A multi-fidelity active learning method for global design optimization problems with noisy evaluations

Abstract A multi-fidelity (MF) active learning method is presented for design optimization problems characterized by noisy evaluations of the performance metrics. Namely, a generalized MF surrogate model used design-space exploration, exploiting an arbitrary number hierarchical fidelity levels, i.e., coming from different models, solvers, or discretizations, accuracy. The intended to accurately predict while reducing computational effort required simulation-driven (SDD) achieve global optimum. overall prediction evaluated as low-fidelity trained corrected with surrogates errors between consecutive levels. Surrogates are based on stochastic radial basis functions (SRBF) least squares regression and in-the-loop hyperparameters deal training data. adaptively queries new data, selecting both points level via approach. This lower confidence bounding method, which combines associated uncertainty select most promising regions. levels selected considering benefit-cost ratio their use in training. method’s assessed discussed using four analytical tests three SDD fluid dynamics simulations, namely shape NACA hydrofoil, DTMB 5415 destroyer, roll-on/roll-off passenger ferry. Fidelity provided adaptive grid refinement multi-grid resolution approaches. Under assumption limited budget function evaluations, proposed shows better comparison high-fidelity only.