Meta-model-assisted MGDA for multi-objective functional optimization

Abstract

International audienceA novel numerical method for multi-objective differentiable optimization, the Multiple-Gradient Descent Algorithmm (MGDA), has been proposed in [8] [11] to identify Pareto fronts. In MGDA, a direction of search for which the directional gradients of the objective functions are all negative, and often equal by construction [12], is identified and used in a steepest-descent-type iteration. The method converges to Pareto-optimal points. MGDA is here briefly reviewed to outline its principal theoretical properties and applied first to a classical mathematical test-case for illustration. The method is then ex-tended encompass cases where the functional gradients are approximated via meta-models, as it is often the case in complex situations, and demonstrated on three optimum-shape design problems in compressible aerodynamics.The first problem is purely related to aerodynamic performance. It is a wing shape optimization exercise w.r.t. lift and drag in typical transonic cruis

Gradient based biobjective shape optimization to improve reliability and cost of ceramic components

“…However, in combination with expensive numerical simulations such approaches tend to be inefficient. On the other hand, gradientbased algorithms [17,24,53] require efficient gradient computations and are often applied in the context of adjoint approaches and using weighted sum scalarizations of the objective functions. See [41] for a comparison.…”

Data-Driven Models for Yacht Hull Resistance Optimization: Exploring Geometric Parameters Beyond the Boundaries of the Delft Systematic Yacht Hull Series

“…For this reason, recently, Data-Driven Models (DDMs) are attracting the attention of the industry and academia for their ability to accurately surrogate complex experimental (e.g., EFD) [26]- [29] or numerical (e.g., CFD) [1], [14], [16], [17], [19], [21], [22] procedures based on a historical collection of their inputs and outputs, with a function that is computationally expensive to construct but computationally inexpensive to use. Consequently, DDMs can be included directly both in a human-driven optimization loop reducing the computational requirements (i.e., time) between design iterations or developing a fully automated optimization loop requiring minimal human intervention, enabling the exploration of a wider design space [2], [30].…”