Using multi-objective optimization to balance system-level model complexity
Alexander Pollok, D.J. Bender · 2014
Deciding what to include in a model is the essence of physical modeling. If this kind of decisions is done incorrectly, it results in a model that is both less accurate and offers less performance than others. This paper presents a method to deal with these decision problems and quantify the results. A workflow is proposed, where the modeler first builds a model and includes several sets of replaceable subsystems. A multi-objective optimization algorithm then identifies selections of subsystems that are pareto optimal regarding accuracy and computation time. Finally, the modeler can pick one of these selections based on the modeling intent. An implementation of the proposed method using Modelica and Python is presented and potential pitfalls are explained. Special consideration is given to the quantification of the error or distance between models of different structures. The method is then illustrated by the use of two examples, one of them a complex model of an avionics cooling system cold plate. Suggestions are given regarding the embedding of the method into typical modeling workflows by the use of custom annotations.