FTRL no-regret learners in zero-sum games yield inherent strategic surplus to clairvoyant optimizers, with Omega(N_sub/eta) for fixed strategies and Omega(eta T/poly(n,m)) in random games, plus a steep vs non-steep regularizer dichotomy.
Online convex programming and generalized infinitesimal gradient ascent
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.GT 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
No Coin Left Behind: Maximizing Strategic Surplus Against No-Regret Dynamics
FTRL no-regret learners in zero-sum games yield inherent strategic surplus to clairvoyant optimizers, with Omega(N_sub/eta) for fixed strategies and Omega(eta T/poly(n,m)) in random games, plus a steep vs non-steep regularizer dichotomy.