Search

This Ph.D. Thesis elaborates and analyzes several hybrid Soft-Computing algorithms for optimization and prediction problems in Atmospheric Physics. The core of the Thesis is a recently developed optimization meta-heuristic, the Coral Reefs Optimization Algorithm (CRO), an evolutionary-based approach which considers a population of possible solutions to a given optimization problem. It simulates different procedures mimicking real processes occurring in coral reefs in order to evolve the population towards good solutions for the problem. Alternative modifications of this algorithm lead to powerful co-evolution meta-heuristics, such as theCRO-SL, in which Substrates implementing different search procedures are included. Another modification of the algorithm leads to the CRO-SP, which considers Species in the evolutionof the population, and it is able to deal with different encodings within a single population.These approaches are hybridized with other Machine Learning and traditional algorithms such as neural networks or the Analogue Method (AM), to come up with powerful hybrid approaches able to solve hard problems in Atmospheric Physics...

Trabajo presentado en la European Geosciences Union General Assembly, celebrada en Viena (Austria), del 8 al 13 de abril de 2018, Currently the atmosphere is scrutinized using a number of systems ranging from conventional meteorological stations, to geostationary and polar orbiting satellites. These observations are then assimilated in models for meteorological forecasting or climate reanalysis. The increasing availability of these products should not hide the need of keeping and improving the operation of weather and climate stations, which can be compromised due to constraints in the budget of the operating agencies. In this context, it is highly relevant to identify which are the stations producing the most relevant data for a given variable or process. This can be formulated as an optimization problem, in the aim of finding the ideal subset of points that best reproduces the characteristics of a certain region. Recent developments in Evolutionary Algorithms have validated the performance of biologically-inspired operators to solve an extensive collection of optimization problems. Specifically, the Coral Reef Optimization (CRO) algorithm has provided efficient solutions to the Representative Selection problem, outperforming other well-known methodologies. Here, we combine the CRO algorithm with the widely-used Analog Method in order to identify the most representative set of monthly average temperature time series for Europe from gridded (ERA Interim) and un-gridded (ECA) datasets. Our approach exhibits good performance and shows similar results in both cases. Moreover, the best solutions obtained with this methodology are climatologically consistent, and include points from Scandinavia, Central and Southern Europe, Eastern Europe and the Black Sea. Interestingly, once the number of selected locations reaches a certain threshold, the improvement in the temperature reconstruction is achieved by adding points nearby the previously-identified zones and not by reshaping the number of sub-regions. Therefore, a reduced subset of original points over key areas is enough to capture the climate variability of the entire region. Furthermore, our method excels by reducing the reconstruction misfit up to 28% when compared with other metaheuristic procedures, such as the greedy algorithm, based on individual selections. These results are just an example of the multiple applications of this methodology, which range from the identification of best located climate proxies to the design of operational meteorological networks. Besides, they also stress the need for further studies where evolutionary algorithms can shed light on the dynamics of several climate variables by providing the minimum number of sampling locations to obtain spatially-resolved climate fields.

The proliferation of social networks and their usage by a wide spectrum of user profiles has been specially notable in the last decade. A social network is frequently conceived as a strongly interlinked community of users, each featuring a compact neighborhood tightly and actively connected through different communication flows. This realm unleashes a rich substrate for a myriad of malicious activities aimed at unauthorizedly profiting from the user itself or from his/her social circle. This manuscript elaborates on a practical approach for the detection of identity theft in social networks, by which the credentials of a certain user are stolen and used without permission by the attacker for its own benefit. The proposed scheme detects identity thefts by exclusively analyzing connection time traces of the account being tested in a nonintrusive manner. The manuscript formulates the detection of this attack as a binary classification problem, which is tackled by means of a support vector classifier applied over features inferred from the original connection time traces of the user. Simulation results are discussed in depth toward elucidating the potentiality of the proposed system as the first step of a more involved impersonation detection framework, also relying on connectivity patterns and elements from language processing. The presented work has been possible thanks to the funding support of the Basque Government under the CYBERSID project grant and the computing infrastructure of the i2BASQUE academic network.

In this paper we present a methodology for feature selection and clustering over variables describing countries’ economies and ICT indicators to study and identify investment opportunities, based on similarities between European and Latin American countries. We address two different problems. First, the work is based on a feature selection problem carried out with the Coral Reef Optimization algorithm. The CRO is a novel bio-inspired based on the simulation of reef formation and coral reproduction. On the other hand, the K-Means++ method is a high-performance robust tool designed to solve clustering problems. Together, both algorithms are able to successfully identify investment opportunities in Latin America and quantify the potential of the telecommunications industry in both regional areas. The work considers different economical and ICT’s variables from different European and Latin America countries datasets (mainly Agenda 21 and other available and global sources) for the period 2002-2012.

Black-box optimization problems (BBOP) are de ned as those optimization problems in which the objective function does not have an algebraic expression, but it is the output of a system (usually a computer program). This paper is focussed on BBOPs that arise in the eld of insurance, and more speci cally in reinsurance problems. In this area, the complexity of the models and assumptions considered to de ne the reinsurance rules and conditions produces hard black-box optimization problems, that must be solved in order to obtain the optimal output of the reinsurance. The application of traditional optimization approaches is not possible in BBOP, so new computational paradigms must be applied to solve these problems. In this paper we show the performance of two evolutionary-based techniques (Evolutionary Programming and Particle Swarm Optimization). We provide an analysis in three BBOP in reinsurance, where the evolutionary-based approaches exhibit an excellent behaviour, nding the optimal solution within a fraction of the computational cost used by inspection or enumeration methods.

Black-box optimization problems (BBOP) are de ned as those optimization problems in which the objective function does not have an algebraic expression, but it is the output of a system (usually a computer program). This paper is focussed on BBOPs that arise in the eld of insurance, and more speci cally in reinsurance problems. In this area, the complexity of the models and assumptions considered to de ne the reinsurance rules and conditions produces hard black-box optimization problems, that must be solved in order to obtain the optimal output of the reinsurance. The application of traditional optimization approaches is not possible in BBOP, so new computational paradigms must be applied to solve these problems. In this paper we show the performance of two evolutionary-based techniques (Evolutionary Programming and Particle Swarm Optimization). We provide an analysis in three BBOP in reinsurance, where the evolutionary-based approaches exhibit an excellent behaviour, nding the optimal solution within a fraction of the computational cost used by inspection or enumeration methods.

Los Algoritmos Genéticos (GAs) son una clase de técnicas de búsqueda basadas en los principios de la evolución y la genética, que han sido aplicadas con razonable eficacia a múltiples problemas de optimización. Las principales características de los GAs son su gran robustez y versatilidad, lo que permite su empleo en problemas muy dispares. No obstante, se conoce la existencia de tipós de problemas en los que los GAs tienen un rendimiento probadamente bajo (denominados genéricamente problemas GA-hard). Estudios previos muestran la existencia de una relación entre la dificultad de un pro blema para un GA y la estructura del desarrollo en serie de Walsh de la función objetivo. Esta tesis desarrolla dicha relación a través del concepto de espectro, definido a partir de la expansión de Walsh: se presentan así diversas técnicas que mejoran el rendimiento de los GAs en problemas GA-hard y en problemas reales mediante la modificación de los correspondientes espectros. En primer lugar, se realiza un estudio del efecto que tiene la aplicación de una trans formación lineal (TL) en el espectro de una determinada función y, por tanto, en sus pres taciones. Se especifica en qué condiciones una TL puede mejorar el rendimiento de un GA y se propone un algoritmo para la búsqueda de TLs. Incidentalmente, estos desarrollos arrojan luz sobre resultados contradictorios que aparecen en la literatura a cerca de la viabilidad de TLs en esquemas evolutivos. En segundo lugar, se elabora una interpretación del funcionamiento de algoritmos híbridos (AHs) en problemas de optimización con restricciones fuertes. Concretamente, se propone un esquema híbrido -basado en la unión de un GA, como algoritmo global,y de una red de Hopfield, como algoritmo local-, que presenta ventajas en problemas de este tipo; y se estudian sus prestaciones a través de las modificaciones que causa en el espectro del problema. Por último, se aplica el AH diseñado a dosproblemas de optimización que aparecen en ingeniería de telecomunicación: el diseño de la trama de transmisión en una red PRN, y la minimización de la interferencia co-canal en sistemas de satélites mediante la reasignación de canales. En ambos problemas se consiguen ventajas significativas sobre algoritmos previos., Genetic Aigorithms (GAs) are a class of search techniques based on the principies of evolution and genetics, which have been efficiently applied to a wide variety of optimization probiems. The most irnportant characteristics of GAs arerobustness and versatility,which allow their application to very different kinds of problems. However, there exists optimization problems in which GAs are known to perform pooriy (these problems are generically denoted as GA-hard problerns). Previous studies have shown that there exists a reiationship between GA performance and the structure of the Waish series expansion of the problern’s objective function. This thesis eiaborates on this relationship by means of the spectrum concept, defined from the Waish expansion: thus, sorne techniques to improve GA perforrnance in GA-hard problems, and in real problems, are presented. The interpretation for these techniques rests on the modifications they cause to probiems’ espectra. Firstiy, we provide sorne new results about the effect of Linear Transforrnations (LTs) in function spectra, and therefore in the corresponding GA performance. We investigate the conditions for a LT to improve GA performance in a given function, and propose an aigorithm to search a feasible LT. Eventually, the obtained resuits shed light on sorne apparently contradictory resuits in the literature about the viability of using LTs to enhance evolutionary algorithms. Secondiy, we elaborate an explanation of Hybrid Aigonthms (HAs) performance in hard constrained optimization probierns. A HA based on mixing GAs and Hopfield neural network is proposed (HNN-GA) and its features studied by means of spectra modifica tions. Finaliy, the HNN-GA is applied to two optimization problems in telecommunications: The design of a Packet Radio Network transmission frarne, and the minimization of co-channel interference in satellite systems using channel reassignment. In both problems HNN-GA outperforms existing algonthms.