- 作者: 林巍聳; 周祥敏; 王吉祥
- 作者服務機構: 國立台灣大學電機工程研究所
- 中文摘要: 本論文創設一種遞迴式動態類神經網及其學習法則,並用以實現未模式化非線型系統的直接控制,遞迴式動態類神經網的主要特點有三,(一)有線型動態單元可以使類神經網持續的輸出信號、(二)有本體權值可以調整非線型激發函數的增益值、(三)有遞迴單元可以產生關聯記憶的功能。若與前饋式類神經網相比較,遞迴式動態類神經網不需要將先前的輸入、輸出信號,透過外加的元件回饋到類神經網,因此,信號連結線大為減少,類神經網的應用架構也大為簡化。本論文所提出的學習法則,包含一個無監督者的離線學習法則和一個有監督者的線上學習法則,離線學習法則是Hebbian法則的一種應用,目的是將遞迴式動態類神經網內的主要權值,由原始的亂數改進到適當值,使後續的線上學習法則有較好的起始點,因而可以加快學習的收斂速率,也可以避免在啟動直接控制時,因控制信號偏差太大而造成損毀。習用的間接類神經網控制,大都在學習機制中,另加一個受控體鑑定單元,以取得逆向傳遞的學習誤差信號,本論文所提出的學習法則都不需要用到受控體的模式(或Jacobian),因此無需鑑定受控體,也因此可用以解決未模式化非線型系統的直接控制問題。例題的電腦模擬中,以四種不同的控制結構,測試用遞迴式動態類神經網直接控制非線型系統的性能,其中擾動結構的控制模擬,顯示出線上學習法則的韌性,而離線學習法則對設定類神經網起始權值的效果,也由線上學習過程的加速收斂得到印證。
- 英文摘要: The recurrent dynamic neural network (RDNN) and its learning algorithm is proposed torealize direct control of unmodelled nonlinear systems. The RDNN has linear dynamics which sustainthe output, somatic weight to adjust the gain of the nonlinear activation function, and a recurrent com-ponent to perform the function of associative memory. Compared with feedforward neural networks,an RDNN does not need to explicitly feed past inputs and past outputs of the plant to the networkmodel. The proposed learning algorithm includes an unsupervised off-line method to initialize theweights and a supervised on-line method to update the weights. Weight initialization is used toimprove the weights from their random settings so that smooth starting of direct control and fast con-vergence of the on-line learning are achieved. The proposed on-line learning rule is basically agradient method without use of the plant Jacobian. As a result, no a priori information, such as themodel structure, of the controlled plant is required. RDNN-based direct control has been investigatedby means of computer simulations for four different structures to realize inverse control of a non-linear dynamic system. The robustness of the on-line learning rule is examined using a disturbedstructure, and the effectiveness of weight initialization is confirmed by improvement of convergence inlearning.
- 中文關鍵字: direct control by RDNN; recurrent neural networks; nonlinear systems
- 英文關鍵字: --