Improved Population-Based Incremental Learning of Bayesian Networks with partly known structure and parallel computing

Abstract

Bayesian Network is a frequently-used model for fault detection and diagnosis in industrial processes. In this article, some modifications are made to Population-Based Incremental Learning and the improved algorithm is applied to structure learning of Bayesian networks. A pre-training step with K2 algorithm is added to the Population-Based Incremental Learning process to obtain an initial probability vector. Then, an elitist strategy is introduced into this method, providing a better way to update the probability vector. Individuals generated in every iteration, and elites in history are utilized to update the vector. The nature of this method makes it possible to learn the bayesian network whose structure is partly known, for sometimes we can specify some parts of the structure with prior process knowledge. A benchmark network Alarm and an industrial process are provided for performance evaluation and comparisons. Furthermore, we parallelize the algorithm to make it more efficient to learn Bayesian Networks. The speed of Improved Population-Based Incremental Learning has been improved significantly after parallelization.