A novel learning cloud Bayesian network for risk measurement

Abstract

Bayesian network (BN) is a popularly used approach for risk analysis. Because it is a graphic model being able to deal with randomness yet unable to model ambiguity, the fuzzy set theory is often combined with it to create a so-called fuzzy BN. Instead of using the classical fuzzy set theory, this paper intends to combine a normal Cloud model with the BN. In the normal Cloud model, an element belonging to a certain qualitative concept is not certain and precise as well. The Cloud BN is a generalization of the fuzzy BN. It is more adaptive for the uncertainty description of linguistic concepts, for example, the risks. Using the normal Cloud model, the following numerical characteristics of the variables can be estimated: the expectation, the dispersion degree compared with the expectation, and the dispersion degree of entropy. Consequently, the risk assessment contains a richer set of analytical information. Cloud BNs attract growing research interests. Compared to its precedents, the Cloud BN in this paper has a learning capability. Since the risk factors may have a combined effect, the causal relationships among the variables can be very complex, and hidden variables may exist. The learning mechanism allows for automatic structure discovery from data, giving rise to a dynamically evolving network. The proposed learning Cloud BN is able to represent the real risk situation better than its precedents. Its effectiveness and applicability are demonstrated by an illustrative case for risk prediction of the face instability in an underground tunnel construction project.