Grey-Box Bayesian Optimization via Adaptive Linearization of Composite Functions
A Bayesian-optimization method for systems that are only partly a black box: the objective is written as a composition of known analytical functions and Gaussian-process pieces, so structural knowledge such as physics and sparse component interconnections can be exploited. Because the resulting composite distribution is generally intractable, the method (BOIS) applies adaptive first-order linearizations to derive closed-form mean and variance of the composite objective, letting the acquisition function use that structure without expensive sampling. Students learn how partial structural knowledge is folded into a probabilistic surrogate to make expensive optimization cheaper.
On the Implementation of a Bayesian Optimization Framework for Interconnected Systems Leonardo D.
A grey-box variant of Bayesian optimization (BO) for optimizing expensive-to-evaluate systems whose internal structure is partly known. Standard BO fits a single Gaussian-process surrogate to a black…