Olivier Wintenberger (LSTA, UPMC)

Optimal online learning in stochastic environment with exponential weights

vendredi 5 mai 2017, 9h30 - 10h30

Salle du conseil, espace Turing

In this talk, we solve the two statistical aggregation problems named Model Selection (MS) and Subset selection (S) after Bunea et al. (2007). The first problem is solved thanks to a new algorithm called Bernstein Online Algorithm (BOA) in Wintenberger (2016), a regularized version of the exponential weights. The second problem is solved using BOA as a subroutine in acceleration routines using localization called Sparse Accelerated Exponential Weights (SAEW) in Gaillard and Wintenberger (2017). This is a joint work with Pierre Gaillard.