Online Autotuning of Algebraic Multigrid via Contextual Bandits and Reinforcement Learning

DOI10.48550/ARXIV.2611.01193MaRDI QIDQ7414792arXivFDO

Authors Xudong He, Rui Peng Liu, Yuanzhe Xi, Ruiyu Wang, Yilang Ding, Rui Peng Li

Publication date 31 October 2026

Published in arXiv

Tldr We develop an online autotuning method for algebraic multigrid that reduces cumulative time-to-solution over sequences of related linear systems using contextual bandits and reinforcement learning. Our joint autotuner adapts setup parameters within AMG hierarchy construction and relaxation weights between cycles based on problem descriptors, reducing total runtime by 19.8-43.3% over default configuration.

Full work available at URL https://doi.org/10.48550/ARXIV.2611.01193

arXiv classification

cs.LGmath.NA

This page was built for publication: Online Autotuning of Algebraic Multigrid via Contextual Bandits and Reinforcement Learning

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7414792)