A general scheme for minimising Bayes risk and incorporating priors applicable to supervised learning systems

Abstract

BARTIN (Bayesian real-time network) is a general structure for learning Bayesian minimum-risk decision schemes. It comprises two unspecified supervised learning nets and associated elements. The structure allows separate prior compensation and risk minimization and is thus able to learn Bayesian minimum-risk decision schemes accurately from training data and priors alone. The design provides a new mechanism (the prior compensator) for correcting for discrepancies between class probabilities in training and recall. The same mechanism can be adapted to bias output decisions. The general structure of BARTIN is described and the enumerative and Gaussian specific form are presented. The enumerative form of BARTIN was applied to a visual inspection problem in comparison with the multilayer perceptron.<>