Bayesian Inductively Learned Modules for Safety Critical Systems

J.E. Fieldsend, T.C. Bailey, R.M. Everson, W.J. Krzanowski, D. Partridge and V. Schetinin

Abstract

This work examines the use of Bayesian inductively learned classification methods in relation to safety critical systems. Central to our approach to critical software systems is the necessity to generate meaningful confidence measures not normally associated with predicted states of a system. This is achieved in this study by casting the problem in a Bayesian formulation, and is implemented using reversible jump Markov chain Monte Carlo (RJ-MCMC). We compare conventional and novel classification architectures, including generalised linear models, probabilistic k-nn and radial basis functions. Results from these methods are illustrated on real life critical systems, including medical trauma data. We develop a new technique, building on the reject region idea, to generate classification volume envelopes, in order to mark regions of classification uncertainty. We also report results on the trade-off between model complexity and the width of the posterior predictive probability. Finally, we discuss important emergent issues.

PDF  (657 kb)