ARTIFICIAL INTELLIGENCE

Author: MONTANER RIGALL MIQUEL.
Year: 2003.
University: GIRONA [www.udg.es].
Place of defense: ESCUELA POLITECNÍCA SUPERIOR.
Place of preparation: UNIVERSIDAD DE GIRONA.

Author: LEÓN ROJAS JUAN MIGUEL.
Year: 2003.
University: EXTREMADURA [www.unex.es].
Place of defense: ESCUELA POLITECNICA.
Place of preparation: ESCUELAS POLITECNICA.

Summary: It defines a technical framework for comparing fuzzy sets which are in a space of representation and comparison of the units vaguely profiled. Another is defined framework téorico to work with expressions lingà ¼ isticas allocations probability about events concerning vague units focusing on how these probabilities are updated in line with how it spreads through the vague references in the Bayesian.

Author: GIL VACAS ANTONIO.
Year: 2003.
University: CÓRDOBA [www.uco.es].
Place of defense: E.T.S.I. AGRONOMOS Y MONTES.
Place of preparation: E.T.S.I. AGRONOMOS Y MONTES.

Summary: The demand for water is increasing continuously over the years, while natural resources, namely rainfall, not experiencing this trend. There is a need to assess the demand for irrigation water that an irrigable area can generate. The objective of this thesis is to reproduce, model or predict the behavior of the farmer when acting as regante. The characteristics of this problem makes it difficult to develop a mathematical model, however, the Artificial Intelligence (AI) is being used as a powerful tool in such cases. The proposed model consists of a system neurodifuso based on artificial neural networks (RNA), and fuzzy logic (LD), which has been implemented in the irrigable area of Genil-Cabra. They are considering 11 input variables to the model, of which 7 are numeric variables and come directly to a submodelo based on RNA: surface period and 5 climatic variables, the other 4 variables are called lingüsticas and undergo a transformation into a submodelo based in LD, before entering the RNA. The application of the model for the cultivation of olive grove predicts water demand of the farmer with an error of 14.70%, while if we consider further cultivation of the plot as a variable demand is predicted with an error of 19.38 %.

Author: ALEGRE MARTHA ANALIA.
Year: 2003.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: Facultat de Informatica.
Place of preparation: EDIFICI C6 Campus NORD.

Summary: There is a general agreement that the syntactic structure of a sentence providing important information for semantic interpretation, and this is one of the main reasons why it is investigated in parsing. It is assumed that the hierarchical clustering behind the syntactic structure allows adequately represent grammatical relations between phrases and words. Roles such as syntactic subject or object are derived from these relationships and once these are identified syntactic relations, relations predicado-argumento can be 'read' directly syntactic tree. Reviewing the literature on analyzers extensive coverage, it's easy to realize that the trees, graphs and other similar objects (typical of the methods of learning symbolic) can not be easily represented by vectors traits. For this reason, in order to use methods that operate with vectors (networks neuroanles, suport vector machines or certain statistical methods) to process natural language complex mechanisms are needed to translate the information contained in these classes of objects (trees, graphs) to vectors features. In this thesis we believe that the numerical methods or statistical methods and especially for use with vector operations, such as neural networks or support vector machines, the use of trees to represent information (syntagmatic classical structures) is an obstacle a solution to obtain compositional and efficient systems. To get efficient and compositional systems, we have focused on the problems of invariance (within the meaning of this word in the model object recognition in machine vision) rather than as a solution to power structures represented by vectors syntagmatic. The system we are proposing is composed of two modules. The first module is a semantic parser used by the POS of each word input to build structures proposicionales of prayer. The structures are built proposicionales directly without having to compute the structure sintagmática. The second module receives input and the output of the first module, the propositional structure, and replace the words by their semantic representations (semantic classes) and traits subcategorizacón With all this information as input, Release 2 supervises the correct location of each element in propositional structure. Each new element that adds to the propositional structure should meet all the requirements of subcategorización and related restrictions on selection. The módulo1 requires a relatively very little training. You can try an easy and efficient way coordinations, punctuation and syntactic ambiguities. These phenomena have traditionally been very difficult to treat because the method used classical structures prayers. The type of structure that Release 1 computes can take into account both syntactic and semantic units: including local and long-distance units. The efficiency of the model has been tested with real phrases texts (prayers PTBII) and the results have been very promising. Regarding the Release 2, we tested the effectiveness of various aspects also using prayers PTBII. With regard to ambigà ¼ ages syntactic we tested the model with two especially difficult problems: the ambigà ¼ age in the allocation of phrases preposicionales and ambigà ¼ age in the allocation of relative sentences. The results are comparable to the best achieved so far in these phenomena in the lingüstica computer. With regard to the sense disambiguation, we tested the model in the disambiguation of the names of argument in positions of verbs and the results are very promising.

Author: ROURE ALCOBE JOSE.
Year: 2003.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: Dep. llenguatges i sistemes informàtics.
Place of preparation: EDIFICI C6 Campus NORD.

Summary: The incremental learning approach was firstly motivated as the human capability for incorporating knowledge from new experiences worth being programmed into artificial agents. However, nowadays there exist other practical (i.e. industrial) reasons which increase the interest in incremental algorithms. Nowadays, companies from a very wide range of activities store huge amounts of data every day. One-shot algorithms are not easily able to process and incorporate to a knowledge base this great amount of continuously incoming instances in a reasonable amount of time and memory space. We believe that, in this environment, incremental learning becomes particularly relevant since this sort of algorithms are able to revise already existing models of data without beginning from scratch and without re-processing past data. We present two different and general heuristics in order to convert batch hill-climbing searchers into incremental ones. We believe that the heuristic that we call Traversal Operators in Correct Order (TOCO) is the most novel and original contribution. This heuristic states that, given a learned knowledge structure and the learning path used to obtain the structure where the traversal operators are ordered in decreasing contribution of quality, the structure will be revised only when the order of traversal operators is changed in the light of new data and also that the structure will be rebuild from the first unordered operator of the path. So, the benefit of the TOCO heuristic is twofold. First, the model will only be revised when it is invalidated by new data, and second, in the case that it must be revised, the learning algorithm will not begin from scratch. The second heuristic of our work, that we called Reduced Search Space (RSS) heuristic, uses the knowledge gathered from previous learning steps and states that structures that had very low quality in past learning steps will still have low quality with respect to the new dataset in the current learning step. We formally justify the correctness of these two heuristics. In order to do so, first, we introduce the concept of continuous quality functions. Roughly speaking, we say that a quality function is continuous over the space of datasets when given a knowledge structure and two arbitrarily similar datasets (i.e. there is a short Kullback-Leibler distance between them), the function returns very similar quality values for the structure measured with respect to both datasets. From this definition we will justify the TOCO heuristic noting that if the order of the traversal operators of a given learning path changes when new data instances are added to the dataset it means that the new dataset is significantly different from the former one and thus it is worth revising the structure. Similarly, We justify the RSS heuristic noting that if the order of the traversal operators of a given learning path does not change, it means that both new and old datasets are not significantly different and thus structures that used to be of very low quality will still keep low quality values. Our heuristics need to store the sufficient statistics in order to avoid scanning datasets multiple times. We identify from literature AD-trees as an approach to efficiently store in memory sparse sufficient statistics. This structure will allow us to store in memory the sufficient statistics necessary to potentially perform a search among the entire space of Bayesian networks without going through already seen data. We also propose two additional heuristics that are coupled to the field of Bayesian networks. Both heuristics are concerned in avoiding to store sufficient statistics that are unlikely to be useful for learning future structures.

Author: CERQUIDES BUENO JESUS.
Year: 2003.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: LSI.
Place of preparation: CAMPUS NORD UPC, FIB.

Author: BLANCO GÓMEZ ROSA.
Year: 2004.
University: PAÍS VASCO [www.ehu.es].
Place of defense: FACULTAD DE INFORMÁTICA.
Place of preparation: FACULTAD DE INFORMÁTICA.

Summary: The work of the thesis makes a contribution in two related areas: learning from automatic data networks Baysianas and classifiers Bayesianos. In learning networks Bayesianas, approximations "score + search" looking for the best Bayesian network for a measure and a search space given. In this area, the contributions of the thesis focuses on the use of methods floating, GRASP and algorithms for estimating distributions as engines in the search for networks Bayesianas. While results are not as good as expected, they are competitive with those obtained by other algorithms proposed by the literature. The supervised classification is to build a model from a set of d atos labeled, so that in future such a model predict the class of an instance unlabeled. Related to the supervised classification is the selection of variables, where only selected variables contributing information for the class. In this thesis, ranking elected paradigms are a subset of classifiers Bayesianos; naive Bayes, semi-naive Bayes, tree augmented naive Bayes and k dependence Bayesian classifier. Contributions in this area focus of the proposal an approach for filtering and wrap these classifiers. In experiments conducted shows that the proposed classifiers Bayesianos improve outcome of naive Bayes when data variables include redundant and irrelevant. Finally, we have implemented the filtering and approximations wrapper for classifiers Bayesianos three real problems of biomedical environments, obtaining very promising results.

Author: PÉREZ GONZÁLEZ FEDERICO.
Year: 2004.
University: PAÍS VASCO [www.ehu.es].
Place of defense: E.T.S. DE INGENIERÍA DE BILBAO.
Place of preparation: E.T.S DE INGENIERÍA DE BILBAO.

Summary: This thesis presents a new understanding imaging system based on the use of wavelet transform reversible. The system has been designed to combine both understood without understanding losses as a loss. Presents a new mechanism codificacicón components adapted to the use of sub-levels. Also included is an algorithm alignment of the sub allowing its readiness to adapt to the needs of the application. Additionally, it uses a novel encoder entrópico resulting in a structure of fabric easily manageable. The system of understanding can be applied to both images in gray levels, and color images. For the latter case presents additional mechanisms to continue to maintain the conditions of reversibility. It outlines different mechanisms climbed as extensions of system understanding. It has developed a system climbed as highly configurable. In addition, it also provides a coding system escalation with multiple narrative that allows adapting the information generated little robust communication channels.

Author: GÓMEZ HIDALGO JOSÉ M..
Year: 2004.
University: COMPLUTENSE DE MADRID [www.ucm.es].
Place of defense: FACULTAD DE CIENCIAS MATEMÁTICAS.
Place of preparation: FACULTAD DE CIENCIAS MATEMÁTICAS.

Summary: With the growing amount of information available in electronic format in the modern information society, it is important to provide users effective ways to prevent overloading of información.En many environments (Internet, libraries, businesses and many more) All data continues remain available mostly in the form of texto.Por therefore effectiveness in various tasks classification of text, such as filtering and Recovery Information, and others, is critical to success in education and business, and even in hobbies or travel. The Automatic Text Categorization (CAT) automatic allocation of documents to predefined class plays a key role in this context Access to Información.La CAT is used to help catalogers (or reemplazarles) in the task of properly classify books in libraries, Web pages directories, or to provide a directory structure to the information available on the websites and corporate intranets. In terms Finally, users of the libraries and the Internet, or customers and employees make use of these controlled vocabularies and directory structures to achieve more effective access to information available in these environments. At present, there is a consolidated model for building systems CAT.El model is the use of techniques Retrieval of Information and Learning System to represent manually classified documents (the collection of training). Especially in environments with thematic categories, this model has been proven as effective as the use of human catalogers experts, provided that sufficient documents entrenamiento.Sin however, there are still opportunities to improve their effectiveness due to a number of problems including lack of data The extremely high dimensionality of the representation of text, and ambigà ¼ lexical age. In this thesis, porponemos the integration of information available in Databases Léxicas (BDLS) to improve efectiviadad of systems CAT.Las BDLS contain a lot of information about the lexical elements of one or more languages and can be Used to complement the information available in the collection of training to improve the effectiveness of categorization Texto.Hemos developed two models of integration, based on the ways in which the BDLS used in the recovery Información.El first model is based on query expansion, while the second is based on indexing conceptual.Hemos conducted a series of experiments, the results show that both models have éxito.También we studied the integration of the two models in a more general, which draws from the best of both.

Author: MACHÓN GONZÁLEZ IVÁN.
Year: 2004.
University: OVIEDO [www.uniovi.es].
Place of defense: ESCUELA POLITÉCNICA SUPERIOR DE INGENIERÍA DE GIJÓN.
Place of preparation: ESCUELA POLITÉCNICA SUPERIOR DE INGENIERÍA DE GIJÓN.

Summary: The use of neural networks models generated narrative that lets you view and understand the data, pattern recognition and identify relationships and correlations. You can also create predictive models of key parameters from the data available. These techniques can be integrated into existing software platforms and hardware to work in real time with the process to study. The past two decades have produced successful results in the treatment of wastewater in large part because of stringent policies. Costs related to the aerobic phase are usually the most important biological treatments. In this regard, it is advisable to make changes in the operation and / or configuration of the plant in order to reduce the cost of aeration. However, this would imply the consumption of time and money so the best solution is to use simulation models to evaluate the potential effects of alternative operational strategies. This thesis presents a methodology for working with neural networks both feed-forward (MLP) with auto-organized maps (SOM). Within this point, it is very important validation of the neural network that depend on the type of network used. It takes place through neural networks feed-forward estimating the concentration of ammonia in the effluent from a treatment plant consisting of two biological reactors in series. Moreover, it has developed an independent software tools including data acquisition system of a plant wastewater treatment of coking plant Arcelor in Avilés (Spain) and the algorithms for auto-organized maps as a technique for monitoring process. This algorithm is integrated mapping auto-organized as a technique for monitoring the process. This algorithm will be integrated into the application to automatically train a network SOM. The operator of the plant can visualize the last SOM network that corresponds to the last treatment cycle aerobic batch reactor (SBR) and draws correlations between process variables. It makes the classification of data using clustering algorithms particionales and is calculated estimation of the current state of the process through networks SOM that have been trained and validated according to the topographic and quantization error. You can achieve a saving of operating costs and an increase in plant throughput by estimating the length of the main activation aerobic treatment. The most relevant result is the detection of endpoint of the aerobic phase through which we can finalize it and thus reducing the time of treatment cycle by increasing the operational capacity of the plant and cost savings.

Author: SÁNCHEZ SOLANA ANTONIO MIGUEL.
Year: 2004.
University: MÁLAGA [www.uma.es].
Place of defense: ESCUELA TÉCNICA SUPERIOR DE INGENIERÍA INFORMÁTICA.
Place of preparation: ESCUELA TÉCNICA SUPERIOR DE INGENIERÍA INFORMATICA.

Summary: The loss of oil in the alperujo is around 3% to 4% of oil on wet sample in normal conditions. The importance of improvements in the performance of this industrial process stems from the strong impact of this product in the economy Spanish and especially andaluza.En quality control in industrial tasks by computer vision, has been applied with considerable success in other campos.El quality control in industrial tasks by computer vision, has been applied with considerable success in other campos.La basis for implementing it in alperujos is that 1) the expert (master miller) makes use of tactile and visual information to perform certain tasks control smooth operation of the plant; 2) the experts involved in the chemical analysis of bagasse , familiar with the experience (visually), if marc are refresher (exhausted), or if on the other hand, contain a significant percentage of oil and the percentages of moisture is important; 3) the visual feature related information Touch is the textura.Métodos alternative as Nuclear Magnetic Resonance (NMR) and Soxhlett require pre-drying the sample, which implies at least 6-7 hours waiting to obtain the analytical result (in the case of method Soxheltt is compounded in this time of the extraction of the fat by disolvente.El method Near Infrared (NIR) is faster response time (several minutes) but it requires the manipulation of an operator (specialized) and deuna buens calibration the dispositivos.Con the proposed system aims to improve performance in the production of oil in the mills, through the application of analysis techniques imágenes.Para This has designed a system that feeds pomace, maner continuous and uniform camera for taking imágenes.Las images are sent to a computer (located in the control panels of the system) where they perform a detecicón bubbles or lumps, some features are extracted, maner which then can be done in a classification Based on a neural network for clustering.

Author: LAZKANO ORTEGA ELENA.
Year: 2004.
University: PAÍS VASCO [www.ehu.es].
Place of defense: FACULTAD DE INFORMATICA.
Place of preparation: FACULTAD DE INFORMATICA.

Summary: The robot navigation in environments semi-estructurados is a problem with multiple applications open interest for applications of various kinds as can be: robot guide in museums and public institutions, robots for cleaning large areas, support systems for people with motor disabilities , and so on. The systems approach Based on Performance proposes an incremental design methodology to address the problem of intelligent autonomous systems. This research work can be classified into two areas: 1 .- Automatic Learning: Different skills interesting for navigation robots have been designed and implemented using paradigms of this area. 2 .- Robot Navigation: under the paradigm of systems based on behavior, and following taxonomy of navigation systems biomiméticos, presents the incremental development of a control architecture for robot navigation. We identify the basic modules or behavior to wander, and integrates a set of identifiers enough marks to get a response activated by perception. Moreover, it completes a procedural description of the environment that keeps the location and plan for achieving the goals.

Author: FERNANDEZ GOMEZ DE SEGURA ELSA.
Year: 2004.
University: PAÍS VASCO [www.ehu.es].
Place of defense: FACULTAD DE INFORMATICA.
Place of preparation: FACULTAD DE INFORMATICA.

Summary: The problem of distortions introduced by the light source is present in all applications that use digital images for the detection and recognition in its broadest sense. In this paper we have focused on the images of Nuclear Magnetic Resonance, given its growing diagnostic use. In addition, there, this problem may be formulated in a more affordable solution for the general case, because the images show a clear structure functions as a constant pieces. We have implemented and compared various approaches, which include our propositions based on the minimization of an objective function defined on the classification error and bias lighting. In addition to the magnetic resonance imaging, we have adapted our algorithm for correcting lighting applications for the recognition of faces, with positive results.

Author: Sánchez Maroño Noelia.
Year: 2004.
University: A CORUÑA [www.udc.es].
Place of defense: Facultad de Informática.
Place of preparation: Facultad de Informática.

Summary: This thesis proposes two new algorithms to reduce the size. The first is an algorithm based on dimensional analysis that can reduce the size of entry of a physical problem or engineering. Reducing the size is achieved through the implementation of the theorem Pi, the fundamental theorem of dimensional analysis on the set of input and output variables that make up the problem, so the variables are transformed into monomials adimensionales therefore The proposed algorithm is a method for extracting features. The implementation of the algorithm provides the different sets of monomials adimensionales arising out of the variables of a problem. Each of these sets form the entrances and exits of a functional or neural network, this will get different approaches to solve the same problem. The adequacy of this algorithm is illustrated by its application to three physicists and engineering problems. The performance results obtained are compared with those derived from the direct application, ie, without using the proposed algorithm, functional and neural networks. Moreover, the method of extraction of features compared to other existing methods: PCA and ICA. The second algorithm presented is a alfotirmo learning network based on functional decomposition ANOVA (Analysis Of Variance), which helps determine the topology of the same from the data. So far, and, according to the peer-reviewed literature, is the only existing algorithm allowing the topology of functional network from the data. As this algorithm is based on the decomposition ANOVA also provides a sensitivity analysis, both global and local levels. In this way, not only will you get an approximation to solve a given problem, but you get a model that includes a well-defined structural learning, and a parametric sensitivity analysis. The overall sensitivity indexes are the ones who determine the topology of the network functional, as pointing out the importance of each variable, and every combination of variables in problem solving. The suitability of the algorithm is demonstrated on two problems in implementing engineering ral: resistance from a beam and design of a dam. In some cases, rates of overall sensitivity suggest eliminating certain variables, given her little or no influence in achieving the objective function. It is therefore proposed an extension of the algorithm that allows its application as a method evolvente selection of features. There has been a comparative study with other methods of selection of features, both envelopes as filter, demonstrating that the proposed method achieves similar results, sometimes even better, both in the accuracy achieved in reducing the number of variables.

Author: Penas Centeno Marta.
Year: 2004.
University: A CORUÑA [www.udc.es].
Place of defense: Facultad de Informática.
Place of preparation: Facultad de Informática.

Summary: The Doctoral thesis describes and implements a system for detecting objects in images based on techniques perceptual organization. Traditionally, all approaches to this objective have been based on the results produced by the detectors edges as Canny or Sobel. By contrast, the scheme will be based on results obtained through the decomposition of Gabor. The introduction of decomposition Gabor, a process computationally complex. The final results will improve due to two main reasons: the first is that the decomposition of Gabor provides comprehensive and reliable information about the orientation of the edges on the world stage and the second is that it does not require adjustment of parameters from the user based the input image. The scheme is divided into several stages. The first is the removal of the primitive directional image through a filter bank Gabor distributed in a range of frequency channels (frequencies *) and guidance (* Guidance). Due to the decomposition of Gabor transforms the input image in (* x * frequencies guidelines) images outcome, it is necessary to integrate information from these images in a single result. It will use a set of structures auto-organizativas based networks of artificial neurons, organized hierarchically to result in a single image pseudo-color in which a colored pixel define the orientation of the corresponding edge in the input image. The third stage involves the detection of segments in the picture pseudo-color result of the earlier prosecutions. It uses a detector segments based on the combination of two segments detectors widely used, and the Hough transform detector segments Burns. The last step is the organization of the segments identified through a process of perceptual organization based on the detection of relationships between pairs of segments (colinearity, parallelism, unions, etc.) and generating a set of graphs that reflect the structure of the scene. Finally, the validity of the system has been demonstrated through its application to two distinct areas: the detection of buildings in photographs, and the detection of proteins in medical imaging.

Author: PICÓ VILA DAVID.
Year: 2004.
University: POLITÉCNICA DE VALENCIA [www.upv.es].
Place of defense: Dep. Sistemas Informáticos y ComputaciÓn.
Place of preparation: Universidad Politécnica de Valencia.

Summary: Machine translation is a very active field of natural language technology that has experienced strong development in recent decades. The techniques that have been put in place could be classified, somewhat simplistically, technical "knowledge-based", using information provided by experts lingà ¼ istas, and techniques "based on examples," that automatically extracted information from body text. The work of this thesis is part of the second family of techniques, based on examples. We have focused on the use of a particular model: the transducers stochastic finite state. This model (and similar) is used in many areas of language technologies and different purposes. We use it here for translation between different languages. The thesis deals with the problem of creating transducers stochastic finite state from samples of translation. These samples consist of pairs of phrases in different languages in which a translation of the phrase is another. On the one hand, we bring several algorithms for estimating the probability of transducers, on the assumption that the structure is known already. On the other hand, and as a major contribution at work, outlines a method and a set of algorithms to infer transducers (with structure and probabilities) from sample to group them under the symbols generic GIATI (English Grammatical inference algorithms . This report presents a theoretical first exhibition of these algorithms, and continues with a case study of implementation which explores different ways of using GIATI. were also present experimental studies translation show the real possibilities of implementing these techniques.

Author: Ontañón Villar Santi.
Year: 2004.
University: AUTÓNOMA DE BARCELONA [www.uab.es].
Place of defense: Institut d'Investigació Intelligència Artificial.
Place of preparation: Institut d'Investigació Intelligència Artificial.

Summary: This paper presents a framework for the aprendizajeen an arena of distributed data and control descentralizado.Hemos based our working framework Systems Multi-Agente (MAS) in order to gain control decentralized, and Reasoning Based enCasos (CBR), since his lazy nature of learning what hacenadecuado systems multi-agentes dynamic. In addition, we are interested in autonomous agents operating as (emensembles). An ensemble of actors solves problems lasiguiente way: each actor individually solve the problem actualindividualmente and makes his prediction, then all esaspredicciones are added to form a prediction global.Así therefore, in this paper we are interested in desarrollarestrategias learning based on cases and ensembles parasistemas multi-agente.Concretamente put forward a working framework called Reasoning Based on Cases Multi-Agente (MAC), a aproximaciónal CBR-based agents. Each agent in a single system / mac / is capable of learning and soluciar problems individually using CBR with the basis of their individual cases. In addition, each case basis is owned by an individual, and any information that base cases will be displayed or shared only if people decided so. Therefore, this preserves the framework of data privacy and autonomy of the actors to reveal information. This thesis focuses on developing strategies for individual agents with the ability to learn can enhance their performance when working both individually and when trabjan as an ensemble. Moreover, decisions in a system MACs making in a decentralized manner, with each agent has decision-making autonomy. Therefore, the techniques developed in this work provide a framework for increased performance as a result of individual decisions taken in a decentralized manner. Specifically, we will present three strategies: strategies for creating ensembles of actors, strategies to make retention of cases in systems multi-agentes, and strategies to make redistribution of cases.

Author: GIL GARCIA REYNALDO JOSE.
Year: 2004.
University: JAUME I DE CASTELLON [www.uji.es].
Place of defense: E.S. TECNOLOGIA Y CIENCIAS EXPERIMENTALES.
Place of preparation: E.S. TECNOLOGIA Y CIENCIAS EXPERIMENTALES.

Summary: This thesis focuses on the problem of clustering. At the same are presented and evaluated various clustering algorithms based on graphs, both sequential and parallel and proposes solutions to the three classification problems that may arise in practice; obtaining shares or disjoint groups, obtaining coverage or groups overlapping and building hierarchies of groups. We propose three sequential clustering algorithms and four parallel. It is presented to others a general framework capable of generating various algorithms hierarchical aglomerativos, both static and dynamic. All algorithms proposed in the thesis can be used as a routine coverage within this framework. The various sequential and parallel algorithms developed are applied to solving a specific problem, the grouping of documents. The experiments conducted with several collections of documents shows that our algorithms obtained groups with a quality comparable to the best algorithms proposed in the literature. This is accomplished with additional advantages as not to restrict the space of representation of objects and the role of similarity between them, have only one parameter, be independent of the order, among others. On the other hand, parallel algorithms achieve good acceleration and scalability isotemporal. Despite the fact that we use it in the group of documents is not restricted to this area, which can be used in any problem of Pattern Recognition where needed grouping objects of any kind.

Author: BESADA JUEZ JESUS MANUEL.
Year: 2004.
University: PONTIFICIA COMILLAS [www.upcomillas.es].
Place of defense: ALBERTO AGUILERA 25.
Place of preparation: ESCUELA TECNICA SUPERIOR DE INGENIERIA.

Summary: This thesis presents a comprehensive methodology and closed loop for the automatic refinement of knowledge on the basis of a story or experience available examples belonging to a particular domain of knowledge. The thesis falls within the field of automatic refinement of knowledge, presenting a procedure that can refine or learn new knowledge, improve and learn from the experience automatically as they are available examples and all of this with a high accuracy and interpretability of knowledge gained. The methodology is presented consists of four main algorithms or steps to achieve its goal. The main features and innovations brought about by this argument are as follows: 1. Extraction of knowledge: The objective of this stage is to extract knowledge internally coded in a 'Perceptrón Multicapa' already trained and validated, and put it in a comprehensible form, with a pledge that such knowledge is as faithful as possible to the original source. This algorithm is able to work with any type of 'Perceptrón Multicapa' without imposing restrictions on their topology. It is also capable of working with both problems of classification and functional approach, and does not suffer from problems of combinatorial explosion. 2. Automatic Training: The purpose of this phase is to obtain the best possible model from examples, ie propose the most appropriate topology 'Perceptrón Multicapa' that suits some examples given automatically. This algorithm is capable of working with both real and discrete attributes, being able to increase or decrease the complexity of the model, either increasing or decreasing the number of neurons and / or the number of entries in the model, although the fixed number hidden Layers' Perceptrón Multicapa '3. Comparison of Knowledge: With the knowledge gained from previous phases, this algorithm performs compared with the prior knowledge that this was the problem if it exists, to determine the need for updating that prior knowledge. This algorithm does not depend on the semantics used on previous knowledge, and makes the comparison using much evidence as certainties. 4. Conversion of knowledge: The objective of this phase is to obtain the equivalent of prior knowledge that can be a problem (expressed in any semantics), a modular structure of the type networks' Perceptrón Multicapa ', so that these networks serve as a good starting point for conducting training with the examples available. The ability to use any type of semantics, and the conversion to a modular structure of networks, are its main features. The thesis has used public data to test and measure the benefits of different algorithms, conducted comparisons with the last similar methodologies found in the literature. In addition, it presents a real case where full implementation methodology, which illustrates well the entire cycle proposed.

Author: SAPENA VERCHER OSCAR.
Year: 2004.
University: POLITÉCNICA DE VALENCIA [www.upv.es].
Place of defense: Dep. Sistemas Informaticos y Computacion.
Place of preparation: Universidad Politécnica de Valencia.

Summary: Research in planning independent domain has long focused on the development of efficient search techniques, usually aimed at finding an optimal sequencing (or near-optimal) actions leading to the system from the current state to state goal . The problem of planning separate domain, however, is a very complex problem. Therefore, although the latest planners, especially those based on the application heuristics, they are very fast, still needed several minutes to solve many problems of size medium / large. There are, however, many real-world applications, such as control of mobile robots or artificial intelligence agents in computer games and simulations, in which an excessive computation time is not acceptable. A response time acotable is not the only requirement for this type of applications. It is not uncommon, for example, that the planner could not access all the information environment. Something even more likely that the world constantly changes, as the planner is not only agent capable of acting on the world and change it. These are, among others, the qualities that make planners classics are not the best choice for this kind of problem. The new planning techniques capable of coping with the difficulties of this kind of domains (or at least some of them) are part of the planning practice. In this paper, thesis, which is part of this practical aspect of the planning, proposing an approach based on the integration of planning and implementation processes. Thus, the scheduler can incorporate into their plans of the information acquired during the execution environment. The algorithm proposed planning is based on several planning techniques well known classic, as is the calculation of heuristics and decomposition of goals, but combined in a novel way. This algorithm provides a number of useful features for planning in many real environments: first offers a behavior very similar to the algorithms anytime, providing an initial response within a limited period of time, and refinándola while there is time available. This behavior allows you to react quickly to unexpected events and changes in the goals during the execution. Permits also solve problems with incomplete information using shares sensorización, and supports the use of numeric variables and the definition of functions optimization.