Predictive performance of “highly complex” artificial neural networks

摘要: Abstract The effectiveness and the “indeterminacy” of artificial neural networks (ANNs) are discussed. In way determination parameters different models a clear distinction should be made between two approaches: (i) parameter estimation (classical approach) (ii) ANN approach with aim prediction (mainly interpolation). classical regression procedures used, while in latter one so-called “training” using back propagation algorithm is applied. widely employed combinatorial materials science for “information mining” purposes. Parameters obtained have definite physical meaning. It also means that absence presumed hypothetical model cannot advantage usage ANNs it does not require casual system investigated. However, by no Generally, number to determined significantly higher than data training set. has emphasized good predictive ability superior importance meaning parameters.