Proceedings available:
http://www.aaai.org/Press/Reports/Conferences/cf-00-01.php


An Implementation of a Parallel Machine Learner

Lee A. Adams, Kevin M. Livingston, and Jennifer Seitzer

We present an implementation of speculative parallelism in the realm 
of deductive and inductive reasoning systems. Machine learning, a form
 of induction, can be thought of as a search for a generalizing rule 
that summarizes a collection of data. In this work, we broach the 
search for such a rule by simultaneously traversing the search space 
at ten different starting points; ten ranking algorithms are executed 
simultaneously on a distributed architecture. Additionally, we present
 a data parallel design where each learning algorithm, itself, is 
distributed among many processors. We exemplify these concepts of 
speculative and data parallel parallelism by transforming a sequential 
knowledge-based deduction and induction system, INDED, into a 
speculatively parallel system where several processors simultaneously 
search for an accurate rule obtained in a supervised learning 
environment. Moreover, a data parallel implementation of the 
fundamental operation of each learning algorithm, ranking, is 
parallelized among the cluster nodes. We present algorithms for work 
delegation as well as final rule assessment used in selection of the 
superior learned rule.