| 摘要: | 在當今數據驅動的時代，時間序列數據的重要性日益突顯，但隨著問題複雜化，傳統的時間序列預測方法顯得不足夠使用。本論文使用一種能夠針對多個複數型態的目標進行預測的模型，稱作球形複數型態的模糊類神經推理系統(Sphere complex neuro fuzzy inference system, SCNFIS)。透過球形複數模糊集(Sphere complex fuzzy sets, SCFSs)的複數向量輸出能力，進行多目標的預測，以達到降低資料維度的問題。其次，我們使用多目標特徵挑選演算法，針對目標挑選最重要的特徵，並使用複數型態的減法分群法（Subtractive clustering for complex-valued data, SCC）、規則挑選(Rule selection)進行模型大小的最佳化調整，挑選適當模糊規則數建立。在參數學習部分，本研究提出了一種新的混合式演算法，稱作MEGWO-RLSE，以多群進化式狼群演算法(Multiple evolving grey wolf optimizer, MEGWO)更新模型的前鑑部參數，遞迴最小平方估計法(Recursive least squares estimator, RLSE)進行後鑑部的參數更新。整體而言，透過對模型規則進行特定比例的篩選，以及演算法參數空間切割，都有助於模型的預測。本研究中，進行了三個實驗針對提出模型和方法進行效能驗證，包括使用Mackey-Glass時間序列進行預測，驗證SCNFIS的模型效能、透過股票收盤價進行多目標預測，驗證模型具有多目標預測的能力，最後使用不同的股票標的，針對機器學習中的參數空間切割進行研究。從實驗結果上，SCNFIS在使用MEGWO-RLSE進行參數訓練上，均具有不錯的效能。;In today′s data-driven era, the importance of time series data is becoming increasingly prominent, playing a crucial role in various fields. However, with the complexity of problems and the exponential growth of data volume, traditional time series forecasting methods are proving inadequate. Therefore, this paper use a model capable of predicting multiple complex targets, which we call the Sphere complex neuro fuzzy inference system (SCNFIS). Utilizing the complex vector output capability of Sphere complex fuzzy sets (SCFSs), this model aims to perform multi-target predictions, thereby addressing the issue of high data dimensionality.
Moreover, we introduce a multi-target feature selection algorithm to identify the most important features for each target. The model′s size is optimized using Subtractive clustering for complex valued data (SCC) and rule selection methods to determine the appropriate number of fuzzy rules. In terms of parameter learning, this study presents a new hybrid algorithm named MEGWO-RLSE, which updates the If-parts parameters of the model using the Multiple evolving grey wolf optimizer (MEGWO) and the then-parts parameters using the Recursive least squares estimator (RLSE). Additionally, specific proportions of model rule selection and the segmentation of algorithm parameter spaces contribute to the model′s predictive capabilities.This research includes three experiments to validate the performance of the model and methods. The experiments involve predicting the Mackey-Glass time series to verify the effectiveness of SCNFIS, conducting multi-target predictions using stock closing prices to demonstrate the model′s multi-target prediction capability, and studying parameter spaces segmentation in machine learning using different stock targets. Experimental results show that SCNFIS, when trained with MEGWO-RLSE, achieves satisfactory performance. |
