Empirical Investigation of MOEAs for Multi-objective Design of Experiments

Abstract

Many machine learning algorithms require the use of good quality experimental designs to maximise the information available to the model. Various methods to create experimental designs exist, but the solutions can be sub-optimal or computationally inefficient. Multi-objective evolutionary algorithms (MOEAs), with their advantages of being able to solve a variety of problems, are a good method of creating designs. However, with such a variety of MOEAs available, it is important to know which MOEA performs best at optimising experimental designs. In this paper, we formulate experimental design creation as a multi-objective optimisation problem. We compare the performance of different MOEAs on a variety of experimental design optimisation problems, including a real-world case study. Our results show that NSGA-II can often perform better than NSGA-III in many-objective optimisation problems; RVEA performs very well; results suggest that using more objectives can create better quality designs. This knowledge allows us to make more informed decisions about how to use MOEAs when creating metamodels.