Learning Canonical Models from Data

Probabilistic reasoning, among methodologies used within the domain of artificial intelligence, is recognized as one of the most promising foundations for machine learning techniques. Bayesian networks – one of the examples of a probabilistic modeling techniques are widely acclaimed and used by both industrial and scientific communities. Apparatus for extracting information from ever–increasing amounts of data, was one of the key factors for the successes of companies like Google, IBM or Facebook. Nowadays, empirical data for many problem domains are freely accessible and, in many cases, growing at an exponential rate. There are fields of research, however, that are not privileged with an immensity of available data. This document addresses the problem of eliciting probabilistic information from small samples of data. Areas where acquisition of empirical data is either costly or difficult, may benefit from any advancement in that regard. We will focus on learning parameters for the so–called canonical models of Bayesian networks as, in our opinion, currently available methods leave room for improvements