倒傳遞神經網路雖然可以建構準確的非線性模型，但它屬於黑箱模型，無法定量衡量每個輸入變數的重要性。本研究修改了傳統倒傳遞網路演算法，在輸入層的每一個輸入單元加上「變數重要性指標」來定量衡量該輸入單元對模型的重要程度，並以最陡坡降法推導出其修正量公式，使得在學習過程中能調整其值，而能正確衡量輸入變數的重要性。經由三個模擬的數值例題顯示本方法確實能正確衡量線性函數、二次函數、交互作用函數的輸入變數重要性。此外，一個真實
的應用實例証實，依本方法得到的輸入變數的重要性之排序來選取變數，可以用最少量的變數建構準確的分類模型，顯示本法所得到的輸入變數的重要性之排序是正確的。

Although back-propagation neural networks (BPN) may construct accurate
non-linear model, they belong to black box models, and can not quantitatively measure the importance of each input variable. This study modified the conventional BP algorithm, and Variable Importance Index (VII) was added on each input unit of input layer to quantitatively measure its importance to model, and the learning rule for the
indexes was deduced with the gradient steepest descent method, which can adjust the indexes in the learning process, and then they can correctly measure the importance of input variable. By way of three artificial numerical examples, it is demonstrated that this method can correctly measure the importance of input variable of linear, quadratic,
and interactive function. In addition, a real application example confirmed that selecting the input variables according to the input variable importance indexes can build the accurate classification model with the fewest number of input variables, which demonstrated that the sort of importance of input variable obtained by the method is correct.