Restricted Global Optimization for QAOA

TL;DR

This paper demonstrates that restricted global optimization methods outperform local optimizers in tuning QAOA parameters, leading to improved performance in solving combinatorial problems with fewer function evaluations.

Contribution

The study introduces the effectiveness of restricted global optimization techniques over local methods for QAOA parameter tuning, addressing a key challenge in quantum algorithm optimization.

Findings

Restricted global optimizers outperform local optimizers in QAOA.

Global methods improve QAOA performance across diverse instances.

Restricted global optimization requires fewer function evaluations than unrestricted methods.

Abstract

The Quantum Approximate Optimization Algorithm (QAOA) has emerged as a promising variational quantum algorithm for addressing NP hard combinatorial optimization problems. However, a significant limitation lies in optimizing its classical parameters, which is in itself an NP hard problem. To circumvent this obstacle, initialization heuristics, enhanced problem encodings and beneficial problem scalings have been proposed. While such strategies further improve QAOA's performance, their remaining problem is the sole utilization of local optimizers. We show that local optimization methods are inherently inadequate within the complex cost landscape of QAOA. Instead, global optimization techniques greatly improve QAOA's performance across diverse problem instances. While global optimization generally requires high numbers of function evaluations, we demonstrate how restricted global optimizers still show better performance without requiring an exceeding amount of function evaluations.