Core Concepts
We design parsimonious algorithms for caching and metrical task systems that use a reduced number of predictions while achieving strong consistency, robustness, and smoothness guarantees.
Abstract
The paper proposes algorithms for caching and metrical task systems (MTS) that use a reduced number of predictions while maintaining strong performance guarantees.
For caching:
The algorithm, called F&R, consists of two parts: Follower and Robust.
Follower is 1-consistent (optimal with perfect predictions) but lacks robustness and smoothness.
Robust is used when Follower detects an incorrect prediction, providing the desired robustness and smoothness.
The algorithm uses O(f(log k)) OPT predictions, where f is an increasing convex function, and achieves consistency 1, robustness O(log k), and smoothness O(f^-1(η/OPT)).
For general MTS:
The algorithm queries the predictor only once every a time steps, making at most T/a queries in total.
It achieves consistency O(a) and smoothness O(a)(1 + 2η/OFF), where OFF is the cost of an arbitrary offline algorithm and η is the prediction error.
The algorithm can be combined with any online algorithm for MTS to achieve robustness comparable to that algorithm.
The paper also provides lower bounds, showing that the number of predictions used by the algorithms is close to optimal and that the smoothness cannot be improved beyond a certain threshold when the predictions are queried only periodically.
Stats
The number of page faults incurred by the offline optimal solution (OPT) is used to bound the number of predictions required by the caching algorithm.
The cost of an arbitrary offline algorithm (OFF) is used to bound the performance of the MTS algorithm.
The total prediction error (η) is used to quantify the smoothness of both algorithms.
Quotes
"Producing these predictions might be a costly operation – this motivated Im et al. (2022) to introduce the study of algorithms which use predictions parsimoniously."
"We design parsimonious algorithms for caching and MTS with action predictions, proposed by Antoniadis et al. (2023), focusing on the parameters of consistency (performance with perfect predictions) and smoothness (dependence of their performance on the prediction error)."