Canonical Probability Distributions for Model Building, Learning, and Inference

摘要

The performed project focused on three major issues: (1) development and application of parametric conditional probability distributions, (2) improvements of stochastic sampling algorithms based on importance sampling, and (3) practical applications of our general purpose decision modeling environment to diagnosis of complex systems. We have proposed a new class of parametric probability distributions, named Probabilistic Independence of Causal Interaction (pICI) models. We have shown that this class of models leads to improvements in learning of and inference in Bayesian networks. We have booked considerable advances in stochastic sampling algorithms for Bayesian networks based on importance sampling, preserved our leading role, and gained recognition of the community. Finally, we have developed a special module of GeNIe and SMILE(TM), the systems developed in the framework of our project, that supports diagnostic applications, and fielded the module in practical industrial settings.