ARTIFICIAL INTELLIGENCE (3)

Author: VICENTE RIPOLL MARIA ASUNCION.
Year: 2005.
University: MIGUEL HERNÁNDEZ DE ELCHE [www.umh.es].
Place of defense: E.POLITECNICA SUPERIOR DE ELCHE.
Place of preparation: ESCUELA POLITECNICA SUPERIOR DE ELCHE.

Summary: This thesis focuses on the recognition of objects freely through artificial vision. Specifically, it is oriented towards the recognition of objects based on the appearance, that is, making use of all the information of the object in images more representative. The methods for recognizing objects based on the appearance use feature extraction techniques to compress models appearance of the objects to recognize. One objective of this thesis is to clarify a number of contradictions in the present, on which method of extracting linear features is more suitable in certain circumstances. Specifically, the aim is to bring some light to the lack of consistency between the results obtained by analyzing separate components (ICA) and those obtained by the classical method of principal component analysis (PCA). Another area of work of the thesis focuses on the analysis of new methods of extraction of features, such as the random projection (RP) or non-negative matrix factorization in (MFN). In this thesis is stated under what conditions these new methods can improve the properties of traditional techniques such as principal component analysis and linear discriminant analysis (LDA). Thus, it proposes a new architecture (subspace changed object) for the construction of models appearance of the objects, which improves recognition results using non-negative matrix factorization on. The detailed study of all these techniques for extracting features concludes with the development of a novel system for detecting objects freely in informal settings, based on the use of multiple channels of information (multichannel appearance), which allows detection objects in various scales (multiscale). The new detection method proposed the following features: search combines two stages in parallel, makes use of two methods of extracting features (principal component analysis and non-negative matrix factorization restricted nature (NMFSC)), makes a automatic adjustment of the influence of each of the channels of information depending on the object to be detected and generates, on the basis of information training object of interest, a threshold automatic detection. The search system objects proposed is implemented on a real world environment, besides presenting a detailed analysis of the detection capability of this system.

Author: AZNAR GREGORI FIDEL.
Year: 2005.
University: ALICANTE [www.ua.es].
Place of defense: ESCUELA POLITÉCNICA SUPERIOR DE ALICANTE.
Place of preparation: UNIVERSIDAD DE ALICANTE.

Summary: Humans and animals have evolved to develop the ability to use their senses to survive. The sensory fusion, which is one of the cornerstones of this development is done in a natural way for animals and humans to get a better interaction with the surrounding environment. The emergence of new sensors, advanced processing techniques, hardware and process improvements have made viable fusion of many types of data. Currently merger sensory systems have been used extensively for tracking objects, automatic identification, reasoning, etc.. Apart from many other areas of application (such as the monitoring of complex systems, the automatic control of industrial manufacture?) Fusion techniques are also used in the field of artificial intelligence and robotics. This thesis provides the model FSR-BAY for sensory fusion robotics. This model takes into account certain aspects which from our point of view has been handled by secondary most of the current fusion architectures: incomplete and uncertain information, skills and learning to use a uniform representation of the information, independent the level of merger. It also provides two case studies of the proposed model applied to an autonomous agent. The first case concerns the merger cooperative information used to merge data from multiple sensors of the same type. The second case merges so competitive both heterogeneous and homogeneous information.

Author: Samper Márquez Juan José.
Year: 2005.
University: GRANADA [www.ugr.es].
Place of defense: Escuela Técnica Superior de Ingeniería Informática.
Place of preparation: Departamento de Inteligencia Artificial - UNED.

Summary: In this thesis develops a new system of information retrieval and filtering, called NectaRSS, recommends that information to a user based on the interests of the latter. The method automatically makes the task of acquiring user preferences, avoiding explicit feedback. A review of all concepts related to the system, showing different approaches from which the scientific community has addressed the problem, with particular impact in the context of the Web, which will apply initially. Lastly, check the effectiveness of the proposed method by applying it to implement a smart aggregator used by multiple users heterogeneous, demonstrating their ability to deliver personalized information according to the interests of each individual.

Author: TORRES GOTERRIS FRANCISCO.
Year: 2005.
University: POLITÉCNICA DE VALENCIA [www.upv.es].
Place of defense: UNIVERSIDAD POLITÉCNICA DE VALENCIA.
Place of preparation: UNIVERSIDAD POLITÉCNICA DE VALENCIA.

Summary: In this thesis, entitled "Systems dialogue based on stochastic models," describes the state of the art in the area of systems for dialogue and presents the work done in the design and implementation of the modules of a system of dialogue determined . The thesis focuses on the study of the management of dialogue from a statistical approach. The thesis provides the development of a system of comprehensive dialogue (with input and output text, in English, and for a task semantic domain restricted, as defined in the research project BASURDE). This system consists of modules natural language understanding, dialogue management and generating responses in natural language. Since the central objective of the thesis, development of the module loader dialogue has been the main job, and therefore is exposed to the maximum extent in this report. The limited size of the corpus of dialogue with the task BASURDE has resulted in a severe difficulty in developing a dialogue manager based solely on statistical models. The module loader dialogue finally implemented determines its strategy of dialogue by combining multiple sources of knowledge: some of stochastic nature, the models learned from the body, others a heuristic rules incorporating semantic and pragmatic knowledge, either generic or specific to the task. Finally, it seemed the simulation of users as an alternative technique for purposes such as assessing the behavior of the dialogue, the expansion of the corpus through dialogues synthetic, or learning dynamic stochastic models of dialogue. We have designed and implemented the corresponding modules simulators user explored the possibilities of this technique central aim of the thesis, development of the module loader dialogue has been the main job, and therefore is exposed to the maximum extent in the present memory. The limited size of the corpus of dialogues with the task BASURDE has resulted in a severe difficulty in developing a dialogue manager based solely on statistical models. The module loader dialogue finally ends implemented its strategy of dialogue by combining multiple sources of knowledge; some stochastic in nature, the models learned from the body, others a heuristic rules incorporating semantic and pragmatic knowledge, either generic or specific to the task. Finally, it seemed the simulation of users as an alternative technique for purposes such as assessing the behavior of the dialogue, the expansion of the corpus through dialogues synthetic, or learning dynamic stochastic models of dialogue. We have designed and implemented the corresponding modules simulators user explored the possibilities of this technique.

Author: ESCUDERO BAKX GERARD.
Year: 2005.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: SALA D'ACTES-EUETIB.
Place of preparation: EDIFICI C6 Campus NORD.

Author: SANCHEZ FIBLA MARTI.
Year: 2005.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: SALA DEL LLAC-CAMPUS NORD.
Place of preparation: EDIFICI C6 Campus NORD.

Summary: This thesis presents a collection of new algorithms for solving Constraint Satisfaction Problem (CSP) and Weighted Constraint Satisfaction Problem (WCSP). We pursue two main objectives: enhancing solving methods for WCSP, which are of recent development, and narrowing the gap between search and inference methods. The first part of the thesis is devoted to search methods for solving WCSP. In a branch-and-bound context, the lower bound computation in each node of the search is of great importance and has a serious impact in the practical efficiency of algorithms. We start from an algorithm called Russian Doll Search (RDS) that has a costly, yet very powerful lower bound and we develop three enhancements of it: Specialized RDS (SRDS), Full SRDS and Opportunistic SRDS. We then tackle the problem of exploiting the global structure of the problem inside search. An algorithm exists for CSP that is able to exploit what is called pseudo-tree structure extend it to WCSP, obtaining algorithm Pseudo Tree Partial Forward Checking (PT-PFC). This algorithm has a source of inefficiency mainly related to bad quality local lower and upper bounds. We suggest a solution to this problem by combining pseudo-tree search for WCSP with the RDS techniques that we previously developed obtaining algorithms PT-RDS and PT-SRDS. In all this first part the aim is enhancing the practical efficiency of existing search algorithms with respect to the time spent in solving several benchmarks. The second part of the thesis is devoted to complete inference methods for solving CSP and WCSP. Complete inference solves the problem by a sequence of transformations that obtain an equivalent problem. In these transformations variable elimination plays an important role. We present some new inference operations that permit us to factorize a constraint into a set of smaller size constraints. We then introduce factorization into variable elimination. The result is algorithm {it Adaptive Consistency with negative factorized constraints ADC-_factor. With these operations we define also an alternative method to eliminate a binary domain variable. Next, we introduce the idea of Filtering which consists in anticipating tuples of constraints that will become inconsistent when joined with other constraints of the problem. One could say that we are doing the equivalent to look-ahead during search but the goal is to reduce memory storage instead of pruning branches and reduce time. We introduce filtering in the main complete inference algorithms for CSP and WCSP producing Delayed Variable Elimination with filtering ADC (ADC-DVE-F) and Bucket Elimination-DVE-F algorithms. Still in the complete inference context, we generalize filtering to tree decomposition methods that yield us to Cluster Tree Elimination CTE-F} algorithm. We also present an iterative pure inference algorithm that performs a sequence of more accurate approximations to solve the problem, we call it Iterative Mini Cluster Tree Elimination (IMCTE). All contributions to the complete inference part of the thesis are devoted to enhance the memory spent by these methods.

Author: ALCALÁ FERNÁNDEZ JESÚS.
Year: 2005.
University: GRANADA [www.ugr.es].
Place of defense: ESCUELA TÉCNICA SUPERIOR DE INGENIERIA INFORMÁTICA.
Place of preparation: ESCUELA TÉCNICA SUPERIOR DE INGENIERIA INFORMÁTICA.

Summary: The main aim of the thesis is to develop models lingüsticos through learning techniques and adjustment to improve the balance between interpretability and precision, using the representation of 2-tuplas lingüsticas which is a more flexible representation of the functions of belonging that allows make a lateral move for them. This we can decompose in the following specific objectives: 1. Extending the CDM using an approach that allows us to make a complete BC learning through learning a priori from the data base and based on the representation of 2-tuplas lingúísticas. 2. Expanding the representation of 2-tuplas lingüsticas a representation 3-tuplas lingüsticas that allows us to make an adjustment and lateral extent of the responsibilities of membership. 3. Improving the balance between interpretability / combining precision adjustment based on the representation of 3-tuplas lingüsticas with selection rules

Author: Alonso Burgos Sergio.
Year: 2005.
University: GRANADA [www.ugr.es].
Place of defense: E. T. S. de Ingeniería Informática.
Place of preparation: Escuela Técnica Superior de Ingeniería Informática.

Summary: The decision-making group is a field of study that recently was taking a lot of relevance. However, most of the models and processes that have been studied not contemplate the possibility that the information available to solve the problem of decision-making is incomplete, ie, the experts who provide the information are not able to express any preferences that they require. This thesis presents a process for selecting alternative and a process of consensus that allow you to work with incomplete information (preferably relations diffuse incomplete), thus solving the problem of incomplete information.

Author: Prados Suárez Ma. Belén.
Year: 2005.
University: GRANADA [www.ugr.es].
Place of defense: E.T.S. de Ing. Infor. y Teleco..
Place of preparation: E.T.S. Ingeniería Informática y Telecomunicaciones.

Summary: The segmentation of images is to draw the different regions existing in a given image. Regional means similar sets of pixels and connected. Most of the techniques detailed proposals so far have not taken into account the vagueness that may exist in these regions due to Shines, transparencies, fuzzy edges and degraded color or light. Techniques diffuse if they address this problem, but using the same measure to get all regions of the image, without taking into account the particularities of each. This thesis has proposed a methodology to represent and be tailored on an individual basis to the imprecision in each region of the image. It has defined a measure of similarity between diffuse pixels of the image and measures of connectivity based on the aggregation of similarities, along with an algorithm to calculate the diffuse region corresponding to a region of the image. The study of the different functions of aggregation for the calculation of connectivity (in particular, T-normas), provides various measures that allow individually tailored to the imprecision inherent in each region. To make this adjustment automatically, it has been proposed a technique for estimation of the imprecision in the region and has been linked with the value of precision the extent to be used for that region, establishing a functional relationship with the parameter of the T-norma Weber. Applying this methodology on a package of seeds representing the regions of interest, you can get a fuzzy image segmentation. To determine the initial set of seeds is not a trivial problem, which has proposed an algorithm that part of a sobre-siembra on the image and select a set of seeds representing regions. To make this selection has been defined a measure of goodness of the seeds. Moreover, in analyzing an image it can be viewed at different levels of detail. To look at these different levels of precision, a methodology has been proposed to construct a hierarchy of nested segmentations diffuse from a diffuse initial segmentation. Each level of the hierarchy corresponds to a fuzzy segmentation of the image, so that at the lower levels are segmentations with a large number of small regions and in the higher regions with few large, formed from the union the regions of unprecedented levels. This hierarchy has been built through the definition of a relationship of similarity between regions diffuse diffuse, which takes into account both the similarity between regions such as the inaccuracy in their transition. The various levels of the hierarchy are given by different alfa-cortes that can be done about this relationship of similarity.

Author: PÉREZ RODRÍGUEZ EDUARDO JAVIER.
Year: 2005.
University: MÁLAGA [www.uma.es].
Place of defense: E.T.S INGENIERÍA DE TELECOMUNICACIÓN.
Place of preparation: E.T.S. INGENIERÍA DE TELECOMUNICACIÓN.

Summary: This thesis has developed basic infrastructure that allows the capacity of navigation in dynamic unstructured environments to an autonomous mobile generic, allowing its subsequent adaptation to different physical agents and easy extension of their capabilities. This has been developed and employed a new control architecture called DLA (Distributed and Layered Architecture), which allows you to combine algorithms through cooperative interaction processes freely distributed. This architecture has been designed with the objective of transparency in its use, consolidated by the simplicity and portability of the same. It also presents the possibility of distributing processes on several machines depending on the network load and processors involved, controlling the speed of response of the modules. The basic infrastructure of navigation developed using a hybrid structure, which combines properly and how to conduct synchronous fast reactive processes deliberate planning. This way you can respond quickly to environmental stimuli captured while undergoing a planning system operation for the development of complex tasks. The behavior of the reactive system has been developed through a scheme based on learning able to acquire kinematic constraints during the training process. This facilitates the adaptation of behavior reactive to different players with different dynamic characteristics without the need for a specific analysis. For its implementation has been used paradigm of reasoning based on cases, which allows for a fast learning both from our own experience as a training agent. The deliberate processes built into the system help increase the efficiency and capabilities of navigation on the system operation. We have developed two different levels of planning: a planner of roads that makes it possible to calculate a path free of obstacles to reach the final destination, and route planner which allows calculating a path formed by the regions to cross to reach the final destination. To carry out these processes of planning has generated a kind of probabilistic metric representation on the basis of information received from sensory environment, and a topological representation from representation metrics through a hierarchical pyramid structure. Both remain constantly connected through a structure of links, which allows plans spread quickly from one level to another. The architecture implemented has proven to be very robust operating correctly in real dynamic unstructured environments for different physical agents after a short training.

Author: NAVARRETE TERRASA ANTONIO.
Year: 2005.
University: POMPEU FABRA [www.upf.edu].
Place of defense: DEPARTAMENTO DE TECNOLOGÍA.
Place of preparation: UNIVERSIDAD POMPEU FABRA.

Summary: The geographic datasets represent reality through a series of thematic entities that are often not defined in a precise manner and that can understand different subjects in different ways. In this context, the integration of geographic information from different sources presents a major challenge in terms of semantics. This thesis proposes a solution to this problem based on ontologies and logic description. It has defined a framework semantic whose core is an ontology representing the thematic concepts in a repository dataset, as well as the relationships between these concepts. The ontology is built through a process of fusion (merging) of ontologies for implementing the datasets that have been inserted into the repository. This paper proposes a semi-automatic method of merging, which have developed three different algorithms in order to generate a list of suggestions for mapping operations that an expert may accept or modify. This marcosemántico allows semantic definition of services that go beyond the functionality that current catalogs offer geographic information. In particular, one of the three services semantic defined in this dissertation is the integration into a new dataseet thematic information from different sources. Finally, the semantic framework and its services will be used in a system of indexing and retrieval of geo-referenced multimedia elements (still images and video clips) from its geographical thematic content.

Author: GONZÁLEZ MORGADO M. CRISTINA.
Year: 2005.
University: PAÍS VASCO [www.ehu.es].
Place of defense: FACULTAD DE INFORMÁTICA.
Place of preparation: FACULTAD DE INFORMÁTICA.

Summary: This thesis focuses on the study of theoretical algorithms ADDs (Estimation of Distribution Algorithms). The ADDs optimization algorithms are valid both in space as a container that can be framed in discrete within the Evolutionary Computation. They are based on heuristic stochastic populations of individuals, where each individual in the population encodes a possible solution to the problem of optimization. So far, much work has been done both in the creation of new ADDs as in the implementation of the same, but this effort has not been accompanied by a theoretical analysis of the same. This thesis want to be in the theoretical analysis of ADDs, providing information about mathematical modeling and performance. This paper addresses the major issues that measure the performance of any optimization algorithm: convergence and complexity temporary. On the one hand it is important to know under what conditions can be guaranteed that the algorithm reaches an optimal solution (convergence). Moreover, the expected number of steps needed to achieve an optimal solution (known as computation time) is an important measure of an efficient algorithm. Related to this, in this thesis explores the relationship between the computational time and the size of the problem (complexity temporary). The two mathematical tools used along this thesis to analyze and model ADDs have been the Markov chains and discrete dynamic systems. The main contributions of this thesis has been: * Construction of an analytical framework based on Markov chains for the study of the convergence of a generic EDA. Using the previous condition has been analyzed discrete ADDs more common. To ensure that it can not be convergence across the condition had been imposed conditions in the estimation of genetic parameters to ensure convergence. * Studying the behavior of PBIL applied to the minimization of the role OneMax in two dimensions, showing the dependence of the behavior of the algorithm some initial parameters of the same. * Analyze the algorithm PBIL applied to the minimization of the role OneMax in two dimensions, showing the dependence of the behavior of the algorithm some initial parameters of the same. * Analyze the algorithm PBIL with discrete dynamic systems. A joint system with the dynamic interactions of PBIL has proven that the iterations of PBIL follow the dynamic system for a long time when a certain parameter (alpha) of the algorithm is close to zero. Analyzing the dynamic stability of the system has concluded that PBIL only converge to the best local and converges to the global optimum unimodal functions. * Analyze the algorithm UMDA in continuous spaces. Modeling the behavior of the algorithm in linear functions such behavior has been studied far from the optimum, while modelándolo for quadratic function has been studied near optimal behavior. * Analyze the worst time expected to reach the optimum for the first time. This analysis is based on modeling of ADDs using Markov chains, and is valid for any function pseudo-booleana. It has been shown that the time but expected to reach the optimum for the first time in UMDA, MIMIC, TREE and EBNABIC is exponential (in the size of the individual). The study also provides ideas for marking the average time expected to reach the optimum for the first time. * Calculate empirically time can be expected, on average to achieve the optimum for the first time for certain ADDs applied to certain functions pseudo-booleanas. Concluding that, in the cases studied greater complexity of probabilistic model used does not imply lower mean time to achieve the optimum for the first time. 8 As well co 227 mo that the fact that the algorithm mimics the structure of the objective function does not imply lower mean time to achieve the optimum for the first time.

Author: SANTANA HERMIDA ROBERTO.
Year: 2005.
University: PAÍS VASCO [www.ehu.es].
Place of defense: FACULTAD DE INFORMÁTICA.
Place of preparation: FACULTAD DE INFORMÁTICA - UNIVERSIDAD DEL PAÍS VASCO.

Summary: The thesis introduces a number of properties on the kind of approach that uses Kikuchi descomposiciones based cliques. An approximation algorithm that learns from this data is introduced and evaluated in different types of problems approximation. Markov properties of approximate Kikuchi in the product approximations Kikuchi local defined in a decomposition of the graph. Additionally, the proposition clarifies the place of descomposiciones based cliques in conjunction with other techniques inspired by methods of statistical physics and discusses the implementation of the outcome introduced in designing learning algorithms approximations Kukuchi. An algorithm based on "search + assessment" that learns approximations Kikuchi from data is entered. The thesis presents the results in the use of this algorithm in the approximate probability distributions generated from networks Bayesianas. An algorithm to estimate widespread distributions is introduced. The thesis deals with the problem of optimization methods based graphical models to problems in computational biology and bioinformatics. The results in the application of different variants of estimation of distribution algorithms for problem solving computations protein.

Author: PÉREZ DE LA FUENTE JESÚS M..
Year: 2005.
University: PAÍS VASCO [www.ehu.es].
Place of defense: FACULTAD DE INFORMÁTICA.
Place of preparation: FACULTAD DE INFORMÁTICA.

Summary: This paper presents a new paradigm of automatic learning: the tree algorithm construction consolidated. This algorithm has been developed to solve classification problems where necessary change the original distribution of the classes, and, in addition, is seeking an explanation of why the classification. To do so, the algorithm is based on a set of samples taken from the original database with the proper sampling technique and builds a single classification tree, we called tree consolidated, as consensus induced knowledge of all of them. Several experiments have been carried out to analyze the behavior of the trees consolidated in different contexts of implementation and from different perspectives. The results, in general, the trees are well suited to consolidated problems which modifies the distribution of the classes, but also when they are used as a paradigm of classification capabilities with more explanatory, reducing the error rate of classification and getting a greater structural stability, ie greater stability in explaining the classification. Finally, it has developed a parallel implementation of the algorithm to deal with problems large volumes of data efficiently.

Author: PEDRINACI GODOY CARLOS.
Year: 2005.
University: PAÍS VASCO [www.ehu.es].
Place of defense: FACULTAD DE INFORMÁTICA.
Place of preparation: FACULTAD DE INFORMÁTICA.

Summary: This thesis investigates the development of advanced services based on knowledge using the tools and technologies available in the context of semantic web. The main contribution of this thesis lies in what we have called a Platform for Opportunistic Reasoning. This is an infrastructure for the development of knowledge-based services are capable of thinking on the information distributed on the web. The platform consists of an introduction of the model-based reasoning on the blackboard (Blackboard Model) adapted to work on the web. The platform has been evaluated through the development of two applications. The first is a system of aid for the design of events (conferences, meetings, symposia) and the second is responsible for managing the payment of copyright by issuing music on the internet. Both applications show the applicability of the Platform for Opportunistic Reasoning.

Author: TRELLA LÓPEZ MÓNICA.
Year: 2005.
University: MÁLAGA [www.uma.es].
Place of defense: ESCUELA TÉCNICA SUPERIOR DE INGENIEROS EN INFORMÁTICA.
Place of preparation: ESCUELA TÉCNICA SUPERIOR DE INGENIERÍA INFORMÁTICA (UNIVERSIDAD DE MÁLAGA).

Summary: The Guardians Intelligent Systems (STI) are educational applications based on instructive learning, guiding students adapted and monitor their learning. The birth of the Web has led to a new generation of STI are known as Intelligent Learning Environments for the Web (WILE, Intelligent Web-based Learning Environment (. The process of developing these systems is very expensive, taking into account his multidisciplinary (involved technicians, teachers, experts in the domain of education and educators, who normally are built from scratch for a domain and some specific educational objectives. The results are often very specific systems in which the knowledge and how to present it they are closely related. The international community calls for the construction of learning environments, for the reuse of what has been termed LO. LO, as defined for the IEEE, is a "non-entity digital or digital it can be used, reused or referenced during technology-based learning (technology-supported learning). "Many organizations standardizing impact on the importance of creating reusable materials and therefore independent of the platform that creates and maintains. however WHAT are the development-oriented systems "not smart" who lack the benefits of an individualized learning provided by WILE. This thesis aims to combine the advantages of both approaches defining a methodology for the development of intelligent learning environments for the web based on the integration of education systems that already exist. also has built a framework that supports and facilitates the construction of WILE by non-specialist users, so as to allow the use of educational programs such as LO environments applied to various smart domains. This proposal aims to reduce the development costs of increasing the supply of educational resources without sacrificing the intelligent application of techniques. MEDEA, as a methodology, articulates the development process, depending on the type of user involved in an educational system and needs arising in its context, choosing the integration of previous developments. The proposed methodology is independent of the implementation. As a result, their application is not dependent on domains, teaching strategies and methods of diagnosis and is not influenced by any technology development. Est methodology is a bet on integration allowing to be used previous developments completed (both the content and form of submitted), ie allows reuse not only educational materials but "intelligence" of other systems.

Author: MOLINA VILAPLANA JAVIER.
Year: 2005.
University: POLITÉCNICA DE CARTAGENA [www.upct.es].
Place of defense: E.T.S. DE INGENIEROS INDUSTRIALES.
Place of preparation: ESCUELA SUPERIOR DE INGENIEROS INDUSTRIALES.

Summary: In this Doctoral Thesis proposed and developed new models of neural inspiration for biological control and learning tasks grip by anthropomorphic robotic devices. In the first part of the thesis is being exhausted out a review of the most relevant aspects of human behavior and animal movements during grip objects which highlights the invariant features of this movement, established through numerous experiments psicofídicos with humans and primates. Below is a look at the current state of knowledge regarding the neurobiology underlying driving behaviors described above. On this basis, the thesis presents a model for organizing the movement of grip that mimics the interactions between different areas of the cortex and basal ganglia during the planning and execution of movement grip under normal conditions and under conditions of parkinsonian motor deficits. The model generates realistic trajectories of grip over the computer and continues to update signals encoding the difference between the engines that are established to carry out the task, and the current state of the late effects of the movement involved in the execution that task. The main assumptions of the model are: 1 - The control of the transport of the hand and the opening of the fingers takes place through the action of signals passing talámicas whose modulation is provided by the neural circuits of the lymph basal. These signals allow the coordinated implementation of the various sub-task of composing a grip. 2 - The disruption of the program driving detected in Parkinson's disease, is due to a change in the functionality of the network and interneuronas cholinergic of striated before a defection of doparían is triatal. Under these conditions, the network of interneuronas loses the ability to maintain segregated information loops crotico-processed in parallel and basal ganglia occur as a result of matings between different channels coritco-subcorticales affecting patterns of kinematic prototype movement grip. 3 - The application of this model to a system in which the end effectors of the movements are antropomorficante realistic, involves the development, based on the results of experiments designed expressly mentally in this Doctoral Thesis, a control scheme biologically plausible to reduce the dimensionality of the problem of coordinating the gesture of the hand, during the movement of grip. This control scheme is what is defined as at the Library of Theses gestures. 4-Learning that allows the establishment of programs engines after the perception of the object is performed via a novel architecture neural network based on multi-cortical connectivity between areas of the cortex and posterior parietal cortex promoter who, after number of stages of learning, it is capable of generating movements correct grip for a set anthropomorphic robot arm-hand when this system is presented with objects of different shape and size, regardless of their location or orientation in space. The thesis presents numerous results for the simulation of models in different situations and outcomes related to the introduction of such models on a platform robotics antromoporfa oriented grip objects. These results support theoretical assumptions underlying this investig 8 ation and 344 on the other hand show the capabilities of the models developed to act as high-level controllers in the task of guiding grip humanoid robotic manipulators.

Author: MÉNDEZ REBOREDO JOSÉ RAMÓN.
Year: 2005.
University: VIGO [www.uvigo.es].
Place of defense: ESCUELA SUPERIOR DE INGENIERÍA INFORMÁTICA.
Place of preparation: ESCUELA SUPERIOR DE INGENIERÍA INFORMÁTICA.

Summary: This paper presents a hybrid system of Artificial Intelligence capable of detecting and filtering spam messages. In this context, he noted the need for tools capable of acquiring new knowledge about the domain knowledge discarding obsolete in a dynamic way. In these situations, it has been verified that a hybrid model that implements a system of reasoning based on cases may provide an effective method, that provided by other classical techniques of Artificial Intelligence. The proposed model is based on the use of a reasoning system based on the request which incorporates a structure indexing messages, a strategy vote, a mechanism for calculating the quality of the solution and created a technique for the elimination of knowledge irrelevant, with the aim of obtaining a high level of precision and allow for rapid adaptation to changes in the environment. Each of these techniques are used in a different stage in the life cycle of reasoning system based on past experiences to recover bodies adapt to the present problem, review the proposed settlement and acquire new knowledge. The rationale for the hypothesis put forward in this work is done on an experimental basis, using three different body of evidence. The results obtained from experiments on the model developed, are compared with those generated by the use of different techniques classic Artificial Intelligence, which allows the realization of a qualitative and quantitative analysis of the effectiveness and efficiency of the proposed system . Finally, and in view of the results obtained, it is concluded that the application of the hybrid model is defended special interest in the scope of the problem studied. In this sense, the model developed generates adequate results and is able to dynamically adapt to the changing characteristics of the environment. Finally, it provides a set of key ideas to deal with the construcicón systems for filtering spam to be derived from the work done.

Author: GÓMEZ MEIRE SILVANA.
Year: 2005.
University: VIGO [www.uvigo.es].
Place of defense: ESCUELA SUPERIOR DE INGENIERÍA INFORMÁTICA.
Place of preparation: ESCUELA SUPERIOR DE INGENIERÍA INFORMÁTICA.

Summary: This thesis establishes a new algorithm to transcribe polyphonic music for piano. The initial analysis of the problem is based on the decomposition of traditional spectrum in the frequency domain, however, it introduces a number of new terminologies and concepts designed specifically to discuss the inherent complexity of polyphony. In particular, establishing the rule of transcription with two purposes: 1-Eliminate the need to determine a precise time of onset (onset) of each note, as other researchers have arisen on this issue. 2-Allow the introduction of a series of other new concepts -grupo notes polyphonic and resolution group notas- which represent the basis for the algorithm that we developed. The rationale for the hypothesis put forward in this work is done on an experimental basis, presenting the implementation of the proposed system to a set of test cases grouped on the basis of different criteria, extensive enough to check on a statistical reliability of the algorithm, and comprised of small segments of music chosen from among a large number of recordings and sheet music for piano. It has defined a methodology for assessing the complexity of the test cases and get a more objective measure of performance based on a set of statistical defined purpose.

Author: DÍAZ HERMIDA FÉLIX.
Year: 2005.
University: SANTIAGO DE COMPOSTELA [www.usc.es].
Place of defense: ESCOLA TÉCNICA SUPERIOR DE ENXEÑERÍA.
Place of preparation: FAUCLTAD DE FÍSICA.