Design, Implementation, and Analysis of a Parallel Description Classifier

A classifier is a central reasoning component of modern knowledge representation systems. Classifiers provide such fundamental intelligent services as concept categorization, instance recognition, and query processing. Unfortunately, as the size of the knowledge base grows, classifiers become less useful because the classifier must process a significant fraction of the knowledge base to perform any given inference. This paper investigates the extent to which parallel processing may be applied to the classification problem. We describe a MIMD implementation of a parallel classifier which uses a message-passing paradigm to effect interprocessor communications. Simulations and analysis of a local-area network implementation of the parallel classifier indicate that very large speedups may be obtained, and that speedups are limited only by the depth of the knowledge base. Preliminary results indicate that graph partitioning algorithms that cluster interdependent portions of the knowledge base may help to improve the efficiency of the parallel classifier. * Kevin Knight and Milind Tambe provided comments on an earlier draft which greatly improved this paper. Jeff Koller gave us useful advice on parallel machines and introduced us to PVM. Support for this work was provided by the Advanced Research Projects Agency under contract no. DABT63-91-C-0025. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of ARPA or the U.S. Government.