| 摘要: | 隨著資料的快速增加，如何有效分析隱藏在大量資料中的價值日益重要。在資料分析的領域中，時間序列的分析與預測是一主要的研究方向。本研究提出一個複數模糊類神經模型(Complex neuro-fuzzy model)，其結合複數模糊集(Complex fuzzy set)、T-S模糊系統(T-S fuzzy system)形成該模型。在參數學習，以多群連續型蟻群演算法(Multi-group continuous ant colony optimization, MGCACO)與遞迴最小平方演算法(Recursive least squares estimator, RLSE)結合，成為MGCACO-RLSE複合型演算法，進行參數的搜尋與最佳化。多群連續型蟻群演算法在演化過程中，加入資料流通、淘汰、繼承等特性，能夠減少演算法落入區域最佳解以及加速進行參數的最佳化。在資料進入模型前，利用特徵選取(Feature selection)的方式，擷取其中較為有影響力的資料進行預測，減少模型負擔。此外，模型的歸屬程度(Membership degree)是複數型態，並且能拆解成多個不同的歸屬程度，使模型達到預測多目標的效果。本研究以三個實驗來驗證模型的效能與研究理論。實驗一的單目標證實本研究的理論，實驗二與實驗三的多目標實驗，個別證明模型的複數形態輸出以及利用多個歸屬程度達到的多目標輸出方法。個別實驗結果皆與過往文獻比較，實驗顯示本研究模型在時間序列預測上有良好效能，更加證實本研究的可行性。;In the age of information, it is increasingly important to deal with big data effectively for both scientific researches and applications of data science. In this study, we have proposed a complex neuro-fuzzy system to data prediction, using complex fuzzy sets and logic and T-S neural fuzzy modeling. Complex fuzzy set is an advance fuzzy set whose membership degrees are complex-valued and defined with the unit disc of the complex plane, in contrast to regular fuzzy set whose membership degrees are real-valued and defined within [0,1]. For model construction, we have used feature selection to get influential data for the proposed model. For the optimization of the proposed model, we have developed a hybrid method for machine learning, denoted as MGCACO-RLSE, integrating the proposed multi-group continuous ant colony optimization (MGCACO) and the well-known recursive least squares estimator (RLSE) method. The MGCACO-RLSE has been applied to optimize the parameters of the proposed model. For MGCACO, we have added some properties, such as data exchange, elimination, and inheritance to increase its search efficacy, in the sense of reducing the possibility of being trapped at a local optimum and thus increasing the chance of finding the optimization solution. In this study, we have conducted three experiments to verify the effectiveness and rationale of the proposed approach. In experiment one, the proposed approach was tested for data prediction with single target to see the feasibility of the research thought. In experiments two and three, the proposed approach was tested for multi-target prediction. With the experimental results, the proposed approach has shown good performance, through performance comparison to other methods in literature. |
