台灣近期之自有住宅率高達82%，受到傳統文化住者有其屋以及遺贈稅調降的影響，使得台灣出現高房價和低租金的現象。政府為了緩解此狀況開辦住宅補貼方案，此補貼方案截至目前，常見資料中有許多申請戶因子相同，然而最後卻得出不同之核准結果。本研究之目的即為，以納入粒子群演算法與模糊理論之多維矩形複合式類神經網路PSO-based Fuzzy Hyper-Rectangular Composite Neural Networks(PFHRCNN)對住宅補貼資料進行分類，並提供一項快速篩選工具給使用者自行操作，PFHRCNN整合類神經網絡與以規則為基礎之運算，採納監督式學習導向學習演算法(Supervised Decision-Directed Learning SDDL)，以實現100％訓練資料正確識別率，在分類中納入模糊理論，並用最佳化方案來協調各規則。取用之資料為101年住宅補貼資料，選定重要分析項次共12項，將中文格式之資料數值化並剔除衝突資料後得36086筆。分類結果中，最高正確率高達98.6%，最低之正確率也有近乎90%，明顯可看出此分類器之選擇有相當優異的成效。預測之結果顯示，單親家庭與曾為租金補貼戶兩項因子同時擁有，有極高的可能性通過審核；又或是新竹以南之鄉鎮，若曾為租金補貼戶，再次通過的可能性亦很大。;Thousands of social vulnerable families apply for rental subsidy every year in Taiwan. For the past years before the change of the review criteria takes place, there are, for example, 89,750 applications in 2012. To develop a rapid and accurate tool for fast screening out unqualified applications and to seek proper review rules in a consistent viewpoint throughout the entire Taiwan cities and townships, the research objectives are to build up a tool to fast screen out unqualified applications so as to save enormous manpower and, to provide rules for reviewers using PSO-based Fuzzy Hyper-rectangular Composite Neural Networks (PFHRCNN). By looking into 89,750 datasets collected from the government data bank, each application contains 88 features where only 12 features are available for further analysis. Data trimming was performed to eliminate dilemmatic data that include exactly the same features but have different outcomes. A high proportion of 59.79% or 53,664 sets of the total involve such the circumstance. As a result, the rest of data or 36,086 sets were used to develop the fast screen-out tool and to obtain the review rules. The tool eventually yields a high accuracy rate at 98.6% and 66 rules suggested for review criteria. The findings support that single parent families who had rental subsidy living in the central or southern Taiwan have the highest chances to be granted again.
