Algorithms for Caching and MTS with reduced number of predictions
Karim Abdel Sadek, Marek Elias
TL;DR
This paper introduces parsimonious algorithms for caching and MTS that reduce the reliance on predictions, maintaining strong performance guarantees and adapting to the number of available predictions.
Contribution
The authors develop algorithms for caching and MTS that are parsimonious in their use of predictions, with performance scaling based on the number of available predictions.
Findings
Caching algorithm is 1-consistent and robust.
MTS algorithm's consistency and smoothness scale linearly with available predictions.
Both algorithms match previous guarantees when no prediction restrictions are applied.
Abstract
ML-augmented algorithms utilize predictions to achieve performance beyond their worst-case bounds. Producing these predictions might be a costly operation -- this motivated Im et al. '22 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. '20, focusing on the parameters of consistency (performance with perfect predictions) and smoothness (dependence of their performance on the prediction error). Our algorithm for caching is 1-consistent, robust, and its smoothness deteriorates with the decreasing number of available predictions. We propose an algorithm for general MTS whose consistency and smoothness both scale linearly with the decreasing number of predictions. Without the restriction on the number of available predictions, both algorithms match the…
Peer Reviews
Decision·ICLR 2024 poster
Strengths: 1. The paper is well-written and well-structured. I appreciate that the authors provide sufficient intuitions on appropriate places, which significantly improves the readability of the paper. 2. I particularly like the model where the prediction cannot be obtained by at least some fixed time step. This idea seems new and can capture many applications in practice. 3. The technical contribution is solid. Although the high-level picture of the proposed algorithm follows a typical framew
Weaknesses: I don't see any major weaknesses in the paper. My main concern is that due to its technical nature, one would really be required to check all proofs for correctness. I am unfortunately not able to do that due to the time limit for reviewing. I only checked a few lemmas and I believe they are correct. If all proofs are correct, then I think this is a good paper.
This is a technically solid paper. Online caching and MTS are very classical and important problems in online optimization, and can capture many real-world applications. The paper advances in the direction of augmenting online algorithms by restricted learning. The two problems are good targets since the existence of a robust framework for them alleviates the concern about robustness. The authors show non-trivial trade-offs between the query complexity and the smoothness, which may influence fut
One weakness is that the action prediction setting still looks a little bit weird. To me, particularly in online caching, predicting the next arrival time of the current element makes more sense and is more learnable. It seems hard to learn the optimal actions from historical data in practice. But I understand that this is a setting proposed in the previous work, and the local information predictions (e.g. next arrival time) may not help in the restricted-learning setting.
- I’m impressed by the results and techniques. - The paper is very clearly written, and the comparison to previous works is adequate. - This paper has a very comprehensive discussion of how to potentially implement the predictor and the ways to tune the algorithm for better practical performance — I find this very helpful.
- The MTS part does not seem to connect well enough to the rest of the paper which mainly focuses on caching; especially I cannot see the tight relation in terms of algorithms/techniques. - The experiments only show a marginal advantage over baselines on real datasets.
Videos
Taxonomy
TopicsCaching and Content Delivery · Optimization and Search Problems · Advanced Data Storage Technologies
MethodsMatching The Statements