Distributed tree decomposition with privacy

  • Authors:

  • Vincent Armant;Laurent Simon;Philippe Dague

  • Affiliations:

  • Laboratoire de Recherche en Informatique, LRI, Univ. Paris-Sud & CNRS, Orsay, France;Laboratoire de Recherche en Informatique, LRI, Univ. Paris-Sud & CNRS, Orsay, France;Laboratoire de Recherche en Informatique, LRI, Univ. Paris-Sud & CNRS, Orsay, France

  • Venue:
  • CP'12 Proceedings of the 18th international conference on Principles and Practice of Constraint Programming
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

Tree Decomposition of Graphical Models is a well known method for mapping a graph into a tree, that is commonly used to speed up solving many problems. However, in a distributed case, one may have to respect the privacy rules (a subset of variables may have to be kept secret in a peer), and the initial network architecture (no link can be dynamically added). In this context, we propose a new distributed method, based on token passing and local elections, that shows performances (in the jointree quality) close to the state of the art Bucket Elimination in a centralized case (i.e. when used without these two restrictions). Until now, the state of the art in a distributed context was using a Depth-First traversal with a clever heuristic. It is outperformed by our method on two families of problems sharing the small-world property.