Bayesci eniyilemede gürbüz yeterlilik

This thesis explores the use of a recent optimization framework called "robust satisficing", in the context of Bayesian optimization. Robust satisficing (RS) is an alternative to optimizing, where the goal is to find a solution to a problem that achieves a predefined threshold robustly, under environmental uncertainties. On the other hand Bayesian optimization (BO) is a well-established framework for optimizing difficult to evaluate black-box functions. Previously, the BO literature has mainly focused on optimization, robust optimization or pure satisficing. The goal of this work is to introduce RS in to the BO framework. We analyze the problem in the contextual GP setting where distribution shifts on the contextual variable introduces uncertainties, and also in an adversarial setting where actions chosen are subjected to a perturbation from an adversary. We develop novel algorithms in both settings, using both a known RS approach and a new modification of it. We prove regret bounds for our algorithms in both settings and do extensive simulations that show the advantages of our approach compared with other state-of-the-art methods such as distributionally robust optimization.