本研究針對時間序列預測之問題，提出了一個改良式的複數模糊類神經系統。改良的地方在於本研究根據模型輸出與目標間的誤差提出了一種誤差回饋的方式，藉由將誤差回饋到模型的後鑑部以改善模型預測時間序列的準確度。在模型輸入的選擇上，本研究基於夏農資訊熵(Shannon Entropy)的理論提出了多目標特徵挑選的方法，其主要是計算特徵對於目標所提供的資訊量，並以一套特徵選取策略針對特徵提供的資訊量來進行特徵的挑選。在模型建立上是使用結構學習(Structure learning)的方式，以數據本身客觀地建構模型，改變傳統以主觀的方式決定模型的建立。複數模糊集合(Complex fuzzy sets, CFSs)比以往傳統的模糊集合具有更佳的解釋能力，運用於模糊類神經系統(Neural fuzzy system)中能夠傳遞更大量的資訊，增加模型預測的效能，並且能夠讓模型同時預測多個目標。在模型參數學習階段，將粒子群最佳化演算法(Particle swarm optimization, PSO)與遞迴最小平方估計法(Recursive least squares estimator, RLSE)結合形成PSO-RLSE複合式演算法。PSO用於模型前鑑部之參數優化，而RLSE則負責模型後鑑部之參數更新，此種分而擊之的概念能夠降低模型參數的維度，提升找到模型參數最佳解的機率，並能降低模型整體的訓練時間。本篇研究使用財經市場的時間序列進行多目標預測的實驗，並從實驗結果顯示本論文所提出之模型與其他文獻相比有較好的預測結果。;This study proposes a modified complex fuzzy neural system for the problem of time series prediction. The modified way is that this study proposes an error feedback method according to the errors between the model outputs and the targets. This paper feeds errors back to the model′s consequent part to increase the model prediction accuracy. In the selection of model input, this study proposes a multi-targets feature selection method based on the theory of Shannon entropy. The main purpose is to select features by calculating the amount of information provided by the features for the target, and using a set of feature selection strategies to do the multi-targets feature selection. In the establishment of the model, this paper changes the traditional subjective way to objective way by using the training data itself to construct the model setting. Complex fuzzy sets (CFSs) have better interpretive capabilities than traditional fuzzy sets. They can deliver a larger amount of information than traditional fuzzy sets in neural fuzzy systems, increase the effectiveness of model prediction, and let the model predict multi-targets at the same time. In the model parameter learning phase, Particle swarm optimization (PSO) and Recursive least squares estimator (RLSE) are combined to form a PSO-RLSE hybrid algorithm. PSO is used to optimize the parameters of the model′s premise part, while RLSE is responsible for the parameters of the model′s consequent part. This concept of divide-and-conquer can reduce the dimension of the model parameters, increase the probability of finding the best solution of the model parameters, and reduce the overall training time of the model. This study uses time series of the financial market to conduct the multi-targets forecasting experiments. The experimental results show that the model proposed in this paper has better forecasting abilities than other literatures.
