Catálogo de publicaciones - libros

Multiple Classifier systems have been developed in order to improve classification accuracy using methodologies for effective classifier combination. Classical approaches use heuristics, statistical tests, or a meta-learning level in order to find out the optimal combination function. We study this problem from a Reinforcement Learning perspective. In our modeling, an agent tries to learn the best policy for selecting classifiers by exploring a state space and considering a future cumulative reward from the environment. We evaluate our approach by comparing with state-of-the-art combination methods and obtain very promising results.

The role and importance of methods for lexical knowledge elicitation in the area of multilingual information processing, including machine translation, computer-aided translation and cross-lingual information retrieval is undisputable. The usefulness of such methods becomes even more apparent in cases of language pairs where no appropriate digital language resources exist. This paper presents encouraging experimental results in automatically eliciting bilingual lexica out of Greek-Turkish parallel corpora, consisting of international organizations’ documents available in English, Greek and Turkish, in an attempt to aid multilingual document processing involving these languages.

In this paper feed-forward neural networks are introduced where hidden units employ orthogonal Hermite polynomials for their activation functions. The proposed neural networks have some interesting properties: (i) the basis functions are invariant under the Fourier transform, subject only to a change of scale, and (ii) the basis functions are the eigenstates of the quantum harmonic oscillator, and stem from the solution of Schrödinger’s diffusion equation. The proposed neural networks demonstrate the particle-wave nature of information and can be used in nonparametric estimation. Possible applications of neural networks with Hermite basis functions include system modelling and image processing.

Coalition formation is an important area of research in multi-agent systems. Computing optimal coalition structures for a large number of agents is an important problem in coalition formation but has received little attention in the literature. Previous studies assume that each coalition value is known a priori. This assumption is impractical in real world settings. Furthermore, the problem of finding coalition values become intractable for even a relatively small number of agents. This work proposes a distributed branch-and-bound algorithm for computing optimal coalition structures in linear production domain, where each coalition value is not known a priori. The common goal of the agents is to maximize the system’s profit. In our algorithm, agents perform two tasks: ) deliberate profitable coalitions, and ) cooperatively compute optimal coalition structures. We show that our algorithm outperforms exhaustive search in generating optimal coalition structure in terms of elapses time and number of coalition structures generated.

The innovative feature of the system presented in this paper is the use of pattern-matching techniques to retrieve translations resulting in a flexible, language-independent approach, which employs a limited amount of explicit a priori linguistic knowledge. Furthermore, while all state-of-the-art corpus-based approaches to Machine Translation (MT) rely on bitexts, this system relies on extensive target language monolingual corpora. The translation process distinguishes three phases: 1) pre-processing with ‘light’ rule and statisticsbased NLP techniques 2) search & retrieval, 3) synthesising. At Phase 1, the source language sentence is mapped onto a lemma-to-lemma translated string. This string then forms the input to the search algorithm, which retrieves similar sentences from the corpus (Phase 2). This retrieval process is performed iteratively at increasing levels of detail, until the best match is detected. The best retrieved sentence is sent to the synthesising algorithm (Phase 3), which handles phenomena such as agreement.

Bayesian Networks are proven to be a comprehensive model to describe causal relationships among domain attributes with probabilistic measure of appropriate conditional dependency. However, depending on task and context, many attributes of the model might not be relevant. If a network has been learned across multiple contexts then all uncovered conditional dependencies are averaged over all contexts and cannot guarantee high predictive accuracy when applied to a concrete case. We are considering a context as a set of contextual attributes, which are not directly effect probability distribution of the target attributes, but they effect on a “relevance” of the predictive attributes towards target attributes. In this paper we use the Bayesian Metanetwork vision to model such context-sensitive feature relevance. Such model assumes that the relevance of predictive attributes in a Bayesian network might be a random attribute itself and it provides a tool to reason based not only on probabilities of predictive attributes but also on their relevancies. According to this model, the evidence observed about contextual attributes is used to extract a relevant substructure from a Bayesian network model and then the predictive attributes evidence is used to reason about probability distribution of the target attribute in the extracted sub-network. We provide the basic architecture for such Bayesian Metanetwork, basic reasoning formalism and some examples.

The prediction of the translation initiation site (TIS) in a genomic sequence is an important issue in biological research. Several methods have been proposed to deal with it. However, it is still an open problem. In this paper we follow an approach consisting of a number of steps in order to increase TIS prediction accuracy. First, all the sequences are scanned and the candidate TISs are detected. These sites are grouped according to the length of the sequence upstream and downstream them and a number of features is generated for each one. The features are evaluated among the instances of every group and a number of the top ranked ones are selected for building a classifier. A new instance is assigned to a group and is classified by the corresponding classifier. We experiment with various feature sets and classification algorithms, compare with alternative methods and draw important conclusions.

As is widely known, the popular Black & Scholes model for option pricing suffers from systematic biases, as it relies on several highly questionable assumptions. In this paper we study the ability of neural networks (MLPs) in pricing call options on the S&P 500 index; in particular we investigate the effect of the hidden neurons in the in- and out-of-sample pricing. We modify the Black & Scholes model given the price of an option based on the no-arbitrage value of a forward contract, written on the same underlying asset, and we derive a modified formula that can be used for our purpose. Instead of using the standard backpropagation training algorithm we replace it with the Levenberg-Marquardt approach. By modifying the objective function of the neural network, we focus the learning process on more interesting areas of the implied volatility surface. The results from this transformation are encouraging.

The Relevance Vector Machine(RVM) is a widely accepted Bayesian model commonly used for regression and classification tasks. In this paper we propose a multikernel version of the RVM and present an alternative inference algorithm based on Fourier domain computation to solve this model for large scale problems, e.g. images. We then apply the proposed method to the object detection problem with promising results.

In this paper we present an artificial cortical network, inspired by the Human Visual System (HVS), which extracts the salient contours in color images. Similarly to the primary visual cortex, the network consists of orientation hypercolumns. Lateral connections between the hypercolumns are modeled by a new connection pattern based on co-exponentiality. The initial color edges of the image are extracted in a way inspired by the double-opponent cells of the HVS. These edges are inputs to the network, which outputs the salient contours based on the local interactions between the hypercolumns. The proposed network was tested on real color images and displayed promising performance, with execution times small enough even for a conventional personal computer.