Message Tracing and Model Debugging
Message Tracing
Models aren't the real world, they are imperfect descriptions of the
real world which are good enough for the purpose at hand. Often they
are good enough to make useful predictions, but sometimes they exhibit
surprising behavior. In those cases, we would like to isolate where
that behavior comes from. Models, like computer programs, often need
to be "Debugged".
In the Shafer and Shenoy algorithm, used by
Graphical-Belief, probabilities (or beliefs) are
calculated by passing messages through the junction tree.
Graphical-Belief allows you to select any node or
edge in the tree and see the messages passing through the
corresponding node in the junction tree. This has proved an
invaluable tool for debugging models. (Almond [1995a] provides a detailed
example in which an anomalous behavior is found to be cause by
counting the same evidence twice.)
Evidence Flows
Madigan, Mosurski and Almond [1996] suggest
showing evidence flow by coloring the arrows. In this scheme, the
inner width of the arrow represents the actual or expect weight of
evidence each node provides for the next from
that node, while the outer width represents the potential weight of
evidence (the maximum weight of evidence over all possible
instantiations of that node). The inner bar is colored red or blue
according to whether the evidence is positive or negative.
The figure above illustrates the difficulty with this scheme. Look
closely at the two evidence chains (in both cases X-5 supplies the
evidence and X-1 is the target) and try and determine where the
difference lies. As we felt that interpreting evidence chains would
require too much analyst training, we never fully incorporated them into
Graphical-Belief.
Where to go from here
If you have been faithfully following the linear path through this web
site, you have now reached the end of the example section. If you
still want more information about Graphical-Belief,
you can download some technical
reports or look into getting a beta test
license. For any questions which are not covered there, you can
send me email at almond@acm.org.
Thank you for your perserverance.
Explanation Return to the beginning of the
explanation examples.
Return to
the main example page.
Back to overview of Graphical-Belief.
View a list
of Graphical-Belief in publications and downloadable technical
reports.
The Graphical-Belief user
interface is implemented in Garnet.
Get more
information about obtaining Graphical-Belief (and why
it is not generally available).
get
the home page for Russell Almond , author
of Graphical-Belief.
Click
here to get to the home page for Insightful (the company that StatSci
has eventually evolved into).
Russell Almond, <lastname> (at) acm.org
Last modified: Tue Aug 20 11:39:02 1996