模糊類神經系統結構之自我組織技術

Abstract

在設計模糊類神經系統時，一個最基本的問題是系統結構的制定。要制定 一個兼具符合系統需求、並且同時擁有良好效能的結構是非常困難的。在 本篇論文中，我們將提出一種新的方法來解決此一問題。由於遺傳演算法 在尋找最佳解的應用方面擁有良好的效力，因此，我們利用此一特點，來 搜尋一個模糊類神經系統的最佳結構。然而，由於遺傳演算法的演化過程 僅僅利用先天所遺傳的資訊來求得最佳解，所以它花費相當多的時間在搜 尋的過程中。鑑於此一缺點，我們提出了一種結合先天遺傳資訊，與後天 學習知識的演化方法，稱作拉瑪克遺傳演算法。此一方法是藉由將後天所 學得的知識寫回基因中，以達到加速遺傳演算法搜尋最佳解的目標。我們 將所提出的方法應用在糖尿病以及肌肉萎縮症的診斷方面，實驗結果顯示 我們所提出的方法確實能加速遺傳演算法演化的速度，並且具有良好的效 能。 A basic problem, structure identification, in the design of a neuro-fuzzy system is explored in this thesis. In general, it is hard to identify a proper topology for a neuro-fuzzy system to achieve the best performance. We propose an approach to solve this problem. It is known that genetic algorithms (GAs) are an effective search method to find an optimal solution. We use it to accomplish the goal of structure identification of a neuro-fuzzy system. However, a GA is usually time-consuming due to that the evolution of the GA depends solely on its innate knowledge encoding. To cope with this problem, we propose a Lamarckian genetic neuro-fuzzy model which enables the inheritance of acquired knowledge in the evolution of GA. Taking advantage of feeding new knowledge back to genes, the GA search can speed up to find out the global optimum. We apply our approach to medical diagnosis problems: diabetes, dystrophy, and urodynamics. Simulation results show that the proposed model is an efficient method for achieving high performance.