Exploring the Impact of Distance Metrics on Alternative Generation in a Multiobjective Problem
Garrett M. Foster, Scott Ferguson · 13th AIAA/ISSMO Multidisciplinary Analysis Optimization Conference · 2010
Multi-objective optimization relies on model accuracy to generate solutions that are indeed optimal. When confidence in the model is not absolute, it is beneficial to explore designs near the solution to gain additional insight. Doing so without direction would be ill-advised as the quantity of designs that could be evaluated are too numerous. Instead it is suggested that a tool to help the designer locate a few designs that perform similarly, but are as unique as possible with regards to their design variables, be utilized. Modeling to generating alternatives (MGA) is a family of techniques that meet this requirement. The goal of MGA is maximizing the uniqueness of a design with respect to an original point of interest. Uniqueness in this work is defined by the distance between two designs in the design space. By focusing on the designs that have the greatest difference between them, the designer can efficiently explore the design space. Such flexibility can be useful when aspects of the model change or when the goal is differentiating a product from a crowded field of similar products. This paper explores the effect that different distance metrics, within an MGA approach, have on the solution to a multi-objective optimization problem. The goal is identifying metrics most effective at capturing uniqueness.