一种实现结构风险最小化思想的结构自适应神经网络模型

Structure self-adaptive neural network model realizing structural risk minimization principle

本文提出了一种实现结构风险最小化思想的结构自适应神经网络学习模型,该方法运用遗传算法进行神经网络结构参数的学习,运用BP算法进行神经网络内部权值学习,有效地实现了结构风险最小化思想。与传统的基于经验风险最小的神经网络模型相比,它具存更强的自适应能力,能够弥补学习方法本身的缺陷,充分保证了模型的泛化能力。最后,将本文方法应用于非线性时间序列预测和模式识别,并与基于结构风险最小原则的支持向量机学习模型进行了比较,算例充分表明了本文方法的正确有效性。

In this paper, the structure self-adaptive neural network learning model, which can realize SRM principle, is put forward. In the model, GA is used to implement neural network structure parameter learning, and back propagation algorithm （BP） is used to carry out inner weight learning, and the SRM principle is realized effectively. Comparing with traditional neural network model based on ERM, structure self-adaptive neural network model possesses stronger self-adaptive ability, and it can remedy the shortcomings of single learning method and fully assure model generalization. In the end, the proposed new method is applied to non-linear time series forecast and pattern recognition, and is compared with support vector machine （SVM） learning model that is based on SRM. Examples show the correctness and validity of the proposed method.