# Learning to Learn with Quantum Optimization via Quantum Neural Networks

**Authors:** Kuan-Cheng Chen, Hiromichi Matsuyama, Wei-Hao Huang  
**arXiv:** [2505.00561](https://arxiv.org/abs/2505.00561)  
**Date:** 2025-05-01  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2505.00561v1/

## Key points

- Quantum Approximate Optimization Algorithm (QAOA) is a promising approach for solving combinatorial optimization problems efficiently using shallow-depth quantum circuits.
- Performance and scalability of QAOA depend on effective parameter tuning, which can be challenging due to rugged energy landscapes and hardware noise.
- A novel approach introduced in the work combines Quantum Long Short-Term Memory (QLSTM) optimizers with QAOA, leveraging LSTM architectures.
- By training the QLSTM optimizer on smaller graph instances, the approach generalizes rapidly to larger and more complex problems, reducing the number of iterations needed for convergence.
- Extensive benchmarks on Max-Cut and Sherrington-Kirkpatrick model instances show that QLSTM-based optimizers converge faster and achieve higher approximation ratios compared to classical baselines.
- The key innovation is using QLSTM optimizer to recommend parameter updates for QAOA based on historical observations of cost values and parameter adjustments.
- While the study focuses on applying this approach to QAOA for Max-Cut problems, it can be extended to other combinatorial optimization problems by substituting the relevant cost Hamiltonian.

## Summary

In the realm of quantum computing,  have emerged as a promising avenue for efficiently solving classically intractable combinatorial optimization problems. This is achieved by leveraging shallow-depth quantum circuits. However, the performance and scalability of QAOA often rely heavily on effective , which can be challenging due to rugged energy landscapes and hardware noise. To address this issue, a novel  has been introduced in this work. The framework combines , specifically  architectures, with QAOA. By training the QLSTM optimizer on smaller graph instances, the approach is able to rapidly generalize to larger and more complex problems. This results in a significant reduction in the number of iterations required for convergence. Through extensive benchmarks on Max-Cut and Sherrington-Kirkpatrick model instances, it has been demonstrated that QLSTM-based optimizers converge faster and achieve higher approximation ratios compared to classical baselines. The key innovation lies in utilizing the QLSTM optimizer to recommend parameter updates for QAOA based on historical observations of cost values and parameter adjustments. While the focus of this study is on applying this approach to QAOA for Max-Cut problems, the same strategy can be extended to other . This can be done simply by substituting the relevant cost Hamiltonian.