Shareable and Inheritable Incremental Compilation in iOOBN

Abstract

Object-oriented Bayesian networks (OOBNs) allow modellers to construct compositional and hierarchical models, using an inheritance hierarchy of classes ad subclasses, enabling reuse and supporting maintenance. Reasoning with both ordinary Bayesian networks (BNs) and OOBNs requires the important computational task of inference, the computing of new posterior probability distributions given a set of evidence. A widely used inference technique in ordinary BNs involves compiling the BN into a so-called junction tree (JT) before performing the inference; the compilation step is only performed when the model changes. In current OOBN software, the OOBN is first transformed into the underlying BN, so-called flattening, then the standard inference is performed. Researchers have proposed methods for incremental compilation of BNs, rather than recompiling from scratch for each network modification; these can apply to OOBNs also after flattening. Here, we propose a new incremental compilation technique that reuses existing compiled JTs of both embedded components and superclasses, and does not require flattening. We demonstrate through experimental analysis that this can reduce compilation time, and produces compact JTs that are cost-effective for inference.