We review and extend the qualitative relationships about the informational relevance of variables in graphical decision models based on conditional independencies revealed through graphical separations of nodes from nodes representing utility on outcomes. We exploit these qualitative relationships to generate non-numerical graphical procedures for identifying partial orderings over chance variables in a decision model in terms of their informational relevance. We describe an efficient algorithm based on a consideration of local properties of a property we refer to as u-separation. Finally, we present results of computational efficiencies gained via the application of the new policies, based on analyses of sample networks with different degrees of connectivity.
Published Date: May 2001
Registration: ISBN 978-1-57735-133-7
Copyright: Published by The AAAI Press, Menlo Park, California.