In order to enhance computational efficiency and precision of response surface methods

the global response surface method is improved. Using RBF neural network substitutes BP network as the approximation of implicit performance function

and the response surface is formed in the sphere of hyper-pyramid at iteration step. Compared with other response surface methods

numerical case studies show that the proposed method not only markedly reduces the times of iterations and finite element analysis to shorten the processing time

but also benefits to better accuracy. Moreover

the engineering examples of shallow foundation and retaining wall are analyzed on their structural reliability. The calculation results show that the proposed method may gain reliability index satisfyingly by quicker convergence rate and less machine time

and to be suitable for reliability analysis in geotechnical engineering.