論文提要內容： 早期求解結構最佳化問題時，大多採用數學規劃法，此方法需要計算繁雜的梯度函數，而且求得的解答往往僅為局部最佳解；而類神經網路法有別於此類方法，是一種平行分散處理的計算模式，其所建構的分析模型具有非線性特性，可獲得較一般迴歸分析更準確的結果。為此，本文以類神經網路法進行鋼結構最佳化設計，一方面藉此瞭解類神經網路法在求解結構最佳化問題之適用性，一方面藉以建立梁柱構件及梁柱構架等鋼結構之最佳化設計模式。 本文針對梁柱桿件及梁柱構架等鋼結構，首先利用鋼結構設計手冊內的常用H型鋼斷面，以及由均勻隨機亂數產生的可變H型鋼斷面建立兩種不同測試集，再以自編軟體進行測試樣本之結構強度計算，以ETABS軟體進行結構分析，以類神經網路法建立預測模型，最後以CAFE程式進行結構最佳化設計。本文所採用的CAFE程式是一種基於類神經網路法與實驗計畫法的最佳化設計系統。研究成果顯示，運用本文之最佳化設計模式搭配CAFE程式確實可以獲得比以往文獻更輕的設計成果。本文研究成果大幅提升類神經網路法應用於結構最佳化之實用性。 In the past, the mathematical programming method is often used to solve the structural optimization problems. This method requires calculating the complex gradient function and sometime its answer is only a local optimum. The artificial neural network differs from this one. It is a parallel distributed processing mode of calculation. It can obtain more accurate results than those of regression analysis because its analytical model that has characteristics of nonlinear. Therefore, this thesis uses the artificial neural network method to optimize the design of steel structures. On the one hand, it is used to understand artificial neural network method in solving optimization of structures problems of applicability, on the other hand it is used to establish the optimal design patterns for beams, columns and frame structures of steel. Firstly, for beams, columns and frame structures of steel, this thesis uses two types of H-beams section to establish two different test sets. One type of the H-beams section is commonly used in the steel-structure design manual and the other is generated by the uniform random number. Next, this thesis composes the software to calculate the structural strength of the test samples, then using ETABS software for structural analysis and artificial neural network to build predictive models. Finally, the optimum structural design results are obtained by using CAFE software. The CAFE software used in this thesis is an optimization design system based on artificial neural network and design method of experiments. The result of the thesis have shown that using the design pattern in this article together with the CAFE software will result in getting a lighter structural design than the previous literature result. This research has significantly improved the practicality of using the artificial neural network on structural optimization.