隨著兩岸交流頻繁，人民幣兌台幣匯率對企業乃至個人的影響與日俱增。本研究應用單隱含層GABP與PSOBP神經網路建立人民幣兌台幣匯率之預測模型。輸入期數設定2期、3期、4期、5期與6期五種，隱含層傳遞函數設定tansig與logsig兩種，共建立10個GABP模型與10個PSOBP模型。從對測試樣本的預測誤差分析得知，GABP模型之預測誤差小於PSOBP模型。輸入期數2期、傳遞函數為tansig之GABP模型具有最小的預測誤差。由於匯率是隨時間動態變化之資料，沒有一種預測模型是絕對有效的，因此同時建立多個預測模型並從中擇優作為最後實際應用之模型是必要的。

With the frequent cross-strait exchanges, the impact of the RMB against the Taiwan dollar exchange rate on enterprises and individuals is increasing. In this study, single hidden layer GABP and PSOBP neural network prediction models of the exchange rate of RMB against the Taiwan dollar are created. The number of input periods has five types: two, three, four, five and six. The transfer function in hidden layer has two types: tansig and logsig. Therefore, 10 GABP models and 10 PSOBP models are established. According to the test sample prediction error analysis, the prediction errors of GABP models are less than those of PSOBP models. The GABP model provided with tansig transfer function and 2 input periods has the smallest prediction error. Since the exchange rate changes dynamically over time, no forecast model is absolutely valid. Thus, creating multiple forecasting models simultaneously and selecting the best model from all created models is necessary.