A Model Selection Algorithm of Influence Diagrams Based on Structural Decomposition
Yao Hong · 2007
The data dependency, computation complexity and non-probability relation problems are faced by model selection of Influence diagrams. Based on the decomposition of Influence diagrams,a PS-EM algorithm is presented for learning the probability structures of IDs, and a BP Neural Network is introduced to use learning local utility functions of the utility part. The SEM algorithm is improved by PS-EM algorithm. PS-EM algorithm presents a new MDL scoring which includes the prior knowledge of network structures for reducing the dependency on data, and learning parameters and scoring structure are separated for improving the computation efficiency. On the oil wildcatter model, the experiment results show that PS-EM algorithm is better the time performance and less the data dependency than criterion SEM algorithm, and the model selection of the utility part is easy to achieve.