论文部分内容阅读
针对全连接BP网络在解决大规模复杂问题时存在的收敛速度缓慢等问题,提出一种功能分区的BP网络结构模式.利用RBF神经元的物理特性对输入样本空间进行分解,并将分解后的样本送给不同的子BP网络学习.与全连接BP网络相比,降低了网络在学习过程中的权值搜索空间,提高了学习速度,改善了网络泛化性能,体现了人脑在学习过程中的知识积累特征.对三维墨西哥草帽函数逼近和双螺旋分类的实验结果表明,该网络能够解决全连接BP网络不能有效解决的问题.
Aiming at the slow convergence speed of fully connected BP network in solving large-scale and complex problems, this paper proposes a structural model of BP network with functional zoning. The input sample space is decomposed by the physical properties of RBF neurons, and the decomposed The samples are sent to different sub-BP network learning.Compared with the fully connected BP network, the network reduces the weight search space in the learning process, improves the learning speed, improves the network generalization performance, and reflects the human brain in the learning process In the three-dimensional Mexican straw hatch function approximation and double helix classification experimental results show that the network can solve the fully connected BP network can not effectively solve the problem.