A Local Approach to Optimal Queries

Abstract

We examine the problem of self-directed learning using queries in contrast to learning from random examples. We argue that, to provide any significant improvement over learning from random examples, a self-directed learning algorithm must make use of sequential queries, that is, new query points must be selected based on the answers to previous queries. Additionally, to perform well in a complex domain, such an algorithm must distinguish between degrees of the relative utility of querying various points in the domain. We present an algorithm that implements these two guidelines to produce “optimal” queries for a single linear threshold unit, in the sense that these queries minimize the bound on possible error. The algorithm is generalized to networks of such neurons using a local approach where each unit attempts to minimize its perceived error.