Learning to Learn with Quantum Optimization via Quantum Neural Networks
Kuan-Cheng Chen 2 3 1 , Hiromichi Matsuyama 4 , Wei-hao Huang 4 · 2025
Read Learning to Learn with Quantum Optimization via Quantum Neural Networks by Kuan-Cheng Chen 2 3 1 , Hiromichi Matsuyama 4 , Wei-hao Huang 4 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-00561.txt.
Introduction
Kuan-Cheng Chen, Hiromichi Matsuyama, and Wei-Hao Huang teach a quantum optimizer to learn how to optimize. Quantum circuits that tackle hard sorting and choice problems need their settings tuned, and that tuning is slow because the energy landscape is rugged and real hardware is noisy. The team pairs the optimizer with a quantum neural network that remembers past runs.
Their network uses a quantum version of long short-term memory, a design built for sequences. Trained on small problems, the learned optimizer then tunes circuits for larger ones. On two standard testbeds, a graph-cutting task and a classic spin model, it converged in fewer rounds and reached better answers than classical tuning baselines. The gains held as problem size grew, which is where naive tuning usually stalls. The authors present meta-learning as a practical route to scalable quantum optimization on today's noisy machines.
Catalog record
| CREATOR | Kuan-Cheng Chen 2 3 1 , Hiromichi Matsuyama 4 , Wei-hao Huang 4 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-00561
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Learning to Learn with Quantum Optimization via Quantum Neural Networks Thanks: Corresponding Author: [email protected] Kuan-Cheng Chen 2 3 1 , Hiromichi Matsuyama 4 , Wei-hao Huang 4 Abstract Quantum Approximate Optimization Algorithms (QAOA) promise efficient solutions to classically intractable combinatorial optimization problems by harnessing shallow-depth quantum circuits. Yet, their performance and scalability often hinge on effective parameter optimization, which remains nontrivial due to rugged energy landscapes and hardware noise. In this work, we introduce a quantum meta-learning framework that combines quantum neural networks, specifically Quantum Long Short-Term Memory (QLSTM) architectures, with QAOA. By training the QLSTM optimizer on smaller graph instances, our approach rapidly generalizes to larger, more complex problems, substantially reducing the number of iterations required for convergence. Through comprehensive benchmarks on Max-Cut and Sherrington-Kirkpatrick model instances, we demonstrate that QLSTM-based optimizers converge faster and achieve higher approximation ratios compared to classical baselines, thereby offering a robust pathway toward scalable quantum optimization in the NISQ era. I Introduction Optimization problems arise in numerous scientific and engineering contexts, such as logistics [ 1 ] , finance [ 2 ] , smart city [ 3 ] and communication [ 4 , 5 , 6 ] . Although classical optimization algorithms have made considerable progress in addressing these problems [ 7 , 8 , 9 ] , they often prove inadequate for large-scale or highly complex tasks due to limited computational resources and intrinsic algorithmic constraints. Quantum computing has emerged as a transformative computational paradigm that harnesses quantum mechanical effects to tackle problems widely believed to be intractable on classical machines. Within this realm, the Quantum Approximate Optimization Algorithm (QAOA) [ 10 ] has attracted considerable attention for its capacity to solve combinatorial optimization problems, particularly NP-hard ones [ 11 ] , using comparatively shallow quantum circuits suited to near-term quantum devices [ 12 , 13 , 14 ] . Despite these advantages, the quest to optimize QAOA parameters remains a formidable challenge due to its non-convex optimization landscape, the presence of numerous local minima, and the practical impact of noise and measurement uncertainties on quantum hardware. In addition, recent work by Bittel et al. demonstrates that the classical training procedure central to variational quantum algorithms (VQAs) is itself NP-hard; furthermore, under the assumption that \bm{P\neq NP} , no polynomial-time algorithm can achieve an optimization error \Delta<1 for all problem instances, highlighting a fundamental limitation in VQA-based hybrid quantum computations [ 15 ] . Recent advancements in quantum machine learning (QML) have provided new opportunities to overcome existing challenges in optimization. Quantum machine learning algorithms utilize quantum properties such as entanglement, superposition, and interference to augment classical learning capabilities [ 16 , 17 ] . Quantum-enhanced recurrent neural networks [ 18 , 19 , 20 ] , particularly Quantum Long Short-Term Memory (QLSTM) networks [ 21 ] , have emerged as effective tools capable of learning sophisticated optimization strategies by harnessing quantum parallelism and intrinsic quantum correlations [ 22 , 23 ] . In this work, we propose a novel hybrid quantum-classical framework integrating Quantum Neural Networks (QNNs) with quantum-enhanced learning optimizers to significantly enhance the performance of quantum optimization algorithms. Specifically, we utilize a QLSTM-based optimizer designed to dynamically optimize the parameters of QAOA circuits. This QLSTM-driven optimization scheme enables the learning of generalized strategies that effectively address a broad range of optimization problems. By embedding the QLSTM within a hybrid quantum-classical architecture, our approach realizes a “learning to learn” paradigm [ 24 ] , markedly accelerating convergence and improving the quality of solutions obtained by quantum optimization methods. This work makes three primary contributions. First, it introduces a hybrid quantum-classical optimization framework that couples QNNs and quantum machine learning optimizers to improve the performance of quantum algorithms. Second, it proposes a QLSTM-based optimizer that learns parameter update rules for QAOA, achieving faster convergence and superior solution quality relative to standard classical methods. Third, it presents an extensive benchmarking procedure against established classical optimizers, including exact solutions to Max-Cut and SK Model instances, thereby establishing a robust reference for evaluating the efficacy of the proposed QLSTM strategy. The structure of this paper is as follows. Section II provides a critical review of recent developments in QAOA, with particular emphasis on hybrid quantum-classical optimization strategies and their inherent limitations. Section III presents the theoretical foundations of QAOA, establishing a basis for the methodological contributions that follow. We also examine the computational bottlenecks associated with variational parameter training and motivate the need for learning-based optimizers. Section IV details our proposed QLSTM-driven meta-learning framework, including its architectural design, mathematical formulation, and training protocol. Section V reports numerical results on both Max-Cut and SK model instances, empirically validating the generalizability, scalability, and performance gains afforded by our method. Finally, Section VI concludes with a summary of contributions and outlines future research directions, particularly toward broader applications within the landscape of variational quantum algorithms. II Related Works QAOA has emerged as a promising class of VQAs designed to address a variety of combinatorial optimization problems. Initially proposed by Farhi et al . [ 10 ] , QAOA integrates classical optimization techniques with quantum operations to variationally minimize a cost function. Despite its conceptual attractiveness and preliminary empirical successes, the original QAOA formulation frequently encounters challenges such as barren plateaus, slow convergence, and limited approximation ratios at shallow circuit depths. Consequently, extensive research efforts have been dedicated to designing more expressive and efficient ansätze that enhance QAOA performance across diverse problem instances. A notable research direction involves multi-parameter or multi-angle approaches, which assign distinct parameters to individual qubits or terms in the cost Hamiltonian. For example, ma-QAOA [ 25 ] introduces a separate parameter for each element within the cost and mixer Hamiltonians, enabling precise control over local interactions. Similarly, QAOA+ [ 26 ] supplements traditional QAOA layers with an additional, problem-independent yet multi-parameterized layer, improving approximation ratios on random regular graphs. Alternative strategies seek to decrease required circuit depth by incorporating counterdiabatic driving or problem-dependent modifications. Notably, DC-QAOA [ 27 , 28 ] utilizes counterdiabatic terms to enhance convergence rates and approximation quality while maintaining shallow circuit depth. Incorporating local fields or modifying the mixer Hamiltonian also provides effective avenues for reducing runtime and addressing hardware constraints. For instance, ab-QAOA [ 29 ] applies local fields to expedite computation in combinatorial optimization tasks, while ADAPT-QAOA [ 30 ] iteratively selects mixer Hamiltonians using gradient-based criteria, generating problem-specific circuits dynamically. Recursive QAOA [ 31 ] progressively reduces problem size through non-local qubit elimination strategies, overcoming limitations arising from locality constraints. Broadening the operator space is another compelling approach. For instance, QAOAnsatz [ 32 ] generalizes conventional cost and mixer operators, facilitating flexible incorporation of both hard and soft constraints. Variants such as GM-QAOA [ 33 ] incorporate Grover-inspired selective phase shifts tailored for vertex covers and traveling salesperson problems, while Th-QAOA [ 34 ] replaces the traditional phase separator with a threshold function suitable for Max-Cut and Max Bisection problems. Constraint-preserving mixers [ 35 ] directly integrate hard constraints within the ansatz, ensuring feasible solutions remain consistently within the solution space. Another significant research trajectory emphasizes improved initialization and feedback mechanisms. FALQON [ 36 ] bypasses classical optimization loops by adjusting parameters based on real-time measurement feedback,