中文摘要：

提出一种蜂群-神经网络集成方法,与一般的神经网络集成方法不同的是：（1）集成个体的生成首先利用蜂群算法优化三层BP神经网络的结构和连接权,并以优化后的网络结构和连接权作为新的神经网络结构和初始连接权,再进行新一轮BP神经网络训练后生成;（2）为提高集成个体间的差异度,首先对个体进行分类,其次利用ABC算法对每一类个体进行最优组合搜索,选取相关系数最低的一个组合的均值作为该类的代表,最后对不同类别的代表作平均集成。在西太平洋热带气旋强度的预测试验中,所提出的蜂群-神经网络集成方法的泛化能力不仅明显优于单个神经网络,也优于Bagging和AdaBoost这两种集成方法。是一种具有较高应用价值的神经网络集成预测方法。

英文摘要：

An artificial bee colony （ABC）neural network ensemble method is presented in this paper. It differs from traditional NNE method in following aspects ： （ 1 ） the generation of individuals of neural network ensemble is realised in such a way, that the artificial bee col ony algorithm is used to optimise the structure and connection right of 3tier BP neural network at first, then the optimised network structure and connection right will be used as the structure and primary connection right of the new network respectively, at last the generation com pletes from a new round of BP neural network training. （2） In order to increase the difference between the individuals of the ensemble, all the individuals are classified first, and then ABC algorithm is employed for optimal combinatorial searching on every individual, and the mean of a combination with lowest correlation coefficient will be selected as the representative of that class, finally the average ensemble is derived based on the representatives of different classes. In prediction test on the intensity of Western Pacific tropical cyclone, the generalisation ca pability of the artificial bee colonyneural network ensemble method proposed in this paper is significantly better than the single neural net work, and is also superior to the ensemble methods of Bagging and AdaBoost. It is a neural network ensemble prediction method with high ap plied value.