Minimizing communication cost in a distributed Bayesian network using a decentralized MDP

Abstract

In complex distributed applications, a problem is often decomposed into a set of subproblems that are distributed to multiple agents. We formulate this class of problems with a two layer Bayesian Network. Instead of merely providing a statistical view, we propose a satisficing approach to predict the minimum expected communication needed to reach a desired solution quality. The problem is modelled with a decentralized MDP, and two approximate algorithms are developed to find the near optimal communication strategy for a given problem structure and a required solution quality.