霍普菲爾類神經網路在強健控制器上的設計應用與實作(I)

Abstract

本計畫的主要目的在探討霍普菲爾神經網路在強健控制器設計上的離線學習作業和及時應用的理論與實際問題。系統原有的控制器因長久使用和損壞導致調校不易，需要被迫更換，然而除了原有控制器和系統的輸出輸入數據外，原有的控制器和系統的設計方法卻難以獲得。因此，本計劃採用霍普菲爾神經網路提出一個新的設計控制器方法。我們使用現存控制器的輸出輸入數據來訓練霍普菲爾神經網路，透過學習後的霍普菲爾神經網路可以取代現存的控制器，並且達到更強健的控制效能。在第一年的計畫中，我們將發展學習功能的演算法及強健程度的分析，並以電腦模擬驗證。然而，如果需要同時使用多個霍普菲爾神經網路時，錯綜複雜的回饋線路將不利於硬體實作，所以在第二年的計畫中，將試著藉由細胞式神經網路架構，來嘗試減低霍普菲爾神經網路的回饋複雜度；以上的要點將在本計畫的第一、二年以電腦模擬的方式完成。本計劃的第三年將以倒單擺和平衡球的硬體實作來驗證本計劃的理論推導結果。

This project is to explore the theoretical and practical issues of off-line training and real-time applications of Hopfield neural network (HNN) as a robust controller for control systems. The controller for a control system could have malfunctioned due to the aging problem, where the details of the controller design are usually not available. Considering the difficulties in reconstructing the controller, we prefer to replace the existing controller by a new one trained by the input and output data of the control system. Hence, this project proposes a new approach of the design of robust controller by using HNN. We will train a HNN using the training data from the existing controller. This trained HNN can then be used to replace the existing controller with better robust performance than that of the existing controller. Both training algorithm with robustness analysis will be performed in the first year by extensive computer simulations. However, if more HNN neurons are required for a complex system, its messy feedback architecture will be infeasible for hardware implementation. Thus a Cellular Neural Network (CNN) architecture for HNN as a robust controller will be investigated. The above issues will be conducted in the first and second years using computer simulations. Finally, the hardware implementation of the trained HNN will be realized by controlling several plants, such as an inverted pendulum system and a ball-and-beam system, in the third year.