Covariance-Adaptive Least-Squares Algorithm for Stochastic Combinatorial Semi-Bandits

We address the problem of stochastic combinatorial semi-bandits, where a player can select from P subsets of a set containing d base items. Most existing algorithms (e.g. CUCB, ESCB, OLS-UCB) require prior knowledge on the reward distribution, like an upper bound on a sub-Gaussian proxy-variance, which is hard to estimate tightly. In this work, we design a variance-adaptive version of OLS-UCB, relying on an online estimation of the covariance structure. Estimating the coefficients of a covariance matrix is much more manageable in practical settings and results in improved regret upper bounds compared to proxy variance-based algorithms. When covariance coefficients are all non-negative, we show that our approach efficiently leverages the semi-bandit feedback and provably outperforms bandit feedback approaches, not only in exponential regimes where P $\gg$ d but also when P $\le$ d, which is not straightforward from most existing analyses.