A1739
Title: Adaptive Bayesian optimization with consistent smoothness estimation
Authors: Saifei Sun - City University of Hong Kong (Hong Kong) [presenting]
Abstract: A novel algorithm, Adaptive Matern Kernel Bayesian Optimization (AMKBO), is presented to address the challenges of hyperparameter uncertainty in Gaussian Process (GP)-based Bayesian Optimization. AMKBO accelerates convergence by adaptively estimating the smoothness parameter of the Matern Kernel, improving the hyperparameter adjustment strategy, and incorporating an opposition-based learning mechanism into acquisition function Optimization. Specifically, AMKBO generates initial sampling points using random curves and constructs an estimator for the smoothness parameter of the covariance Kernel, allowing the Kernel to model functions with varying degrees of smoothness. During Optimization, a refined length-scale adjustment strategy is employed to achieve a more balanced trade-off between exploration and exploitation while preventing excessively aggressive exploration. In parallel, the opposition-based learning mechanism enhances search coverage and computational efficiency in acquisition function Optimization. Experimental results on synthetic benchmark functions and real-world problems demonstrate that AMKBO outperforms existing methods by achieving faster convergence and avoiding local optima.
