類神經網路學習演算法收斂之改良及在結構工程之應用

Abstract

本文主要藉由最佳化方法改良倒傳遞網路學習演算法,利用其決定搜尋方 向,藉由黃金分割法和四點三次內插多項式混合求取最佳步長.提供使用者 便利工具,只需決定架構,不須任何參數調整.並由數學範例,非破壞性檢測 範例分別給 NW2 plus和本文改良方法學習和測試,獲得本文方法學習效益 較 NW2 佳.其中最佳化方法,又以 BFGS和CGM較佳.測試成果亦較原作者為 佳.而且在此範例學習和測試情況皆相當好. This reserach applied four different optimal methods to decide the direction for searching .Golden Section Method and Cubic Polynomial Interpolation were used to find the optimal step length to improve the learning efficiency of Backpropagation (BP) learning alogrithm. Once the architecture of BP is decided, no adjustment of any parameters have to be done.We found that the efficiency of learning process by the suggested method is better than BP by NW2 plus. Among all the above optimial methods, the CGM and BFGS are better. It has a good performance on the case of structure analysis, both learning and predicition of sections.