Rule-based update methods for a hybrid rule base

摘要

In this paper, we present methods for efficient updates of a hybrid rule base. The hybrid rule base consists of neurules, a type of hybrid rules combining symbolic rules and neural networks. A neurule base, called the target knowledge, is produced by conversion from a symbolic rule base, called its source knowledge. The presented methods concern modifications to the target knowledge, due to insertion of a new rule in or removal of an old rule from its source knowledge. The methods (a) require as little re-conversion as possible and (b) preserve the number of neurules as small as possible. This is achieved by storing information related to the conversion process in a tree, called the splitting tree. Experimental results demonstrate the benefits of using the splitting tree.