TSK模糊神经网络及其约束优化学习算法

针对非线性动态系统特点,提出了一种基于TSK(Takagi-Sugeno-Kang)模糊模型的动态回归模糊神经网络DRFNN(Dynamic Recurrent Fuzzy Neural Network),该模糊神经网络由静态网络和动态网络两部分组成,其中静态网络用来实现规则的条件部分和解模糊部分的计算,由FIR动态滤波器实现的内反馈回归网络用来实现规则的结论部分,为了加快网络收敛速度,给出了基于约束优化算法的网络参数迭代算法,把网络结构优化和参数学习作为一个约束优化问题来解决.应用于非线性系统的辨识和控制仿真试验说明了DRFNN网络及其算法对解决非线性系统问题的有效性.

TSK-DRFNN and its constrained optimization algorithm

A novel DFNN(dynamic fuzzy neural networks )based on TSK(Takagi-Sugeno-Kang) fuzzy model was presented to the nonlinear dynamic system. The DFNN was constitutive of static networks and dynamic networks. The static networks realized premise and defuzification part. The recurrent dynamic networks realized by FIR filter was used for realizing consequence part. Beside this a new algorithm, constrained optimization method-FUNCOM(fuzzy neural constrained optimization method)was suggested for reducing the convergence time of networks parameter. The network training task was formulated as a constrained optimization problem. The proposed dynamic model equipped with the learning algorithm was applied in a nonlinear dynamic system. Comparisons with other FNN(fuzzy neural network) and DFNN (dynamic fuzzy neural network)were given and discussed, indicating the effectiveness of the DRFNN(dynamic recurrent fuzzy neural network) and the algorithm.