TR2011-087

Message-Passing Algorithms for Optimization and Inference: Belief Propagation and Divide & Concur


    •  Yedidia, J.S., "Message-Passing Algorithms for Optimization and Inference: Belief Propagation and Divide & Concur", Journal of Statistical Physics, DOI: 10.1007/​s10955-011-0384-7, Vol. 145, No. 4, pp. 860-890, November 2011.
      BibTeX TR2011-087 PDF
      • @article{Yedidia2011oct,
      • author = {Yedidia, J.S.},
      • title = {Message-Passing Algorithms for Optimization and Inference: Belief Propagation and Divide & Concur},
      • journal = {Journal of Statistical Physics},
      • year = 2011,
      • volume = 145,
      • number = 4,
      • pages = {860--890},
      • month = oct,
      • doi = {10.1007/s10955-011-0384-7},
      • issn = {0022-4715},
      • url = {https://www.merl.com/publications/TR2011-087}
      • }
Abstract:

Message-passing algorithms can solve a wide variety of optimization, inference, and constraint satisfaction problems. The algorithms operate on factor graphs that visually represent and specify the structure of the problems. After describing some of their applications, I survey the family of belief propagation (BP) algorithms, beginning with a detailed description of the min-sum algorithm and its exactness on tree factor graphs, and then turning to a variety of more sophisticated BP algorithms, including free-energy based BP algorithms, "splitting" BP algorithms that generalize "tree-reweighted" BP, and the various BP algorithms that have been proposed to deal with problems with continuous variables.
The Divide and Concur (DC) algorithm is a projection-based constraint satisfaction algorithm that deals naturally with continuous variables, and converges to exact answers for problems where the solution sets of the constraints are convex. I show how it exploits the "difference-map" dynamics to avoid traps that cause more naive alternating projection algorithms to fail for non-convex problems, and explain that it is a message-passing algorithm that can also be applied to optimization problems. The BP and DC algorithms are compared, both in terms of their fundamental justifications and their strengths and weaknesses.