「世大智科/天才家居」-我們創業囉
Contact Professor: Yeh-Liang Hsu (徐業良)

八十二學年度元智大學機械工程研究所陳重光碩士論文

Master thesis by C. G. Chen, Mechanical Engineering Department, Yuan Ze University, 1994

82碩士論文:無微分式最佳設計化法

輕量化設計是現今航太、汽車等機械工業追求的重要目標之一,最佳化設計方法在機械設計上的重要性不可言喻。然而傳統的最佳化設計著重於數值方法解最佳化模型,且在求取最佳解時,常需要函數(目標函數與限制條件)的微分值。可是就工程設計問題而言,由於常常會有內隱式函數限制條件的存在,且函數值以及一次微分值取得不易,因此傳統數學規劃法在這方面沒有多大的實用性。在本文中,我們提出兩種無需求取函數微分值的無微分式最佳化設計法—『單調性分析』與『類神經網路模擬近似法』。
單調性分析法主要是利用大部分機械性質、行為的單調性。對設計模型做一定性分析。在本論文中我們將介紹並證明單調性原則,同時並介紹將單調性分析過程自動化的電腦程式MONO。 至於類神經網路模擬近似法則是一種結合倒傳遞神經網路與新搜尋準則的最佳化方法,其主要作法是利用網路來模擬可行區域的邊界,並進而求得設計問題的近似最佳解。

Non-Differentiation Optimization Methods

Minimum weight design has become one of the most important goals of many mechnical designs, such as automobiles and airplanes. So design optimization is an important tool in mechanical design. Traditional numerical optimization algorithms often haveto evaluate the function values and first derivative of objective function and constraints. however, in engineering optimization problems,there often exists implicit constrains, whoses first derivatives are very expensive to evaluate. Therefore traditional mathematical programming methods may not be practical in such applications. In this thesis, we study two design optimization methods which do not need the first derivatives-"Monotonicity Analysis Method" and "Artificial Neural Networ Simluation Approximation Method".
Monotonicity Analysis utilizes the monotonicity of most mechnical properties and behaviors to preform qualitative analysis on the optimization model. This thesis introduces and proves the monotonicity principles. Alogic program MONO, which automatically generates rigorous Monotonicity Analysis steps and global facts about the optimization model, is also developed.
????????Aritifical Neural Network Simluation Approximation Method (ANNSAM) is a framework which combines back-propagation neural networks with a search algorithm for discrete-variable engineering optimization problems. This method uses neural networks to simluate the boundary of feasible domain of the optimization model, and find the approximate discrete optimum in an iterative manner.