Construction of confidence intervals for neural networks based on least squares estimation

摘要: We present the theoretical results about construction of confidence intervals for a nonlinear regression based on least squares estimation and using linear Taylor expansion model output. stress assumptions which these are based, in order to derive an appropriate methodology neural black-box modeling; latter is then analyzed illustrated simulated real processes. show that output also gives tool detect possible ill-conditioning network candidates, estimate their performance. Finally, we approach compares favorably with other analytic approaches, it efficient economic alternative nonanalytic computationally intensive bootstrap methods.