Beyond Static Mini-Bucket: Towards Integrating with Iterative Cost-Shifting Based Dynamic Heuristics

Abstract

We explore the use of iterative cost-shifting as a dynamic heuristic generator for solving MPE in graphical models via Branch and Bound. When mini-bucket elimination is limited by its memory budget, it may not provide good heuristics. This can happen often when the graphical model has a very high induced width with large variable domain sizes. In addition, we explore a hybrid setup where both MBE and the iterative cost-shifting bound are used in a combined heuristic. We compare these approaches with the most advanced statically generated heuristics.