Incomplete Relationship in Sampling Inference Algorithm of Dynamic Uncertain Causality Graph
Hao Nie, Qin Zhang · 2021 3rd International Conference on Applied Machine Learning (ICAML) · 2021
Dynamic Uncertain Causality Graph (DUCG) is a new generation of probability graph model (PGM). It uses virtual random events and casual relationship intensity to represent the uncertain causality, instead of the conditional probability tables (CPTs), so combination explosion problems of CPTs can be avoided in DUCG. Systems based on DUCG have been deployed in many areas such as nuclear power plants, spacecrafts and medical diagnoses. A basic sampling algorithm based on recursive reasoning method and cut-off estimation has been proposed to solve the combination explosion problem in DUCG inference process. In this paper, the sampling algorithm is adapted to support the incomplete expression, which is an importance characterize for DUCG. Because incompleteness expression can reduce the complexity of both construction and inference by ignoring some states of variables. Two new detailed schemes are proposed to provide this new feature. The first one is based on the numerical approximation and the second one is based on the mechanism of DUCG. A practical example based on practical viral hepatitis C is presented to verify the discussions.