Using SMILE Wrappers

Diagnosis is one of the most successful applications of Bayesian networks. The ability of probabilistic knowledge representation techniques to perform a mixture of both predictive and diagnostic inference makes it very suitable for diagnosis. Bayesian networks can perform fusion of observations such as patient (or equipment) history and risk factors with symptoms and test results.

A Bayesian network built with SMILE represents various components of a system, possible faulty behaviors produced by the system (symptoms), along with results of possible diagnostic tests. The model essentially captures how possible defects of the system (whether it is a natural system, such as human body, or a human-made device, such as a car, an airplane, or a copier) can manifest themselves by error messages, symptoms, and test results.

To work with diagnosis, some nodes in the network should be designated as diagnostic faults, and some other nodes should become diagnostic observations. It is also possible to provide observation costs. The output of SMILE diagnostic algorithms is two-fold: (1) the posterior marginal probability of diagnostic faults, and (2) ranking of possible observations from the most to the least informative from the point of view of the current diagnostic focus. Depending on the selected diagnostic algorithm, the ranking is based on one of the following measures:

•an information-theoretic measure known as cross-entropy and expresses, for each observation node X individually, the expected reduction in entropy of the probability distribution over the specified subset of fault states after observing X. Cross-entropy is a utility-free measure of value of information and it gives a good idea about the value of the observations for diagnosing the disorder in question.

•a distance between two vectors representing the probabilities of pursued faults. SMILE can use multiple distance definitions, including Euclidean, cosine or cityblock.

Tutorial 9 uses a HeparII diagnostic model to show the diagnostic results for single- and multi-fault diagnosis.