Flickering Multi-Armed Bandits

The paper introduces Flickering Multi-Armed Bandits (FMAB), a model for sequential decision-making in environments with dynamically changing action availability, represented through stochastically evolving graphs. It presents a two-phase lazy random walk algorithm that achieves high-probability sublinear regret bounds and demonstrates near-optimality, addressing the challenges of information acquisition and navigation in constrained settings. These findings are significant for practitioners as they provide insights into learning efficiency in environments with restricted action accessibility, applicable in fields like robotics and disaster response.