**Shielding Against Conditioning Side-Effects in Graphical Models **

*by Mark Crowley *

Abstract:

In a directed graphical model it is often convenient to specify a prior distribution on a set of variables by adding a descendant variable that is conditioned on to induce the desired distribution. For example, by conditioning on an ``or'' of some variables, we can specify the distribution that one of the variables is true, but given that one is true, the others are independent and have low probability. Such conditioning is an easier way to specify some distributions than other methods and arises very naturally for those familiar with directed models.

Unfortunately, there are side effects of this method that may be undesirable, such as tying together the distributions of ancestors that should remain independent. In this talk I will outline this problem and present the result of my master's thesis which a method to counteract these side-effects while maintaining the useful properties of conditioning and avoiding an increase in complexity of the inference algorithm.