Sequential Approximate Multi Objective Optimization with Limited Data Samples

Abstract

Today, multi-objective optimization using simulation is widely used. Multi-objective sequential approximation optimization is used to perform multi-objective optimization problems with a small number of simulations. In the past, a large amount of simulation data was required because a uniform method of adding training data was used, despite the fact that the dominant effects were different. This study simplifies the problem by dividing Pareto solutions with different dominant effects. Pareto solutions with the same dominant effect are assumed to be close in distance in the design space, and the Pareto solution is divided by clustering to create a Pareto solution with a simple shape. We then describe how to evaluate and augment it with a small number of data. A multi-objective successive approximation optimization method using these methods is proposed, and the multi-objective optimization is demonstrated using a power device as an example. The results show that clustering by distance in the design space is equivalent to dividing the Pareto solution by the dominant effect.