从受训神经-模糊模型中提取知识

本文对用于模式分类、函数逼近、参数估计的多层感知器 (MLPs)给出 1个清晰的关于内部行为的解释。作者以单隐层的 MLP为例 ,论述了关于 MLP的内部行为的半线性分析理论。对受训的MLP,将隐层单元的输出分别定义为网络输出的正、负“内部分量”;定义内部分量的连接权重集为给定问题的“内部判别模式”;建立了 MLP和模糊集相结合的新模型 ;分析了 MLP的结构为 N- 2 - 1和N- H- 1 ,给出权重初始化的方法 ;提出了 1种从受训神经 -模糊模型 (NFMs)中提取知识的全新的具有实用价值的方法。

Knowledge Extracting from Trained Neuro-Fuzzy Models

This paper describes a clear interpretation of internal behaviors of multi-layer perceptrons (MLPs) used for pattern classifications, functional approximations, and parameter estimations. Taking a MLP with a single hidden layer for an example, a semi-linear analysis theory of internal behavior of MLPs is presented. For a trained MLP, some of outputs of hidden units are defined as positive or negative "internal components" of a net-output; the distributions (sets) of connection weights of the internal components are defined as the "internal decision patterns" for given problems. A model combining MLP with fuzzy-set approach, the MLPs with N-2-1 or N-H-1 architecture, and the method of initializing weights are proposed. A new practical approach for knowledge extracting from trained neuro-fuzzy models (NFMs) is introduced, which is useful for real-world problems.