| Summary: | 博士 === 淡江大學 === 電機工程學系 === 88 === In this dissertation, we investigate how to construct the appropriate fuzzy system for different design purposes in which only the input-output data of the considered system are given. In general, the fuzzy system design is composed of two steps: structure identification and parameter identification. In the process of structure identification, the structure of the constructed fuzzy system is determined by the input-space partition. The following three types are considered to partition the input space.
Grid partition: In this type, the genetic algorithm (GA) is applied to select an appropriate grid partition of the input space. Each individual in the GA approach is defined to represent a grid partition of the input space. Then, an appropriate fitness function is defined to guide the search procedure to select an appropriate grid partition of the input space for the input-output data of the considered system.
Hybrid partition: In this type, some fuzzy regions are generated by the different grid partitions of the input space. Then, the SVD-QR method is applied to select significant fuzzy regions so that the selected fuzzy regions effectively describe the input-output data of the considered system.
Scatter partition: In this type, a clustering algorithm is proposed to group the input-output data of the considered system into some clusters. Then, we map one cluster to one fuzzy region so that the fuzzy regions can effectively describe the behavior of the input-output data of the considered system.
Therefore, in the process of structure identification, three types of the input-space partition are proposed to determine the number of fuzzy rules and the fuzzy sets in the premise part from the input-output data.
In the process of parameter identification, the gradient descent and recursive least-squares methods are applied to tune the parameters of the fuzzy system from the input-output data so that the fuzzy system can effectively approximate the considered system.
Besides, the proposed clustering algorithm in the scatter partition is applied to construct a fuzzy system directly from the training data of the considered system so that the constructed fuzzy system can actually classify the training data. Moreover, a rule elimination method is proposed to remove the redundant fuzzy rules so that the constructed fuzzy system has a small number of fuzzy rules under the situation that all the training data are correctly classified. In each proposed method, several examples are considered to illustra
|
|---|
