非對稱因果式類神經模糊系統於時間序列預測之研究;Time-series forecasting using neuro-fuzzy system with asymmetric causality

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/81257

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/81257

Title:	非對稱因果式類神經模糊系統於時間序列預測之研究;Time-series forecasting using neuro-fuzzy system with asymmetric causality
Authors:	陳郁晴;Chen, Yu-Ching
Contributors:	資訊管理學系
Keywords:	多目標特徵挑選;人工神經網路;球型複數模糊集;球型複數神經模糊系統;混合式機器學習;multi-target feature selection;artificial neural networks (ANN);sphere complex fuzzy sets (SCFS);sphere complex neuro-fuzzy system (SCNFS);hybrid machine learning algorithm
Date:	2019-07-16
Issue Date:	2019-09-03 15:40:43 (UTC+8)
Publisher:	國立中央大學
Abstract:	時間序列資料的變化有眾多變因，在預測上一直是具有挑戰性的問題和研究。最常應用於股市上的股價變化，從時間的推移中找出股票之間的關係。本篇設計一多目標時間序列預測模型，應用於股票指數的預測。模型結合兩種模型架構，人工神經網路（Artificial Neural Networks, ANN）及球型複數神經模糊系統（Sphere complex neuro-fuzzy system, SCNFS），以進行多目標運算。本篇在SCNFS中加上箭靶層（Aim object layer），形成非對稱因果式的SCNFS。資料前處理使用多目標特徵挑選，以亂度熵（Entropy）的概念，從龐大的資料集中挑選出對目標具有貢獻的輸入資料。機器學習的部分使用混合式機器學習演算法，結合無導數最佳化演算法（Derivative-free optimization algorithm）及遞迴最小平方估計法（Recursive least squares estimation, RLSE），有效訓練模型的參數。本篇共進行三個實驗，實驗一以相同資料集同時進行單目標預測及多目標預測，驗證模型預測多個目標之能力。實驗二透過不同的機器學習演算法，進行相同資料集與模型架構的多目標預測，研究不同演算法之訓練效能。實驗三以相同的資料集及機器學習演算法，進行不同模型架構下之多目標預測，研究模型架構的優劣。;Time-series forecasting is a challenge problem in research and application because there are many affecting factors in it. The most common application of time-series is the stock market. This research designed a kind of time-series forecasting model for multi-target to predict the stock index. In order to support multi-target operation, this model combines two kinds of structure: artificial neural networks (ANN) and sphere complex neuro-fuzzy system (SCNFS). The different part of SCNFS in this research is aim object layer (AOL). It can make SCNFS with asymmetric causality. In the part of data preprocessing, we use multi-target feature selection which extends from the concept of the entropy. It can select the features which contribute to the whole target from the large datasets. Moreover, we use hybrid machine learning algorithm, which efficiently trains the parameters through the derivative-free optimization algorithm and recursive least squares estimation (RLSE). At the end of this research, we did three experiments. The first experiment is to compare the results of single-target and multi-target forecasting trained by the same datasets, and to investigate whether our model can predict multiple target effectively. The second experiment uses the same datasets and model structure with different machine learning algorithms, and research the performance of different algorithms. The third experiment is to understand the performance of different model structure with the same datasets and machine learning algorithms.
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	145	View/Open

社群 sharing

Loading...