查詢結果分析
來源資料
相關文獻
- 模糊可微小腦模型類神經網路在函數逼近器之應用
- Development of a Sensory Data Glove Using Neural-Network-Based Calibration
- Neural Network Procedures for Taguchi's Dynamic Problems
- A Fast and Efficient Competitive Learning Design Algorithm Based on Weight Vector Training in Transform Domain
- 專家系統振動訊號圖型判別之研究
- 反傳遞模糊類神經網路於流量推估之應用
- 類神經網路(Neural Networks)的種類及其在影像處理上的應用
- C++Fuzzy類神經網路物件導向發展系統之建立
- 臺灣汽保費率之估計--對數線性費率模式與類神經網路之比較
- 運用類神經網路於股價指數之套利--以日經225指數為例
頁籤選單縮合
題名 | 模糊可微小腦模型類神經網路在函數逼近器之應用=Fuzzy Differentiable CMAC Neural Network and it's Application on Function Approximator |
---|---|
作者 | 蔡樸生; 林盈灝; 胡厚楨; Tsai, Pu-sheng; Lin, Ying-hao; Hwu, Hou-jen; |
期刊 | 中華技術學院學報 |
出版日期 | 20071200 |
卷期 | 37 2007.12[民96.12] |
頁次 | 頁99-116 |
分類號 | 448.5 |
語文 | chi |
關鍵詞 | 類神經網路; 函數逼近器; 模糊可微分小腦模型; 學習機制; 最陡梯度法; Neural network; Function approximator; Fuzzy differentiable cerebellar model articulation controller; Learning mechanism; Steepest descent algorithm; |
中文摘要 | 本文提出一個新型類神經網路的結構,稱為模糊可微小腦模型。在小腦模型的結構下,結合模糊邏輯系統以及最陡梯度法的學習機制來逼近一個非線性的函數或系統。小腦模型網路是屬於一系列對映查表法結構,具有快速的學習收斂、良好的逼近效果、極佳的類化特性、結構簡單、容易以硬體或FPGA晶片實現等特性。由於傳統小腦模型中的超立方塊是嵌入均勻分佈函數,使得類神經網路的輸出在每一個分割區塊間均為等值,具有不可微分的特性。本文所提的模糊可微小腦模型是將可微函數(如高斯函數)嵌入感受場空間的超立方塊中,形成非定值且可微分的區塊。此外,網路結構中引入模糊推論引擎來取代聯想記憶體的位址索引矩陣,並且以模糊歸屬函數來描述超立方塊與真實記憶體的對映關係。新型的網路結構不僅保留原有小腦模型的特點,藉由可微分的特性,傳統的權重均分法已經不是學習機制的唯一方法,取而代之是以最陡梯度法來推導最佳化參數或權重記憶體的內容。透過模擬結果證實,本文所提的網路結構不僅具有行性,對於非線性函數也具有很好的逼近效果以及收斂性能。 |
英文摘要 | In this paper, a modified structure of fuzzy differentiable cerebellar model articulation control (FD_CMAC) is proposed for dealing with the function approximation problems. A FD_CMAC artificial neural networks that merges fuzzy logic system and steepest descent algorithm with fast learning characteristic, good generalization capability, and convergence property can be established. Its simple structure also enables the A FD_CMAC to be easily realized by hardware of FPGA. Based on the learning mechanism of the human cerebellum, a trained CMAC can approximate nonlinear function in a generalized lookup table style over a domain to any desired accuracy such that the satisfactory performance can be obtained. In contrast, the traditional CMAC uses rectangular basis function in the receptive field, its output keeps always constant within each quantized state and not differentiable. To overcome this problem, the A FD_CMAC adopts a differentiable Gaussian basis function embedded in each hypercube structure such that the derivative information of input and output variables is validation. Due to the differentiable property of A FD_CMAC, the steepest descent method can be applied to derive the learning algorithm. By such a new learning algorithm, the learning speed and converge rate is promoted apparently. To demonstrate the performance of the proposed A FD_CMAC, it is shown that the performance is better than traditional CMAC and simple structure is more suitable in practical implementation. |
本系統之摘要資訊系依該期刊論文摘要之資訊為主。