Open Collections

Shielding against conditioning side effects in graphical models

University of British Columbia

When modelling uncertain beliefs with graphical models we are often presented with "natural" distributions that are hard to specify. An example is a distribution of which instructor is teaching a course when we know that someone must teach it. Such distributions over a set of nodes can be easily described if we condition on a child of these nodes as part of the specification. This conditioning is not an observation of a variable in the real world but by fixing the value of the node, existing inference algorithms perform the calculations needed to achieve the desired distribution automatically. Unfortunately, although it achieves this goal it has side effects that we claim are undesirable. These side effects create dependencies between other variables in the model. This can lead to different beliefs throughout the model, including the constrained variables, than would otherwise be expected if the constraint is meant to be local in its effect. We describe the use of conditioning for these types of distributions and illuminate the problem of side effects, which have received little attention in the literature. We then present a method that still allows specification of these distributions easily using conditioning but counterbalancing side effects by adding other nodes to the network.

Text

2009-12-11

For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use.

http://hdl.handle.net/2429/16514

Computer Science

University of British Columbia

UBCV

DSpace