Sham M. Kakade, Matthias W. Seeger, Dean P. Foster
We present a competitive analysis of some non-parametric Bayesian al- gorithms in a worst-case online learning setting, where no probabilistic assumptions about the generation of the data are made. We consider models which use a Gaussian process prior (over the space of all func- tions) and provide bounds on the regret (under the log loss) for com- monly used non-parametric Bayesian algorithms — including Gaussian regression and logistic regression — which show how these algorithms can perform favorably under rather general conditions. These bounds ex- plicitly handle the inﬁnite dimensionality of these non-parametric classes in a natural way. We also make formal connections to the minimax and minimum description length (MDL) framework. Here, we show precisely how Bayesian Gaussian regression is a minimax strategy.