Best-of-Both-Worlds Algorithms for Partial Monitoring

This paper considers the partial monitoring problem with $k$-actions and $d$-outcomes and provides the first best-of-both-worlds algorithms, whose regrets are bounded poly-logarithmically in the stochastic regime and near-optimally in the adversarial regime. To be more specific, we show that for non-degenerate locally observable games, the regret in the stochastic regime is bounded by $O(k^3 m^2 \log(T) \log(k_{\Pi} T) / \Delta_{\mathrm{\min}})$ and in the adversarial regime by $O(k^{2/3} m \sqrt{T \log(T) \log k_{\Pi}})$, where $T$ is the number of rounds, $m$ is the maximum number of distinct observations per action, $\Delta_{\min}$ is the minimum optimality gap, and $k_{\Pi}$ is the number of Pareto optimal actions. Moreover, we show that for non-degenerate globally observable games, the regret in the stochastic regime is bounded by $O(\max\{c_{\mathcal{G}}^2 / k,\, c_{\mathcal{G}}\} \log(T) \log(k_{\Pi} T) / \Delta_{\min}^2)$ and in the adversarial regime by $O((\max\{c_{\mathcal{G}}^2 / k,\, c_{\mathcal{G}}\} \log(T) \log(k_{\Pi} T)))^{1/3} T^{2/3})$, where $c_{\mathcal{G}}$ is a game-dependent constant. Our algorithms are based on the follow-the-regularized-leader framework that takes into account the nature of the partial monitoring problem, inspired by algorithms in the field of online learning with feedback graphs.