Structure Learning in Bayesian Networks with Parent Divorcing
Ulrich von Waldow, Florian Röhrbein · Conference Cognitive Science · 2015
Bayesian networks (BNs) are an essential tool for the modeling of cognitive processes. They represent probabilistic knowledge in an intuitive way and allow to draw inferences based on current evidence and built-in hypotheses. In this paper, a structure learning scheme for BNs will be examined that is based on so-called Child-friendly Parent Divorcing (CfPD). This algorithm groups together nodes with similar properties by adding a new node to the existing network. The updating of all affected probabilities is formulated as an optimization problem. The resulting procedure reduces the size of the conditional probability tables (CPT) significantly and hence improves the efficiency of the network, making it suitable for larger networks typically encountered in cognitive modelling.