On the role of procrastination in machine learning

Abstract

The traditional model of inductive inference is enhanced to allow learning machines to procrastinate about how many trials they will need to complete an inference or about how accurate their solution will be. Hierarchies of classes of learnable phenomena (represented by sets of recursive functions) isomorphic to the structure of the constructive ordinals are revealed. The existence of such hierarchies indicates a potential advantage of procrastination as a learning technique. Tradeoffs between the two types of procrastination are shown not to exist in general. One of the new classes of sets of inferrible functions introduced in this paper turns out to have the somewhat unusual property of being closed under finite unions. A new technique, based on a hardest to lear set (relative to a class), is used. © 1993 Academic Press, Inc.