類神經網路機械學習模式發展:房屋結構之初步設計

Abstract

電腦由於計算效率高，因此工程師常常使用它處理高計算量的問題。隨著相關電腦技術的進步，工程師開始期待電腦能更佳有效的輔助設計。基於朝向自動化設計的遠程目標，人工智慧技術正逐漸為土木工程人士所重視，並且期望電腦能處理經驗導向的問題。現階段的人工智慧技術雖尚不足以完全模擬工程師的邏輯思考方式，但是將人工智慧應用於結構設計中已證實能穫得相當程度效率提升。 過去的文獻可得知，倒傳遞類神經網路(Back-propagation Neural Network)為土木工程中最常被討論到的機械學習模式，其優點為演算法簡單且具有良好的一般化能力。然而其學習策略是採取數學最佳化的方式強記經驗，因此局部最小化成為無法避免的問題，尤其是對於需要大量案例的複雜問題，想達到良好學習收斂的困難度相形提高。除此之外，耗時的學習過程也是造成業界不願嘗試的理由之一。 因此本篇論文的主旨是要發展全新的機械學習模式，以期能讓人工智慧技術在結構設計過程中發揮更大的功效。在本論文將重心放在房屋結構設計中的初步設計階段，發展了兩個完全不同架構的機械學習模式:(1)IFN (Integrated Fuzzy Neural Network)類神經網路;(2)MS_CMAC (Macro Structure CMAC)類神經網路。IFN由主網路-UFN (Unsupervised Fuzzy Neural Network)推理模式和輔助網路-監督式類神經網路所組成，其中UFN推理模式採用局部資訊來推理，因此隨著蒐集到的案例增加其預測準確性將不斷的提升。除此之外模式本身具有自我調整系統參數的能力，可隨著案例庫的資料異動自動去修正參數達到最佳的預測能力。CMAC為一種極為快速的監督式類神經網路，MS_CMAC則是串連了許多單自由度的CMAC成樹狀結構，然後再以time inversion技巧執行計算的工作。其概念上近似於將多自由度的問題切割成一群單自由度的子問題，因此降低了神經網路學習收斂的困難度，間接得也穫得較高精確的預測結果。 為了驗證IFN和MS_CMAC的能力， IFN將應用於初始設計問題和鋼結構粱的設計問題，前者是一極為複雜的經驗導向問題，後者則為需要反覆計算的工作。MS_CMAC則將應用於評估鋼結構設計中所需要用到的設計參數，複雜的數值計算解析過程將由類神經網路所取代。論文中的結果將證實，IFN和MS_CMAC確實優於已被廣汎使用的BPN。

Computer programs are widely used to assist engineers in solving problems by shifting the burden of numerical computation to the machine. Furthermore, new methods and tools encourage civil engineers to use numerical computation in creative and imaginative ways. Despite the completely automatic structural design is presently not available; the efficiency of the conventional structural design is significantly improved by adopting some techniques of artificial intelligence (AI). Applying neural network computing, one of the artificial intelligence techniques, to structural engineering is currently an active subject in computer-aided design. Most of the previous researches concentrated on the back-propagation neural network (BPN) because BPN has a good generalization. However, the BPN sometime performs poor learning convergence when a large number of instances are used for a complicated problem, owing to its global optimization learning scheme. In addition, a long computational time in learning stage is another drawback for engineers to use BPN in structural design. The goal of this dissertation is to develop novel machine learning models to make structural design system more powerful, especially for preliminary design in building structural design. Based on the information flow in building structural design, two different kind of neural networks, integrated fuzzy neural network (IFN) and macro structure CMAC (MS_CMAC), are developed. (1) The IFN learning model combined an unsupervised fuzzy neural network (UFN) reasoning model with a supervised neural network as an assistant network. The UFN reasoning model adopts local information scheme to interpret a large number of instances for complicated problems within an acceptable computing time. Meanwhile, a self-organized learning is developed to refine the UFN reasoning. (2) The CMAC is a supervised learning model used mainly in control due to its rapid learning. The MS_CMAC is a tree-based structure of one-dimensional CMACs, where the ensemble is trained by the time inversion technique. The main feature of MS_CMAC is to decompose a multi-dimensional problem into a set of one-dimensional sub-problems so as to improve the learning convergence and prediction. For verifying the feasibility of IFN and MS_CMAC in structural design, the IFN learning model is utilized to model the initial design of building structure and steel beam design problems. The initial design is an experience-oriented problem, and the steel beam design is an iterative process under LRFD specification. Also, the MS_CMAC is employed to determine some design coefficients which are generally obtained by numerical approaches. The results indicate that the two neural networks are useful tools for engineers to solve structural design problems.