Bilkent University
Department of Computer Engineering


Trouble Shooting by Graphical Models


Prof. Hans-J. Lenz
Freie Universität

We consider problems in industry or in services where "products" or "services" are completed and delivered, but cause specific kinds of troubles at the customer's site during usage. Using Bayesian Belief Networks ("Causal Models") as a proper subclass of graphical models we introduce a real life example "Gas stove explosion caused by Company X", which is a good representative of this trouble shooting examples class. Then we recollect the basic definitions, the underlying model, and the famous Global Markov Property Theorem of DAGs, cf. Pearl, Wermuth, Spiegelhalter, Lauritzen et al., which substitutes global knowledge by a sufficient set of 'chunks' of partial knowledge. This idea is called "Local Computation" and is one of some few ways to fuse data in a mathematically sound way. Propagation, i.e. risk assessment in the sense of abductive reasoning, is enabled by the Junction Tree Algorithm, see Jordan (1998). This algorithm is known to be efficient for propagation in proper DAGs, but, of course, can become extremely inefficient for all kind of networks with high complexity (large number of arcs and nodes) etc. We close with risk assessment of the gas stove problem under various environmental conditions concerning the values of the main influential factors, cf. "scenarios". Note that in such cases there exist no data at hand, even when the first trouble cases will come in and are recorded! As all probability distributions – marginal as well as conditional ones – are fixed by fusing all hypothetical expert knowledge at hand inside the frame of discernment, we finally show some flair of sensitivity analysis on top of our What-if-Analyses ("scenarios"). Moreover, the report makes clear how to detect and locate the crucial trouble making ovens installed in the households.


DATE: 9 July, 2009, Thursday@ 13:40