Online Learning of Bayesian Network Parameters

[摘要] The paper introduces Voting EM, an online learning algorithm of Bayesian network parameters that builds on the EM(n) algorithm suggested by (Bauer et al., 1997). We prove convergence properties of the algorithm in the mean and variance, and demonstrate the algorithm's behavior on synthetic data. We show the relationship between Maximum-Likelihood (ML) counting and Voting EM. We demonstrate that Voting EM is able to adapt to changes in the modelled environment and to escape local maxima of the likelihood function. Voting EM also handles both the complete and missing data cases. We use the convergence properties to further improve Voting EM by automatically adapting the learning rate n. The resultant enhanced Voting EM algorithm converges more quickly and more closely to the true CPT parameters; further, it adapts more rapidly to changes in the modelled environment. 8 Pages

[发布机构] HP Development Company

[学科分类] 计算机科学（综合）