Design for TSK-Type fuzzy neural networks based on MSC-GA and BP
Liang Zhao · 2010
In this paper, a hybrid learning algorithm which automatically trains the TSK-Type fuzzy neural network is proposed. It consists of two stages: structure identification and parameter optimization. 1) structure identification stage (the construction of fuzzy if-then rules base) comprises the mean shift clustering (MSC) and the genetic algorithm (GA). 2) parameter optimization stage is based on the back-propagation algorithm with the momentum term (BP). The MSC is used to partition the input vector space for performing initial structure learning. Then, the GA is adopted to prune redundant fuzzy if-then rules. After the structure identification is completed, the BP is applied to obtain the optimal means and variances of the membership functions of each input variable and the output weights connecting the fuzzy rule layer and output layer. The simulation experiment verifies that the hybrid learning algorithm achieves good performance in learning accuracy than those of some traditional methods.