This document discusses Gaussian process bandit optimization, which is a method for adaptively sampling an unknown function to maximize its value. It proposes using an upper confidence bound (UCB) approach, where samples are selected to maximize an upper bound on the function value while also exploring uncertain regions. The key points are: 1) It proves regret bounds for UCB in this setting that depend on how quickly information can be gained about the function from samples, known as the maximal information gain. 2) This connects Gaussian process bandit optimization to Bayesian experimental design, which aims to maximize information gain. 3) Experiments on temperature and traffic data show the UCB approach performs comparably to existing heuristics while providing the