Optimistic No-regret Algorithms for Discrete Caching

We take a systematic look at the problem of storing whole files in cache with limited capacity context optimistic learning, where caching policy has access to prediction oracle (provided by, e.g., Neural Network). The successive file requests are assumed be generated by an adversary, and no assumption is made on accuracy oracle. In this setting, we provide universal lower bound for prediction-assisted online proceed design suite policies range performance-complexity trade-offs. All proposed offer sublinear regret bounds commensurate Our results substantially improve upon all recently-proposed policies, which, being unable exploit predictions, only O(?T) regret. pursuit, design, best our knowledge, first comprehensive Follow-the-Perturbed leader policy, which generalizes beyond problem. also study different sizes bipartite network Finally, evaluate efficacy through extensive numerical experiments using real-world traces.