Catálogo de publicaciones - libros

This paper proposes a novel model of support function machine (SFM) for time series predictions. Two machine learning models, namely, support vector machines (SVM) and procedural neural networks (PNN) are compared in solving time series and they inspire the creation of SFM. SFM aims to extend the support vectors to spatiotemporal domain, in which each component of vectors is a function with respect to time. In the view of the function, SFM transfers a vector function of time to a static vector. Similar to the SVM training procedure, the corresponding learning algorithm for SFM is presented, which is equivalent to solving a quadratic programming. Moreover, two practical examples are investigated and the experimental results illustrate the feasibility of SFM in modeling time series predictions.

In this paper we study the application of Bayesian network models to classify multispectral and hyperspectral remote sensing images. Different models of Bayesian networks as: Naive Bayes (NB), Tree Augmented Naive Bayes (TAN) and General Bayesian Networks (GBN), are applied to the classification of hyperspectral data. In addition, several Bayesian multi-net models: TAN multi-net, GBN multi-net and the model developed by Gurwicz and Lerner, TAN-Based Bayesian Class-Matched multi-net (tBCM) (see [1]) are applied to the classification of multispectral data. A comparison of the results obtained with the different classifiers is done.

In group decision making with linguistic information, the decision makers (DMs) usually provide their assessment information by means of linguistic variables. In some situations, however, the DMs may provide fuzzy linguistic information because of time pressure, lack of knowledge, and their limited attention and information processing capabilities. In this paper, we introduce the concepts of triangular fuzzy linguistic variable and its member function, and introduce some operational laws of triangular fuzzy linguistic variables. We propose a formula for comparing triangular fuzzy linguistic variables, and develop some operators for aggregating triangular fuzzy linguistic variables, such as the fuzzy linguistic averaging (FLA) operator, fuzzy linguistic weighted averaging (FLWA) operator, fuzzy linguistic ordered weighted averaging (FLOWA) operator, and induced FLOWA (IFLOWA) operator, etc. Based on the FLWA and IFLOWA operators, we develop a practical method for group decision making with triangular fuzzy linguistic variables, and finally, an illustrative example is given to verify the feasibility and effectiveness of the developed method.

Prediction and classification of particular faults in rotating machinery, based on a given set of measurements, could significantly reduce the overall costs of maintenance and repair. Usually the vibration signal is sampled with a very high frequency due to its nature, thus it is quite difficult to do considerably long forecasting based on the methods, which are suitable for e.g. financial time series (where the sampling frequency is smaller). In this paper new forecasting and classification technique for particular vibration signal characteristics is proposed. Suggested approach allows creating a part of control system responsible for early fault detection, which could be used for preventive maintenance of industrial equipment. Presented approach can be extended to high frequency financial data for the prediction of “faults” on the market.

The dependence of the classification error on the size of a bagging ensemble can be modeled within the framework of Monte Carlo theory for ensemble learning. These error curves are parametrized in terms of the probability that a given instance is misclassified by one of the predictors in the ensemble. Out of bootstrap estimates of these probabilities can be used to model generalization error curves using only information from the training data. Since these estimates are obtained using a finite number of hypotheses, they exhibit fluctuations. This implies that the modeled curves are biased and tend to overestimate the true generalization error. This bias becomes negligible as the number of hypotheses used in the estimator becomes sufficiently large. Experiments are carried out to analyze the consistency of the proposed estimator.

This paper investigates the application of novelty detection techniques to the problem of drug profiling in forensic science. Numerous one-class classifiers are tried out, from the simple k-means to the more elaborate Support Vector Data Description algorithm. The target application is the classification of illicit drugs samples as part of an existing trafficking network or as a new cluster. A unique chemical database of heroin and cocaine seizures is available and allows assessing the methods. Evaluation is done using the area under the ROC curve of the classifiers. Gaussian mixture models and the SVDD method are trained both with and without outlier examples, and it is found that providing outliers during training improves in some cases the classification performance. Finally, combination schemes of classifiers are also tried out. Results highlight methods that may guide the profiling methodology used in forensic analysis.

Generative Topographic Mapping (GTM) is a non-linear latent variable model that provides simultaneous visualization and clustering of high-dimensional data. It was originally formulated as a constrained mixture of distributions, for which the adaptive parameters were determined by Maximum Likelihood (ML), using the Expectation-Maximization (EM) algorithm. In this paper, we define an alternative variational formulation of GTM that provides a full Bayesian treatment to a Gaussian Process (GP)-based variation of GTM. The performance of the proposed Variational GTM is assessed in several experiments with artificial datasets. These experiments highlight the capability of Variational GTM to avoid data overfitting through active regularization.

Recently researchers have introduced methods to develop reusable knowledge in reinforcement learning (RL). In this paper, we define simple principles to combine skills in reinforcement learning. We present a skill combination method that uses trained skills to solve different tasks in a RL domain. Through this combination method, composite skills can be used to express tasks at a high level and they can also be re-used with different tasks in the context of the same problem domains. The method generates an abstract task representation based upon normal reinforcement learning which decreases the information coupling of states thus improving an agent’s learning. The experimental results demonstrate that the skills combination method can effectively reduce the learning space, and so accelerate the learning speed of the RL agent. We also show in the examples that different tasks can be solved by combining simple reusable skills.

This paper introduces a new approach, called recurring multistage evolutionary algorithm (RMEA), to balance the explorative and exploitative features of the conventional evolutionary algorithm. Unlike most previous work, the basis of RMEA is repeated and alternated executions of two different stages i.e. exploration and exploitation during evolution. RMEA uses dissimilar information across the population and similar information within population neighbourhood in mutation operation for achieving global exploration and local exploitation, respectively. It is applied on two unimodal, two multimodal, one rotated multimodal and one composition functions. The experimental results indicated the effectiveness of using different object-oriented stages and their repeated alternation during evolution. The comparison of RMEA with other algorithms showed its superiority on complex problems.

An application of cluster analysis to identify topics in a collection of posters abstracts from the Society for Neuroscience (SfN) Annual Meeting in 2006 is presented. The topics were identified by selecting from the abstracts belonging to each cluster the terms with the highest scores using different ranking schemes. The ranking scheme based on log-entropy showed better performance in this task than other more classical TFIDF schemes. An evaluation of the extracted topics was performed by comparison with previously defined thematic categories for which titles are available, and after assigning each cluster to one dominant category. The results show that repeated bisecting k-means performs better than standard k-means.