Fooling Algorithms in Non-Stationary Bandits using Belief Inertia
The episode discusses Mendelson and Tadmor's paper on 'Fooling Algorithms in Non-Stationary Bandits using Belief Inertia.' The hosts explain how an algorithm's historical beliefs create momentum that resists new evidence when the environment changes. This inertia can cause standard algorithms to suffer linear regret even with a single reward change point, suggesting a need for more robust AI design.