A Real-World Application: The Boston Housing Data

Abstract

An ensemble of GM-RVFL networks is applied to the prediction of housing prices in the Boston metropolitan area on the basis of various socio-economic explanatory variables. The ARD scheme is tested and found to succeed in identifying and effectively switching off two redundant dummy inputs added to the data. The employment of a network committee leads to significantly better results than achieved with an individual network. A simple Bayesian regularisation scheme is applied, but found to decrease only the generalisation ‘error’ of the single-model predictor. For a committee, the best generalisation performance is achieved when employing over-complex, under-regularised models that, individually, overfit the training data.