基于神经网络的交互式物理规划及在机械设计中的应用研究

NEURAL NETWORKS BASED INTERACTIVE PHYSICAL PROGRAMMING AND ITS APPLICATION IN MECHANICAL DESIGN

在给定Pareto解附近，用神经网络建立了Pareto曲面的近似模型，以探索新的Pareto解。在给定Pareto解附近随机产生一组Pareto解，利用可视化工具将它们展示给设计者，并用定性和定量相结合的方法评定它们的分值。利用神经网络，建立Pareto解到评分值的映射，以表达设计者在给定Pareto解附近的局部偏好。然后用遗传算法进行优化，找到最佳符合设计者偏好的Pareto解。以这个Pareto解为期望点，求解折衷规划，从而得到最终的优化设计方案。

Interactive physical programming is based on physical programming，a new effective and computationally efficient approach for multidisciplinary design optimization．It takes into account the designer's or the decision maker's (DM's) preferences during the optimization process，and allows for design exploration at a given Pareto design．The approximate model of Pareto surface at a given Pareto design is developed using neural networks for design exploration．At the given Pareto design，a group of Pareto designs are generated randomly．They're brought to the decision maker using a Pareto visualization tool，and are evaluated with both qualitative and quantitative analysis．A map from Pareto designs to their corresponding evaluation value is established using a neural networks model，it illustrates the decision maker's local preference at the given Pareto design．Then genetic algorithms is used in optimization to find the Pareto design which mostly accords with the decision maker's local preference．The obtained Pareto design is used as the aspiration point in compromise programming，and the final design solution can be obtained．