Links
Other Project sites
Source code and Task tracking repository
Team Foundation Server project documents
TME's Smart Diagnostics site (obsolete)
Related Intel sites
Distributed Detection & Inference
External sites
Web application
This screen shot shows the start page for the Smart Diagnostics web application, partially hidden behind a diagram of a model built with the application. Sandwiched in between, in yellow and hardly visible is the link editor pop-up. The pop-up allows sophisticated users to edit the causal links created by observing the user's workflow as they solve a problem. The model diagram is not shown to the user.
Project
Smart Diagnostics is an approach to solving diagnostic problems that is based on the concept of query-based diagnostics. It creates a diagnostic Bayes network by interpreting the sequence of actions of the user's work-flow while they are solving a problem, effectively interleaving diagnostic reasoning actions and model elicitation. Additionally as an editing feature, the user can both add and remove variables and causal links, freeing the user to employ their own conceptual model. When new variables are created, the user can do so in their current vocabulary, taking suggestions from existing variable names, to avoid possible repetition of terms. By marrying a work-flow structured application with a diagnostic Bayes network based on capturing the user's causal understanding, the system allows for continuous refinement of the model.
As knowledge-engineering practitioners are well aware, elicitation of causal relationships and model structure places large cognitive demands on domain experts. Part of the challenge is the accessibility of knowledge; during an elicitation session, the domain expert is faced with a set of hypothetical circumstances, that often consist of questions about rare occurrences (since descriptions of common occurrences can often be derived from data). One would like to elicit this knowledge when it is current with the activities of the expert. Why not gather this knowledge at the time it is used? During diagnosis the expert is immersed in the problem and the causal relationships have a cognitive immediacy not available generally.
A primary intent of building the application around work-flow is to hide the Bayes network from the user. In doing so we make several simplifying assumptions on model structure: The cause and observable variables form a bi-partite network. Provision is made for contextual variables that condition the causes; however this feature is not implemented in the current version. All variables are dichotomous, with states corresponding to “normal” and “abnormal”. We believe that these constraints make the right tradeoff for this domain between complexity and versatility.
Contact
John Mark Agosta RNB6–61
408.765–0429
john.m.agosta at intel.com
John Mark's Homepage
Bryan Pollard HD1–442
508.612–3165
bryan.pollard at intel.com
Matthias H. Giessler
480.552–7781
matthias.h.giessler at intel.com
Thomas R. Gardos HD2–330
978.553–6200
thomas.r.gardos at intel.com
Courtesy BlueWave, Open Web Design
Thanks to Dubai Hotels