构造型神经网络双交叉覆盖增量学习算法

研究了基于覆盖的构造型神经网络(cover based constructive neural networks,简称CBCNN)中的双交叉覆盖增量学习算法(BiCovering algorithm,简称BiCA).根据CBCNN的基本思想,该算法进一步通过构造多个正反覆盖簇,使得网络在首次构造完成后还可以不断地修改与优化神经网络的参数与结构,增加或删除网络中的节点,进行增量学习.通过分析认为,BiCA学习算法不但保留了CBCNN网络的优点与特点,而且实现了增量学习并提高了CBCNN网络的泛化能力.仿真实验结果显示,该增量学习算法在神经网络初始分类能力较差的情况下具有快速学习能力,并且对样本的学习顺序不敏感.

An Incremental BiCovering Learning Algorithm for Constructive Neural Network

The algorithm of incremental learning in cover based constructive neural networks (CBCNN) is investigated by using BiCovering algorithm (BiCA) in this paper. This incremental learning algorithm based on the idea of CBCNN can set up many postive-covers and negative-covers, and can modify and optimize the parameters and structure of the neural networks continuously, and can add the nodes according to the need and prune the redundant nodes. BiCA algorithm not only keep the advantages of CBCNN but also fit for incremental learning and could enhance the generalization capability of the neural networks. The simulational results show that the BiCA algorithm is not sensitive to the order of the sample and could learn quickly and steady even if the performance of initial CBCNN is not very good.