Machine learning algorithms for pattern visualization in classification tasks and for automatic indoor temperature prediction

Abstract

This thesis explores aspects in the field of machine learning, and specifically of pattern classification and regression or function approximation. Although there are many methods of classification for multi-dimensional patterns, in general, they all behave like "black boxes" where the explanation of their operation is difficult or impossible. This thesis develops methods of reducing the dimensionality of data to project multi-dimensional classification problems over a two-dimensional space (a plane). The classifiers can thus be used to learn the projected data and to create two-dimensional maps of classification problems whose graphic nature makes intuitive and easy to understand, helping to explain the classification problem. After a review of the existing techniques for dimensionality reduction, several methods are proposed to project the multidimensional data on the plane, minimizing the overlap between classes. These methods allow to project new patterns not used during the projection learning process. Eight types of linear, quadratic and polynomial projections are proposed and combined with four overlapping measures between classes. These projections are compared with another 34 dimensionality reduction methods existing in the literature on a wide collection of 71 benchmark classification problems. The best results have been obtained by the Polynomial Kernel Discriminant Analysis of degree 2 (PKDA2), which creates visual and selfexplanatory maps of the classification problems on which a reference classifier (the support vector machine, or SVM) fails only slightly less than on the original multi-dimensional data. A web interface and a local standalone application are also provided, developed using the PHP and Matlab programming languages, respectively, which allow to apply these projections in order to visualize the 2D maps of any classification problem. In the scope of regression, a wide collection of regressors has been applied for the automatic prediction of temperatures in air conditioning systems (HVAC). These systems have a direct impact on both energy consumption and the comfort of buildings, so an accurate and reliable modelling of the temperature behavior constitutes the starting point for the development of energy efficiency plans. The use of regressors to predict the evolution of indoor temperature of buildings based on internal and external (climatic) conditions allows to evaluate the impact of the modifications in the HVAC systems from a comfort perspective. With the aim of developing an efficient model for HVAC systems, this thesis has evaluated 40 regressors, which belong to 20 different regressor families, using real data generated by an intelligent building, namely the Centro Singular de Investigación en Tecnoloxías da Información (CiTIUS) of the University of Santiago de Compostela (USC). In addition, different models based on neural networks which allow automatic re-training and on-line learning of new data have been developed and compared to the previous 20 off-line regressors. The ability of on-line learning provides robustness to the neural models and allows them to: 1) face circumstances never seen in training due to exceptional climatic situations; and 2) support alterations in the components of the systems produced by errors or changes in the sensor systems.