Francesco Emanuele Stradi

Optimal regret in adversarial MDPs with stochastic hard constraints is finally achieved, bridging the gap between existing upper and lower bounds by achieving $\widetilde{\mathcal{O}}(\sqrt{T}/\rho + 1/(d\rho))$ regret.

Finally, a constrained MAB algorithm that gracefully degrades under adversarial constraint drift, achieving near-optimal regret when constraints are stochastic and smoothly transitioning to adversarial robustness.

Replicable algorithms can now achieve the same performance as non-replicable ones in constrained multi-armed bandit problems, opening the door to more reliable and reproducible online learning experiments.