標題: | 具線上學習能力之自我建構類神經模糊推理網路 On-Line Self-Constructing Feedforward/Recurrent Neural Fuzzy Inference Networks |
作者: | 莊家峰 Juang, Chia-Feng 林進燈 Lin Chin-Teng 電控工程研究所 |
關鍵字: | 類神經網路;模糊網路;遞迴式網路;基因演算法;監督式褲習;加強式學習;neural network;fuzzy network;recurrent network;genetic algorithm;supervised learning;reinforcement learning |
公開日期: | 1996 |
摘要: | 本論文提出基於監督式及加強式學習之新的類神經模糊推理網路的 建構法。首先,基於監督式學習,我們提出一具有線上學習能力的自我建 構前向類神經模糊推理網路 (SONFIN) 。此SONFIN網路本身為一修正式的 TSK型模糊系統。初始時,網路本身並無法則的存在。法則的產生與調整 乃是由線上同時進行的結構與參數學習來完成。就結構的學習而言,網路 的前件部乃是根據對正型的分群法來作彈性分割。後件部的學習,起初是 依據分群法來給定每條法則的單值 。其後,在必要時,再依據投影相關 量測法來依序加入較重要的元素(輸入變數),這些元素並以線性組合的 型式存在於後件部中。前件部與後件部的學習可產生一有效率、動態自我 增長的網路。此為SONFIN網路的一主要特徵。至於參數調整,後件部可由 最小平方法或遞迴式演算法調整,前件部參數則由倒傳遞演算法調整。結 構與參數學習同時進行的結果,使本網路具快速的學習能力。此為網路的 另一特點。此外,為了加強SONFIN的知識表達能力,可對輸入變數作線性 轉換,如此可減少法則數的使用數目,或提高精確度。這些線性轉換參數 也可在參數學習過程中做動態調整。 其次,為了處理動態的監督式學 習,我們提出了一遞迴式自我建構類神經模糊推理網路 (RSONFIN)。 RSONFIN本身為一遞迴式多層聯結網路並可用來實現動態糢糊推理,因此 可視為由一連串的動態法則所構成。網路的動態關係乃是經由加入表示記 憶元素的迴授聯結到前向類神經模糊網路中而成。初始時,RSONFIN本身 並無任何隱藏結點(即無歸屬函數與模糊法則)的存在。隱藏結點的產生乃 是經由線上同時進行的結構辨別(負責動態模糊法則的建構)與參數辨別 (歸屬函式可調參數的調整)來完成。結構辨別與參數辨別的結合,可產生 了一學習快速、結構小的動態類神經模糊網路。 最後,針對加強式學 習的問題,我們提出以基因演算法為基礎的模糊法則建構法。並以此法來 解決模糊控制器的設計問題。我們所用的基因演算法乃是以共生法為基礎 。當此法應用在模糊系統的設計時,可與模糊法則的區域對應關係緊密結 合。使用所提出的共生演算法為基礎的模糊控制器 (SEFC) 設計法,所花 的控制錯誤嘗試及CPU時間,均比其它基因加強式演算法佳。此外,相對 於一般前件部均採用格子狀切割的基因–模糊系統設計法,本法採用彈性 切割。因此,所須要的法則數較少,且不須事先對輸入變數作切割。模糊 法則的後件部可採用不同的型式,如單值或TSK型模糊法則。 以上所 提的SONFIN 、RSONFIN與 SEFC法均經由電腦模擬加以驗證與比較。其中 SONFIN已被用來做系統辨別、通訊通道等化器設計、水槽溫度控制、渾沌 訊號預測、噪音語言辨識。RSONFIN已被用來做動態訊號預測、適應性噪 音消除、動態系統辨識、動態控制器設計。 SEFC已被用來做倒單擺控制 、磁浮系統控制、水槽溫度控制。這些模擬的結果均驗證了所提方法的效 率與能力。 New methodologies for constructing neural fuzzy inference networksbased upon supervised or reinforcement learning are proposed in this thesis.First, based upon supervised learning,a Self-cOnstructing Neural Fuzzy Inference Network (SONFIN)with on-line learning ability is proposed.The SONFIN is inherently a modified TSK-type fuzzy rule-based modelpossessing neural network's learning ability.There are no rules initially in the SONFIN.They are created and adapted as on-line learning proceedsvia simultaneous structure and parameter identification. In the structure identification of the precondition part,the input space is partitioned in a flexible wayaccording to an aligned clustering-based algorithm.As to the structure identification ofthe consequent part,only a singleton value selected by a clustering methodis assigned to each rule initially.Afterwards, some additional significant terms(input variables) selectedvia a projection-based correlation measure for each rulewill be added to the consequent part(forming a linear equation of input variables)incrementally as learning proceeds.The combined precondition and consequentstructure identification scheme can set upan economic and dynamically growing network,a main feature of the SONFIN.In the parameter identification,the consequent parameters are tuned optimallyby either least mean squares (LMS)or recursive least squares (RLS) algorithms,and the precondition parameters are tuned by backpropagation algorithm.Both the structure and parameter identification are done simultaneouslyto form a fast learning scheme,which is another feature of the SONFIN.Furthermore,to enhance the knowledge representation ability of the SONFIN,a linear transformation for each input variablecan be incorporated into the networkso that much fewer rules are needed or higher accuracy can be achieved.Proper linear transformations are also learned dynamicallyin the parameter identification phase of the SONFIN. Second, a Recurrent Self-cOnstructing Neural Fuzzy Inference Network (RSONFIN)is proposed for dealing with dynamic supervised learning problems.The RSONFIN is inherently a recurrent multilayered connectionist networkforrealizing the basic elements and functions of {\em dynamic} fuzzy inference, and may be considered to be constructed from a series of dynamic fuzzy rules.The temporal relations embedded in the networkare built by adding some feedback connectionsrepresenting the memory elements to a feedforward {\em neural fuzzy} network.There are no hidden nodes (i.e., no membership functions and fuzzy rules) initially in the RSONFIN.They are created on-line via concurrent structure identification(the construction of dynamic fuzzy if- then rules)and parameter identification(the tuning of the free parameters of membership functions).The structure learning together with the parameter learningforms a fast learning algorithm for building a small, yet powerful,dynamic neural fuzzy network.Two major characteristics of the RSONFIN can thus be seen:1) The recurrent property of the RSONFIN makes it more suitable fordealing with temporal problems.2) No predetermination, like the number of hidden nodes, must be given,since the RSONFIN can find its optimal structure and parametersautomatically and quickly. Finally, an efficient genetic reinforcement learning algorithmfor designing fuzzy controllers is proposed in this thesis.The genetic algorithm (GA) adopted in this thesis isbased upon {\em symbiotic evolution} which,when applied to fuzzy controller design,matches well with the local mapping property of a fuzzy rule.Using this Symbiotic-Evolution-based Fuzzy Controller (SEFC) design method,the number of control trials as well as consumed CPU timeare reduced considerablyas compared to traditional GA-based fuzzy controller design methodsand other types of genetic reinforcement learning schemes.Moreover, unlike the traditional fuzzy controllerswhich partition the input space into grids,the SEFC partitions the input space in a flexible wayresulting in fewer fuzzy rules.In the SEFC, different types of fuzzy ruleswhose consequent parts are singletons, fuzzy sets,or linear equations (TSK-type fuzzy rules) are allowed,and the free parameters (e.g., centers and widths of membership functions)and precondition-to-consequent mapping are all tuned automatically. Especially, for the TSK-type fuzzy rule,only the significant terms (input variables) are selectedby the proposed learning algorithmto participate in the consequent of each rule. Capabilities and performances of the proposed SONFIN, RSONSIN, and SEFC methods are verified and comparedthrough various computer simulations.SONFIN hasbeen applied to system identification,communication channel equalization, water bath temperature control,chaotic time series prediction, and noisy speech recognition problems.RSONFIN has been applied to dynamic sequence prediction,adaptive noise cancellation, dynamic identification,and dynamic control problems.SEFC has been applied to cart-pole balancing system,magnetic levitation system,and water bath temperature control problems.Capabilities and performances of the proposed methodsare verified from these simulations. |
URI: | http://140.113.39.130/cdrfb3/record/nctu/#NT850327064 http://hdl.handle.net/11536/61722 |
顯示於類別: | 畢業論文 |