Scalable Exact MAP Inference in Graphical Models

This paper presents parallel dovetailing in a distributed-memory environment for exact MAP inference in graphical models. Parallel dovetailing is a simple procedure which performs multiple searches in parallel with different parameter configurations. We evaluate empirically the performance of parallel dovetailing with three state-of-the-art AND/OR search algorithms in solving various MAP inference benchmarks. Our results clearly show that parallel dovetailing is effective, yielding considerable speedups and improving the solving abilities of these state-of-the-art baseline methods.