| 摘要: | 股票的波動是一種時間序列的資料。時間序列的預測是一個重要的研究議題，人工智慧計算模型目前正被廣泛使用於該議題，例如:類神經模糊系統等。本文提出複數型模糊類神經系統 (Complex neuro-fuzzy system)並應用於多目標時間序列預測，此模型具有多組複數型態輸出，其中，每一組複數型態的輸出，其實部和虛部可分別針對兩個不同實數型態目標進行預測。有關特徵挑選，本研究採用多目標特徵挑選，篩選出針對所有目標有利的特徵，並以此作為模型輸入，以降低模型整體運算負擔及提高資料運用效率。在模型方面，由輸入層、複數模糊集合神經層 (Complex fuzzy sets layer)、前提式神經層 (Premise neural layer)、T-S神經層 (Takagi-Sugeno neural layer)及輸出層，建構出多層式類神經網路。在參數學習方面，訓練模型時我們採用分治原則(Divide-and-conquer principle)。複數模糊集合神經層的參數使用不同的演算法優化，像是粒子群演算法 (Particle swarm optimization, PSO)、人工蜂群演算法 (Artificial bee colony optimization, ABCO); T-S神經層的參數使用遞迴式最小平方演算法 (Recursive least-squares estimation, RLSE)進行優化; 其他的神經層沒有參數需要優化。在實驗方面，我們設計三個實驗檢驗模型的效能，將PSO-RLSE及ABCO-RLSE實驗結果結合投資策略，計算模型利潤互相比較也與不同的文獻方法比較。本研究提出新的投資策略，與過去做利潤的比較，經由效能及利潤比較結果，本文提出多目標預測的研究方法表現出優秀效能以及投資效果。;Stock fluctuations are time series data. The prediction of time series is an important research topic. Artificial intelligence models are currently being widely used in this topic, such as neuro-fuzzy systems. This paper proposes a complex neuro-fuzzy system and applies it to multi-target time series prediction. This model has multiple complex-valued outputs, every output can have real and imaginary parts for two different real-valued targets, respectively. With regard to feature selection, this study uses multi-target feature selection to filter out features that are beneficial to all targets and use this as the model inputs to reduce the overall computational burden and improve data utilization efficiency. In terms of model, multi-layer neural network is constructed from input layer, Complex fuzzy set layer (CFS layer), Premise neural layer, Takagi-Sugeno neural layer (T-S neural layer), and output layer. For parameter learning, we use the divide-and-conquer principle when training the model. The parameters of the complex fuzzy set neural layer are optimized using different algorithm, such as particle swarm optimization (PSO), artificial bee colony optimization (ABCO); the parameters of the T-S neural layer are optimized using recursive least-squares estimation (RLSE), other neural layers have no parameters to optimize. In terms of experiments, we use three experiments to test the performance of the model. We combine investment strategy with PSO-RLSE and ABCO-RLSE experimental results, respectively, and calculate model profit to compare with each other and the different literature methods. This study proposed a new investment strategy. Through the results of the performance comparison and the profits comparison, this paper presents a multi-target prediction method showing excellent performance and investment effect. |
