目的関数のニューラルネットワーク推定による衝撃最適設計法の開発 : 第1報, 簡易非線形モデルでの基礎検討(<小特集>材料・構造の衝撃問題)

概要

論文の詳細を見る
In the process of optimization, expensive finite-element method (FEM) software in often applied to solve engineering design problems. To obtain the optimal solution requires an extreme numbers of iterations between the analysis software and optimization program. The work is time-consuming due to the amount of computer time required. Therefore, instead of using the FEM analysis, an approximate method using sensitivity analysis has been proposed to solve this problem. However, it is difficult to obtain an explicit form of sensitivity for the nonlinear dynamic system. While the difference method is usually used for the sensitivity analysis, it is not the most efficient method to greatly reduce the calculation. In this paper, the authors propose an efficient approximation method to simulate the objective function using the holographic neural network (HNN). This neural network exhibits a large learning capability and dynamic features of memory, of a two-degree-of-freedom nonlinear dynamic system. The optimization approach is simplified and the tremendous calculation expense is decreased, because the objective function in design areas can be approximately calculated using the trained neural network.
一般社団法人日本機械学会の論文
1997-12-25