USE OF METHODOLOGICAL DIVERSITY TO IMPROVE NEURAL-NETWORK GENERALIZATION

Littlewood and Miller [4] present a statistical framework for dealing with coincident failures in multiversion software systems. They develo p a theoretical model that holds the promise of high system reliabilit y through the use of multiple, diverse sets of alternative versions. I n this paper, we adapt their framework to investigate the feasibility of exploiting the diversity observable in multiple populations of neur al networks developed using diverse methodologies. We evaluate the gen eralisation improvements achieved by a range of methodologically diver se network generation processes. We attempt to order the constituent m ethodological features with respect to their potential for use in the engineering of useful diversity. We also define and explore the use of relative measures of the diversity between version sets as a guide to the potential for exploiting inter-set diversity.