時間序列分析與機器學習是當今重要議題，隨著大數據時代來臨，資料與特徵維度的增加使得模型在運算更具挑戰性。因此，本次研究將會專注在如何以有限資源減緩資料維度與參數量對模型的風險，提出了一種新的全複數模糊模型（SCFM），利用複數型態結合輸入維度來達到降維的目的，進而減少模型計算複雜度與計算成本，其中，以投影式球形複數模糊集的複數輸出達到雙目標預測。其次，本研究基於傳遞熵的概念改良多目標特徵選擇演算法，為模型提供更有力之特徵，透過輸入空間最適區塊挑選法（GSIS）挑選可用區塊降低模型參數與規則數量，並於機器學習中提出一個新的混合式演算法，稱作GD-WOA-RLE，混和兩種演算法，以高斯鯨群演算法（GD-WOA）更新模型前鑑部參數及以遞迴最小平方估計法（RLSE）更新後鑑部參數。本次研究也透過三個實驗來驗證模型效能，包含兩種單目標預測：以函數逼近檢驗GSIS效能、透過預測台灣積體電路製造（TSMC）的股價資訊驗證所提出之特徵選取的有效性與優勢，最後同時預測貨幣市場中USD2EUR與CNY2EUR的匯率。從實驗結果，驗證了SCFM配合GD-WOA-RLSE在單、雙目標的預測上均有不錯的效能，並且能夠透過減少參數與模型大小使得模型有更好的表現。;Time series analysis and machine learning are currently important issue. With the rise of the big data, the increase in the number of data and feature dimensions presents challenges in model computation. Therefore, this study focuses on mitigating the risks of high input dimensionality and parameter quantity on models with limited resources. We propose a novel approach called the Sphere Complex Fuzzy Model (SCFM), which combines complex-valued inputs to achieve dimensionality reduction to reduce model computational complexity and costs. The SCFM incorporates the use of Sphere Complex Fuzzy Sets with Projection as the fuzzy sets for dual-target prediction. Furthermore, we enhance the multi-target feature selection algorithm based on transfer entropy, providing more powerful features for the model. We introduce the grid-type selection of input space (GSIS) to select suitable blocks within the input space, reducing the number of model parameters and rules. Additionally, we propose a hybrid algorithm, GD-WOA-RLE, which combines the gaussian distribution-based whale optimization algorithm (GD-WOA) for updating the parameters of If-parts and the recursive least squares estimator (RLSE) for updating the then-parts parameters. We conduct three experiments to evaluate the performance of model, including function approximation using GSIS, predicting stock prices of TSMC, and forecasting exchange rates of USD2EUR and CNY2EUR. The experimental results demonstrate the effectiveness and advantages of the proposed feature selection and the performance of the SCFM with GD-WOA-RLE in single- and dual-target predictions. By reducing parameters and model size, the proposed approach achieves improved model performance.
