Unsupervised Evolutionary Algorithm for Dynamic Bayesian Network Structure Learning
Jingguo Dai, Jia Ren · Lecture notes in computer science · 2015
The introduction of temporal dimension makes it difficult and complex to learn dynamic Bayesian network (DBN) structure for huge search space, hence many studies focus on some particular types of DBN, such as dynamic Naive Bayesian Classifier (DNBC). In order to overcome the limited applicability of DBN structure learning methods, this paper proposes an unsupervised evolutionary algorithm in which the selection of initial population has been implemented by means of mutual information to reduce the search space. Furthermore, in view of the poor performance of traditional encoding scheme and the recount of Bayesian information criterion (BIC) score when calculating the individual fitness, we provide a new structure representation without a necessity of the acyclicity test and an updating algorithm for BIC scores with the help of family inheritance to improve the efficiency. Simulations on synthetic data demonstrate that the proposed unsupervised evolutionary algorithm is effective in DBN structure learning.