模糊回声状态网络

摘要

针对基于梯度下降的模糊递归神经网络训练效率低、容易陷入局部极小的缺点,本文基于回声状态网络(Echo State Networks,ESNs)和TS模型提出一种新的模糊模型结构——模糊回声状态网络(Fuzzy Echo State Networks,FESNs).FESNs由多条TS类型的模糊规则组成,规则后件采用ESNs网络.研究表明,TS模型和ESN都可以看做是FESN模型的某种特例,而且FESNs具有较强的非线性映射能力、局部反馈以及学习算法稳定等特点.同时,其模型参数确定方法与经典TS模型以及ESN一样可以归结为一个线性回归问题,大大减少了网络训练的计算量.仿真实验表明,与经典TS模型相比,FESNs在不显著增加建模时间情况下可有效提高建模精度.%Recurrent fuzzy neural networks, which is usually trained by gradient descent, have some inherent shortcomings such as inefficiency of training process, local minimum. In this paper, we proposed a novel fuzzy neural network based on Echo state network and TS fuzzy model,called fuzzy echo state network (FESN), which is a generality of both TS modal and ESN. FESNs consist of several fuzzy IF-THEN rules,each of which has a ESN as consequent part. We illustrate that FESNs have some interesting characteristics, such as better nonlinear mapping capacity, local feedback and stable learning, which results in that FESNs can deal with dynamics of nonlinear system. Furthermore, similar to the TS model and ESN,parameters of the FESNs can be determined by solving a linear regression problem, which dramatically reduce the computing burden of the training process. Experiments shows that FESN can effectively enhance the accuracy of modeling dynamical system at the expense of not using excessive additional time compared with TS model and ESN.