Catálogo de publicaciones - libros

Stochastic logic programs (SLPs) provide an efficient representation for complex tasks such as modelling metabolic pathways. In recent years, methods have been developed to perform parameter and structure learning in SLPs. These techniques have been applied for estimating rates of enzyme-catalyzed reactions with success. However there does not exist any method that can provide statistical inferences and compute confidence in the learned SLP models. We propose a novel approach for drawing such inferences and calculating confidence in the parameters on SLPs. Our methodology is based on the use of a popular technique, the bootstrap. We examine the applicability of the bootstrap for computing the confidence intervals for the estimated SLP parameters. In order to evaluate our methodology we concentrated on computation of confidence in the estimation of enzymatic reaction rates in amino acid pathway of Saccharomyces cerevisiae. Our results show that our bootstrap based methodology is useful in assessing the characteristics of the model and enables one to draw important statistical and biological inferences.

Multiple disorders are a daily problem in medical diagnosis and treatment, while most expert systems make an implicit assumption that only single disorder occurs in a single patient. In our paper, we show the need for performing multiple disorders diagnosis, and investigate a way of using inductive rules in our Case-based Reasoning System for diagnosing multiple disorder cases. We applied our approach to two medical casebases taken from real world applications demonstrating the promise of the research. The method also has the potential to be applied to other multiple fault domains, e.g. car failure diagnosis.

A Point Distribution Model (PDM) of the prostate has been constructed and used to automatically outline the contour of the gland in transurethral ultrasound images. We developed a new, two stages, method: first the PDM is fitted, using a multi-population genetic algorithm, to a binary image produced from Bayesian pixel classification. This contour is then used during the second stage to seed the initial population of a simple genetic algorithm, which adjusts the PDM to the prostate boundary on a grey level image. The method is able to find good approximations of the prostate boundary in a robust manner. The method and its results on 4 prostate images are reported.

An integrated method using the adaptive segmentation of brain tissues in Magnetic Resonance Imaging (MRI) images is proposed in this paper. Firstly, we give a template of brain to remove the extra-cranial tissues. Subsequently, watershed algorithm is applied to brain tissues as an initial segmenting method. Normally, result of classical watershed algorithm on gray-scale textured images such as tissue images is over-segmentation. The following procedure is a merging process for the over-segmentation regions using fuzzy clustering algorithm (Fuzzy C-Means). But there are still some regions which are not partitioned completely, particularly in the transitional regions between gray matter and white matter. So we proposed a rule-based re-segmentation processing approach to partition these regions. This integrated scheme yields a robust and precise segmentation. The efficacy of the proposed algorithm is validated using extensive experiments.

In this paper we describe a human organ selection process in which agents argue over whether a given donor’s organ is viable for transplantation. This process is framed in the CARREL System; an agent-based organization designed to improve the overall transplant process. We formalize an argumentation based framework that enables CARREL agents to construct and assess arguments for and against the viability of a donor’s organ for a given potential recipient. We believe that the use of argumentation has the potential to increase the number of human organs that current selection processes make available for transplantation.

The market demand for dental implants is growing at a significant pace. In practice, some dental implants do not succeed. Important questions in this regard concern whether machine learning techniques could be used to predict whether an implant will be successful and which are the best techniques for this problem. This paper presents a comparative study on machine learning techniques for prediction of success of dental implants. The techniques compared here are: (a) constructive RBF neural networks (RBF-DDA), (b) support vector machines (SVM), (c) k nearest neighbors (kNN), and (d) a recently proposed technique, called NNSRM, which is based on kNN and the principle of structural risk minimization. We present a number of simulations using real-world data. The simulations were carried out using 10-fold cross-validation and the results show that the methods achieve comparable performance, yet NNSRM and RBF-DDA produced smaller classifiers.

This work presents an infant cry automatic recognizer development, with the objective of classifying three kinds of infant cries, normal, deaf and asphyxia from recently born babies. We use extraction of acoustic features such as LPC (Linear Predictive Coefficients) and MFCC (Mel Frequency Cepstral Coefficients) for the cry’s sound waves, and a genetic feature selection system combined with a feed forward input delay neural network, trained by adaptive learning rate back-propagation. We show a comparison between Principal Component Analysis and the proposed genetic feature selection system, to reduce the feature vectors. In this paper we describe the whole process; in which we include the acoustic features extraction, the hybrid system design, implementation, training and testing. We also show the results from some experiments, in which we improve the infant cry recognition up to 96.79% using our genetic system. We also show different features extractions that result on vectors that go from 145 up to 928 features, from cry segments of 1 and 3 seconds respectively.

A prospective paradigm, named Growing Functional Modules (GFM) has been introduced in a recent publication. Based on the epigenetic approach, the GFM paradigm is conceived to automatically generate "artificial brains" that are able to build, through interaction, their own representation of their environments. The present application consists in designing an artificial brain for a simple virtual mushroom shaped robot named hOnGo. This paper describes this initial implementation and its preliminary results.

In highly dynamic environments with uncertainty the elaboration of long or rigid plans is useless because the constructed plans are frequently dismissed by the arrival or new unexpected situations; in these cases, a “second-best” plan could rescue the situation. We present a new real-time planning method where we take into consideration the number and quality of future options of the next action to choose, in contrast to most planning methods that just take into account the intrinsic value of the chosen plan or the maximum valued future option. We apply our method to the Robocup simulated soccer competition, which is indeed highly dynamic and involves uncertainty. We propose a specific architecture for implementing this method in the context of a player agent in the Robocup competition, and we present experimental evidence showing the potential of our method.

This paper shows the performance of randomized low-discre-pancy sequences compared with others low-discrepancy sequences. We used two motion planning algorithms to test this performance: the expansive planner proposed in [1], [2] and SBL [3] . Previous research already showed that the use of deterministic sampling outperformed PRM approaches [4], [5], [6]. Experimental results show performance advantages when we use randomized Halton and Sobol sequences over Mersenne-Twister and the linear congruential generators used in random sampling.