the final part we discuss possible variations of the proposed logic.

This paper presents the problem of the coordination of actions in a multi-agent system. The main difficulty in resolving this problem is the limited information. Each agent is provided only with partial information about the state of the team of agents. This paper presents a hybrid technique that combines the game theory tool and the voting schema, which is applied to create the method of coordination that deals with this problem. Appropriate simulation results of the proposed techniques are presented.