ANN集成的训练误差收敛和泛化误差有界的证明  

Convergence and Generalization Proof of ANN Integration

作　　者：黄红梅[1] 罗兵[2] 

机构地区：[1]广东水利电力职业技术学院,广东广州510635 [2]广东工业大学自动化学院,广东广州510006

出　　处：《广东水利电力职业技术学院学报》2008年第4期58-64,共7页Journal of Guangdong Polytechnic of Water Resources and Electric Engineering

摘　　要：按AdaBoost自适应提升算法构建多个简单ANN然后集成,简化了ANN设计,改善了ANN性能。该理论证明:在单个ANN具有一定的要求较低的性能条件下,即误差小于50%时,该ANN集成对两分类问题的训练误差具有收敛性,对测试集的泛化误差有界。从而从理论上可以解释其对ANN性能的改进。ANN integration on AdaBoost algorithm can get high performance while simplifying ANN design and improving its performance.This theory shows that genetic algorithm,which uses minimum marginal as fitness function and where error is less than 50%,can have convergence influence on training error regarding two kinds of problems and have boundary on the general-ization error of testing integration.Thus,the improvement of ANN performance can be achieved and explained theoretically.

关 键 词：ANN集成 ADABOOST 机器学习 训练误差 泛化性能 

分 类 号：TP11[自动化与计算机技术—控制理论与控制工程]