Abstract: To improve the prediction precise of support vector machine model, an adaptive multiple kernels learning (AMKL) method is proposed. In this method, a tree structure is used to screen the kernels. Additionally, this processing can be implemented with growing and cutting branches manipulation for adding and multiplying the kernels in each layer. This would enhance the nonlinear and diversity characteristics of multi-kernels. Grid traversal and particle swarm optimization method are applied to solve the optimization problem of kernel parameters, weight coefficient and model parameters. It could offset the assignment deviation of parameters which occurs in the lack of a prior knowledge of the training samples. AMKL method is used to predict the settlement of metro tunnel. Comparisons between experimental results of AMKL and the ones of single kernel functions show that AMKL effectively improves the accuracy and generalization.

Key words: tunnel settlement, support vector machine, adaptive multiply kernels, prediction