一种启发式选择性神经网络集成设计方法  

Heuristic picking approach to selective neural network ensemble

作　　者：崔路阳 魏海坤[2] 

机构地区：[1]东南大学自动化学院,南京210096 [2]东南大学复杂工程系统测量与控制教育部重点实验室,南京210096

出　　处：《东南大学学报（自然科学版）》2010年第S1期297-301,共5页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金资助项目(60875035);江苏省自然科学基金资助项目(BK2008294)

摘　　要：针对集成中各子网准确率与子网间差异度的权衡难题,提出一种启发式子网挑选法(HP-BGP).不同于传统以正确率或差异度为指标的选择方法,该方法采用神经网络集成在验证集上泛化性能最优为子网的选取准则.首先,在不同的初始条件下生成足够的备选子网,然后,根据能使集成泛化性能提高最快的标准,每次选取能与集成中已有网络组合泛化性能最好的子网,逐个挑选加入集成,直至选择到合适的网络个数或达到一定的误差要求.以LIC1问题为平台,与传统的子网挑选法PB、PAH、GA进行比较,仿真结果表明HPBGP法在测试集上集成泛化效果与选择所用时间的综合指标上优于常规的子网挑选方法.Aimed at the difficulty in balancing the subnets' accuracy and the diversity among the subnets,the HPBGP ( heuristic picking based on generalization performance) method is proposed. Unlike the traditional evaluation criterion which uses subnet's accuracy or diversity among subnets,in this method the optimal generalization performance on the validation set of NNE ( neural network ensemble) is used as the evaluation criterion of subnet selection. First,enough subnets are generated by different initial conditions of neural network. Then,the subnet that can most improve the generalization performance of NNE is selected one by one. Selection stops when the number of the subnet is enough or the requirement of error is met. Based on the platform of LIC1( launch interceptor condition 1) ,simulation results show that the HPBGP outperforms in composite index of time and generalization compared with the conventional subnet selection approaches such as PB( picking the best) ,PAH( picking according to heuristic) and GA( genetic algorithm) .

关 键 词：选择性神经网络集成 平均正确率 差异度 启发式挑选 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]