| 摘要: | 本研究使用複數的模糊集合(Complex fuzzy set)取代模糊類神經網路(Fuzzy neural network)中的模糊集合，在演算法上以粒子群最佳化演算法(Particle Swarm Optimization, PSO)為基礎提出了改良的多群粒子群演算法(Multi-group particle Swarm Optimization, PSO)，並且結合遞迴最小平方估計法(Recursive least squares estimator, RLSE)成一個複合式的演算法，此外在模型輸入的選擇上基於夏農資訊熵(Shannon Entropy)提出了特徵選取方法，在本篇論文中將運用特徵選取、複數模糊類神經網路以及混合式的機器學習演算法於股票以及匯率上時間序列的預測。本篇論文使用的特徵選取方法是透過計算特徵對於目標所提供的資訊量，並以一套選取策略針對特徵提供的資訊量多寡進行特徵的挑選。複數模糊集合比傳統一般的模糊集合具有更佳的解釋能力，運用於模糊類神經網路中能夠傳遞更大量的資訊，增加模型預測的效能，並且能夠讓模型能夠同時預測多達6個目標。在模型學習階段，改良式的多群粒子群最佳化演算法能夠比原本的粒子群最佳化演算法更快速的收斂，並且增加找到全局最小化的機率，另外與遞迴最小平方估計法結合能夠減少多群粒子群最佳化演算法需要學習的參數數量，增加多群粒子群最佳化演算法的效能，並且使用遞迴最小平方估計法以計算的方式找出最佳的近似解，而不是透過長時間的訓練，能夠減少模型整體的訓練時間。本篇研究使用股票以及匯率做為多目標的實驗，從實驗結果顯示本論文所使用的特徵選取、模型以及機器學習演算法都有良好的結果。;In this study, complex fuzzy sets are used to replace fuzzy sets in neural fuzzy systems. Based on the particle swarm optimization(PSO) algorithm, an improved multi-group particle swarm optimization(MGPSO) algorithm is proposed, and combined with the well-known recursive least squares estimation (RLSE) into a hybrid algorithm, called the MGPSO-RLSE learning method. In addition, a feature selection method based on Shannon entropy is presented to select useful information by significant features which will be used as model inputs in modeling. In this study, the feature selection, complex neural fuzzy system (CNFS) with Takagi-Sugeno (T-S) If-Then rules and hybrid machine learning algorithm are used for finance time series prediction of stock price and exchange rate. For the CNFS, the parameters of If-parts are evolved by the MGPSO while the parameters of Then-parts are updated by the RLSE. Complex fuzzy sets have better ability to interpret the set-element membership description than regular real-valued fuzzy sets. They can be used in neural fuzzy networks to transmit more information, increasing the prediction performance of model. Moreover, due to CFSs used in the proposed CNFS, the model is capable to deal with up to six targets simultaneously. In the model learning stage, the MGPSO, compared to one single swarm of PSO, can increase the probability of finding the optimal solution, with fast learning convergence. In addition, the combination of the RLSE to the MGPSO can lessen the burden of machine learning by the MGPSO alone. Several real-world data sets of stock prices and exchange rates have been used to test the proposed approach in the experiments for multi-objective prediction. Through the experimental results, the proposed approach has shown good performance. |
