Evaluating partition strategies for mini-bucket elimination

Mini-Bucket Elimination (MBE) is a well-known approximation algorithm for graphical models. It relies on a procedure to partition a set of funtions, called bucket, into smaller subsets, called mini-buckets. The impact of the partition process on the quality of the bound computed has never been investigated before. We take first steps to address this issue by presenting a framework within which partition strategies can be described, analyzed and compared. We derive a new class of partition heuristics from first-principles and demonstrate its impact on a number of benchmarks for probabilistic reasoning.