Integration of Forward and Backward Inferences Using Extended Rete Networks

Abstract

In this paper, it is explained that the forward and backward chaining can be naturally integrated on an extended Rete network via the generation of hypotheses from forward chaining inference. We introduce the concept of the hypothesis objects and augment the production rule format by adding the hypothesis condition elements in the LHS part of a rule, which control the forward and backward chaining invocation of rules. A rule can be used in forward directions, backward directions, or mixed directions. To be applied to the backward chaining, a modified version of Rete has been formulated by adding the hypothesis-and-nodes and backward direction edges to the node network. We then develop a backward chaining algorithm which takes advantage of already existing partial match results obtained by the previous forward chaining in the ERMI. The whole mixed inference process has the potential advantages over purely forward and backward chaining strategy.