提出了模糊极大极小神经网络新的隶属函数及新的并集学习算法。算法不受形状因子的影响且与学习顺序无关,各模糊子集的等λ截集中不存在异类训练样本,学习后的隐层节点数较扩充-收缩算法更少,对训练集和检测集的正确识别率更高。 A new membership function of fuzzy maximal minimax neural network and a new algorithm of union learning are proposed. The algorithm is not affected by the shape factor and has nothing to do with the learning sequence. There is no heterogeneous training sample in the equal-λ cutoffs of the fuzzy sub-sets, and the number of hidden-layer nodes after learning is less than that of the augmented-contractive algorithm. The correct recognition rate is higher.