Competitive online generalized linear regression under square loss

Abstract

We apply the Aggregating Algorithm to the problem of online regression under the square loss function. We develop an algorithm competitive with the benchmark class of generalized linear models (our "experts"), which are used in a wide range of practical tasks. This problem does not appear to be analytically tractable. Therefore, we develop a prediction algorithm using the Markov chain Monte Carlo method, which is shown to be fast and reliable in many cases. We prove upper bounds on the cumulative square loss of the algorithm. We also perform experiments with our algorithm on a toy data set and two real world ozone level data sets and give suggestions about choosing its parameters. © 2010 Springer-Verlag Berlin Heidelberg.