Using Belief Change Principles for Evolving Bayesian Network Structures in Probabilistic Knowledge Representations

Abstract

Belief change in Probabilistic Graphical Models in general, and Bayesian Networks in particular, is often thought of as change in the model parameters when data consistent with the graphical model is observed. The assumption is the network structure for the graphical model is a true representation of the knowledge about the domain and therefore it does not change. In dynamic environments, this assumption is not always true. The network structure is bound to change in response to changes in the domain or correction of mistaken propositions. In such domains, the true Bayesian Network structure at any given point in time, and the events that provides an impetus for change in the network structure are unobservable and are not known with certainty. This paper presents, the Unified Belief Change Operator for Bayesian Networks (UBCOBaN). The UBCOBaN effects both belief revision and update on a given Bayesian network structure based on the data emitted from the domain modelled by the Bayesian Network. We present the conceptualization and implementation of the operator, and its evaluation based on synthetic data simulated from the Alarm Network. The operator was found to be more rational, with respect to the principle minimal change, than the classical search-and-score algorithm. The operator was also found to be faster in adapting to necessary changes than the classical search-and-score algorithm.