Core Concepts
This work explores strategies for leveraging warm start algorithms to achieve significantly better performance than competing with a single fixed prediction, by considering stronger benchmarks that account for a set of multiple predictions.
Abstract
The paper considers the problem of using warm start algorithms, which take an instance and a predicted solution as input, and have runtime bounded by the distance between the predicted and true solutions. Previous work has focused on competing with the best fixed prediction in hindsight.
The authors explore settings where warm start algorithms can be used to achieve better performance by competing against stronger benchmarks that consider a set of k predictions. They consider two main settings:
Offline Setting:
The authors show that a simple strategy of running the warm start algorithm in parallel with k predictions can achieve an O(k) approximation to the optimal offline cost, which is the distance from the true solution to the closest of the k predictions.
They then show how to leverage learnable "coarse information" about the instance space, in the form of a k-wise partition, to potentially avoid the O(k) factor.
Online Setting:
The authors formulate an "online ball search" problem to model settings where instances arrive sequentially and are likely to be similar.
They design a competitive algorithm that competes against the best offline strategy of maintaining a set of k moving predictions or "trajectories", where the cost includes both the distance from the true solution to the closest prediction, as well as the total movement of the predictions. This algorithm is deterministic, robust to an adaptive adversary, and oblivious to the value of k.
The key insights are that warm start algorithms are a powerful primitive that can be leveraged in various ways beyond competing with a single fixed prediction, and that structural properties of warm starts, such as running multiple instantiations in parallel, can be exploited to compete with stronger benchmarks.