In this work we study a variant of the well-known multi-armed bandit (MAB) problem, which has the properties of a delay in feedback, and a loss that declines over time. We introduce an algorithm, EXP4-DFDC, to solve this MAB variant, and demonstrate that the regret vanishes as the time increases. We also show that LeCaR, a previously published machine learning-based cache replacement algorithm, is an instance of EXP4-DFDC. Our results can be used to provide insight on the choice of hyperparameters, and optimize future LeCaR instances.