Abstract：The support vector regression(SVR) algorithm has been introduced into parameters identification of numerical model in geotechnical engineering to take advantage of its merits such as small sample，good generalization and global optimization. But，the standard SVR algorithm can only solve one-dimensional output variable regression problem，thus restrict its application in back analysis field. In this paper，an improved SVR algorithm is introduced by decomposing multi-dimension output variables to many one-dimensional output variables，and then the multi-dimensional output variable regression is translated into a multi-layer standard SVR problem. In order to find the optimal parameters of this improved SVR model during sample training course，the genetic algorithm(GA) is combined with it to form the improved GA-SVR algorithm. After the optimal nonlinear mapping between the numerical model parameters and displacement is constructed，GA is used to identify the numerical model parameters within their search interval. In virtue of MATLAB toolbox，GA also is integrated into BP neural network to form the GA-BP algorithm. By comparing the identification results of three-dimensional elastoplastic model parameters in tunnel engineering by the two different algorithms，it can be concluded that the improved GA-SVR algorithm can obtain a higher identification precision and calculation efficiency than the GA-BP algorithm and can be applied to similar calculation parameters identification in geotechnical engineering.
