Particle Swarm Optimisation for learning Bayesian Networks

Abstract

This paper discusses the potential of Particle Swarm Optimisation (PSO) for inducing Bayesian Networks (BNs). Specifically, we detail two methods which adopt the search and score approach to BN learning. The two algorithms are similar in that they both use PSO as the search algorithm, and the K2 metric to score the resulting network. The difference lies in the way networks are constructed. The CONstruct And Repair (CONAR) algorithm generates structures, validates, and repairs if required, and the REstricted STructure (REST) algorithm, only permits valid structures to be developed. Initial experiments indicate that these approaches produce promising results when compared to other BN learning strategies.