Finding Most Likely Solutions

Paper in proceeding, 2007

As a framewrok for simple but basic statistical inference problems we introduce a genetic Most Likely Solution problem, a task of finding a most likely solution (MLS in short) for a given problem instance under some given probability model. Although many MLS problems are NP-hard, we propose, for these problems, to study their average-case complexity under their assumed probability models. We show three examples of MLS problems, and explain that “message passing algorithms” (e.g., belief propagation) work reasonably well for these problems. Some of the technical results of this paper are from the author’s recent work [WY06, OW06].