Multi-Objective Online Learning
This paper presents a systematic study of multi-objective online learning. We first formulate the framework of Multi-Objective Online Convex Optimization, which encompasses a novel multi-objective dynamic regret in the unconstrained max-min form. We show that it is equivalent to the regret commonly used in the zero-order multi-objective bandit setting and overcomes the problem that the latter is hard to optimize via first-order gradient-based methods. Then we propose the Online Mirror Multiple Descent algorithm with two variants, which computes the composite gradient using either the vanilla min-norm solver or a newly designed $L_1$-regularized min-norm solver. We further derive regret bounds of both variants and show that the $L_1$-regularized variant enjoys a lower bound. Extensive experiments demonstrate the effectiveness of the proposed algorithm and verify the theoretical advantage of the $L_1$-regularized variant.
PDF Abstract
MultiMNIST
