Fusión de redes Bayesianas Gaussianas

Las redes Bayesianas constituyen un modelo ampliamente utilizado para la representación de relaciones de dependencia condicional en datos multivariantes. Su aprendizaje a partir de un conjunto de datos o expertos ha sido estudiado profundamente desde su concepción. Sin embargo, en determinados escenarios se demanda la obtención de un modelo común asociado a particiones de datos o conjuntos de expertos. En este caso, se trata el problema de fusión o agregación de modelos. Los trabajos y resultados en agregación de redes Bayesianas son de naturaleza variada, aunque escasos en comparación con aquellos de aprendizaje. En este documento, se proponen dos métodos para la agregación de redes Gaussianas, definidas como aquellas redes Bayesianas que modelan una distribución Gaussiana multivariante. Los métodos presentados son efectivos, precisos y producen redes con menor cantidad de parámetros en comparación con los modelos obtenidos individualmente. Además, constituyen un enfoque novedoso al incorporar nociones exploradas tradicionalmente por separado en el estado del arte. Futuras aplicaciones en entornos escalables hacen dichos métodos especialmente atractivos, dada su simplicidad y la ganancia en compacidad de la representación obtenida.---ABSTRACT---Bayesian networks are a widely used model for the representation of conditional dependence relationships among variables in multivariate data. The task of learning them from a data set or experts has been deeply studied since their conception. However, situations emerge where there is a need of obtaining a consensuated model from several data partitions or a set of experts. This situation is referred to as model fusion or aggregation. Results about Bayesian network aggregation, although rich in variety, have been scarce when compared to the learning task. In this context, two methods are proposed for the aggregation of Gaussian Bayesian networks, that is, Bayesian networks whose underlying modelled distribution is a multivariate Gaussian. Both methods are effective, precise and produce networks with fewer parameters in comparison with the models obtained by individual learning. They constitute a novel approach given that they incorporate notions traditionally explored separately in the state of the art. Future applications in scalable computer environments make such models specially attractive, given their simplicity and the gaining in sparsity of the produced model.

A method to incorporate the effect of taxonomic uncertainty on multivariate analyses of ecological data

Researchers in ecology commonly use multivariate analyses (e.g. redundancy analysis, canonical correspondence analysis, Mantel correlation, multivariate analysis of variance) to interpret patterns in biological data and relate these patterns to environmental predictors. There has been, however, little recognition of the errors associated with biological data and the influence that these may have on predictions derived from ecological hypotheses. We present a permutational method that assesses the effects of taxonomic uncertainty on the multivariate analyses typically used in the analysis of ecological data. The procedure is based on iterative randomizations that randomly re-assign non identified species in each site to any of the other species found in the remaining sites. After each re-assignment of species identities, the multivariate method at stake is run and a parameter of interest is calculated. Consequently, one can estimate a range of plausible values for the parameter of interest under different scenarios of re-assigned species identities. We demonstrate the use of our approach in the calculation of two parameters with an example involving tropical tree species from western Amazonia: 1) the Mantel correlation between compositional similarity and environmental distances between pairs of sites, and; 2) the variance explained by environmental predictors in redundancy analysis (RDA). We also investigated the effects of increasing taxonomic uncertainty (i.e. number of unidentified species), and the taxonomic resolution at which morphospecies are determined (genus-resolution, family-resolution, or fully undetermined species) on the uncertainty range of these parameters. To achieve this, we performed simulations on a tree dataset from southern Mexico by randomly selecting a portion of the species contained in the dataset and classifying them as unidentified at each level of decreasing taxonomic resolution. An analysis of covariance showed that both taxonomic uncertainty and resolution significantly influence the uncertainty range of the resulting parameters. Increasing taxonomic uncertainty expands our uncertainty of the parameters estimated both in the Mantel test and RDA. The effects of increasing taxonomic resolution, however, are not as evident. The method presented in this study improves the traditional approaches to study compositional change in ecological communities by accounting for some of the uncertainty inherent to biological data. We hope that this approach can be routinely used to estimate any parameter of interest obtained from compositional data tables when faced with taxonomic uncertainty.

Identifying wind power ramp causes from multivariate datasets: a methodological proposal and its application to reanalysis data.

Forecasting abrupt variations in wind power generation (the so-called ramps) helps achieve large scale wind power integration. One of the main issues to be confronted when addressing wind power ramp forecasting is the way in which relevant information is identified from large datasets to optimally feed forecasting models. To this end, an innovative methodology oriented to systematically relate multivariate datasets to ramp events is presented. The methodology comprises two stages: the identification of relevant features in the data and the assessment of the dependence between these features and ramp occurrence. As a test case, the proposed methodology was employed to explore the relationships between atmospheric dynamics at the global/synoptic scales and ramp events experienced in two wind farms located in Spain. The achieved results suggested different connection degrees between these atmospheric scales and ramp occurrence. For one of the wind farms, it was found that ramp events could be partly explained from regional circulations and zonal pressure gradients. To perform a comprehensive analysis of ramp underlying causes, the proposed methodology could be applied to datasets related to other stages of the wind-topower conversion chain.

Optimizing Functional Network Representation of Multivariate Time Series

By combining complex network theory and data mining techniques, we provide objective criteria for optimization of the functional network representation of generic multivariate time series. In particular, we propose a method for the principled selection of the threshold value for functional network reconstruction from raw data, and for proper identification of the network's indicators that unveil the most discriminative information on the system for classification purposes. We illustrate our method by analysing networks of functional brain activity of healthy subjects, and patients suffering from Mild Cognitive Impairment, an intermediate stage between the expected cognitive decline of normal aging and the more pronounced decline of dementia. We discuss extensions of the scope of the proposed methodology to network engineering purposes, and to other data mining tasks.

Spatial variability patterns of some Vertisol properties at a field scale using standardized data

Spatial variability of Vertisol properties is relevant for identifying those zones with physical degradation. In this sense, one has to face the problem of identifying the origin and distribution of spatial variability patterns. The objectives of the present work were (i) to quantify the spatial structure of different physical properties collected from a Vertisol, (ii) to search for potential correlations between different spatial patterns and (iii) to identify relevant components through multivariate spatial analysis. The study was conducted on a Vertisol (Typic Hapludert) dedicated to sugarcane (Saccharum officinarum L.) production during the last sixty years. We used six soil properties collected from a squared grid (225 points) (penetrometer resistance (PR), total porosity, fragmentation dimension (Df), vertical electrical conductivity (ECv), horizontal electrical conductivity (ECh) and soil water content (WC)). All the original data sets were z-transformed before geostatistical analysis. Three different types of semivariogram models were necessary for fitting individual experimental semivariograms. This suggests the different natures of spatial variability patterns. Soil water content rendered the largest nugget effect (C0 = 0.933) while soil total porosity showed the largest range of spatial correlation (A = 43.92 m). The bivariate geostatistical analysis also rendered significant cross-semivariance between different paired soil properties. However, four different semivariogram models were required in that case. This indicates an underlying co-regionalization between different soil properties, which is of interest for delineating management zones within sugarcane fields. Cross-semivariograms showed larger correlation ranges than individual, univariate, semivariograms (A ≥ 29 m). All the findings were supported by multivariate spatial analysis, which showed the influence of soil tillage operations, harvesting machinery and irrigation water distribution on the status of the investigated area.

MBIS: Multivariate Bayesian Image Segmentation Tool

We present MBIS (Multivariate Bayesian Image Segmentation tool), a clustering tool based on the mixture of multivariate normal distributions model. MBIS supports multi-channel bias field correction based on a B-spline model. A second methodological novelty is the inclusion of graph-cuts optimization for the stationary anisotropic hidden Markov random field model. Along with MBIS, we release an evaluation framework that contains three different experiments on multi-site data. We first validate the accuracy of segmentation and the estimated bias field for each channel. MBIS outperforms a widely used segmentation tool in a cross-comparison evaluation. The second experiment demonstrates the robustness of results on atlas-free segmentation of two image sets from scan-rescan protocols on 21 healthy subjects. Multivariate segmentation is more replicable than the monospectral counterpart on T1-weighted images. Finally, we provide a third experiment to illustrate how MBIS can be used in a large-scale study of tissue volume change with increasing age in 584 healthy subjects. This last result is meaningful as multivariate segmentation performs robustly without the need for prior knowledge.

Proactive and reactive thermal aware optimization techniques to minimize the environmental impact of data centers

Los Centros de Datos se encuentran actualmente en cualquier sector de la economía mundial. Están compuestos por miles de servidores, dando servicio a los usuarios de forma global, las 24 horas del día y los 365 días del año. Durante los últimos años, las aplicaciones del ámbito de la e-Ciencia, como la e-Salud o las Ciudades Inteligentes han experimentado un desarrollo muy significativo. La necesidad de manejar de forma eficiente las necesidades de cómputo de aplicaciones de nueva generación, junto con la creciente demanda de recursos en aplicaciones tradicionales, han facilitado el rápido crecimiento y la proliferación de los Centros de Datos. El principal inconveniente de este aumento de capacidad ha sido el rápido y dramático incremento del consumo energético de estas infraestructuras. En 2010, la factura eléctrica de los Centros de Datos representaba el 1.3% del consumo eléctrico mundial. Sólo en el año 2012, el consumo de potencia de los Centros de Datos creció un 63%, alcanzando los 38GW. En 2013 se estimó un crecimiento de otro 17%, hasta llegar a los 43GW. Además, los Centros de Datos son responsables de más del 2% del total de emisiones de dióxido de carbono a la atmósfera. Esta tesis doctoral se enfrenta al problema energético proponiendo técnicas proactivas y reactivas conscientes de la temperatura y de la energía, que contribuyen a tener Centros de Datos más eficientes. Este trabajo desarrolla modelos de energía y utiliza el conocimiento sobre la demanda energética de la carga de trabajo a ejecutar y de los recursos de computación y refrigeración del Centro de Datos para optimizar el consumo. Además, los Centros de Datos son considerados como un elemento crucial dentro del marco de la aplicación ejecutada, optimizando no sólo el consumo del Centro de Datos sino el consumo energético global de la aplicación. Los principales componentes del consumo en los Centros de Datos son la potencia de computación utilizada por los equipos de IT, y la refrigeración necesaria para mantener los servidores dentro de un rango de temperatura de trabajo que asegure su correcto funcionamiento. Debido a la relación cúbica entre la velocidad de los ventiladores y el consumo de los mismos, las soluciones basadas en el sobre-aprovisionamiento de aire frío al servidor generalmente tienen como resultado ineficiencias energéticas. Por otro lado, temperaturas más elevadas en el procesador llevan a un consumo de fugas mayor, debido a la relación exponencial del consumo de fugas con la temperatura. Además, las características de la carga de trabajo y las políticas de asignación de recursos tienen un impacto importante en los balances entre corriente de fugas y consumo de refrigeración. La primera gran contribución de este trabajo es el desarrollo de modelos de potencia y temperatura que permiten describes estos balances entre corriente de fugas y refrigeración; así como la propuesta de estrategias para minimizar el consumo del servidor por medio de la asignación conjunta de refrigeración y carga desde una perspectiva multivariable. Cuando escalamos a nivel del Centro de Datos, observamos un comportamiento similar en términos del balance entre corrientes de fugas y refrigeración. Conforme aumenta la temperatura de la sala, mejora la eficiencia de la refrigeración. Sin embargo, este incremente de la temperatura de sala provoca un aumento en la temperatura de la CPU y, por tanto, también del consumo de fugas. Además, la dinámica de la sala tiene un comportamiento muy desigual, no equilibrado, debido a la asignación de carga y a la heterogeneidad en el equipamiento de IT. La segunda contribución de esta tesis es la propuesta de técnicas de asigación conscientes de la temperatura y heterogeneidad que permiten optimizar conjuntamente la asignación de tareas y refrigeración a los servidores. Estas estrategias necesitan estar respaldadas por modelos flexibles, que puedan trabajar en tiempo real, para describir el sistema desde un nivel de abstracción alto. Dentro del ámbito de las aplicaciones de nueva generación, las decisiones tomadas en el nivel de aplicación pueden tener un impacto dramático en el consumo energético de niveles de abstracción menores, como por ejemplo, en el Centro de Datos. Es importante considerar las relaciones entre todos los agentes computacionales implicados en el problema, de forma que puedan cooperar para conseguir el objetivo común de reducir el coste energético global del sistema. La tercera contribución de esta tesis es el desarrollo de optimizaciones energéticas para la aplicación global por medio de la evaluación de los costes de ejecutar parte del procesado necesario en otros niveles de abstracción, que van desde los nodos hasta el Centro de Datos, por medio de técnicas de balanceo de carga. Como resumen, el trabajo presentado en esta tesis lleva a cabo contribuciones en el modelado y optimización consciente del consumo por fugas y la refrigeración de servidores; el modelado de los Centros de Datos y el desarrollo de políticas de asignación conscientes de la heterogeneidad; y desarrolla mecanismos para la optimización energética de aplicaciones de nueva generación desde varios niveles de abstracción. ABSTRACT Data centers are easily found in every sector of the worldwide economy. They consist of tens of thousands of servers, serving millions of users globally and 24-7. In the last years, e-Science applications such e-Health or Smart Cities have experienced a significant development. The need to deal efficiently with the computational needs of next-generation applications together with the increasing demand for higher resources in traditional applications has facilitated the rapid proliferation and growing of data centers. A drawback to this capacity growth has been the rapid increase of the energy consumption of these facilities. In 2010, data center electricity represented 1.3% of all the electricity use in the world. In year 2012 alone, global data center power demand grew 63% to 38GW. A further rise of 17% to 43GW was estimated in 2013. Moreover, data centers are responsible for more than 2% of total carbon dioxide emissions. This PhD Thesis addresses the energy challenge by proposing proactive and reactive thermal and energy-aware optimization techniques that contribute to place data centers on a more scalable curve. This work develops energy models and uses the knowledge about the energy demand of the workload to be executed and the computational and cooling resources available at data center to optimize energy consumption. Moreover, data centers are considered as a crucial element within their application framework, optimizing not only the energy consumption of the facility, but the global energy consumption of the application. The main contributors to the energy consumption in a data center are the computing power drawn by IT equipment and the cooling power needed to keep the servers within a certain temperature range that ensures safe operation. Because of the cubic relation of fan power with fan speed, solutions based on over-provisioning cold air into the server usually lead to inefficiencies. On the other hand, higher chip temperatures lead to higher leakage power because of the exponential dependence of leakage on temperature. Moreover, workload characteristics as well as allocation policies also have an important impact on the leakage-cooling tradeoffs. The first key contribution of this work is the development of power and temperature models that accurately describe the leakage-cooling tradeoffs at the server level, and the proposal of strategies to minimize server energy via joint cooling and workload management from a multivariate perspective. When scaling to the data center level, a similar behavior in terms of leakage-temperature tradeoffs can be observed. As room temperature raises, the efficiency of data room cooling units improves. However, as we increase room temperature, CPU temperature raises and so does leakage power. Moreover, the thermal dynamics of a data room exhibit unbalanced patterns due to both the workload allocation and the heterogeneity of computing equipment. The second main contribution is the proposal of thermal- and heterogeneity-aware workload management techniques that jointly optimize the allocation of computation and cooling to servers. These strategies need to be backed up by flexible room level models, able to work on runtime, that describe the system from a high level perspective. Within the framework of next-generation applications, decisions taken at this scope can have a dramatical impact on the energy consumption of lower abstraction levels, i.e. the data center facility. It is important to consider the relationships between all the computational agents involved in the problem, so that they can cooperate to achieve the common goal of reducing energy in the overall system. The third main contribution is the energy optimization of the overall application by evaluating the energy costs of performing part of the processing in any of the different abstraction layers, from the node to the data center, via workload management and off-loading techniques. In summary, the work presented in this PhD Thesis, makes contributions on leakage and cooling aware server modeling and optimization, data center thermal modeling and heterogeneityaware data center resource allocation, and develops mechanisms for the energy optimization for next-generation applications from a multi-layer perspective.