Abstract

A significant volume of data has been produced in this era. Therefore, accurately modeling these
data for further analysis and extraction of meaningful patterns is becoming a major concern in a
wide variety of real-life applications. Smart buildings are one of these areas urgently demanding
analysis of data. Managing the intelligent systems in smart homes, will reduce energy consumption
as well as enhance users’ comfort. In this context, Hidden Markov Model (HMM) as a learnable
finite stochastic model has consistently been a powerful tool for data modeling. Thus, we have been
motivated to propose occupancy estimation frameworks for smart buildings through HMM due to
the importance of indoor occupancy estimations in automating environmental settings. One of the
key factors in modeling data with HMM is the choice of the emission probability. In this thesis, we
have proposed novel HMMs extensions through Generalized Dirichlet (GD), Beta-Liouville (BL),
Inverted Dirichlet (ID), Generalized Inverted Dirichlet (GID), and Inverted Beta-Liouville (IBL)
distributions as emission probability distributions. These distributions have been investigated due
to their capabilities in modeling a variety of non-Gaussian data, overcoming the limited covariance
structures of other distributions such as the Dirichlet distribution. The next step after determining
the emission probability is estimating an optimized parameter of the distribution. Therefore, we
have developed a recursive parameter estimation based on maximum likelihood estimation approach
(MLE). Due to the linear complexity of the proposed recursive algorithm, the developed models can
successfully model real-time data, this allowed the models to be used in an extensive range of
practical applications.