“Optimization with gaussian processes via chaining” by Emile Contal
Abstract – In this talk, we consider the problem of sequential optimization of Gaussian processes. We generalize the GP-UCB algorithm [Srinivas and al., 2012] to arbitrary kernel and search space. We introduce the notion of covering tree and provide a novel optimization scheme based on covering numbers to automatically calibrate the exploration-exploitation tradeoff. We show how to build efficiently the algorithm and demonstrate that it performs at least as good as GP-UCB in empirical assessment.
“On the complexity of derivative free methods” by Cédric Malherbe
Abstract – We adress the problem of maximizing an unkown function f over a compact set by sequentially observing (noisy) observations of the unkown function. In the first part, we will analyse the case where the function is supposed satisfy some smoothness conditions (e.g. lipschitzian functions…). We will introduce an algorithm for lipschitzian functions that present good theoretical properties. In the second part, we will introduce an adaptative version of the algorithm that estimate the lipschitzian constant online.