Comparing approaches for probabilistic model-based diagnosis

Marcos L. P. Bueno1,2, Arjen Hommersom1,3, Peter J. F. Lucas1,4, Pedro Pereira Rodrigues1,5

1Radboud University Nijmegen (NL), 2Faculdade de Medicina, CINTESIS, Universidade do Porto (PT),  3Open University (NL), 4Leiden University (NL),  5MEDCIDS, Faculdade de Medicina, Universidade do Porto (PT)

Objective: there are several methods for diagnosing faults in consistency-based diagnosis with uncertainties.
Comparisons of their outcomes is scarce in the literature, and often a non-trivial task. In many scenarios only partial observations are available, which makes the problem even more challenging.

Methods: we consider three approaches for diagnosing faulty components. The first one is the conflict measure, which measures how strong inputs and outputs are correlated by means of comparing the chances of seeing inputs isolated and outputs isolated with the chances of seeing inputs and outputs together. The second one is the likelihood of the observations, which corresponds to the denominator of the conflict measure. Finally, we consider a new conflict measure, which we call expected conflict, that merges ideas from the likelihood with expected outputs in order to provide an informative normalization for better guiding the decision on conflicting observations.

Results: preliminary results indicate that the conflict measure is unable to satisfactory deal with some situations where there are missing observations. In particular, components that are likely faulty might not be detected, which can be caused by the lack of sensitiveness to priors on inputs. The likelihood, in turn, seems to suffer from the problem of deciding when a low probability is not conflicting. From our preliminary examples, it appears that the expected conflict can solve the problems elicited so far to a satisfactory extent.

Discussion: the suitability of different measures is also dependent on whether the probabilistic parameters are correct.

keywords: diagnosis, probabilistic reasoning, Bayesian networks

Poster: Comparing approaches for probabilistic model-based diagnosis