Towards Optimal Robustness in Learning-Augmented Paging

Learning-augmented paging has been extensively studied in recent years. A key advantage over naive ML-based approaches is bounded robustness, which guarantees worst-case performance even when predictions are inaccurate, making these algorithms valuable for real-world systems. Prior work achieves robustness bounds of $2H_k + O(1)$ in the randomized setting, leaving a gap to the optimal competitive ratio $H_k$. In this paper, we study how to close this gap. We begin by reviewing online optimality and proving a new property of the latest $H_k$-competitive algorithm, which facilitates our analysis in the learning-augmented setting. Then, we review existing learning-augmented paging algorithms and introduce a unifying primitive, the relative prediction budget, which captures the essence of establishing robustness and reveals that prior algorithms either overuse or underutilize predictions. Guided by the above analysis, we develop a new framework that achieves the best-possible robustness up to an additive constant for learning-augmented paging: $H_k + O(1)$. Experiments further demonstrate strong practical performance.

Modern computer systems rely on a small fast memory to keep recently or soon-to-be-needed data close at hand. When this memory is full, the system must decide what to remove, and a wrong choice can slow everything down. Machine-learning predictions can help by guessing which data will be needed later, but such guesses can fail after deployment, so a system should benefit from good predictions without being badly hurt by bad ones. This paper studies how to make that tradeoff as strong as possible. We identify a simple way to measure how much an algorithm should “spend” its trust in predictions, and use it to design a new caching method. When predictions are accurate, the method behaves like an ideal strategy with future knowledge. When predictions are inaccurate, it still keeps the strongest possible worst-case safety guarantee, up to a small fixed overhead. Experiments on standard computer workload traces show that this added safety also leads to strong practical performance.