Data-Defined Problems and Multiversion Neural-Net Systems

Abstract

Data-defined problems are not restricted to any particular problem domain, and are common.Data-defined problems, as the name suggests, are defined by a set of input-output mappings, and for the problems of particular interest the full details of how the inputs are related to the outputs are unknown.Such problems present the traditional programmer, whether using AI techniques or not, with a difficult task.This is because programming requires a prerequisite understanding of the mechanisms relating input and output such that an algorithmic solution can be devised.Automatic induction techniques demand no such prerequisite information.Neural computing is one such induction technique.And, moreover, it is one that can outperform more traditional AI induction techniques, such as IF-THEN rule systems.Neural computing can do better because it can follow a welldefined path to an optimal result, and because multiple, alternative versions can be cheaply generated to permit exploitation of certain properties of the resultant version set -in particular, 'diversity'.We define a valuable extension to previous definitions of this quantity, and exploit it to produce significant system performance enhancements.In addition, because our neural computing is an approximating technique, it is also amenable to exploitation of diversity through averaging multiple versions.An example of letter recognition is used to illustrate these ideas.