標題: 模糊神經系統之合作式模糊規則轉換機制
A Collaborative Fuzzy Rule Transfer Mechanism for Neural Fuzzy Inference Systems
作者: 木克思
mukesh prasad
林進燈
Lin, Chin-Teng
資訊科學與工程研究所
關鍵字: Preprocessed Collaborative Fuzzy Clustering;Fuzzy Rule Transfer;Neural Network;Fuzzy C-Means;Big Data;Privacy and Security;Self-constructive Neural Fuzzy Inference Network;Preprocessed Collaborative Fuzzy Clustering;Fuzzy Rule Transfer;Neural Network;Fuzzy C-Means;Big Data;Privacy and Security;Self-constructive Neural Fuzzy Inference Network
公開日期: 2015
摘要: 在本論文中,提出一種前處理協同模糊群聚(Preprocessed Collaborative Fuzzy Clustering, PCFC)之技術來完善從模糊C均值群聚和捲入協作流程前所獲得的原型和分割矩陣的不一致性。前處理協同模糊群聚對協同模糊群聚和提供較佳的數據可視化間產生的問題提供了一種解決方案。因為基於隱私和安全問題,一般在協同的過程中一個組織不能直接與其他組織共享數據資訊。但是前處理協同模糊群聚卻幫助分割矩陣和原型之形式的不同組織間分享數據資訊,並在維持數據的隱私和安全的同時依然可以得到一個滿意的結果。 本論文亦針對類神經模糊和模糊推論系統提出了一種新穎協同模糊規則轉移機制。首先使用模糊C-均值產模糊規則生,然後使用前處理協同模糊群聚來適應和更新。第二步,更新規則常被用來決定類神經模糊系統的結構學習面向和模糊推論系統的知識庫子系統。第三步,在參數學習的表現上則透過不用選擇類神經模糊系統的初始化參數來更新模糊規則和使用If-Then型式的模糊規則將模糊輸入轉換至模糊輪出的模糊推論子系統。本論文提出的方法能夠處理巨大的數據集,並同時保留數據集的隱私和安全。完整的數據集被分割成幾個相等的數據集的子集給前處理協同模糊群聚這步驟,其中每個數據集的子集都是單獨群聚。針對本文提出的方法,最初完整數據集被組織成兩個獨立的數據集的子集並且透過協同技術來決定原型(聚類中心)的知識和數據集的子集之分區矩陣。本論文提出的方法可以實現在前處理協同模糊群聚的集體知識出現的一致性,並且透由神經模糊推理網絡的參數學習能力來提昇系統建模過程。在進行時間序列預測問題上,本論文提出的方法也優於其它現有的方法。 更進一步,本論文展現一種使用前處理協同模糊群聚來結合Mamdani型式與TSK(Tagaki-Sugeno-Kang)型式模糊推論系統的新系統建模範例。在本論文提出的方法中,前處理協同模糊群聚通常用提取一組規則去取代Mamdani型式模糊推論系統中模糊C-均值分群的使用和TSK型式模糊推論系統中刪減分群法的使用來模組化資料行為。本論文提出的方法結合了前處理協同模糊群聚的知識學習能力與含有Mamdani型式與TSK型式模糊推論系統建模實力的模糊C-均值之規則學習能力來對於給定的問題集合提供精確的系統模型。 本論文所提出的方法幫助了解與想像系統結構建模的資料分析。 此外,本論文展現一種改進過的自我建構類神經模糊推論網路(Self-constructing neural fuzzy inference network, SONFIN),稱之為溫和提昇(Soft-boosted) 自我建構類神經模糊推論網路 (SB-SONFIN)。該網路輕輕地提昇自我建構類神經模糊推論網路的學習過程來達到較低錯誤率的高速學習。因為模糊規則的數目和初始化權重是自我建構類神經模糊推論網路兩個重要的因素,溫和提昇自我建構類神經模糊推論網路透過兩種方式增強自我建構類神經模糊推論網路的學習能力: (1)用模糊集合的寬度而採用隨機數值來初始化權重; (2) 使用學習過的模糊規則數量來溫和地提昇參數學習率。本論文提出的溫和提昇方案的有效性已在各種現實世界和基準數據集進行驗證。實驗結果顯示在給定的數據集中溫和提昇自我建構類神經模糊推論網路優於其它已知方法.
In this study, a preprocessed collaborative fuzzy clustering (PCFC) technique is proposed to refine the inconsistency of the prototype and partition matrices obtained from fuzzy c-means clustering before getting involved in the collaboration process. The PCFC provides a solution to the problems related to the collaborative fuzzy clustering and gives better data visualization. In the collaboration processes, one organization cannot share data information directly with other organization, due to the privacy and security issues. However, the PCFC helps to share data information within different organizations in the form of partition matrix and prototypes while maintaining the privacy and security of the data along with the ability to yield a satisfactory result. This study also proposes a novel collaborative fuzzy rule transfer mechanism for neural fuzzy and fuzzy inference systems. First, the fuzzy rules are generated facilely by the fuzzy c-means (FCM), and then adapted and updated with preprocessed collaborative fuzzy clustering. Second, the updated rules are used to decide the structure of the learning phase of neural fuzzy systems and knowledge-based sub-system of the fuzzy inference system. Third, the parameter learning is performed with the updated fuzzy rules without selecting initial parameters for the neural fuzzy systems and consequently the inference sub-system uses If-Then type fuzzy rules to convert the fuzzy input to the fuzzy output for fuzzy inference systems. The proposed method is capable of dealing with immense datasets while preserving the privacy and security of these datasets. The entire dataset is divided into several equally-sized subsets of datasets for the PCFC procedure, where each of the subsets of the dataset is clustered separately. For the proposed method, initially the entire dataset is organized into two individual subsets and the knowledge of prototype (cluster centers) and the partition matrix of these subsets are deployed through the collaborative technique. The proposed method is able to achieve consistency in the presence of the collective knowledge of the PCFC and boosts the system modeling process by parameter learning ability of the neural fuzzy inference networks. The proposed method outperforms other existing methods for time series prediction problems. Further, this study presents a new system modeling paradigms for Mamdani type and Tagaki-Sugeno-Kang (TSK) type fuzzy inference systems (FIS) with combination of PCFC. In this proposed method, PCFC is used to model the data behavior by extracting a set of rules instead of using fuzzy c-means clustering for Mamdani type FIS and subtractive clustering for TSK type FIS. The proposed method combines the knowledge learning capability of the preprocessed collaborative fuzzy clustering and rule learning ability of the FCM with the modeling strength of Mamdani type fuzzy inference system and Tagaki-Sugeno-Kang type fuzzy inference systems to provide an accurate model of system for given sets of problems. The proposed method helps to understand and visualize the data analysis for structural modeling of systems. Also, this study presents the Soft-boosted self-constructing neural fuzzy inference network (SB-SONFIN), which is an improved version of the self-constructing neural fuzzy inference network. The SB-SONFIN softly boosts up the learning process in order to reach a lower error rate with higher learning speed. Since the number of the fuzzy rules and initial weights are two important factors for the SONFIN, the SB-SONFIN enhances the learning power of the SONFIN by two ways: (1) it initializes weights with widths of the fuzzy set rather than just with random values; (2) it softly boosts the parameter learning rates with the number of learned fuzzy rules. The effectiveness of the proposed soft boosting scheme is validated on various real world and benchmark datasets. The experimental results show that the SB-SONFIN outperforms other known methods on given datasets.
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT079955855
http://hdl.handle.net/11536/127583
顯示於類別:畢業論文