Adaptive Online Learning of Bayesian Network Parameters
Ira L. Cohen, Alexandre Bronstein, Fábio Gagliardi Cozman · 2001
The paper introduces Voting EM, an adaptive online learning algorithm of Bayesian network parameters. Voting EM is an extension of the EM( ) algorithm suggested by [1]. We show convergence properties of the Voting EM that uses a constant learning rate. We use the convergence properties to formulate an error driven scheme for adapting the learning rate. The resultant algorithm converges with the optimal rate of near a maximum while retaining the ability to increase the learning rate in the vicinity of a local maximum or due to changes in the modelled environment.