BARTIN: minimising Bayes risk and incorporating priors using supervised learning networks

BARTIN (BAyesian Real-Time Network) is a general structure for learning Bayesian minimum risk decision schemes. It comprises two user-specified supervised learning nets (an observer and a utility network) and associated elements. This two stage structure allows separate minimisation of risk and compensation for changes in prior probabilities. It is 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 together with its enumerative and Gaussian forms. The enumerative form of BARTIN is applied to a visual inspection problem and compared with the MLP. The value of taking both priors and risk into account is demonstrated