Title: Achieving Optimal Robustness in Learning-Augmented Paging

Recent years have seen significant research into learning-augmented paging. A primary benefit of this approach, as opposed to basic machine learning-based methods, is its bounded robustness. This feature ensures that the algorithm maintains worst-case performance guarantees even if predictions are flawed, a trait that makes these algorithms highly suitable for deployment in real-world systems. While previous studies have established robustness bounds of $2H_k + O(1)$ within the randomized framework, a discrepancy remains when compared to the optimal competitive ratio of $H_k$. This paper investigates strategies to bridge this gap.

We start by examining online optimality and establishing a novel property of the most recent $H_k$-competitive algorithm, which serves as a foundation for our analysis in the learning-augmented context. Furthermore, we survey existing learning-augmented paging algorithms and propose a unifying concept called the relative prediction budget. This concept encapsulates the core mechanism for ensuring robustness and highlights that earlier algorithms tend to either over-rely on or underuse predictive information.

Building on these insights, we introduce a new framework for learning-augmented paging that attains the best possible robustness, up to an additive constant: $H_k + O(1)). Our experimental results further confirm the strong practical efficacy of this approach.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC