Learning to Cache and Caching to Learn: Regret Analysis of Caching Algorithms

Crucial performance metrics of a caching algorithm include its ability to quickly and accurately learn popularity distribution requests. However, majority work on analytical analysis focuses hit probability after an asymptotically large time has elapsed. We consider online learning viewpoint, characterize the “regret” in terms finite difference between hits achieved by candidate with respect genie-aided scheme that places most popular items cache. first Full Observation regime wherein all requests are seen show Least Frequently Used (LFU) is able achieve order optimal regret, which matched efficient counting design we call LFU-Lite. then Partial only for currently cached cache, making it similar problem related multi-armed bandit problem. how approaching this “caching bandit” using traditional approaches yields either high complexity or but simple exploits structure can ensure regret. conclude illustrating our insights numerical simulations.