| 摘要: | 本論文研究中之預測模型為複數模糊類神經模型，藉由複數模糊集(Complex fuzzy set)取代傳統模糊類神經模型(Fuzzy neural network)中使用的傳統模糊集，並採用混合式機器學習進行模型參數學習，包含提出以粒子群最佳化(Particle Swarm Optimization, PSO)與隨機最佳化(Random Optimization, RO)並行運作之演算法，以及遞迴最小平方估計法(Recursive least squares estimator, RLSE)。另外，資料進入模型學習前，先經由基於夏農資訊熵(Shannon Entropy)的特徵選取方法，選出對目標有影響力之特徵作為模型輸入。特徵選取方法是藉由計算特徵對於目標提供的資訊量多寡，來進行特徵的挑選。複數模糊集合比傳統的模糊集合具有更多的空間以附載更多的資訊，運用於模糊類神經網路時，使在神經網路內部傳遞訊息時，能夠包含更大量的資訊，提升模型預測準確度，且藉用複數的性質，模型能夠進行多目標的處理。在機器學習階段，藉由粒子群最佳化與隨機最佳化的並行運作，並透過競爭與學習的策略，增加找到更佳解的機率，再加上與遞迴最小平方估計法作結合，來提升模型運算效率。在模型預測方面，本論文以股票與匯率作為實驗對象，並從實驗結果顯示本論文提出之混合式機器學習、特徵選取與模型都有良好的表現。;In the study, the predictive model is a complex fuzzy neural model. The complex fuzzy sets are used to replace the traditional fuzzy set used in the traditional fuzzy neural network. Based on parallel operation with the particle swarm optimization (PSO) algorithm and the random optimization (RO) algorithm, an improved algorithm is proposed, and combined with the recursive least squares estimation (RLSE) into a hybrid machine learning algorithm, called the RoPso-RLSE learning method. In addition, a feature selection method based on Shannon entropy is presented to select useful features which will be used as model inputs in modeling. In this study, the feature selection, complex neural fuzzy system and hybrid machine learning algorithm are used for time series prediction of stock price and exchange rate. The feature selection selects features by calculating the information provided by the features for the targets. Complex fuzzy sets (CFSs) have better description for set-element relationship than tradition fuzzy sets in membership. They can be used in neural fuzzy networks to transmit more information and increasing the prediction performance of model. Moreover, due to the property of CFSs, the model can perform multi-target forecasting simultaneously. In the machine learning stage, the hybrid algorithm RoPso, compared to use single PSO or RO only, can increase the probability of finding the optimal solution, with fast learning convergence. In addition, combining the RLSE with RoPso can reduce the loading of machine learning by the RoPso 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. |
