由於以高階演算法進行結構最佳化設計時，往往會因參數設定因素或收斂速度太快與太慢，導致最終結果不一或僅找到局部最佳解，一般常需另外加入其他策略方可減少計算量及獲得較佳結果。類神經網路法有別於這些方法，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，它已廣泛應用於各個領域。 本文針對桁架與構架結構，以SAP2000軟體，配合具有交叉驗證法與訓練測試法之類神經網路法CAFE軟體，進行結構最佳化設計。本文對桁架結構提出以滿載應力法進行零桿篩選，且利用中心斷面快速估算式求得搜尋空間的中心位置，再於該中心位置訂定適當範圍隨機取樣分析；對構架結構則提出以試誤法估得鄰近限制條件之斷面參數範圍，再於該斷面參數範圍中進行隨機取樣分析，藉以減少樣本數目、迭代次數與計算時間。 研究結果顯示，利用CAFE軟體及本文所建議的隨機取樣方法，確實能夠大幅減少結構最佳化計算時間，並得到與文獻相近似甚至較佳的最佳解；而由CAFE軟體所得影響限制條件輸入變數之權重，顯示此二種樣本設計方式極為正確，大幅提升類神經網路法應用於結構最佳化之實用性。 The advanced algorithms for structural optimization design often leads to problems which the final result is not unique and the result is local optium due to the influence of parameter setting or too fast and too slow convergence speed. In general, adding other strategies in analyzing is often necessary in order to reduce the amount of calculation and obtain better results. Artificial neural network (ANN), a parallel distributed processing computing model which differs from advanced algorithms can construct nonlinear characteristics, which creates more accurate results of regression analysis and it has been widely used in various fields. CAFE software has been used with SAP2000 in this study for optimum design of truss and frame structures. There are two methods of samples collecting design recommended in this research for both truss and frame structures. First method proposes sifting out all members from the zero bars by FSD, and manifesting a quick central section evaluation formula by unit load method. Central sections for other members except zero bars are calculated with the formula. As for truss structure, the central position in the search space is first established, then an appropriate range from central position, which it finds samples randomly. The second approach is a trial and error method, which a range of parameters near to constraint criteria is used, and random sampling is taken as well. A number of procedures were done in order to reduce the number of samples, the number of iterations and computation time. The results have shown that the use of both the CAFE software and the random sampling method can indeed significantly reduce the structural optimization calculation time and get better results compared to the literature. Due to the CAFE software we can find the significances of input variables that impacts constraints, this shows that two kinds of sample design approach is most favorable, significantly increasing the neural network method to structural optimization of practicality.