Adaptive control mechanisms in gradient descent algorithms

TL;DR

This paper introduces a systematic approach to designing adaptive stepsize rules for gradient descent on convex functions, leveraging control theory to improve convergence and robustness.

Contribution

It presents a novel adaptive control framework for gradient descent stepsize selection using Lyapunov analysis, combining past and predicted information.

Findings

Lyapunov-based adaptive stepsize rules improve convergence.

Closed-loop adaptation enhances robustness to errors.

Theoretical and numerical results validate the approach.

Abstract

The problem of designing adaptive stepsize sequences for the gradient descent method applied to convex and locally smooth functions is studied. We take an adaptive control perspective and design update rules for the stepsize that make use of both past (measured) and future (predicted) information. We show that Lyapunov analysis can guide in the systematic design of adaptive parameters striking a balance between convergence rates and robustness to computational errors or inexact gradient information. Theoretical and numerical results indicate that closed-loop adaptation guided by system theory is a promising approach for designing new classes of adaptive optimization algorithms with improved convergence properties.