Abstract: In practice, there always is small variance (perturbation) between patterns obtained and objective patterns for fuzzy neural networks, sequentially, there may be big variance (oscillation) between the output of the neural networks and the objective true output for some input. Hereby, the authors propose robustness concept of Fuzzy Associative Memory (FAM for short) with perturbation of sample patterns in the paper. further, concretely analyze such robustness and the controlling method of FAM using a new weight learning algorithm advanced in the paper. In the end , the authors proved by experiment that the robustness of FAM is not good when used the new weight learning algorithm advanced in the paper.

Key words: Fuzzy associative memory, Learning algorithm, Sample pattern, Perturbation, Robustness

摘要： 在实际问题中，所获取的模糊神经网络的训练模式对总与客观真实的模式对存在一定的小幅误差（摄动），从而可能导致对某些输入网络的实际输出与期望输出有很大的误差．为此，本文提出了训练模式集摄动对模糊联想记忆网络(FAM)的鲁棒性概念，并具体讨论了采用一种新的权值学习算法时FAM的这种鲁棒性及其控制方法。最后通过实验证明了采用这种新的权值学习算法时，FAM对模式摄动不会拥有好的鲁棒性。

关键词: 模糊联想记忆, 训练模式, 学习算法, 摄动, 鲁棒性