作者: Veit Elser , Jonathan S. Yedidia , Nate Derbinsky , José Bento
DOI:
关键词: Local consistency 、 Algorithm 、 Belief propagation 、 Convex optimization 、 Constraint satisfaction 、 Mathematical optimization 、 Computer science 、 Message passing 、 Circle packing
摘要: We describe how the powerful "Divide and Concur" algorithm for constraint satisfaction can be derived as a special case of message-passing version Alternating Direction Method Multipliers (ADMM) convex optimization, introduce an improved based on ADMM/DC by introducing three distinct weights messages, with "certain" "no opinion" weights, well standard weight used in ADMM/DC. The messages allow our to implement propagation case, while speed convergence some problems making focus only active constraints. three-weight give greatly performance non-convex such circle packing solving large Sudoku puzzles, retaining exact ADMM problems. also advantages compared other algorithms upon belief propagation.