Scale Free Adversarial Multi Armed Bandits
This addresses the problem of robust decision-making under uncertainty in adversarial environments for researchers in online learning and bandit algorithms, representing a significant advance with novel bounds.
The paper tackles the scale-free adversarial multi-armed bandits problem, where the player lacks prior knowledge of loss scales or rounds, and achieves regret bounds of $ ilde{\mathcal{O}}(\sqrt{nL_2} + L_\infty\sqrt{nT})$ with non-adaptive exploration and $ ilde{\mathcal{O}}(\sqrt{nL_2} + L_\infty\sqrt{nL_1})$ with adaptive exploration, which are the first to adapt to $\|\cdot\|_2$ and $\|\cdot\|_1$ norms of losses.
We consider the Scale-Free Adversarial Multi Armed Bandits(MAB) problem. At the beginning of the game, the player only knows the number of arms $n$. It does not know the scale and magnitude of the losses chosen by the adversary or the number of rounds $T$. In each round, it sees bandit feedback about the loss vectors $l_1,\dots, l_T \in \mathbb{R}^n$. The goal is to bound its regret as a function of $n$ and norms of $l_1,\dots, l_T$. We design a bandit Follow The Regularized Leader (FTRL) algorithm, that uses an adaptive learning rate and give two different regret bounds, based on the exploration parameter used. With non-adaptive exploration, our algorithm has a regret of $\tilde{\mathcal{O}}(\sqrt{nL_2} + L_\infty\sqrt{nT})$ and with adaptive exploration, it has a regret of $\tilde{\mathcal{O}}(\sqrt{nL_2} + L_\infty\sqrt{nL_1})$. Here $L_\infty = \sup_t \| l_t\|_\infty$, $L_2 = \sum_{t=1}^T \|l_t\|_2^2$, $L_1 = \sum_{t=1}^T \|l_t\|_1$ and the $\tilde{\mathcal{O}}$ notation suppress logarithmic factors. These are the first MAB bounds that adapt to the $\|\cdot\|_2$, $\|\cdot\|_1$ norms of the losses. The second bound is the first data-dependent scale-free MAB bound as $T$ does not directly appear in the regret. We also develop a new technique for obtaining a rich class of local-norm lower-bounds for Bregman Divergences. This technique plays a crucial role in our analysis for controlling the regret when using importance weighted estimators of unbounded losses. This technique could be of independent interest.