Comparative Study on Gradient and Hessian Estimation by Kriging and Neural network approximation for Optimization

摘要

This paper discusses accuracy of estimated results of gradient and Hessian components of an original function. Kriging method and hierarchical neural network are used for estimation. Those methods will give a global approximated response surfaces, and gradient and Hessian components can be also estimated directly without using finite differences of estimated function values. However, those components may often include large estimation errors even if an approximated surface for objective function values can be constructed well. In this paper, therefore, accuracy of estimated results of gradient and Hessian components are investigated when objective function values are well estimated.