基于GA、BP神经网络和多元回归的集成算法研究

摘要

遗传算法、BP神经网络和多元回归是目前应用比较广泛的数据挖掘算法，它们各俱优点，同时也存在诸多无法避免的缺陷。该文在前三者的基础上，提出一种BP网络与多元回归模型融合的杂合BP网络，并采用遗传算法优化杂合BP网络的初始权值，有效地避免几种方法在单独使用时存在的缺陷。验证实验结果表明：新方法所建立的模型在收敛速度、精度和泛化能力上都明显优于GA、BP神经网络和多元回归，并且较当今比较热门的ELM、SVRKM和SVM也有较显著的改进。%Genetic Algorithm, BP neural network and multiple regression are used widely in data mining algorithms, each of them have their benefits. Simultaneously, they have some inevitable flaws. On the basis of previous three, I made some improvements in the structure of them. First, I propose a hybrid BP network based on the integration of BP Network and multiple regression models. Then I used the hybrid genetic algorithm to optimize the initial weights of hybrid BP network. In that way, I effectively avoid the inevitable flaws when they alone. Validation results show, in convergence speed accuracy and generalization ability, the model of new methods is better than Genetic Algorithm, BP neural network and multiple regressions. In addition, the model of new methods has significant improvements compared with ELM, SVRKM and SVM.