Swarm-Based Descent, Efficiency, and Posterior Accuracy with Compressed NAT-Modelled Bayesian Networks

Abstract

Non-Impeding Noisy AND trees (NATs) reduce the space to model Bayesian networks (BNs) to be linear on the number of parents, by NAT modelling a BN. We present an algorithm based on swarm techniques to speed up the compression of a BN into a NAT modelled BN. This algorithm has proven to allow for significant speed up in compression. Next, we investigated the inference performance of multiplicative factorized (MF) NAT modelled BNs for a range of sparse BNs. We show that MF-NAT modelled BNs allow for significant speed up in inference for a range of sparse BN structures. Lastly, an empirical study is performed on the inference accuracy of the compressed NAT modelled BNs. The study concluded that the posterior probabilities from inference with NAT modelled BNs had a better accuracy than the NAT modelled CPTs, which shows that compression errors were attenuated and not amplified.