基于灰色关联分析的分层模糊神经网络

为解决“模糊规则爆炸”问题,提出一种基于模糊神经网络从试验样本抽取模糊规则的方法。首先,根据灰色关联分析的结果,将输入变量进行两两组合建立分层模糊子系统。其次对每个模糊子系统设计分层参数、结构优化算法。在权值学习过程中,模糊进化规划与分层方法相结合,网络的各层权值独立优化,并且各层权值优化问题简化为二次型问题,降低了权值优化过程中的计算复杂性。最终能够实现整个模糊神经网络的分层优化,各层神经元单独训练且训练结果互不影响。与常规的前向进化神经网络方法相比较,该方法通过对神经元的部分解群体的进化,缩短了个体的编码长度,显著地减少了计算量。同时这种方法不但能够很大程度上简化适应值的计算,更重要的是能够降低适应值空间的复杂性,从而能够加速进化算法收敛到全局最优点。

Layered Fuzzy Neural Network Based on Gray Correlative Analysis

A fast approach for automatically generating fuzzy rules from samples using fuzzy neural networks was proposed for 'fuzzy rules explosion'.Firstly,according to the results of gray correlative analysis,many layer fuzzy sub-systems were built by two input variables combining together.Secondly,in the method,the architecture and weight evolution were carried out layer by layer.Combing fuzzy evolutionary programming with the layer-wise method,the weights of the each layer of a neural network were treated independently;as a result,weights can be updated layer by layer.The weight optimization problem of each layer can be reduced to one with a quadratic form.By evolving a population of neurons instead of neural networks,the length of coding is decreased and the cost of computation is alleviated.Meanwhile the method not only simplifies the computation of the fitness but also decreases the complexity of the fitness space.