FTRL dynamics with constant step size are inherently exploitable in zero-sum games, yielding surplus of order Omega(N/eta) for fixed optimizers and Omega(eta T / poly(n,m)) for alternating ones in random games.
Behavior based learning in identifying high frequency trading strategies
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
-
On the Exploitability of FTRL Dynamics
FTRL dynamics with constant step size are inherently exploitable in zero-sum games, yielding surplus of order Omega(N/eta) for fixed optimizers and Omega(eta T / poly(n,m)) for alternating ones in random games.