Research on compound learning algorithm of Bayesian networks structures
Xiao Liu, Haijun Li · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2006
Today, there are more mature and relative perfect means of how to learn structures or parameters from completed data and learn parameters of fixed structure from uncompleted data. But it is a more difficult thing that learning structures of Bayesian Networks from uncompleted data. A compound learning algorithm is proposed; it combines the EM algorithm, Monte Carlo sampling algorithm and evolution algorithm together, uses EM algorithm to learn parameters of networks in uncompleted data, then samples the best network, converts the uncompleted data to completed data, and then evolves the structure using evolution algorithm. This algorithm could get over the defect of EM algorithm that frequently gains local maximum. Because data processing is based on posterior networks structures, structures of Bayesian Networks is optimizing and optimizing with evolution computing, the reliability of complementary data is higher. Learning rate is high and performance of this algorithm is good.