Abstract
Traditional online algorithms encapsulate decision making under uncertainty,and give ways to hedge against all possible future events, while guaranteeing anearly optimal solution as compared to an offline optimum. On the other hand,machine learning algorithms are in the business of extrapolating patterns foundin the data to predict the future, and usually come with strong guarantees onthe expected generalization error. In this work we develop a framework for augmenting online algorithms with amachine learned oracle to achieve competitive ratios that provably improve uponunconditional worst case lower bounds when the oracle has low error. Ourapproach treats the oracle as a complete black box, and is not dependent on itsinner workings, or the exact distribution of its errors. We apply this framework to the traditional caching problem -- creating aneviction strategy for a cache of size $k$. We demonstrate that naivelyfollowing the oracle's recommendations may lead to very poor performance, evenwhen the average error is quite low. Instead we show how to modify the Markeralgorithm to take into account the oracle's predictions, and prove that thiscombined approach achieves a competitive ratio that both (i) decreases as theoracle's error decreases, and (ii) is always capped by $O(\log k)$, which canbe achieved without any oracle input. We complement our results with anempirical evaluation of our algorithm on real world datasets, and show that itperforms well empirically even using simple off-the-shelf predictions.