一种RBF网络结构调整的稳健增量学习方法  

A Robust Incremental Learning Algorithm Based on Growing and Pruning Strategy

作　　者：刘建军[1] 胡卫东[1] 郁文贤[1] 

机构地区：[1]国防科技大学电子科学与工程学院,湖南长沙410073

出　　处：《计算机仿真》2009年第7期192-194,227,共4页Computer Simulation

基　　金：重点实验室基金资助课题(9140C8001020603)

摘　　要：以实现RBF网络的增量学习能力和提高其增量学习的稳健性为目的,给出了一种RBF网络增量学习算法。算法首先对初始数据集进行聚类得到初始的RBF网络结构,然后采用GAP-RBF算法中的隐层节点调整策略动态调整网络结构实现RBF网络增量学习。RBF网络的初始化降低了初始数据集样本训练顺序对RBF网络性能的影响,增强了其增量学习的稳健性。IRIS数据集和雷达实测数据集仿真实验表明,算法具有较好的增量学习能力。This paper presents an incremental learning approach based on RBF network aiming at making it learn new information without forgetting old knowledge and improving its robustness. The proposed algorithm combines a RBF network initialized using a clustering algorithm which alleviates the influence of the training order of the original patterns with the growing and pruning strategy of the GAP - RBF algorithm. The initialization of the RBF network improves the robustness of the algorithm. The incremental learning performance of the growing and pruning strategy is evaluated based on iris data sets and real radar data sets and the results indicate that the algorithm has excellent incremental learning ability.

关 键 词：模式识别 径向基函数网络 隐层节点调整 增量学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]