中文摘要 |
不動產價格的高低與土地徵收補償、土地開發的成本及收益、房地產投資等密切相關,因此,如何準確地估算不動產的價格一直是地政相關領域所關注之焦點。以往已有不少結合不動產大量估價與人工神經網路之研究,但所建構的模式大多以倒傳遞神經網路為主,甚少考量其他的網路模式,而且,其網路架構多採隨機指定,未能有系統地比較和分析不同網路架構的差異。因此,本研究透過不同的人工神經網路模式與多元迴歸來建構不動產估價的分析預測模式,並透過台北市006、007、008年不動產實際交易案例的實證分析,來比較多元迴歸與不同人工神經網路模式間之差異,以及進一步地比較人工神經網路中不同網路架構之優劣,以提供作為未來估價實務的參考之用。實證結果顯示,由倒傳遞神經網路(BPN)、輻狀基底函數網路(RBFN)、多層函數連結網路(MFLN)所建構的人工神網路模式,在模式適合度指標與模式預測度指標的表現上,皆遠優於多元迴歸模式,其中又以多層函數連結網路模式(MFLN)表現最佳,該模式的預測準確度除超過一般實務要求水準外,更優於以往研究所建構的人工神經網路模式。另外,在網路架構的分析比較上,顯示隱藏層與其單元數的數目愈多,會使模式趨於複雜,進而使模式收斂較慢;而在模式適合度與預測度的表現上,除倒傳遞神經網路模式(BPN)顯示二層隱藏層的模式表現較佳外,輻狀基底函數網路(RBFN)與多層函數連結網路(MFLN),皆顯示一層隱藏層之模式表現較佳。
Real estate prices affect the compensation of land acquisition, the cost and benefit of land development, and the investment of real estate. Thus, how to evaluate and predict the price of real estate precisely plays an important role in land economics research. This study uses both hedonic multiple regression method (MRA) and different artificial neural networks (ANN) to build models for evaluating and predicting on housing prices. We used the Year 2006 to 2008 data of housing transactions in Taipei City. The empirical results reveal that ANN can be a better alternative for predicting of housing prices. Among the different ANN housing prices models, the best predicting performance show at Multilayer Functional-Link Network (MFLN). In comparing network architecture, it indicates that more hidden layers and more attributes make the model more complicated and make the procedure converge slowly. In Back-Propagation Network (BPN), 2-layer model performs better than other network models in fitted-modeling and forecast accuracy, whereas it shows the performance of 1-layer model is better than 2-layer hidden model for both Multilayer Functional-Link Network (MFLN) and Radial Basis Function Network (RBFN). |