Diffusing Inference: An Inference Method for Distributed Problem Solving

Abstract

This paper presents the diffusing inference method, which is an inference method for distributed problem solving, and discuss its properties. Diffusing inference differs from other methods such as task -sharing and result-sharing, and is appropriate for applications that inherently require distributed search techniques. To discuss the algorithm and its properties rigorously, a formulation of distributed problem solving based on the state space graph is introduced. Local knowledge of each agent is represented as a partial graph in this formulation. Diffusing inference works as follows: an agent that received an initial state begins to search using only local knowledge for the goal state as far as it can. If it cannot achieve the goal, then it allocates the rest of the search to the other available agents and they continue to search likewise. Inference diffuses, as a result, among agents as distributed search proceeds. Furthermore, the property of completeness, i.e. the diffusing inference algorithm finds a solution and terminates whenever it is given a problem with a solution, is discussed and proved. Generally speaking, the performance of a diffusing inference system depends on the relation between its inference speed and its communication speed. Therefore a performance evaluation from this viewpoint is presented by using simulation techniques. It shows a guideline where the performance of diffusing inference by multiple agents is superior to that of a centralized search by a single agent.

This research was supported in part by a Grant-in-Aid for Scientific Research of the Japanese Government under grant no. 01780051 and in part by International Information Science Foundation under grant no. 90-3-2-287. The authors would like to thank Dr. Mark Klein for reading drafts of this paper and supplying helpful comments.