本研究發展了一個改良的最佳化演算法，並將之用在複數模糊類神經系統建模之參數學習，應用於時間序列之預測。在許多最佳化演算法中，差分進化演算法(Differential evolution, DE)是一種著名且正在發展中的演算法，具有相當之潛力。在本研究中，提出一個改良的 DE演算法結合了自我適應差分進化(Self-adaptive DE, SaDE)及相對點差分進化(Opposition-based DE, ODE) 。新演算法稱為自適應相對點差分進化(Self-adaptive opposition-based differential evolution, SaODE)，同時具有SaDE的變異策略投票機制及自我調整參數機制以及ODE的相對點搜尋機制。之後，再將SaODE與遞迴最小平方估計法結合而形成SaODE-RLSE複合式學習演算法，用來進行複數模糊類神經系統建模之參數學習，並應用於真實世界之時間序列預測。在時間序列預測中，SaODE-RLSE複合式學習演算法之效能與其他研究方法之效能作比較，實驗結果顯示SaODE-RLSE複合式學習演算法在複數模糊類神經時間序列預測有不錯的效能。In this thesis, an improved hybrid-learning algorithm has been developed to apply on the modeling of complex neuro-fuzzy system (CNFS) for the problem of time series forecasting. Differential evolution (DE) is a noted optimization method that is still in progress for search efficiency. It has great potential in parameter estimation for the purpose of modeling. In this study, an enhanced DE algorithm called self-adaptive opposition-based differential evolution (SaODE) has been studied, combining the ideas by two state-of-the-art DE methods: the self-adaptive DE (SaDE) and the opposition-based DE (ODE). The proposed method has the advantages of both ODE and SaDE in the search capability of opposite points by ODE as well as the selection capability of different mutation strategies with self-adaptive search parameters by SaDE. For the parameter learning of CNFS, the proposed SaODE has combined further with the well-known method of recursive least squares estimation (RLSE) to become the so-called SaODE-RLSE hybrid learning method for parameter estimation. A number of examples for time series forecasting have been used to test the proposed approach, whose results are compared with those by other approaches. The experimental results indicate that the proposed approach shows promising performance.
