ARTIFICIAL INTELLIGENCE (2)

Author: GIRET BOGGINO ADRIANA SUSANA.
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: In the area of Intelligent Manufacturing Systems is the ultimate need for methodologies for systems Holónicos Manufacturing (HMS-Holonic Manufacturing Systems), based on principles of software engineering, to attend a software engineer in all phases of development and guides that provide analysis and design clear and unambiguous. We are convinced that methodologies in the area of Multi-Agent Systems (GHS) are good candidates for the model of HMS. Some reasons are: the similarities between the approach holónico and approach of agents, the extensive use of agents, as the tool for implementing HMS and the availability of comprehensive methodologies SMA. However, there are some areas that we must include a methodology for SMA which is capable of modeling requirements HMS adequately: recursive structure of holones, levels of abstraction in the system, modeling specific guides for HMS and an analytical approach design and mixed (ascending and descending). In this thesis we propose Agent Abstract entity modeling for as autonomous entities with recursive structures. The definition of Agent Abstract extends the traditional definition of agent adding a structural perspective to the concept of agent "... an agent Abstract can be an agent, or can be a GHS, which in turn is composed of agents Abstract ...". Agent Abstract is an attempt to unify the concepts of agent and holón and to simplify the analysis and design stages. Thus, it will be easier to translate the products obtained from modeling methodologies for HMS executable elements for the implementation of the system holónico. The main contribution of this dissertation is a methodology for Multi-Agent Systems Holónicos Manufacturing. This methodology is based on the notion of Agent Abstract and requirements modeling HMS. Our methodology defines a mixed development process, and provides guidance specific to HMS that help the software engineer to identify and implement holones. In nuestroenfoque HMS specified breaking out more concrestos to different views of systems: model agent, organization model, a model of interaction, environment model and model of tasks and objectives. The manner in which these views are defined based on the methodology INGENIAS. The extensions we have made to meta-modelos of INGENIAS deal with the notion of Agent Abstract, redefining some of relationships to adapt to the new entities modeled, and the inclusion of features for modeling real-time methodology RT-MESSAGE . The process of developing our methodology provides the software engineer clear and specific guidelines for HMS, and development phases for the complete life cycle of HMS. Phase Analysis System Requirements and second Identification and specification Holones define the analysis phase of our approach. The objective of the analysis phase is to provide a high-level specification of HMS from the system requirements. The analysis phase takes a top-down approach recursive. One advantage of a recursive analysis is that their results, this is the Módelos Analysis, defines a set of basic elements and rules of composition. The next step in the development process is the design Holones which is a bottom-up process to produce Architecture System from Módelos Analysis of the previous phase. The objective now Implementation Holones is to produce the executable code for phase Installation and Configuration. Finally maintenance activities are conducted in the phase of Operation 8 and Mante 2cd nimiento.

Author: BUENAPOSADA BIENCINTO JOSE MIGUEL.
Year: 2004.
University: POLITÉCNICA DE MADRID [www.upm.es].
Place of defense: FACULTAD DE INFORMATICA.
Place of preparation: FACULTAD DE INFORMACIÓN.

Summary: The facial expressions are an essential component of the communication process between humans. It is not surprising, therefore, that the quantification of facial expressions is a topic of interest in the field of computer vision by its usefulness in the construction of advanced communications interfaces with the computer. It is also a topic of interest in the field of animation, an area in which the confluence graphic techniques and vision. The face is a difficult subject by analyzing techniques Computer Vision presents because vast areas with little texture and its regions more expressive (eyes, eyebrows and mouth) suffer deformities not rigid. If we add to this the difficulty of modeling and predicting the movement of the head along with the existence of or with occlusions (jiciones changing lighting, understand why the analysis of the face by Computer Vision is a difficult problem that today open continues. In this thesis we develop an analysis system that will allow us to find facial and follow the human face and quantify their facial expressions. order to monitor use an architecture based on the "Phased Approach to Care." This architecture consists A set of followers with different levels of accuracy and computational load. follower less precise means searching randomly face in the picture, paying attention to regions with a color similar to the skin. region Once we have a nominee, it employing the texture of the face of the person to continue to locate in a more precise way his head. A follower based on the appearance, trained with the side of the user, will enable us to deal with the movement of the face and rigid, while estimating the rigid movement of the whole. The resulting system controls the computational cost and accuracy in monitoring using a follower less precise and lower costs when conditions are degraded environment, and increasing accuracy when improves. Eventually, as a practical application, movement will use the parameters of a model to encourage estimated figure 3D. order to tackle the problem of changes in lighting, has proposed a standardization process color based on a well-known algorithm record colored static scenes, the algorithm Gray World, but extended to the case of image sequences. With the new development can follow objects from their color, with a good degree of robustness to changes in illumination. addition, it has developed an efficient follower based on differences in the levels of gray (SSO) which can estimate the position and orientation in 3D on a plane using a standard design movement. is based on the extension to the case design of the idea of factoring of the Jacobian of Hager and Belhumeur. has also been developed a procedure for the election of all pixels more information template monitoring to further increase performance follower. Lastly, it has developed a technique for factoring of the Jacobian to solve the problem of tracking deformable objects. using our algorithm is the movement is not possible to follow rigid areas of the face in real time. The resulting algorithm is interesting not only because of its computational efficiency, but also because it is easier to train the well-known models assets Appearance (AAM).

Author: ROMERO MERINO ENRIQUE.
Year: 2004.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: Llenguatges i sistemes informàtics.
Place of preparation: EDIFICI C6 Campus NORD.

Author: DAUDE VENTURA JORDI.
Year: 2004.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: DEPARTAMENTO LLENGUATGES I SISTEMES INFORMÁTICS.
Place of preparation: EDIFICI C6 CAMPUS NORD.

Author: ESPINOSA SOTO JESÚS ANTONIO.
Year: 2004.
University: ALICANTE [www.ua.es].
Place of defense: FACULTAD DE CIENCIAS.
Place of preparation: FACULTAD DE CIENCIAS.

Summary: Currently there are many groups working qu using the program Geodyn to determine the precise calculation of satellite orbits, and in particular of LAGEOS. We need to ascertain the presence of a parameter to determine a weight to the comments of each season. This report presents a porcedimiento to find the weight to the comments of each station monitored by laser telemetry, which optimize the contribution of the observed data on the problem of precise calculation of orbits. To perform this task the memory is divided into three parts. The first, comprising the first and second chapters, introduced the grading method to be used. In this case the classification diffuse. Here there is a study on the current state of classification diffuse. In the second segment, which includes the third and fourth chapters, develops the core of the thesis. The third chapter explains the principles of a critical look at the algorithms more widely diffused in classification of a sample of data. It will demonstrate why this feature is convex. Besides introducing a new procedure to accelerate the convergence of classification algorithms diffuse. In the latter part applied the new role gained in the ranking algorithm diffuse, to get the weights to be assigned to the observations of the seasons of SLR. This presents techniques for monitoring laser telemetry, and shows the importance of these in achieving a model of the gravitational field of the earth by tracking satellite telemetry laser, and shows the importance of these in the pursuit of a model gravitational of land through seguimietno satellites in different orbits. With these weights calculating the orbit of the satellite LAGEOS is significantly enhanced with ralación the standard procedure for allocating weights.

Author: GONZALEZ BEDIA MANUEL ANDRES.
Year: 2004.
University: SALAMANCA [www.usal.es].
Place of defense: EDIFICIO HISTORICO.
Place of preparation: DEPARTAMENTO DE INFORMATICA Y AUTOMATICA.

Summary: The main objective of the doctoral project is the design of a model agent capable of planning and reschedule in changing contexts of information. The architecture VCBP (planner based Calculation Variacional) proposal, from a base plans and techniques topology on all possible solutions, capable of representing and integrating information available to generate plans to achieve goals, and reschedule in cases those plans currently being interrupted by changes in the environment. The variational calculation techniques make it possible to determine the set of possible solutions to a problem and techniques Fields Jacobi calculate that solution is more easily replanificable in case of being interrupted. The extension E-VCBP model incorporates mechanisms for re-planificación based on the significance of the emotions as mechanisms of transition between plans. The system was validated in different domains, with different and distinct uncertainty dynamism, through a series of simulation experiments. The experimental analysis of the system shows the power of new strategies in the area of planning in dynamic contexts.

Author: MARIN SANCHEZ JESUS.
Year: 2004.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: SALA D'ACTES DE EUETIB.
Place of preparation: EDIFICI C6 Campus NORD.

Summary: This work includes the analysis of a novel stochastic algorithm named Macroevolutionary Algorithm with a theoretical model to make quantitative predictions about its behavior when is applied to optimize globally a function without constraints. It also includes the design and implementation of a general simulation environment to be easily applied to several problems and allowing to make comparisons with other optimization techniques in order to evaluate the balance between solution quality and efficiency. Through this study, we tried to provide to the algorithm the ability to allow an easier and clearer adjustment of its parameters that control such balance between exploration and exploration than standard genetic algorithms. In this way, though depending upon the problem under consideration, the premature falling into local optima could be reduced, as the performed experiments have demonstrated. Moreover, the algorithm performance ---expressed in terms of balance between solution quality and efficiency--- makes it a very competitive optimization method. The thesis gives especial relevance to the theoretical model that describes the macroscopic behavior of its nonlinear stochastic dynamics along the optimization process. The main idea considers the algorithm dynamics as if it would be a biological process to be modelled, obtaining a coupled system of non-linear equations. The model does use of analytical or statistic information about the function, represented as probability distributions according with fitness values and spatial locations of its points. So, the model is able to make predictions about the algorithm behavior as adjusted as good would be the quality of probability distribution. The model considers all algorithm parameter values and, for each time step, it gives quantitative measures as the found-solution quality or the number of function evaluations. As a consequence, this model would be useful in order to get the optima values for these parameters to be applied to a problem, according to desired compromise between efficiency and quality of results

Author: GARCÍA BÁEZ PATRICIO.
Year: 2004.
University: LAS PALMAS DE GRAN CANARIA [www.ulpgc.es].
Place of defense: FACULTAD DE INFORMÁTICA Y MATEMÁTICAS.
Place of preparation: FACULTAD DE INFORMÁTICA Y MATEMÁTICAS.

Summary: One of the major issues that has focused the interest of many researchers throughout all time has been the understanding of the role / structure of the brain and its style of computing. Another major challenge is building on previous knowledge, to obtain artificial systems with skills specific to the brain. Following the huge increase in the availability of information that has occurred with the development of the so-called society of the Information Technology and Knowledge (ICT), have appeared many new problems. The process of the huge amount of data that the company generates, has become a daunting challenge which requires the design of new intelligent system capable of processing and interpreting such data have almost no knowledge about the qualities they present. This directs us towards systems as a matter auto-órganizativo for dotarios the greatest capacity automatically. We believe that the most appropriate approach to the problem will be using approximations based on neural computation. Despite the great benefits that it provides, in recent years the emergence of new neural architecture generalists, able to cope in an efficient and robust to sues raised, it has not been decisive. Our research has been focused mainly on proposing an original solution, effective and high potential, to the problem, enmarcándose this solution and the development of the same, in the approach to neural computing. This thesis covers work performed for the design of a novel adaptive neural architecture, modular, hierarchical, no-supervisada called HUMANN. This network aims to fill the gap within the field application of the lcusterización based on neural computing, as well as the work developed for the validation of the same in different application domain. The neural network is made up of modules HUMANN hierarchically organized in neuronal processing with different neurodinámicas, connection topologies and laws of learning and biological plausibility, both functional and structural. To verify the effectiveness and stability of HUMANN, includes a section that performs an analysis and comparative study, which details the performance of the network using a variety of environments. Finally describes two real-world applications in the fields dela neuroscience and the environment. Both applications are characterized by the use of a large amount of real data and good results.

Author: LUQUE RODRÍGUEZ MARÍA.
Year: 2004.
University: GRANADA [www.ugr.es].
Place of defense: ESCUELA TÉCNICA SUPERIOR DE INFORMÁTICA.
Place of preparation: ESCUELA TÉCNICA SUPERIOR DE INGENIERÍA INFORMÁTICA.

Summary: In a globalized world that is rapidly changing, being constantly informed has become a pressing need, a source of knowledge and money. A recent wave of subscriptions to online news services highlights the importance that society gives to be permanently informed about issues that are of interest to you. This "update" allows both to the individual and the organizations to be competitive and make better decisions. One of the main problems of the Internet is constant and uncontrolled growth of the information that users can access. This growth of the Internet, websites and in documents and Web services, is helping users who have difficult access to the information they need in a simple and efficient. It is necessary, therefore, systems to help them cope with this great mass of information that has become the Internet. As a result, research in areas related to search or access to information, either on the Web or in any other system, has increased significantly in recent years. These investigations are based on different philosophies of work, which may include within the definition of Access to Information (in English, "Seeking information"), a term that denotes a process by which the users are able to obtain information from a system. Some of these processes include: information retrieval (IR), information filtering (EF), access to databases, extracting information and "browsing". In recent years, we are witnessing the growing application of various sciences in these processes of access to information in order to improve them. In particular, methods, concepts and techniques of Artificial Intelligence (AI) are being applied in the process of obtaining information with remarkable success. Therefore, the study of some of the processes of access to information, as well as improvements to them in order to obtain greater efficiency in the search, is shown as a line of very active research. In this report we have focused on the study of IR and FI, in the time that we have proposed combining them with AI techniques for achieving better search for information. In particular, the contribution of this report is to: 1 - The characterization of the user profiles as a classic consultations RI (persistent queries), made under the provisions of the models Boolean or Lingüstico. 2 - The development of algorithms that allow learning to generate automatically, without the user having to interact with the system, taking into consideration Aes multiobjetivo. 3-Use of Information Lingüstica Difusa in defining the query subsystem and its associated subsystem evaluation to provide more expressiveness to the consultations. The main objective of this report has been improving the information needs of users, both permanent and ad hoc, using these techniques for Soft Computing. This objective has been achieved through three proposals: 1-Developing a Genetic Algorithm Programming Multiobjetivo for Learning Automated Queries Persistent. A user can satisfy a need for information that will arise in a given time without more than sitting in front of a type system and the consultation that represents. Then, if the system should have some information about user preferences, it could help in the formulation of the search or use that knowledge to filter 8 informac 1d70 ion, which was originally obtained. However, CRS are limited by the lack of customization in the representation of the user's needs. Therefore, an important issue is the construction of profiles that represent these needs. Belkin and Croft suggested that the techniques RI could be successfully applied in the FI. Thus, the profiles can be represented by consulting classic RI (persistent queries), and also the techniques of formulating queries, such as feedback by relevance or IQBE, can be applied to the process of FI. Accordingly, we have developed an algorithm evolutionary multiobjetivo based Genetic Programming (PG) to enable learning automatically profiles representing the information needs of a user. To provide more expressiveness to the profiles, we have proposed consultations represent classic RI (persistent queries), a model made by RI Boolean, instead of how the classical structure "bag of words". As a basis for the algorithm we used the proposal of Smith and Smith. The authors proposed a process IQBE based on an algorithm to automatically learn PG for consultations Boleanas. In view of what has been suggested by Belkin and Croft, we can interpret it learns profiles are represented as consultations Booleanas, englobándolo well within our framework. In addition, the extension of this proposal with a focus multiobjetivo based on Pareto enables us to obtain multiple profiles with different balance of precision and completeness in a single performance, while comparing the use of AEs multiobjetivo or mono-objetivo. 2-Design of a Model SRI Difuso-Lingüstico based Information Lingüstica Multigranular. The use of variable lingüsticas to represent the input and output information in the recovery process of CRS improves interaction between the user and the system. Most CRS Lingüsticos assume that users provide their information needs through consultations Booleanas whose terms are weighted by values of the variable lingüstica "Importance," while the RSV associated with each document after evaluating a query is a the value of the variable lingüstica "Relevance." In many of these systems, these two variables are valued on the same set of labels S, ie using the same set of labels to express the entry and exit of SRI. However, this way of expressing the entry and exit is not desirable because on the one hand, reduce the possibilities of communication between user and system, and secondly, since both variables lingüsticas represent different concepts, it seems logical to use different sets of tags to model, using, for example, a modeling lingüstico multigranular. We have raised the use of modeling Lingüstico Difuso Multigranular (MLDM) in order to provide more flexibility and provide representation to inquiries snapshots in the process of the IR. To that end, we propose a model of SRI Lingüstico that accepts queries Booleanas weighted with weights lingüsticos and provides as output documents valued for their RSV, expresándolos through information multigranular. Its main features are: A-Consultations Booleanas weighted values and RSV lingüsticos associated with the documents are valued on sets of tags with different granularity and / or semantics. B - The words in the query can be weighed simultaneously with three weights lingüsticos different semantics associated with three different semantics symmetrical threshold, semantics relative importance of quantitative and semantics. C-Boolean operators are modeled using a flexible manner proposed by the operator OWA Yager. D - The documents recovered are classified into classes relevance, identified by values lingüsticos ordinal rated on a set of tags associated with the different weights of the consultations. 3-Development of a Genetic Algorithm for Learning Multiobjetivo Consultation Persistent Represented as Consultations Lingüsticas. The development of proposals for the points earlier shows as models lingüsticos provide more flexibility and ease of instantaneous representation to the consultations in the process of RI, and the characterization of user profiles as a classic consultations RI (persistent queries ) gets greater expressiveness in them. Therefore, finally, we raised combine both proposals seeking to combine the best made by each of them. To this end, we propose to use consultations lingüsticas generated by the CRS Lingüstico raised in the previous section to represent queries persistent. We believe that this form of model profiles as persistent queries flexible materially enhance the interpretability of consultations persistent obtained and efficiency in the recovery. On the other hand, despite the fact that you can use different types of queries to represent persistent queries (profiles), the user will often be very difficult to make the consultation due to the difficulty in selecting the right words to communicate with the system, what classically known as "problems of vocabulary" in the interaction hombre-ordenador. For this reason, they have been applied learning techniques in the construction of automatic "implicit profiling." Specifically, the system will automatically learns the profile from a set of documents provided by the user. Following this philosophy, we have developed a Algortimo Genetic multiobjetivo based on the approach IQBE allowing learn so suatoática, from a set of documents provided by the user, represented as a persistent search queries lingüsticas, solving the above-mentioned "problem of vocabulary ". However, the structure is most common profile of the so-called "bag of words", which consists of a set of keywords that represent the interests of the user. Therefore, it seems necessary to compare our proposal with classical methods of creating profiles, to see how the competition is. As a method of comparison, we used one of the best performing show that at present, based on the vector space model and the theory of probability. Specifically, we believe the value of selecting Robertson (RSV) as a formula for creating profiles and function Okapi BM25 as a function of similarity to the matching profile and documents.

Author: KUTANGILA MAYOYA DAVID.
Year: 2004.
University: GRANADA [www.ugr.es].
Place of defense: ESCUELA TÉCNICA SUPERIOR DE INGENIERIA INFORMATICA.
Place of preparation: ESCUELA TÉCNICA SUPERIOR DE INGENIERÍA INFORMÁTICA.

Summary: The thesis was motivated by the crisis of public transportation in Kinshasa, capital of DR Congo. Foremost among all causes, the city is set up so that all major activities are concentrated in the north, forcing the entire population of Kinshasa to make long and frequent travel from all over town to the north. These movements are not easy to perform. The aim of the thesis is to locate a number of points where establish some large facilities like shopping malls covering the entire metropolitan area of Kinshasa, adding to each of them a bus station to connect these centers, establishing some of them as centers addressing transport flows and determine transmission lines with fixed routes. Attaining these objectives has needed to collect data sobrela city of Kinshasa (Kinshasa socio-economic context), to study issues of localization. Because of the vagueness of the data collected, it has been necessary to use the methodology based on fuzzy sets (softcomputing) that allows process data approximate. This methodology has enabled us to raise fuzzy models with their corresponding algorithms yet been tackled on the problems of P-mediana, center and rotuing are problems locating concerned by the argument. The application of these models to optimize the city of Kinshasa have led to the following results: obtaining 10 sites where best to locate certain facilities (malls), 3 of them serve as centers for addressing transport and identification of 11 transmission lines covering the whole city of Kinshasa. In addition to the introduction and conclusion, the thesis is divided into 4 sections, 2 appendices and a bibliography.

Author: VERGES LLAHI JAUME.
Year: 2004.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: sala d'actes de laFME.
Place of preparation: EDIFICI U DESPATX 518 Campus SUD.

Author: MARTÍNEZ ESTUDILLO ALFONSO CARLOS.
Year: 2004.
University: GRANADA [www.ugr.es].
Place of defense: ESCUELA TÉCNICA SUPERIOR DE INGENIERÍA DE INFORMÁTICA.
Place of preparation: ESCUELA TÉCNICA SUPERIOR DE INGENIERÍA DE INFORMÁTICA.

Summary: This thesis is framed within the field of artificial intelligence, machine learning and neural networks evolutionary. Evolutionary algorithms is designed for learning and the design of neural networks formed by linked product. This algorithm defines two types of mutations to evolve the networks. It is also designed several hybrid methodologies combining evolutionary algorithm, global search, a local search algorithm, the Levenberg-Marquardt, using a process of grouping prior to implement the local search algorithm only to a few individuals in the population , said local search conducted at different times of the evolutionary process, and at the end. Similarly, some fundamental characteristics of this type of neural networks. Different methodologies are used successfully to several evidentiary problems, and even real problem of predictive microbiology related to the growth of lactic acid bacteria Leuconostoc mesenteroides, present in packaged meats cooked. We get more accurate and interpretable models that previously obtained with neural networks units Sigmoid to the problem of predictive microbiology and the problems test results comparable with other methods of regression.

Author: CHÁVEZ GONZÁLEZ ANTONIA MARÍA.
Year: 2004.
University: SEVILLA [www.us.es].
Place of defense: E.T.S INGENIERÍA INFORMÁTICA.
Place of preparation: FACULTAD DE MATEMÁTICAS.

Summary: The memory is part of the field of verification of knowledge bases associated with ontologies, in particular, the application of qualitative spatial reasoning to clean up ontologies so assisted by Automated Reasoning Systems (MRS), all in the context the Semantic Web. The tasks of setting up and managing ontologies are complex, which is compounded if we add the need to evaluate and optimize the quality of the knowledge represented by them. This memory is proposed as a solution to this problem a process of purification ontologies based on the application of qualitative spatial reasoning and formal support for the treatment mereotopológico of spatial representation of ontologies, specifically, the calculations connection Regions (RCC) . They develop each step of the design process, using SRA, thus certifying that the results do not make use of space intuitive properties which, although valid, may not be provable in the theory under consideration. They also do a detailed analysis of the results of each step of the process and the level of generalized solutions proposals, underlining the following: * It automates the space Qualitative Reasoning through SRA, studied extensions calculation RCC and providing a method that allows insertion ontological work with ontologies through its spatial representations. * On the other hand, it gives them a formal semantics to validate changes of ontologies. Thus, it becomes possible translation of an ontology to a spatial representation, reasoning with the tools of calculating space for repair. * Finally, it presents the translation of the effects of these transformations space, giving rise to a new ontology with the assurances it provides us with the mathematical formalism. This report also provides the basis for a classification of anomalies, obtaining a classification and an intuitive and comprehensive resolution delas same method based on the repair of arguments and interpretations using space for the repair. In conclusion, it gets justify the reasoning underlying the proposed method of cleansing, paving the way towards the design and creation of intelligent agents cleaning abnormalities in complex ontologies.

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

Summary: In this thesis focused on the synchronization of chaotic systems is intended, first, to explore this phenomenon from new perspectives such as the concepts of cost and energy sync, concepts that can serve as a tool to analyze and better understand some aspects of synchronization. For example, the cost to maintain a full sync is null for identical systems, while not identical to systems that cost tends towards a finite limit nonzero. In the same vein, the definition of a power synchronization allows us to show that maintaining a regime of complete synchronization between systems not identical requires a flow of energy per unit time between the system and forced an external source of energy. The amount of energy required can be found by averaging the dissipation system guided along the path of the reference system. Furthermore, the thesis contains a more input worked on the synchronization of chaotic systems, such as secure communications, or the timing of adaptive systems not identical. Also, we have developed a scheme that allows dynamic coupling systems coupled to achieve full sync error found in every corner of the state space.

Author: GAHETE DIAZ JOSE LUIS.
Year: 2005.
University: PONTIFICIA COMILLAS [www.upcomillas.es].
Place of defense: ALBERTO AGUILERA, 25.
Place of preparation: ALBERTO AGUILERA, 25.

Summary: This thesis develops the core of the tools that a hypothetical Observatory Internet would need to advise the various Autonomous Systems (hereafter ACE) exchanging traffic on the network, on possible improvements on routes that would benefit them, and for global Internet traffic. The core of the thesis consists of a Genetic Algorithm, which we call the deer Genetic Algorithm (AGC), which is to clone the lives of these societies of animals to obtain routes between ASÂ's better than those achieved by the algorithms are use. The model is based, therefore, on the social behavior and breeding of deer, which in the evolution of species is extremely selective. The AGC need to create routes between two ASÂ's, know the topology of the Internet. For that, we use the information generated by Route Views (RSV), which enables us to chart global Internet routing from routers located in different parts of the world. Basically, these data tell us the routes used in reality to reach a destination from an AS AS origin, driving more than a million routes. The AGC begins requesting data on the number of generations to create, the number of chromosomes (individuals) of each generation, probably crossing, as well as the origin node (number AS origin of the route) and destination node (AS number of destination ) on the route. They are getting the lowest possible number of jumps and our AGC, as has been mentioned, has performed better, giving an overall reduction at around 20%. From the BGP routing tables obtained from servers RV deduce the graph ASÂ's, as well as the logical relationships between them. Relations between ASÂ's are based on heuristics that the size of the AS is proportional to their degree (number of connections neighborhood) in the graph. Of particular analysis on how to make these routes, we infer the real relationships of these ASÂ's. That is, each AS regardless of their physical connection with their neighbors decide who wants to exchange traffic through trade agreements. These relationships will allow us to check whether the routes obtained by the AGC are feasible or should we recommend any change in relations between ASÂ's. The AGC gets, therefore, the best route without considering these policies, seeking only the shortest number of hops to reach a destination, demonstrating as mentioned above, which is more efficient than the traditional algorithms. In comparing the number of hops obtained with our AGC is not considered, when compared with those obtained by RV route, "prepend" (artificial increase in the number of hops). It described above has provided us with two types of results: 1. Routes better for fewer jumps. 2. Nodes that if change contractual relations with its neighbors would get better than the existing routes.

Author: GARRIGA BERGA CARLES.
Year: 2005.
University: RAMÓN LLULL [www.url.edu].
Place of defense: ESCUELA TÉCNICA SUPERIOR DE INGENIERÍA ELECTRÓNICA E INFORMÁTICA LA SALLE.
Place of preparation: ESCUELA TÉCNICA SUPERIOR DE INGENIERÍA ELECTRÓNICA E INFORMÁTICA LA SALLE.

Summary: The fuzzy logic has been implemented efficiently modeling problems since its inception. One of its main advantages is that it provides an understandable representation of knowledge. However, most researchers have focused their efforts on achieving accurate models and omitted capabilities lingüsticas of fuzzy logic. So, today one can find models fuzzy theoretically oriented to humans but it is normally very difficult to understand (or even impossible). This dissertation investigates issues relating to diffuse models intelligible Since that science has shown the application of modeling diffuse in the search for optimal models in terms of the error (in fact a model diffuse is a aproximator Universal), although some researchers have few focusing its efforts on achieving models actually intelligible and even lose some precision. In this paper we propose a comprehensive methodology capable of finding a model of diffuse intelligible locally (rule by rule) from data input. Here we determine the number and position of fuzzy sets required and also define the rules lingüsticas that relate. For this purpose we have developed a process that paulativamente takes into account various steps i techniques, some of which are original contributions. The whole procedure has been studied in detail as one might expect from a thesis, arguing the various solutions that have been considered at every step. The resulting method is very simple and intelligible. Therefore, you get the models end with a low computational cost that also facilitates the adjustment of their different options depending on the nature of the problem and the characteristics of the users. In this thesis we explained the whole methodology and illustrate its benefits (but also its problems) with a number of examples that are related in many cases. At the end we concluded that the goal of obtaining a simple method but intelligible capable of giving intelligible diffuse models has been achieved satisfactorily. KEYWORDS: Artificial Intelligence, Fuzzy Logic, modeling systems.

Author: CARRERAS PEREZ XAVIER.
Year: 2005.
University: POLITÉCNICA DE CATALUÑA [www.upc.edu].
Place of defense: Facultat d'Informàtica de Barcelona.
Place of preparation: EDIFICI C6 Campus NORD.

Author: Suárez Peñaranda Vicente.
Year: 2005.
University: A CORUÑA [www.udc.es].
Place of defense: Facultad de Informática.
Place of preparation: Facultad de Informática.

Summary: In this report elaborates on the analysis of the concept of epistasis. It shows how -conjuntamente with the order of a function-epistasis helps to explain the behavior of some functions whose ends are known (and its difficulty too) and that usually serve as a "laboratory functions" in the study of the difficulty of optimization a function through AG. The good correlation between epistasis and difficulty in order functions under seems to be incomplete as an explanation in the case of high orders. This has led us to generalize the concept of epistasis, showing how this generalization supplements the information provided by the epistasis "classical". The complexity of the mathematical formulation associated with this study is necessary to the use of algebraic tool along with the Walsh transform. The development of this study in this report is structured as follows: it begins with the introduction of AG and the description of its operation. It introduces the concepts of the scheme, order and length of defining a schema and relate to the selection process and genetic operators, leading to the fundamental theorem AG: Theorem of Schemes. It continues with a summary of the results already known on epistasis, understood as estimator of the difficulty. In the field of AG, the initial idea of epistasis enter it intuitively Rawlins, Davidor formula mathematically and then Van Hove rewrites of the compact design and standardizes the concept. The study of the algebraic extreme values of this indicator shows that the standardized epistasis is minimal when the function f is of order 1 (linear functions) optimized for the simple AG. For functions not negative, it is found that the maximum value is achieved for a type of functions on a chain whose value depends on each and every bit of it. The calculation of the standardized epistasis is sometimes very complicated. This task is simplified with the implementation of the Walsh transform, which have shown great efficiency in the calculation of this estimate. The coefficients Walsh to completely characterize any function, and its calculation is performed in a comfortable way through the theorem of Median Value for Hiperplanos. To underscore the simplification achieved in the calculation of epistasis standard with the use of this theorem, as opposed to direct estimation of the epistasis from the definition, include the development done by churches and others to calculate the epistasis standardized functions to Template Based on the definition and we do us using the theorem of Median Value for Hiperplanos. We continue with the characterization of the order of a role through their weightings Walsh, a problem was completely resolved to order functions and a partially order for two. A new estimate of the difficulty of a function, 2-epistasis standardized. Following the ideas of Davidor leading to the standardized epistasis, we propose to establish a measure of distance between any function suitability and space functions of order 2. The reason for the introduction of this new estimator is in the execution of a Genetic Algorithm on two laboratory functions: Template features and functions Royal Road widespread. The epistasis standardized grows at the same time makes the number of generations required for the convergence of the algorithm. However, for high values of the standardized epistasis, surprising the big difference in generations achieved for these d 8 will feature 6c5 ones. By contrast, the values of standard epistasis not reflect this significant difference in the behavior of both functions. It seems that the comparison of an adjustment to the functions of order 1 might be too strong. Linked successfully 2-epistasis standard with the second-order functions: it is precisely those functions that reaches the minimum value for this estimator. Recall that the standard classical epistasis verifies that was nullified if and only if it is calculated on the roles of the first order. It introduces the concept of higher-order epistasis, from a purely algebraic. The mathematical formulation of the concept of epistasis standard developed by Van Hove and obtained for 2-epistasis, the ability to perform a similar development for any value of k. It introduces the concept of k-epistasis standardized and it is estimated its value for some functions. Lastly, relates the algebraic fabric developed with the initial ideas of epistasis described by Davidor.

Author: VALDOVINOS ROSAS ROSA MARIA.
Year: 2005.
University: JAUME I DE CASTELLON [www.uji.es].
Place of defense: ESC. SUP.DE TECN. Y CIENCIAS EXPERIMENTALES.
Place of preparation: ESCUELA SUPERIOR DE TECNOLOGIA Y CIENCIAS EXPERIMENTALES.

Summary: In this Doctoral Thesis, mainly discusses the applicability of Multiple Systems Classification (SMC) under the rule of the nearest neighbor. An essential first line of research focuses on the preprocessing algorithms, in order to solve various problems related to the quality of training shows: the presence of redundant patterns, or atypical noisy databases with a size imbalance between excessive and distributions classes. Another aspect of great relevance refers to the effectiveness of the individual components of SCM in the voting method, which proposes new techniques for static and dynamic balancing of individual decisions. The third focal point refers to the analysis of diverse classifiers using various measures in the existing literature related. Other issues widely discussed throughout this thesis are: sampling techniques (bagging, boosting, arcing and sequential random selection), the size of the SMC, and finally the feasibility of using two models of artificial neural networks (perceptrón multilayer and modular network).

Author: MARTÍNEZ MUÑOZ GONZALO.
Year: 2005.
University: AUTÓNOMA DE MADRID [www.uam.es].
Place of defense: ESCUELA POLITÉCNICA SUPERIOR.
Place of preparation: ESCUELA POLITÉCNICA SUPERIOR - UAM.

Summary: This thesis proposes new methods of generating sets classifiers and several heuristics for improving management and by pruning sets generated with bagging. Specifically, the contributions made in the work are: (1) In this chapter presents three new methods of building joint classifiers that are characterized by using all data without modification training to build each of the set of classifiers . The algorithm basis, submitted Gelfand et al. (SB Gelfand, CS Ravishankar, and EJ Delp. Growing and pruning An iterative algorithm for classification tree design.IEEE Transactions on Pattern Analysis and Machine Intelligence, 13 (2) :138 -- 150, 1991) to build a decision tree of recursively from a set of data that is divided into two subsets. In each iteration, one of the two subsets is used to grow the tree from the decision tree obtained in the previous iteration. Once that has been done to grow the tree until its maximum size it is pruning using another subset of data. The roles of subassemblies are exchanged in each iteration. This process converges to a tree final is stable with respect to the sequence of steps of growth and pruning. To generate a variety of classifiers in all divisions are created so many random examples into two subsets trees as you want to build. Based on this procedure we have proposed three new methods of construction sets classifiers: PGI set, boosting PGI and committees PGI. These methods perform well classification in different databases with a standard computational cost less than sets based CART. (2) This chapter presents a set of classifiers whose members are built from alterations labels kind of a percentage of examples chosen at random from among those who set up the training. Using this method can be achieved great improvements in the classification error when using a high probability of changing class and labels are generated sets with a large number of classifiers. It also shows how classifiers generated following this procedure make mistakes throughout training not statistically correlated. The dependence of the error training sets generated with the size of the package is independent of the problem of classification analysis. In particular, we show how to binary classification problems, this unit can be analyzed in terms of a Bernoulli process. Finally, we show the results of experiments conducted at 15 databases standard that demonstrate improvements to be gained with this procedure. (3) The order in which the classifiers are added in a set can be a useful tool for selecting subsets of classifiers more efficient than the original set completely. Overall, the fallacy of a set of classifiers randomly ordered decreases with increasing the number of classifiers so asymptotic and tends to a constant value. If properly amending the order aggregation classifiers whole, the error of generalization can reach a minimum value is below the mistake of asymptotic complete set. This chapter presents several heuristics using correlations between classifiers generated by bagging to identify a suitable order to select a subset of classifiers with good capabilities generalization. Once all this is sorted pruning to select the first K classifiers according to a predetermined percentage of pruning or other rules of pruning. In this way you can build joint classifiers smaller and less error classification joint test that the original set completely.