| Summary: | 碩士 === 朝陽科技大學 === 資訊工程系碩士班 === 92 === In this thesis, a wavelet-based neuro-fuzzy system (WNFS) is proposed for handling non-linear system applications. The WNFS is a feedforward multi-layer network which integrates traditional Takagi -Sugeno-Kang (TSK) fuzzy model and the wavelet neural networks (WNN). A learning algorithm, which consists of structure learning and parameter learning, is presented. The structure learning is based on the degree measure to determine the number of fuzzy rules and wavelet functions, and the parameter learning is based on the gradient descent method to adjust the shape of membership function and the connection weights of WNN. More notably, in this learning method, only the training data need to be provided from the outside world. Users need not give any a priori knowledge or even any initial information in our proposed model. Hence, there are no fuzzy rules and wavelet functions in the beginning of learning. However, the WNFS is an inherent feedforward network structure. Inefficiency occurs for temporal problems. Therefore, in this thesis, a wavelet recurrent neuro-fuzzy system (WRNFS) is developed for solving temporal problems. The recurrent property comes from feeding the internal variables, derived from membership function matched degree, back to itself. There are also no nodes except the input-output nodes initially in the WRNFS. They are created via concurrent structure learning for the construction of WRNFS and ordered derivative learning for the tuning of free parameters. Finally, various simulations on static and temporal problems are done and performance comparisons with some existing modes. Efficiency of the proposed models (WNFS, WRNFS) is verified from these results.
|
|---|
