A zeroth-order block coordinate descent algorithm for huge-scale black-box optimization.

Published in Proceedings of the 38th International Conference on Machine Learning, 2021

Joint with HanQin Cai, Yuchen Lou and Wotao Yin.

We study ultra-high dimensional black-box optimization (over 1 million variables). We combine the ZORO sparse gradient estimator with a block coordinate descent scheme to obtain an algorithm which is both query and time efficient.

Arxiv version

Conference version