Abstract

Bayesian networks are a successful modeling tool that have become very pop- ular in the last 20 years. Software for constructing models is widely available, but software that combines data and expert knowledge in a principled way to construct networks is rare. We aim on solving this problem by extending GeNIe and SMILE to give the users the possibility to use this feature. To accomplish this we divided the problem into three direrent parts. Firstly, we created a case management system in both GeNIe and SMILE that manages the data of the user, we will call this cases. Managing cases involves editing evidence and target nodes in the network, and more. We take a novel approach of storing the cases in the same file the network is stored (usually data is stored in separate files), to easy the user and to keep the cases consistent with the network.
Secondly, we will use these cases to refine a network created by an expert by applying the Expectation-Maximization (EM) learning algorithm.
We will introduce new canonical gates that, based on some assumptions, require a lower number of parameters to be specified and, consequently, less data is needed to learn these gates than was needed in case of CPTs. Another advantage of these gates is that they have the property that inference algorithms can exploit them. Since learning requires a lot of inference, these gates are also suitable for learning from very large datasets.