Abstract:

The learning performance and the generalization property of support vector machines (SVMs) are greatly  influenced by the suitable setting of some parameters. The parameters selection can be transformed into an optimization problem by defining the root mean square error of a SVM prediction model as an evaluation function. A kind of improved estimation of the distribution algorithm (EDA) with a chaotic-mutation operation was proposed and used to optimize parameters of a ε-SVM including a penalty factor, an insensitive loss coefficient and a width of a Gaussian kernel function. The improved EDA could take advantage of the randomness and ergodicity of chaos, which  could  solve the local minima problem of traditional EDAs. Simulation result of the prediction of a Chebyshev chaotic time series showed that the improved EDA was an effective method of solving the problem for parameters selection of a SVM.

Key words: support vector machine; parameters selection; chaotic-mutation; estimation of distribution algorithm