ARTIFICIAL INTELLIGENCE (4)

Author: FRESNO FERNÁNDEZ VÍCTOR DIEGO.
Year: 2005.
University: REY JUAN CARLOS [www.urjc.es].
Place of defense: ESCUELA SUPERIOR DE CIENCIAS EXPERIMENTALES Y TECNOLOGÍA.
Place of preparation: E.T.S. DE INGENIEROS DE TELECOMUNICACIONES.

Summary: In this PhD thesis represents a proposal for self-representation of Web pages based on combinations of criteria heuristics. It proposes two methods of weighting traits as part of the general definition of a standard representation of documents. With these functions is intended to determine the weight it has a feature on the contents of an HTML document, for it provides a theoretical framework generally supported on a fundamental assumption; reading involves an active process where both the author of a document, such as Reader, contributing their expertise and previous knowledge to informative documentary processes. It is with the primary aim of developing representations solely based on the textual content of the HTML documents. The scope will be the automatic qualification and clustering of web pages. These processes can be used in the creation of Web directories thematic or applied on the results returned after a query to a search engine on an analysis of the structure of hipergrafo which is the Web itself, as well as a study of the content of the text 's own website. The functions proposals seeking to improve the representations based on content found in the literature, and may be used as self-representations or part of representations of mixed type. One of the functions proposed in this thesis, called ACCC (Analytical Combination of Criteria), is based on a linear combination of heuristic criteria derived from the processes of reading and writing texts. The other FCC (Combination of Fuzzy Criteria), is constructed from a combination blurred or fuzzy, the same criteria. One of the advantages offered by PCA and FCC is allowing represent an HTML document without needing to analyze any reference collection. It will not be necessary to extract information on the frequencies of occurrence of the different strands within the collection. This property is interesting in the context of the Web, saw its current size and its growth rate. Moreover, in a context where the heterogeneity of material is a major feature, representations proposals allow the generation of representations independent of the type of page you are considering, as long as they are richer content, so as not to be drawn different heuristics to represent different types of documents.

Author: DOTU RODRÍGUEZ IVÁN JAVIER.
Year: 2005.
University: AUTÓNOMA DE MADRID [www.uam.es].
Place of defense: UNIVERSIDAD AUTÓNOMA DE MADRID. ESUCELA SUPERIOR POLITÉCNICA.
Place of preparation: UNIVERSIDAD AUTÓNOMA DE MADRID.

Summary: Optimization Combinatorics is a branch of applied mathematics and optimization in computer technology related operational research, theory of algorithms and computational complexity theory, which lies at the intersection of several fields such as artificial intelligence, mathematics and software engineering. The combinatorial optimization problems usually involve finding values for a set of variables which are restricted by a set of restrictions, in some cases to optimize a given function (optimization) and others just to find a valid solution (satisfaction). The algorithms for combinatorial optimization problems solved instances of hardship in general thanks to a clever exploration of the search space, partly reducing it in an efficient manner. This thesis is based on the combinatorial optimization algorithms that are considered entered the field of Artificial Intelligence (although it is true that the line which separates the field of operations research is very thin), rather than on operational research algorithm . Thus, the programming methods as a whole or "Branch-and-Bound" not going to be treated. The objective of this thesis is to show that different techniques may be more appropriate for different problems, and hybrid techniques, which include mechanisms for different paradigms can benefit from the advantages and try to minimize the inconvenience for them. All of this is shown in this thesis with the resolution of difficult problems of combinatorial optimization as complenitud of cuasigrupos, golfers social Golomb rulers, using various techniques, which lead to the development of a hybrid algorithm to find Golomb rulers, which incorporates aspects of algorithms genetic local search, rogramación with restrictions and even clustering.

Author: LLOBET AZPITARTE RAFAEL.
Year: 2006.
University: POLITÉCNICA DE VALENCIA [www.upv.es].
Place of defense: Dep. Sistemas Informaticos y Computacion.
Place of preparation: Universidad Politécnica de Valencia.

Summary: To diagnose cancer is performed, among other tests, a test image, such as an x-ray, ultrasound or MRI. Through these tests can spot areas with high suspicion tumor, which must confirm diagnosis finally by conducting a biopsy. Such images, however, are not easy to interpret, leading to the professional designated to analyze, despite his experience, is unable to detect them in a significant percentage of tumors (false negatives). One possibility to improve the diagnosis and reduce the number of false negatives is to use systems of computer-aided diagnosis or computer-aided diagnosis (CAD). A CAD system analyzes medical image and tries to detect suspicious areas to contain any anomaly. These areas are marked on the image with a dual purpose: to draw the attention of professional responsible for analyzing it to the suspect area and provide a second opinion on the diagnosis. This thesis presents and evaluate various techniques of computer vision and recognition of shapes aimed at detecting tumors in medical imaging, with the aim of CAD design systems that would allow for a better diagnosis. The work has focused on the diagnosis of prostate cancer from ultrasound images, and the diagnosis of breast cancer from the X-ray images. We have evaluated various methods for extracting features based on the intensity, frequency, textures or gradients. At the screening stage has been used a classifier based on non - parametric distances (k-vecinos closest) and other parametric models based on Markov. Over the work is evidenced by the different problems that arise in this type of work and proposed solutions to each of them.

Author: MARTÍN YAGÒE LUIS.
Year: 2006.
University: DEUSTO [www.deusto.es].
Place of defense: FACULTAD DE INGENIERÍA.
Place of preparation: FACULTAD DE INGENIERÍA.

Summary: This dissertation deals with the study of the shortcomings that many students face higher technical studies on matters related to mathematics. The main objective is the design and development of an expert system, EVAM (Automated Assessment of Mathematics), which enables automatic detection of those factors might be responsible for the shortcomings in the performance of students and to provide a tool for teachers to correct them. The system collects data on the progress of students in the teaching-learning process of a matter and can provide regular information on the uptake that they have every part of the program to assess specific aspects of the subject matter to guide their teaching process. Therefore, it is a diagnostic tool prior to the actions that allows teachers require information on specific aspects of domain knowledge (mathematics degrees above techniques) and mathematical skills that students possess, store the information, deciding on the process of formal notice of the student information and represent and analyze the results. The knowledge base consists of a set of questions posed to the student to try to determine the status of their knowledge. For its design has devised a three-dimensional structure that allows classify those questions based on the themes and math skills that it is necessary to know its resolution as well as a level of difficulty set from the other two variables. The teacher responsible for the course evaluated, as a user, the system indicates some initial requirements. From them and the answers given by students, the inference engine of EVAM choose the questions that best suit each moment, and creates a model of the student's knowledge that is maintained and updated dynamically. Finally, the system presents the user with the results for analysis and taking the appropriate actions teachers.

Author: MUSEROS CABEDO MARIA LIDON.
Year: 2006.
University: JAUME I DE CASTELLON [www.uji.es].
Place of defense: E.S.DE TECNO.Y CIE. EXPERIMENTALES.
Place of preparation: E.S.DE TECNO.Y CIE. EXPERIMENTALES.

Summary: This thesis is centred on the development and application of a model to reason about shape and about movement in a qualitative way (in a way similar to the human reasoning). The interest of this study originates from the necessity to find solutions for the recognition of objects and the description and reasoning about the movement in situations with high uncertainty, as it is the case of robotic applications, where robots only have limited and vague sensorial information. In these situations the use of a qualitative reasoning, that allows us to handle ambiguities and errors, will be the most suitable. This PhD thesis presents a motion model as a qualitative representational model for integrating qualitatively time and topological information for reasoning about dynamic worlds in which spatial relations between regions and between regions and objects may change. On the other hand, the thesis develops a theory for the recognition of shapes able to describe several types of shapes: regular and non-regular polygons, with or without holes, with or without curved segments and completely curvilinear forms. Each shape is described by a string containing its qualitative distinguished features, which is used to match an object against the others. This theory has been applied, in an industrial domain, for the automatic and intelligent assembly of ceramic mosaics. Moreover, the part of the theory that describes polygonal objects, jointly with the theory of movement has been applied for the simulated navigation of a real robot.

Author: Fernández Leal Angel.
Year: 2006.
University: A CORUÑA [www.udc.es].
Place of defense: Facultad de Informática.
Place of preparation: Facultad de Informática.

Summary: This thesis presents the design of a new computable model for the representation of information and treatment temporary efficient coincidence, called CTCN (Casual Temporary Constraint Networks), and the design, implementation and validation of a framework for the processing of information storm in the domain of application problems clinical diagnosis, which employ biometric records as background information. The scope of this work is the diagnosis of the syndrome in sleep apneas (SAS) is a breathing disorder characterized by the occurrence of 5 or more respiratory pauses (apnea or hipopneas) per hour of sleep. Thus, the work has been integrated into a system of diagnosis of SAS to test their practical application. The model for the representation of information temporary proposed: (a) allows the establishment of temporary restrictions between qualitative and quantitative temporal entities (which may be points or intervals), (b) introduces the representation of causal restrictions that are commonly found in any domain knowledge, and (c) allowing the representation of imprecise knowledge. The representation of causal restrictions is conducted in a manner that allows both the formalization of causality objective (commonly accepted as a public or semi-public knowledge) as of causality subjective (heuristic knowledge or private) Treatment temporary information is structured it , depending on their characteristics and the granularity of time, in different contexts interpretation linked with each other through a mechanism of inference. This mechanism makes vertical abstraction of such information, to finally obtain information with a level of abstraction that allows develop a diagnosis. The inference mechanism is based on the identification of patterns of diagnosis at two levels. The first level consists some time intervals, called reference intervals, which provide temporary minimum conditions to be met for a series of events that can perform some abstraction from them. The second level is the different subsets of events we call intervals inferencia- within intervals of reference from which to infer the occurrence of events relevant to the diagnosis. To make that inference takes into account the timing relationships between events, the degree of inaccuracy of the same information and timeless and static pattern affecting the diagnostic question. The process of information processing temporary realimenta of new events generated by analyzing new patterns of diagnosis, which can be in different contexts. The process ends when there are new patterns to be analyzed

Author: Araúzo Azofra Antonio.
Year: 2006.
University: GRANADA [www.ugr.es].
Place of defense: E. t. s. de Ingen. Infor. y Teleco..
Place of preparation: Escuela Técnica Superior de Ingeniería Informática.

Summary: Performing a review on the problem of selection of features and its state of art, proposes a modular decomposition can implement, extend and combine in an orderly methods for selecting features. We develop a number of improvements for selected characteristics: FOCUS extensions algorithm to expand its scope to continuous values, diffuse and variable lingà ¼ isticas, and new measures for assessing subsets of features for use with various search methods. It conducts an extensive empirical study of a wide variety of methods for selecting features, drawing conclusions on how the methods índican which are the most recommendable to different problems. With the results of the empirical study and the expertise acquired, designing a system for intelligent recommend the method of selection of features most appropriate in each case.

Author: Soler Ruiz Vicente.
Year: 2006.
University: AUTÓNOMA DE BARCELONA [www.uab.es].
Place of defense: Escola Tècnica Superior d'Enginyeria.
Place of preparation: Escola Tècnica Superior d'Enginyeria.

Summary: The problem to solve in this Doctoral Thesis is to find a solution that improves the classification is now to the problem of screening for Down's syndrome in fetuses during the second trimester of pregnancy with non-invasive techniques. The data set used for the detection of Down's syndrome is two classes and type imbalanceado, namely that there is a big difference between the number of cases for fetuses are not affected by Down's syndrome and that it they are. To try to improve the classification is done at present, has developed a new method of Soft Computing based on Fuzzy Logic designed to work with data sets imbalanceados. This method allows us not only to find a good solution, but also extract the knowledge acquired. The method developed is called FLAGID (Fuzzy Logic And Genetic algorithms for Imbalanced Datasets) and is based on the idea that the solution widely as possible, to avoid the effect of Overlearning ( 'overfitting') that occurs in most methods in trying to work with a data set imbalanceado. To provide the tools necessary to generalize the method, it has developed an algorithm called ReRecBF, part of the method FLAGID. This algorithm transforms functions membership drawn from the data by other existing algorithm called DDA / RecBF. This transformation is to convert the functions belonging generated from the cases of clase-menor office triangular, trapezoidal stop functions as belonging to the functions of the clase-mayor and divide the functions belonging to overlap. Finally, because they are generating new features of belonging, a genetic algorithm is used simply to find the rules that most conform to the new features. The results have improved the rate of false positives in the data set for Down's syndrome to 4%, with a true positive rate of 60%. This is the first time that a method be able to lower the 5% false positive rate that hits on the real positive. Moreover, it has drawn knowledge of the outcome, and this has coincided, mostly on the existing knowledge in the field of medicine. Another remarkable fact is that it has been shown that the method is also useful for working with data sets imbalanceados. Finally, the results of this work do new contributions in the field of medicine, such as the importance of the gestational age of the fetus in the detection of positive cases and that the weight of the mother is more important than simply calibrate the two AIA and indicators hormone hCG.

Author: CORNELLA MEDRANO JORDI.
Year: 2006.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: AULA CAPELLA. ETSEIB.
Place of preparation: ETSEIB, Edifici H PLANTA 11 SD.

Summary: Grasping and manipulation of objects are fundamental tasks in robotics, required in many applications ranging from pick and place to assembly operations. These tasks are executed by means of the end-effector of the robot, which is usually designed to accomplish a specific task with specific objects. The current end-effectors used in the industry have reached a high degree of development and sophistication, executing the task for which have been designed with a high reliability in terms of repeatability, precision or execution speed. Nevertheless, the operating flexibility of these end-effectors is still limited and their performances frequently decrease drastically when small variations affect the task, the environment or the object for which they have been designed. Furthermore, the robots are being required to work in more unstructured environments and service applications, such as home or space applications. In these situations, the flexibility of the end-effector acquires more relevance, motivating the development of more general purpose end-effectors. The human hand is the clearest paradigm of high operating flexibility tool, able to grasp and to manipulate a wide range of objects with different shapes and sizes. Then, it is natural the proposal of developments of mechanical devices imitating the human hand performance. The advantages of the mechanical hands with respect to the conventional end-effectors are the result of increasing the number of degrees of freedom of the device, which also increase the complexity of the overall system. Some of the most significant problem are related both with hardware issues (actuation, sensing,...) and software issues (control techniques, grasp synthesis algorithms,...). This thesis deals with the development of grasp synthesis algorithms following the mechanical approach. In this approach, the physical and mechanical properties of the hand, the object and the contact points between them are analyzed, in order to obtain mathematical models of these properties and to identify suitable parameters to quantify them and evaluate the goodness of the grasp. Namely, this thesis focuses on the development of grasp synthesis algorithms able to generate grasps that fulfill the basic properties of disturbance resistance and equilibrium. These two properties take into account the characteristics of the object, the position of the contact points and the contact models, in order to assure that the mechanical hand is able to balance any external force applied on the object by exerting appropriated forces at the contact points. The obtention of grasps with these properties entails to solve two specific problems: the grasp synthesis, which consists in the determination of the contact points between the fingertips and the object, and the force distribution problem, which consists in the determination of the contact forces that the fingers must exert in order to hold the object. In general, a mechanical hand is a high redundant end-effector, implying that these problems may have multiple solutions. Then, the objective of this thesis is the determination of the optimal solution taking into account some of the optimization criteria well established in the literature. The computational cost of the algorithms developed to determine the optimal grasp and the optimal grasping forces is usually a factor that hinders their applications in systems with time constraints. Then, the proposed approaches try to avoid hard iterative procedures and to obtain analytical expressions for the optimal grasp determination whenever is possible. The thesis includes several examples of the proposed methodologies, showing their efficiency and applicability.

Author: HERNANDEZ JIMENEZ ENRIC.
Year: 2006.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: Sala del Llac del Rect. Campus Nord.
Place of preparation: OMEGA DESPATX 240 ND.

Summary: The concept of equality is a fundamental notion in any theory since it is essential to the ability of discerning the objects to whom it concerns, ability which in turn is a requirement for any classification mechanism that might be defined. When all the properties involved are entirely precise, what we obtain is the classical equality, where two individuals are considered equal if and only if they share the same set of properties. What happens, however, when imprecision arises as in the case of properties which are fulfilled only up to a degree? Then, because certain individuals will be more similar than others, the need for a gradual notion of equality arises. These considerations show that certain contexts that are pervaded with uncertainty require a more flexible concept of equality that goes beyond the rigidity of the classic concept of equality. T-indistinguishability operators seem to be good candidates for this more flexible and general version of the concept of equality that we are searching for. On the other hand, Dempster-Shafer Theory of Evidence, as a framework for representing and managing general evidences, implicitly conveys the notion of indistinguishability between the elements of the domain of discourse based on their relative compatibility with the evidence at hand. In chapter two we are concerned with providing definitions for the T-indistinguishability operator associated to a given body of evidence. In chapter three, after providing a comprehensive summary of the state of the art on measures of uncertainty, we tackle the problem of computing entropy when an indistinguishability relation has been defined over the elements of the domain. Entropy should then be measured not according to the occurrence of different events, but according to the variability perceived by an observer equipped with indistinguishability abilities as defined by the indistinguishability relation considered. This idea naturally leads to the introduction of the concept of observational entropy. Real data is often pervaded with uncertainty so that devising techniques intended to induce knowledge in the presence of uncertainty seems entirely advisable. The paradigm of computing with words follows this line in order to provide a computation formalism based on linguistic labels in contrast to traditional numerical-based methods. The use of linguistic labels enriches the understandability of the representation language, although it also requires adapting the classical inductive learning procedures to cope with such labels. In chapter four, a novel approach to building decision trees is introduced, addressing the case when uncertainty arises as a consequence of considering a more realistic setting in which decision maker's discernment abilities are taken into account when computing node's impurity measures. This novel paradigm results in what have been called ``observational decision trees'' since the main idea stems from the notion of observational entropy in order to incorporate indistinguishability concerns. In addition, we present an algorithm intended to induce linguistic rules from data by properly managing the uncertainty present either in the set of describing labels or in the data itself. A formal comparison with standard algorithms is also provided.

Author: RUIZ VEGAS FRANCISCO JAVIER.
Year: 2006.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: SALA D'ACTES DE L'EPSEVG.
Place of preparation: EDIFICI U DESPATX 518 Campus SUD.

Summary: This thesis contributes to the design of specific kernels to be used on machine learning algorithms, such as Support Vector Machines, when the involved patterns are described by interval variables or qualitative labels. With regard to the intervals, two types of kernels are considered in this thesis: explicit kernels and implicit kernels. The first ones are based on the center-radio diagram, allowing an original geometric formalization of the set of closed intervals of the real line. On the other hand, the implicit kernels are based on the concept of interval intersection. One of the most important results in this thesis is the justification that the length of the intersection of intervals satisfies the kernel requirements. The kernel intersection can be improved by the concept of influence functions solving the problem of lack of discrimination presented in some cases by intersection kernels. With regard to the absolute orders of magnitude, using the last mathematical formulations of this structure, a similar formulation is presented showing that the set of qualitative labels supposes a discretization of the set of the intervals in the real line. This point of view allows extrapolating the results obtained in the case of intervals to the qualitative structure of orders of magnitude absolute. In addition, other methodologies are also considered, for example, the decay factor and the vectorial notation of the qualitative labels. The developed kernels will also be able to be used with precise quantitative data by a previous process of discretization. In this thesis a novel method of supervised discretization is proposed. This new method considers the existing order of the output variable (discrete or continuous variable). The developed methodologies are applied to two different situations, the first one on a financial framework and the second one in the industrial field of the painting manufacture for automobile. In the first case, the developed method of discretization and the qualitative kernels are applied to reproduce the rating or measurement of credit risk. In the second case, the hybrid kernels are used (applied simultaneously to data with quantitative and qualitative variables) in the design of a decision aid system for colour adjustment in the automobile industry.

Author: ACOSTA SARMIENTO JESUS ANTONIO.
Year: 2006.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: Sala de Juntes la FME.
Place of preparation: EDIFICI U DESPATX 518 Campus SUD.

Summary: It is commonly established that more intelligent systems can be obtained by the hybridization of Soft Computing methodologies, in order that the weaknesses of some systems be compensated with the strengths of others . Neural Fuzzy Systems (NFSs) and Evolutionary Fuzzy Systems (EFSs) are the most notorious representatives of these hybrid systems. An Evolutionary Fuzzy System is basically a fuzzy system augmented by a learning process based on an evolutionary algorithm (EA), particularly Genetic Algorithms (GAs), which are currently considered as the most well-known employed global search technique. This kind of algorithms have the ability to explore and to exploit complex search spaces, which allows the obtaining of solutions very close to the optimal ones within these spaces. Besides, the genetic codification employed allows to incorporate a priori knowledge in a very simple way and to use it to guide the search. In this PhD. thesis, we propose EFSs that improves a modeling and simulation technique the Fuzzy Inductive Reasoning (FIR). The main goal of the EFSs is to take advantage of the potentialities of GAs to learn the fuzzification parameters of FIR, i.e. the number of classes per variable (granularity) and the membership functions (landmarks) that define its semantics. Due to the fact that it is a methodology based on fuzzy logic, FIR modeling and prediction performance is directly influenced by these discretization parameters. Therefore, the automatic determination of precise fuzzification parameters in the FIR methodology is an interesting and useful alternative to the use of heuristics and/or default values. Moreover, it is expected that the automatic selection of adequate values for these parameters will open up the FIR methodology to new users, with no experience neither in systems modeling nor in fuzzy logic, guaranteeing the best performance of this methodology. Three evolutionary methods of automatic learning of fuzzy partitions are presented: a) The learning of the granularity with uniform membership functions (GA1+EFP), b) The local tuning of the membership functions with a fixed number of classes per variable (GA1+GA2), and c) The learning at the same time of the granularity and the membership functions associated that define its semantics (GA3). The evolutionary methods have been implemented in Matlab and they run in both Windows and Linux environments. The results obtained by the EFSs developed in the four applications studied, i.e. human central nervous system, maintenance costs of electrical medium line in Spanish towns, short-term estimation of ozone concentration in Austria and long-term estimation of ozone concentration in Mexico, were very good. The results obtained by our evolutionary methods have presented higher efficiency in the prediction process than those obtained by other methodologies in previous works, by FIR using default values and, even, by FIR when the fuzzification parameters have been defined by experts in the area. In general, the GA3 and the combination GA1+GA2, in that order, are the ones that have shown better results in all the applications, followed by the GA1+EFP. However, GA3 is the algorithm that presents the greatest computational cost. As general conclusion, we must say that the EFSs designed and implemented in this thesis yielded good results for the task which they were entrusted in FIR methodology. Therefore, the user should decide what EFS turns out to be more convenient for the modeling application at hand in function of time and precision needs.

Author: García Mateos Ginés.
Year: 2006.
University: MURCIA [www.um.es].
Place of defense: Facultad de Infor. Campus de Espinardo.
Place of preparation: Facultad de Informática.

Summary: One of the most active fields of research in the last decade, within the discipline of computer vision, which is known as the area "looking at people." In generic form, it will fall into all systems of image processing, pattern recognition and artificial perception, which aims of the study are human beings. The analysis of faces is one of its major branches, whose applications include: automatic recognition of people, the development of new methods of interaction man / machine coding and labeling video systems for video or multimedia indexing; security video surveillance and access control; systems capture information gestural and helps disabled. This thesis deals with the major problems of visual processing human faces from the point of view of the integral design. Intuitively, a comprehensive design (or simply a projection) is not more than the average of the values of gray an image along the rows or columns of pixels. While the projections are one of the classic techniques of image analysis, its use in the context at hand has been marked by the design of heuristic methods and ad hoc. We raised the need to sustain its use in more formal mechanisms, such as projection models and alignment between one-dimensional signals. These two issues are studied in detail, and suggests several possible ways to model projections and an algorithm for efficient alignment of a signal over a model. Leveraging both elements are designed methods for the detection of faces in static images, the location of facial components, monitoring of the faces in video sequences, facial recognition of people, the analysis of facial expressions, and the estimated the position and orientation 3D face. The extensive experiments conducted demonstrate the advantages of using projections compared to other types of mechanisms: greater capacity to spread; high immunity to noise; and invarianza versus individual factors and facial expressions. The algorithms proposed for the different problems always reach results equal or superior to those of other alternative methods, which are comparable to the state of the art, but with a highly significant improvement in computational efficiency.

Author: Mariño Perez Castor.
Year: 2006.
University: A CORUÑA [www.udc.es].
Place of defense: Facultad de Informática.
Place of preparation: Facultad de Informática.

Summary: This thesis deals with the automatic alignment of medical imaging. Once we have obtained images of a patient at different moments of time or captured using different techniques often necessary to align all images in a single axis of reference in order to extract quantitative measures of the same. To undertake this process manually is a tedious task and prone to errors due to the low quality of the images they provide some of these methods. The system allows developed align sequences SLO (Scanning Laser Ophthalmoscope) through the combination of two different methods for alignment of images, one based on the extraction of features and another based on the intensity of the image. The first techniques developed based its operation in the removal of structures called curved ridge and valley curves, which represents the fund's vascular structures of the eye and is used as a reference in the adjustment process. However, for some of the areas of the streams in which the contrast is very low, the method based on the extraction of features fails, polr what is necessary to implement any alternative alignment. Therefore, it used a technique alignment based on the calculation of an ownership of the images, the mutual information, which is derived from the values of the intensity of the pixels. Once obtained these methods and tailored to our application domain, combined conveniently so that the process would be fully automated, and finally validation was performed in a hospital setting, the Centro Hospitalario Universitario de Santiago. Moreover, once aligned images, can be derived from the sequence a series of measures that make it possible to evaluate the hemodynamic circulation in the retina, as certain anomalies in the movement appear before the neovascularization of the retina, which is what causes the blindness in patients with diabetes retinopática. It has developed a technique to estimate the parameter called hemodynamic time arteria-vena automatically from the aligned sequences SLO allowing contruir called dilution curves, which are some charts that represent the changing level of gray one point of the image to all frames of the sequence.

Author: García Vega Manuel.
Year: 2006.
University: JAÉN [www.ujaen.es].
Place of defense: Escuela Politécnica superior.
Place of preparation: Escuela Politécnica SUperior.

Summary: The disambiguation of the meaning of words (Word Sense Disambiguation) is the problem of assigning a particular meaning to a word polysemic, using their context. This problem has been of interest, virtually since the beginning of computing in the 50's. The disambiguation is an intermediate task and not an end in itself. In particular, it is very helpful, sometimes essential, for many problems of PLN, such as information retrieval, text categorization, the machine translation ... The objective of this thesis is to implement a desambiguador the meaning of words based on the Model Space Vector optimizing the weights of the training vectors using the neural network LVQ (Learning Vector Quantization) monitored Kohonen neural network model and to propose a uniform method of integrating resources used to train the network. The parameters of the network LVQ are optimized for the problem of disambiguation. This work has shown that neural networks, specifically models Kohonen, brilliantly solved the problem of the resolution of the ambigà ¼ age lexical, providing robustness, because the network LVQ is insensitive to small changes observed results homogeneous regardless of training, flexibility because it is easily applicable to any task PLN; scalability, it can be introduced multitude of texts training to fit any domain and effectiveness, because the results are comparable and in many cases outperform the traditional methods used to solve the same problems . We have integrated the corpus SemCor and WordNet lexical database. In addition, it has provided a method for the automatic integration of any corpus. The experiments show good behavior of this network for the specific problem of disambiguation.

Author: LOPEZ PANIAGUA IGNACIO.
Year: 2006.
University: POLITÉCNICA DE MADRID [www.upm.es].
Place of defense: E.T.S. ING. INDUSTRIALES.
Place of preparation: E.T.S. ING. INDUSTRIALES.

Summary: Despite the development of artificial systems in recent years, there is still work that can not be done or that can only be carried out inefficiently and simplified. One of the main lines of current research within artificial systems, focuses on cognitive systems. It is intended that these systems employ some mechanisms of biological origin or inspiration to operate more efficiently and tackle more complex tasks that present in uncontrolled environments. Another motivation for this line is to build machines able to interact socially with humans. These objectives and areas of application involving solve many difficulties, mainly stemming from scenarios operation very high uncertainty and the need to deal with abstract concepts. The cognitive engineering systems at the moment requires a strong theoretical development to provide a conceptual basis and a solid body of methodology to enable analysis and design processes systematic and formal. This thesis is a step in this direction, proposing a semi-formal conceptual basis for analyzing and designing autonomous systems and in particular its subsystems perception. The thesis is formulated from the conceptual framework of the general theory of systems, in the formulation given in An Approach to General Systems Theory, George J. Klir, 1969. This is a reference text in the area. It has been used to build a conceptual framework for analysis of systems regarding their autonomy and structure. This framework serves to precisely analyze the perception, in terms of their relevance to an autonomous system, processes and components involved, and with regard to knowledge and operation of the system. After analyzing surveys widely perceived existing experimental psychological (Gestalt, Inferencial, Ecological, etc.) and technical (Marr), this paper summarizes the perceptual process into two phases, called "sensory" and "directed". The first phase consists of identifying characteristic features of the objects in the environment in the sensor readings. The second, in the interpretation of these characteristics to identify the objects themselves. The theory explains how this process presents an implicit dimension related to the emotions of the system, and other explicit related knowledge (and the transaction with abstract concepts). It also elaborates on how it presents the perception within an autonomous system, seen in the general case as a multithreaded system, distributed and parallel. Over the text provides a number of examples to illustrate the new concepts, and devotes a portion of the text to analyze various systems in terms of the thesis. Finally, it explicitly develops its relationship with other studies on perception, identifying commonalities and differences. The main conclusion is that it has succeeded in developing a conceptual framework semi-formal. However, the initial aim of formalizing it has not been fully achieved, because it is a task of great complexity that requires research. However, it has been exploring possible alternatives to acometerla, concluding: theory classes, fractals, geometry.

Author: IGLESIAS PÉREZ ROSA.
Year: 2006.
University: PAÍS VASCO [www.ehu.es].
Place of defense: FACULTAD DE INFORMÁTICA.
Place of preparation: FACULTAD DE INFORMÁTICA.

Summary: It has been tested in virtual environments monousuarios devices hápticos useful for the simulation of the assembly of mechanical components. These devices significantly improve the performance of tasks and help designers and engineers to evaluate design and digital mock-ups or, for example, decide on the best sequence assembly. Today, on the other hand, are emerging new collaborative virtual environments where multiple users in different locations can works at the same time in various stages of revision: the layout, design detailed analysis of defects, and so on. In this thesis, presents a new collaborative virtual environment called CHAS (Collaborative Haptic Assembly Simulator), which is based on a peer-to-peer network topology. In this environment, two users can simultaneously work together to carry out assembly operations through devices hápticos. Because of the conditions of the network (delay, delay variation, traffic, bandwidth) in virtual environments hápticos distributed there are two major challenges: maintaining the consistency (synchronization of different virtual scenes) and the provision of adequate feeling haptic (touch) to the user. A scheme was designed to ensure consistency and the results were satisfactory. Additionally, a new algorithm was implemented to improve the quality of haptic feeling when the network condition worsened. Different experiments in the laboratory and during several tests between real Labein (Bizkaia) and Queen's University (Northern Ireland), have validated the system CHAS. CHAS provided quality in the haptic sense when both users conducted assembled remote (assembling of an object with another object caught by another user). In addition, when caught collided objects for users (collisions dependent), the haptic perception was satisfactory in most cases.

Author: FERNANDEZ BAYON GUSTAVO.
Year: 2006.
University: OVIEDO [www.uniovi.es].
Place of defense: DEPARTAMENTO DE INFORMATICA.
Place of preparation: CENTRO DE INTELIGENCIA ARTIFICIAL.

Summary: In some real problems learning to deal with information is given in the form of arrangements of objects. In these cases, an alternative may be to adapt the methodology of learning the nature of the data. The goal then is to design learning algorithms can learn from ordinations of objects. Such algorithms are known as methods of learning preferences. Knowledge of departure for such algorithms is formed by comparisons between objects or partial orderings, and the model of knowledge representation varies depending on the approach used. Some examples of the use of learning preferences include risk analysis for strategic decisions in financial fields, navigation systems to aid the user or search engines like Google. Many algorithms learning preferences see as its effectiveness decreases when the number of variables in the input data unrelated to the concept you want to learn skyrockets. Intuitively it can be pointed out that not all the variables of a problem are equally important. Some are related to the concept that you want to learn, and others can be ignored in the process. This degree of connection is known as relevance of variables. The detection of relevant variables, in the purest sense of the term, can lead to inefficient methods. In these cases, the alternative is to focus the search variables useful for the problem and learning algorithm employees. The usefulness of a subset of variables is a concept closely linked to learning method employed. The search and identification of variables useful brings a number of advantages, such as reducing the computational requirements of the system, decreasing the cost of obtaining data entry, or a better understanding of this knowledge into our databases. Our proposal, in general terms, is to adapt mechanisms study of the relevance of learning preferences. The hypothesis raised responds to the idea that systems designed expressly to work with information in the form of preferences must obtain better results than those implemented on other learning paradigms. The report proposes a series of methods for the management of variables, tailored learning preferences. First, there are two systems to work with linear problems: an adaptation of the algorithm Relief, and a version of the method SVM-RFE. It also describes two methods nonlinear because, in practice, it is not always possible to find a solution linear. The first is a direct application of a certain criterion for management, and the second is an extension of SVM-RFE. On the other hand, and as a method of helping the previous describes a filter to detect redundant variables based on the correlation between variables of the problem. Likewise, the report presents a family of algorithms, known as methods based on metric, as an alternative to conventional techniques for selecting models. As an example of such algorithms describes two approaches, ADJ and QADJ, used in combination with the methods of management above variables. All the above methods have been tested on three different real problems. First you have used a scoring problem of live cattle, in order to prove that the variables necessary for the proper valuation of the animals were available in a less expensive way. The other two issues relate to both problems sensory data analysis: the acceptance by consumers of beef and traditional Asturian cider respectively. The results support our hypothesis, and show a great performance of the systems presented, especially if you compare techniques Chosen 8 ion m 32c odelos based on metric with classical techniques. In summary, the methods presented here have demonstrated their ability to improve our processes of learning preferences.

Author: SUAREZ FERNANDEZ MARIA DEL ROSARIO.
Year: 2006.
University: OVIEDO [www.uniovi.es].
Place of defense: SEDE DEPAR. CAMPUS DE VIESQUES. GIJON.
Place of preparation: DEPARTAMENTO DE INFORMATICA. UNIVERSIDAD DE OVIEDO.

Summary: The objective of this report is to present a new definition for exchanging information based on the classic definition. This new definition will apply to the specific problem of optimizing fuzzy partition of random variables blurry. It also shows that in most cases, these partitions optimized offered an error classification based on fuzzy rules. It also aims to further expand the field of study, covering data processing imprecise and demonstrating that this definition is applicable to problems of this type. There will always be a research paper that passes through the search for information on the design of fuzzy partitions, as well as various ways to use the information as a measure of mutual optimization by other authors. It shows the method proposed by us, as well as their application for the optimization of both partitions blurry accurate data with bad data. To be followed by a study of classification algorithms exist in the literature, as necessary to carry out our experiments. And finally implement genetic algorithms for optimization with accurate data with bad data.

Author: García Torres Miguel.
Year: 2006.
University: LA LAGUNA [www.ull.es].
Place of defense: Escuela Técnica Superior de Ingeniería Informática.
Place of preparation: Escuela Técnica Superior de Ingeniería Informática.

Summary: This thesis proposes techniques metaheurísticas for resolving the problem of selecting features that arises in the preprocessing stage of a process of knowledge discovery in databases. We studied and implemented two strategies resolution based on the search dispersed (SS) and the search for environments variables (VNS), respectively. The pilot study conducted with these strategies confirms the effectiveness of the same, given that achieve results that are competitive with those obtained by standard techniques of literature.