動態誤差反向傳遞學習法應用於系統識別與控制

Abstract

雖然目前類神經網路的研究多著重於靜態的前饋式網路上，但未來應用動 態網路於控制相關的問題將愈加重要。所謂的動態網路有二種可能情況： 一是具回饋連接的回歸式網路，另一是動態系統中包含前饋式網路。本論 文重點在探討動態網路的學習法則，稱為動態誤差反向傳遞學習法。主要 的研究內容有：(1) 回歸式與前饋式網路於控制相關應用方法的歸納與分 類；(2) 利用動態最佳化方法推導多種動態誤差反向傳遞學習法則；(3) 應用回歸式與前饋式網路於系統識別的方法和技術； (4) 應用回歸式與 前饋式網路於調節與追蹤的方法和技術；(5) 實證比比較回歸式與前饋式 網路應用於非線性動態系統的效果。 While much of the recent emphasis in the connectionist reseaarch has been on feedforward networks with static back- propagation, it is likely that the use of dynamic networks will be of particular importance in control-related applications. This thesis is focused on a learning methodology for recurrent networks with feedback connections and feedforward networks as subsystems in a dynamic system. Such a learning methodology is termed dynamic back-propagation, which is one of the most prominent learning methods for connectionist networks. A detailed study of dynamic back-propagation is presented to provide an insight of the principal ideas that contributed to the evolution of the concept and the details concerning its practical applications to identification and control.