Parallel Expected Improvement Method for Multi-Fidelity Bayesian Optimization

Multi-fidelity Bayesian optimization has been widely applied in engineering design, which leverages a large number of cheap low-fidelity (LF) samples to assist a small number of costly high-fidelity (HF) samples to improve optimization performance. As the computation of engineering problems becomes more and more complex, parallel sampling strategies have begun to replace single-point sampling strategies in multi-fidelity Bayesian optimization, further reducing human intervention and improving the efficiency of engineering design. To address the demand for multi-fidelity Bayesian parallel optimization, this paper proposes the composite expected improvement method, which utilizes the weighted expected improvement criterion to acquire multiple samples with different search tendencies, and adaptively assigns the fidelity level of samples according to the expected long-term reward. Experimental results from numerical examples and an engineering design problem show that the proposed multi-fidelity parallel optimization method exhibits superior optimization performance.