A comparison of model selection methods for compressive strength prediction of high-performance concrete using neural networks

This paper gives a concise overview of three approaches to nonlinear regression modelling with feed-forward neural networks, involving the use of evidence framework and full Bayesian inference with Markov chain Monte Carlo stochastic sampling. The article then presents an empirical assessment of these approaches using a benchmark regression problem for compressive strength prediction of high-performance concrete. Results on applying various methods to benchmark dataset show that Bayesian approach with the MCMC sampling approximation of learning and prediction gives the best prediction accuracy.