ノイズを含むデータから作成した応答曲面の予測精度

概要

論文の詳細を見る
Response surfaces generated by using an artificial neural network or a Kriging model can approximate multimodal responses and the responses of multi-variable functions, even if the prior information used in creating the surfaces contains noises. To prevent the response surfaces from fitting to the noises in the prior information, regularization methods are applied. The Bayesian regularization method works well with an artificial neural network, and the improved nugget effect method can be effectively applied to the regularization of an ordinary Kriging model. Which response surface model to select depends on the amount of prior information and the number of variables in the response function. The Kriging model generates a more accurate response surface with a smaller amount of prior information than the artificial neural network does. On the other hand, the artificial neural network is more effective for approximating the response of a multivariable function than the Kriging model. As a result, we can apply the response surface method to find the variables that ensure the optimum response of a function. This may increase the opportunities for using optimization techniques in the practical design process.
2008-07-25