Faculty Recruiting Support CICS

Heterogeneous Structures for Online Learning Algorithms

04 Mar
Wednesday, 03/04/2020 12:15pm to 1:15pm
Computer Science Building Room 140
Theory Seminar
Speaker:  Lin Yang

In many online learning paradigms, such as the Expert and Bandit problems, each single decision point (either expert or arm) is evaluated  and learned in a homogeneous way. Motivated by the learning paradigms of some special features, my work introduces a heterogeneous learning structure, which groups similar decision points and then accordingly runs differentiated learning algorithms of different types/parameters. In some cases involving in massive similar or correlated decision points, we prove that this method can speed up the learning process. For example, with a recursive learning structure applied to the Exponential Weighting method, the proposed algorithm can touch the tight regret lower bound for the Lipschitz Expert problem. This is the first optimal algorithm having been reported in the literature.

Faculty Host