Memory-Guided Trust-Region Bayesian Optimization (MG-TuRBO) for High Dimensions

This research paper proposes Memory-Guided Trust-Region Bayesian Optimization (MG-TuRBO), a novel method for efficiently optimizing expensive high-dimensional calibration problems. Evaluated on traffic simulation and digital-twin tuning tasks with up to 84 decision variables, MG-TuRBO outperforms genetic algorithms and standard Bayesian optimization, particularly excelling in higher-dimensional settings when paired with an adaptive acquisition strategy. The work demonstrates significant convergence improvements while maintaining consistency across multiple runs.