Transfer learning in stochastic multi-armed bandits is usually studied under a simplifying assumption: the prior data is either a fixed offline dataset or is generated by a known source. In practice, however, historical observations often come from a previous adaptive policy whose behavior is not fully known. That gap between theory and practice is the motivation for this arXiv preprint.

The authors analyze the case where the source policy is both adaptive and unknown, and they derive problem-dependent lower and upper bounds for transfer performance. These bounds characterize how much can be learned from such historical data, depending on properties of the unknown source policy and its relationship to the target bandit problem.

Because this summary is based on a single source—the paper's abstract—the exact form of the bounds and their practical implications are not fully detailed here. Still, the work points toward a more realistic model of bandit transfer, where prior data is messy and generated by policies we do not fully control or understand.