摘要
本文构造了一种新的基于线性模型、多层前向网络的混合结构神经网络模型,并提出了相应的非迭代快速学习算法。该学习算法能够根据拟合精度要求,运用线性最小二乘法确定相应的最佳网络权值和线性部分的参数,并自动确定最佳的隐层节点数。与BP网络的比较结果表明,本文提出的混合结构前向神经网络的快速学习算法无论在拟合精度、学习速度、泛化能力、还是隐节点数均显著好于BP算法。
In this paper, a new type of neural networks with hybrid structures which consist of a linear model and multiply-layer feedforward neural network are constructed, and the corresponding non-iterative fast learning algorithm is proposed. According to the demand of approach-precision, the best weights of networks and the minimal numbers of hidden nodes are determined using linear least square. Compared with the well-known BP networks, simulation results show that the performance of the networks proposed in this paper is much better whether in approach-precision, learning speed, generalization ability or the numbers of hidden nodes.
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2003年第1期97-101,共5页
Pattern Recognition and Artificial Intelligence
