Cache Replacement as a MAB with Delayed Feedback and Decaying Costs

Abstract

Inspired by the cache replacement problem, we propose and solve a new variant of the well-known multi-armed bandit (MAB), thus providing a solution for improving existing state-of-the-art cache management methods. Each arm (or expert) represents a distinct cache replacement policy, which advises on the page to evict from the cache when needed. Feedback on the eviction comes in the form of a "miss", but at an indeterminate time after the action is taken, and the cost of the eviction is set to be inversely proportional to the response time. The feedback is ignored if it comes after a threshold value for the delay, which we set to be equal to the size of the page eviction history. Thus, for delays beyond the threshold, its cost is assumed to be zero. Consequently, we call this problem with delayed feedback and decaying costs. We introduce an adaptive reinforcement learning algorithm EXP4-DFDC that provides a solution to the problem. We derive an optimal learning rate for EXP4-DFDC that defines the balance between exploration and exploitation and proves theoretically that the expected regret of our algorithm is a vanishing quantity as a function of time. As an application, we show that LeCaR, a recent top-performing machine learning algorithm for cache replacement, can be enhanced with adaptive learning using our formulations. We present an improved adaptive version of LeCaR, called OLeCaR, with the learning rate set as determined by the theoretical derivation presented here to minimize regret for EXP4-DFDC. It then follows that LeCaR and OLeCaR are theoretically guaranteed to have vanishing regret over time.