English

Gradientless Descent: High-Dimensional Zeroth-Order Optimization

Machine Learning 2020-05-20 v4 Optimization and Control Machine Learning

Abstract

Zeroth-order optimization is the process of minimizing an objective f(x)f(x), given oracle access to evaluations at adaptively chosen inputs xx. In this paper, we present two simple yet powerful GradientLess Descent (GLD) algorithms that do not rely on an underlying gradient estimate and are numerically stable. We analyze our algorithm from a novel geometric perspective and present a novel analysis that shows convergence within an ϵ\epsilon-ball of the optimum in O(kQlog(n)log(R/ϵ))O(kQ\log(n)\log(R/\epsilon)) evaluations, for any monotone transform of a smooth and strongly convex objective with latent dimension k<nk < n, where the input dimension is nn, RR is the diameter of the input space and QQ is the condition number. Our rates are the first of its kind to be both 1) poly-logarithmically dependent on dimensionality and 2) invariant under monotone transformations. We further leverage our geometric perspective to show that our analysis is optimal. Both monotone invariance and its ability to utilize a low latent dimensionality are key to the empirical success of our algorithms, as demonstrated on BBOB and MuJoCo benchmarks.

Keywords

Cite

@article{arxiv.1911.06317,
  title  = {Gradientless Descent: High-Dimensional Zeroth-Order Optimization},
  author = {Daniel Golovin and John Karro and Greg Kochanski and Chansoo Lee and Xingyou Song and Qiuyi Zhang},
  journal= {arXiv preprint arXiv:1911.06317},
  year   = {2020}
}

Comments

11 main pages, 26 total pages