Implementation of SRM Principle Based on Wavelet Multi-resolution Approximation
Yinguo Li, Liangfei Zhang, Dongjin Guo, Yong Shi · 2007
In statistical learning theory (SLT), structural risk minimization (SRM) of machine learning is hard to implement. Limitation for implementing SRM principle by computing the Vapnik-Chervonenkis (VC) dimension of function sets is analyzed. A novel idea is proposed to measure the learning capability of function sets by adopting high-frequency spectrum feature in the discrete wavelet base function set. An approach to implementation of SRM learning strategy is given, which is based on wavelet multi-resolution approximation, and a method for smoothing the learning surface are also put forward. Simulations indicate effectiveness of the proposed methods in the approximation of the noise-spoiled nonlinear signals.