Search papers, labs, and topics across Lattice.
2
0
2
6
Optimal oracle complexity for nonsmooth nonconvex composite optimization is finally achieved, matching noncomposite rates by cleverly incorporating the convex regularizer into online learning losses.
Policy optimization for adversarial MDPs now matches state-of-the-art regret bounds by stabilizing local updates with regularized Q-functions, eliminating the horizon gap.
