Abstract:

Selection of the hyper-parameters is critical to the performance of support vector machines (SVM), directly impacting the generalization and regression efficacy of the SVM. An orthogonal experimental design procedure for hyper-parameter selection (ODPS) is clearly desirable given the intractable problem of exhaustive search methods. The authors' previous work in this area involved analyzing the range value of hyper-parameters for SVM of mixed kernel which has been proved and showed a higher convergence rate and a greater flexibility in learning a problem space than single kernel functions, and determining experimental levels for different parameters in order to guide the hyper-parameter selection process. The method selects hyper-parameters optimal composition in terms of orthogonal and interaction effect of hyper-parameters. The results of the performed engineering experiments for the prediction of two typical landslide deformation time series confirmed the reliability and advantage of the proposed approach.

Key words: orthogonal design, support vector machines, hyper-parameters, time series regression, landslide