基于遗传算法的 ART2 神经网络数据集优化

为了提高小样本集情况下自适应谐振(ART)神经网络聚类的可靠性,提出了基于遗传算法的ART2神经网络训练集优化算法,克服了ART1神经网络编码的稳定性尚未完全解决和只能接受二进制模式的缺陷.利用遗传算法的全局寻优能力,通过对训练样本集添加适当的边界样本点,并将边界样本点和原样本集有机结合,以提高ART2神经网络的泛化性能.对ART2神经网络聚类算法的适当变更,以适应样本集的变化情况,并避免ART神经网络在不同训练阶段产生不同的聚类结果.实验证明,采用本算法后,ART2神经网络的聚类准确度可提高30%.

Dataset Optimization of ART2 Based on Genetic Algorithm

To improve clustering reliability of adaptive resonance theory(ART) under situation of small sample set,an optimum algorithm of training set of ART2 neural network based on genetic algorithm was presented in the paper.Such defects that code stability of ART1 neural network cannot be completely solved,and only the binary pattern can be accepted were overcome by the optimum algorithm .It improved generalization performance of ART2 neural network by adding proper boundary sample points to original training set with overall optimization ability of genetic algorithm and by organically combining boundary sample point with original training set.In addition,clustering algorithm of ART2 neural network can be suitable to adapt to the abovementioned change of training sample set and to avoid the different clustering result during the different training stage.Experimental results showed that clustering accuracy of ART2 neural network was improved by 30%.