Catálogo de publicaciones - libros

Training multilayer neural networks is typically carried out using gradient descent techniques. Ever since the brilliant (BP), the first gradient-based algorithm proposed by Rumelhart et al., novel training algorithms have appeared to become better several facets of the learning process for feed-forward neural networks. is one of these. In this paper, a learning algorithm that applies linear-least-squares is presented. We offer the theoretical basis for the method and its performance is illustrated by its application to several examples in which it is compared with other learning algorithms and well known data sets. Results show that the new algorithm upgrades the learning speed of several backpropagation algorithms, while preserving good optimization accuracy. Due to its performance and low computational cost it is an interesting alternative, even for second order methods, particularly when dealing large networks and training sets.

Explaining the causes of infeasibility of Boolean formulas has many practical applications in various fields. A small unsatisfiable subformula provides a succinct explanation of infeasibility and is valuable for applications. In recent years the problem of finding unsatisfiable subformulas has been addressed frequently by research works, which are mostly based on the SAT solvers with DPLL backtrack-search algorithm. However little attention has been concentrated on extraction of unsatisfiable subformulas using stochastic methods. In this paper, we propose a resolution-based stochastic local search algorithm to derive unsatisfiable subformulas. This approach directly constructs the resolution sequences for proving unsatisfiability with a local search procedure, and then extracts small unsatisfiable subformulas from the refutation traces. We report and analyze the experimental results on benchmarks.

Previous clustering ensemble algorithms usually use a consensus function to obtain a final partition from the outputs of the initial clustering. In this paper, we propose a new clustering ensemble method, which generates a new feature space from initial clustering outputs. Multiple runs of an initial clustering algorithm like k-means generate a new feature space, which is significantly better than pure or normalized feature space. Therefore, running a simple clustering algorithm on generated feature space can obtain the final partition significantly better than pure data. In this method, we use a modification of k-means for initial clustering runs named as “Intelligent k-means”, which is especially defined for clustering ensembles. The results of the proposed method are presented using both simple k-means and intelligent k-means. Fast convergence and appropriate behavior are the most interesting points of the proposed method. Experimental results on real data sets show effectiveness of the proposed method.

Independent component analysis (ICA) solves the blind source separation problem by evaluating higher-order statistics, e.g. by estimating fourth-order moments. While estimation errors of the kurtosis can be shown to asymptotically decay with sample size according to a square-root law, they are subject to two further effects for finite samples. Firstly, errors in the estimation of kurtosis increase with the deviation from Gaussianity. Secondly, errors in kurtosis-based ICA algorithms increase when approaching the Gaussian case. These considerations allow us to derive a strict lower bound for the sample size to achieve a given separation quality, which we study analytically for a specific family of distributions and a particular algorithm (fastICA). We further provide results from simulations that support the relevance of the analytical results.

This paper is a report on research towards the development of an abstraction-based framework for decision-theoretic planning. We make use of two planning approaches in the context of probabilistic planning: planning by abstraction and planning graphs. To create abstraction hierarchies our planner uses an adapted version of a hierarchical planner under uncertainty, and to search for plans, we propose a probabilistic planning algorithm based on Pgraphplan. The article outlines the main framework characteristics, and presents results on some problems found in the literature. Our preliminary results suggest that our planner can reduce the size of the search space, when compared with Pgraphplan, hierarchical planning under uncertainty and top-down dynamic programming.

In the paper a method of the correction of handwriting Optical Character Recognition (OCR) based on the semantic similarity is presented. Different versions of the extraction of semantic similarity measures from a corpus are analysed, with the best results achieved for the combination of the text window context and Rank Weight Function. An algorithm of the word sequence selection with the high internal similarity is proposed. The method was trained and applied to a corpus of real medical documents written in Polish.

This paper presents a method for combining classifiers that uses -nearest localized templates. The localized templates are estimated from a training set using -means clustering algorithm, and matched to the decision profile of a new incoming sample by a similarity measure. The sample is assigned to the class which is most frequently represented among the most similar templates. The appropriate value of is determined according to the characteristics of the given data set. Experimental results on real and artificial data sets show that the proposed method performs better than the conventional fusion methods.

Many intermediate program representations are used by compilers and other software development tools. In this paper, we propose a novel representation technique that, unlike those commonly used by compilers, has been explicitly designed for facilitating program element matching, a task at the heart of many software mining problems.

A new approach is proposed for the data-based identification of transparent fuzzy rule-based classifiers. It is observed that fuzzy rule-based classifiers work in a similar manner as kernel function-based support vector machines (SVMs) since both model the input space by nonlinearly maps into a feature space where the decision can be easily made. Accordingly, trained SVM can be used for the construction of fuzzy rule-based classifiers. However, the transformed SVM does not automatically result in an interpretable fuzzy model because the SVM results in a complex rule-base, where the number of rules is approximately 40-60% of the number of the training data. Hence, reduction of the SVM-initialized classifier is an essential task. For this purpose, a three-step reduction algorithm is developed based on the combination of previously published model reduction techniques. In the first step, the identification of the SVM is followed by the application of the Reduced Set method to decrease the number of kernel functions. The reduced SVM is then transformed into a fuzzy rule-based classifier. The interpretability of a fuzzy model highly depends on the distribution of the membership functions. Hence, the second reduction step is achieved by merging similar fuzzy sets based on a similarity measure. Finally, in the third step, an orthogonal least-squares method is used to reduce the number of rules and re-estimate the consequent parameters of the fuzzy rule-based classifier. The proposed approach is applied for the Wisconsin Breast Cancer, Iris and Wine classification problems to compare its performance to other methods.

We propose methods that allow the investigation of local modifications of aerodynamic design data represented by discrete unstructured surface meshes. A displacement measure is suggested to evaluate local differences between the shapes. The displacement measure provides information on the amount and direction of surface modifications. Using the displacement measure in conjunction with statistical methods or data mining techniques provides meaningfull knowledge from the data set for guiding further shape optimization processes.