應用類神經網路於電腦輔助大量估價之研究

Applying the Artificial Neural Network in Computer-assisted Mass Appraisal

政府機關之不動產估價作業主要是提供課稅地價為目的而衍生之行政工作，目前台灣主要以路線估價作業方式處理公部門地價，因此往往需要投入大量的人力、經費。此與歐美等國普遍應用的電腦輔助大量估價作業（computer assisted mass appraisal, CAMA）有極大差異。90年代初期，由於資訊產業的快速成長，利用電腦模擬人類思考模式，而發展出來的類神經網路（artificial neural network, ANN）演算方法被廣泛地運用於各種不同層面的研究。直到90年代後期才慢慢的被運用在不動產估價。本研究將分別運用特徵價格及倒傳遞類神經網路預測高雄市不動產價格，試圖建立一套大量估價模型。經研究實證分析得知，在總體樣本數時，倒傳遞類神經網路預測較特徵價格法之預測能力較佳。但是如果將樣本區分為90%樣本內及10%樣本外資料，特徵價格法之預測能力較佳，而這樣的實證結果說明，在運用各式估價模型時，可以進行交互驗證並且從中找出最適估價模型。

In early times, the land value assessments in Taiwan were made manually and wasted too much manpower, which was very different from the computer-assisted mass appraisal approach adopted in Western countries. In the early 1990's, due to the development of information technology, many researchers imitated the functioning of the human brain to develop the neural network and it was applied in different areas. In the late 1990s, the back-propagation neural network (BPN) was applied to real estate appraisal. This study applies the back-propagation neural network and hedonic price method to predict real estate prices in Kaohsiung city. We evaluate the model performance of the BPN and hedonic price in forecasting Kaohsiung's property prices. Two criteria are used, namely, the mean absolute percentage error (MAPE) and the forecasting error (FE). Regardless of which of the BPN approach or the hedonic price model is used, both are found to have similar forecasting power.