No-Regret Learning in Time-Varying Zero-Sum Games
summary
Proposes a parameter-free algorithm that combines optimistic base learners (satisfying a dynamic RVU property) with an optimistic meta-learner over a log-T grid of learning rates plus dummy experts, under mixture feedback. Claims simultaneous guarantees for three concrete metrics — classical individual regret, a new dynamic Nash regret, and cumulative duality gap — with bounds adaptive to variation measures (V, W, P), recovering best-known rates when the payoff matrix is fixed and showing empirical validation.
AI-written summary · about