Abstract

Choosing a Bayesian network which best explains the dependencies in a dataset is an important problem in scientific areas including medicine, epidemiology, genetic analysis and domino effect analysis. However, inference using Bayesian belief networks is an NP-hard problem, and inference is considered computationally demanding even for feasible problems. With these conditions, it is generally hoped that a small army of algorithms can be developed which will find almost optimal solutions, and these are then all applied to a given problem and the best result taken. We therefore aim to provide an alternative family of methods using continuous evolutionary global optimisation techniques, and quantify its effectiveness compared with the already available hill-climbing method. Various implementations are considered, and it is found that combining available continuous global optimisation packages with the pre-existing score function is computationally unfeasible for even small networks. Potential reasons are discussed, and improvements proposed using custom-built methods, based on the effectiveness of the existing hill climbing algorithm.

The full PDF can be downloaded here, and the presentation PDF can be downloaded here.