模糊類神經網路之自評自調學習法

Abstract

本研究的目的是要發展一套自評自調模糊類神經網路，使機器能透過學習程序自動建立系統模型或控制器。本文以雙啟發規劃法為基礎配合採用高效率之珈克畢恩(Jacobian)解法推導出自評自調學習法則。在控制應用方面，此自評自調模糊類神經網路能夠從混沌開始終而建立合用的控制器，在建立系統模型應用方面，此網路能透過順序最佳化的學習程序逐漸逼近給予之任意時續函數。本設計採用菅野(Sugeno)一階模糊推論作為基本學習模組，再將其轉換並擴展為類神經網路的學習結構，並建立自評自調學習演算法以自動調整前件和後件的網路參數，終而達成自動學習的目標。本文詳述整個設計的細節，並搭配納倫珈(Narendra)基準系統來驗證此自評自調演算法的成效。最後將此自評自調模糊類神經網路應用於控制旋轉倒單擺的運動，電腦模擬結果顯示此旋轉倒單擺系統能夠從混沌開始學習，終而達成上甩、平衡和追隨行進的所有控制動作。

The goal of this research is to develop an adaptive critic neuro-fuzzy inference system (NFIS) for modeling and control. On the backbone of dual heuristic programming (DHP), a DHP adaptive critic learning scheme that utilizes an effective network Jacobian acquisition is proposed. In control applications, the adaptive critic NFIS can learn from scratch to achieve the control objective. In modeling applications, it can approximate arbitrary continuous function through sequential optimization. The learning structure is based on NFIS that contains fuzzy if-then rules of first-order Sugeno fuzzy model. The tuning rules of premise and consequent parameters are derived. Narendra’s benchmark system is used to verify the performance of the proposed adaptive critic learning algorithm. The ability of modeling is demonstrated by approximating a nonlinear continuous function. The proposed design is applied to obtain the control of a rotary inverted pendulum control. Simulation results show that the rotary pendulum system can learn from scratch to obtain swing-up, balancing and trajectory tracking control.