| Summary: | 碩士 === 國立虎尾科技大學 === 電機工程研究所 === 100 === This dissertation proposes knowledge-based cultural differential evolution (KCDE), symbiotic cultural differential evolution (SCDE), and cooperative cultural differential evolution (CCDE) for neural fuzzy inference systems (NFIS). The cultural algorithms acquire the belief space from the evolving population space and then exploit that information to guide the search. The proposed evolutionary algorithms use the five different types of knowledge sources in the belief space to find the global optimal. This dissertation consists of three major parts. In the first part, a KCDE method is proposed for optimizing parameters of the NFIS model. The KCDE adopts five mutation strategies of differential evolution (DE) as the knowledge sources of belief space to influence the population space. These knowledge sources including normative knowledge, situational knowledge, domain knowledge, history knowledge, and topographic knowledge are integrated in belief space. Unfortunately, a fuzzy system must be encoded into an individual in KCDE method that causes the diversity of population space to reduce. Therefore, the second part proposes SCDE method that uses the symbiotic evolution to make the individual represent a partial solution. The individual combines with other partial solutions randomly in the population to build a complete solution, which increases the diversity of population and fast convergence but not premature convergence. However, the process of symbiotic evolution mechanism selects the partial solution to form a complete solution too random that causes the deviation of performance is too high. Therefore, a CCDE is proposed for NFIS model in the third part. The CCDE adopts cooperative coevolution to effectively decompose the fuzzy system into subpopulations, and each individual within each subpopulation evolves separately. Finally, the KCDE, SCDE, and CCDE are applied to implement NFIS model in various nonlinear control system problems. The results of this dissertation demonstrate the effectiveness of the proposed methods.
|
|---|
