Akademik Veri Yönetim Sistemi
Structural and Multidisciplinary Optimization, cilt.67, sa.12, 2024 (SCI-Expanded)
Surrogate-based Bayesian optimization is efficient and useful for global optimization when objective functions are expensive to evaluate. Yet, the commonly used surrogate model, the Gaussian process, faces the scalability challenge for high-dimensional problems due to the high computational cost associated with the inversion of covariance matrices. In this paper, a new exploitation-enhanced sparse Gaussian process (EE-SGP) modeling method is proposed for scalable Bayesian optimization. The proposed EE-SGP strategically selects optimal samples to maximize the likelihood of identifying the global optimum, guided by the Gumbel distribution. This new sampling strategy, coupled with a sparse Gaussian process, significantly reduces the computational burden associated with high-dimensional problems. The optimization process leverages the biogeography-based optimization metaheuristic algorithm, further enhancing the efficiency and effectiveness of the proposed approach. EE-SGP's performance is assessed with analytical benchmark problems and constrained engineering optimization examples. The evaluation criteria include convergence speed and robustness. The studies demonstrate that EE-SGP is a robust, efficient, and scalable algorithm for searching for optimum solutions in high-dimensional spaces.