Enhancing Multi-Fidelity Bayesian Optimization for Mixed-Variable, Multi-Objective Multi-Disciplinary Drone Design Optimization

In a context where fixed-wing drones are increasingly used for diverse applications, it is useful to be able to design drone optimized for their application goals. The corresponding optimization problems involve mixed variables and multiple objectives and constraints. These objectives and constraints are often evaluated using highly complex multi-disciplinary models assimilated to expensive black-boxes. There exists interesting strategies to overcome these difficulties such as Bayesian optimization and particularly methods involving the use of multiple fidelity levels. Indeed the use of fidelity levels with various accuracies and costs usually allows to reduce the overall optimization cost in this context. This article assessed the performance of a developed multi-fidelity, multi-objective Bayesian optimization method able to tackle mixed variables and compared it to the corresponding mono-fidelity method. First comparison was conducted on modified analytical test cases involving mixed variables, followed by a drone design test case with variables related to material choices and battery numbers. Using the same optimization budget for both mono-fidelity and multi-fidelity approaches, our results for the drone design problem showed an improvement in the IGD hypervolume-based performance indicator by a factor of 2 when using multi-fidelity.