以退火神經網路作建築空間配置

Architecture Space Layout Using Annealed Neural Network

建築機能單元的配置是一個重要的設計工作。在複雜的建築物中，好的配置設計在金錢與時間上的節省變得更為明顯。本研究將建築機能單元的配置公式化為一個組合最佳化問題，並採用融合了許多模擬退火與Hopfield 神經網路的特徵的退火神經網路來求解，並以一個具有28 個機能單元的綜合醫院診療大樓作個案研究，此外，也詳細探討退火神經網路的參數對解答品質的影響。研究結果顯示：(1) 退火神經網路對於解決建築機能單元的配置問題相當具有效率；(2) 無論參數組合為何，以30 個隨機的初始狀態所得的合法解中之最佳解差異很小。

Architecture layout design is an important design activity. The impactof good layout practices on money and time saving becomes more obviousin complex architecture. In this study, the layout problem was formulatedas a combinatorial optimization problem. An annealed neural networkmodel, which merges many features of simulated annealing and Hopfieldneural networks, was employed to solve the problem. A case study of ahospital building with 28 facilities was employed to illustrate the practicalapplications. In addition, the effects of various parameters in annealedneural network were examined. Research reported in this paper leads to thefollowing conclusions. (1) An annealed neural network model is rather efficientin solving the architecture layout problem. (2) Whatever the combinationsof the parameters are, the difference of quality between the optimumsolutions among 30 feasible solutions gotten from a random initialstate is rather small.