| 摘要: | 本研究針對時間序列問題，建構一個向量化複數模糊神經系統（A vectorized
complex neuro-fuzzy system, VCNFS）。模型引入自我建構（Self-organization）的概念，
以減法分群法（Subtractive clustering algorithm, SC）利用輸入資料提供的資訊決定初始
參數。模糊系統的前鑑部使用複數模糊集合（Complex fuzzy sets, CFSs）產生複數隸屬
度及複數啟動強度，再將每一個啟動強度擴展為一向量，使模型可進行多目標的預測；
本研究分別使用粒子群演算法（Particle swarm optimization, PSO）、鯨群最佳化演算法
（Whale optimization algorithm, WOA）以及結合 PSO 與 WOA 提出 PSO-WOA，和遞迴最
小平方估計法（Recursive Least Squares Estimator, RLSE）形成混合式的演算法，用於模
型參數學習；引入分治法（Divide-and-conquer）的概念，降低單一演算法需調整的參
數量，減低個別演算法負擔以提升效能將。參數分為兩個集合，分別為模糊系統的前
鑑部參數以及後鑑部參數，前鑑部參數使用 PSO 與 WOA 兩種演算法，相互合作調整；
後鑑部參數由 RLSE 調整，使模型可以快速收斂；在測試階段，利用 RLSE 只需最新資料
對就能更新參數的特性，不需要花費太多計算效能，即可將新資料對納入模型中，優
化系統預測效能，降低產生過度擬合的機率。為了驗證提出模型之預測效能，採用金
融時間序列做為資料集設計三個實驗進行預測。實驗結果表明本研究提出之模型與其
它研究文獻相比擁有良好的預測效能。
;For financial time series problems, this study proposes a vectorized complex neuro-fuzzy
system (VCNFS). The proposed model is based on a neuro-fuzzy framework, where there are
several If-Then rules constructed with complex fuzzy sets (CFSs) in terms of a neural network.
The premise parts of If-Then rules are determined initially by the Subtractive Clustering (SC)
algorithm with the information given by data, while the consequent parts are polynomial
functions. Thus, the model is basically realized by the self-organization and data-driven
concept. The use of complex fuzzy sets (CFSs) enables the proposed model to perform multitarget prediction. For model parameter learning, a novel hybrid learning algorithm is proposed,
called the PSO-WOA-RLSE, which combines the Particle Swarm Optimization (PSO) with the
Whale Optimization Algorithm (WOA) and the Recursive Least Squares Estimator (RLSE). The
PSO-WOA-RLSE algorithm uses the divide-and-conquer principle, where the PSO and WOA
cooperate with each other to adjust the parameters of the premise parts of If-Then rules, and
the RLSE to adjust the parameters of the consequent parts, so that the proposed model can
converge quickly with good accuracy. Parameter optimization by the PSO-WOA-RLSE
algorithm separate the parameters that need to be optimized in a single algorithm, so to
reduce the burden of the algorithm, and improves the performance of the model, in terms of
quick convergence and optimization accuracy. In the testing phase, using RLSE to update
parameters with known data after prediction can reduce the probability of overfitting and
optimize the prediction performance of the system. Three experiments are designed in this
study, all of which use financial time-series datasets to verify the performance of the proposed
multi-target forecasting model. The experimental results show that the model proposed in
this study has good prediction performance compared with other research literatures. |
