On quasi-convex smooth optimization problems by a comparison oracle

TL;DR

This paper introduces a comparison oracle-based optimization method for smooth, strictly quasi-convex functions, requiring only function value comparisons rather than derivatives, with proven convergence rates.

Contribution

It develops a comparison-based gradient descent algorithm that minimizes quasi-convex functions using only comparison oracles, advancing derivative-free optimization techniques.

Findings

Proposed algorithm converges with a rate of O((n D^2/ε^2) log(n D / ε)) comparison queries.

Algorithm effectively minimizes smooth, strictly quasi-convex functions using only pairwise function value comparisons.

Theoretical analysis confirms the efficiency of the comparison-based approach in derivative-free optimization.

Abstract

Frequently, when dealing with many machine learning models, optimization problems appear to be challenging due to a limited understanding of the constructions and characterizations of the objective functions in these problems. Therefore, major complications arise when dealing with first-order algorithms, in which gradient computations are challenging or even impossible in various scenarios. For this reason, we resort to derivative-free methods (zeroth-order methods). This paper is devoted to an approach to minimizing quasi-convex functions using a recently proposed comparison oracle only. This oracle compares function values at two points and tells which is larger, thus by the proposed approach, the comparisons are all we need to solve the optimization problem under consideration. The proposed algorithm to solve the considered problem is based on the technique of comparison-based gradient direction estimation and the comparison-based approximation normalized gradient descent. The normalized gradient descent algorithm is an adaptation of gradient descent, which updates according to the direction of the gradients, rather than the gradients themselves. We proved the convergence rate of the proposed algorithm when the objective function is smooth and strictly quasi-convex in $\mathbb{R}^n$, this algorithm needs $\mathcal{O}\left( \left(n D^2/\varepsilon^2 \right) \log\left(n D / \varepsilon\right)\right)$ comparison queries to find an $\varepsilon$-approximate of the optimal solution, where $D$ is an upper bound of the distance between all generated iteration points and an optimal solution.