Cross-Problem Parameter Transfer in Quantum Approximate Optimization Algorithm: A Machine Learning Approach
Kien X. Nguyen, Bao Bach, Ilya Safro · 2025
Read Cross-Problem Parameter Transfer in Quantum Approximate Optimization Algorithm: A Machine Learning Approach by Kien X. Nguyen, Bao Bach, Ilya Safro (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-2504-10733.txt.
Introduction
Kien Nguyen, Bao Bach, and Ilya Safro borrow solutions across problems for quantum computing in a 2025 paper. The quantum approximate optimization algorithm attacks hard scheduling-type problems, but finding its control settings is slow and traps abound. Borrowing settings from a solved problem can warm-start a new one; the open question was whether borrowing works across different problem families.
The authors trained machine learning models to pick good donor settings from MaxCut, a well-studied problem, and applied them to maximum independent set circuits, untouched or as starting points. Transferred settings cut the number of tuning rounds sharply while reaching about the same answer quality. The result hints that quantum problem families share more structure than expected.
Catalog record
| CREATOR | Kien X. Nguyen, Bao Bach, Ilya Safro |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 30 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-10733
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Cross-Problem Parameter Transfer in Quantum Approximate Optimization Algorithm: A Machine Learning Approach Kien X. Nguyen, Bao Bach, Ilya Safro Abstract Quantum Approximate Optimization Algorithm (QAOA) is one of the most promising candidates to achieve the quantum advantage in solving combinatorial optimization problems. The process of finding a good set of variational parameters in the QAOA circuit has proven to be challenging due to multiple factors, such as barren plateaus. As a result, there is growing interest in exploiting parameter transferability , where parameter sets optimized for one problem instance are transferred to another that could be more complex either to estimate the solution or to serve as a warm start for further optimization. But can we transfer parameters from one class of problems to another? Leveraging parameter sets learned from a well-studied class of problems could help navigate the less studied one, reducing optimization overhead and mitigating performance pitfalls. In this paper, we study whether pretrained QAOA parameters of MaxCut can be used as is or to warm start the Maximum Independent Set (MIS) circuits. Specifically, we design machine learning models to find good donor candidates optimized on MaxCut and apply their parameters to MIS acceptors. Our experimental results show that such parameter transfer can significantly reduce the number of optimization iterations required while achieving comparable approximation ratios. Reproducibility: Our results and code are available at: https://github.com/Nyquixt/Cross-Problem-PT-QAOA . Fig. 1: Overview of the parameter transfer procedure. First, QAOA is run on a quantum circuit to optimize for a donor COP (i.e. MaxCut) and retrieve the optimal parameter set (\boldsymbol{\gamma}^{*},\boldsymbol{\beta}^{*}) . Next, those parameters are directly applied to the acceptor COP (i.e. MIS) without additional tuning. I Introduction The class of quantum approximate optimization algorithm (QAOA) [ 1 ] and its variants is one of the most promising candidates to achieve the quantum advantage in the mathematical optimization domain. Its adaptable structure and relatively shallow circuit depth make it well-suited for near-term quantum devices, unlocking a broad range of potential applications, such as finance [ 2 ] , biology [ 3 ] , and scientific computing [ 4 ] . QAOA is a variational algorithm with a direct application to solve combinatorial optimization problems (COPs). This hybrid quantum-classical algorithm consists of a parameterized quantum circuit whose parameters are iteratively updated by a classical optimizer. Despite recent progress in the field, effectively tuning these parameters remains a significant challenge, as the optimization landscape can exhibit barren plateaus when the depth of the circuit or the number of qubits increases [ 5 , 6 , 7 , 8 ] . Furthermore, scaling QAOA to larger problem instances often entails rapidly increasing circuit depths, which not only complicates parameter optimization but also exacerbates hardware-induced errors on near-term quantum devices [ 9 ] . Such challenges have prompted research on parametric initialization methods, such as linear ramp [ 9 , 10 ] or multistart optimization [ 11 ] . Another emerging strategy to overcome the difficulty of finding good parameters is parameter transfer , wherein parameter sets that have already been optimized for one QAOA instance are reused on another [ 12 , 13 , 14 , 15 ] . For example, Brandão et al. [ 13 ] demonstrated that QAOA parameters optimized for the MaxCut problem on a specific 3-regular graph remain nearly optimal across all other 3-regular graphs. Galda et al. [ 14 ] identified different properties of random graphs as a proxy for parameter transferability. Sureshbabu et al. [ 16 ] introduced a straightforward rescaling method to transfer parameters from unweighted to weighted graphs for MaxCut. While these efforts illustrate the effectiveness of parameter transfer within the same COP, an important next step is to explore whether such transferability extends to fundamentally different COPs [ 12 ] . We explore the following question: given a sufficiently large set of QAOA preoptimized circuits (donors) for one COP, can we build a low resource transferability method for circuits (acceptors) of another COP. Montañez et al. note that it remains unclear why certain COPs (i.e. Bin Packing) exhibit better generalization capability than others, and their empirical studies are based on random instances. Our contribution: To address this open question, we extend beyond purely random instances and adopt a data-driven approach, leveraging machine learning (ML) techniques to systematically identify good donor instances optimized on an unconstrained COP MaxCut, transferring their optimal parameters to a more challenging, constrained COP Maximum Independent Set (MIS). In particular, we use graph low dimensional representations (embeddings) to train a neural network that predicts a transfer score , reflecting how well the parameters from a donor graph will perform on an acceptor graph. By encoding both graphs into low-dimensional embeddings and framing donor selection as a learning task, we capture structural relationships that drive parameter transferability and enable us to discover parameter sets that generalize effectively to new or more complex problems. Our empirical results show that the proposed ML model is capable of identifying suitable donors that could transfer their optimized MaxCut parameters to their corresponding accepter on MIS and achieve an approximation ratio of 0.8385 without further tuning. We also conduct further fast optimization on MIS taken the transferred MaxCut parameters as initialization and show that it only takes 10 steps to improve to an approximation ratio of 0.9500 on which is insufficient otherwise. II Related Work II-A Parameter Transferability in QAOA Finding good QAOA parameters remains difficult due to various challenges, such as barren plateaus [ 6 , 7 , 8 ] . Although deeper QAOA circuits can deliver slight improvements in the approximation ratio, they greatly increase computational overhead. Consequently, rather than optimizing parameters from scratch for every new instance, especially for significantly larger ones, a practical strategy is to reuse parameter sets previously optimized for a “related” instance [ 14 , 12 , 15 ] . The transferred parameters from the donor instance can either be applied directly to the accepter instance, as demonstrated in this work, or be used as a warm start for further optimization. Recent studies have begun examining the transferability of QAOA parameters across different COPs [ 12 ] . Montañez et al. [ 12 ] present a numerical study on transferring QAOA parameters between different COPs. Although their results hint at the potential for cross-problem parameter transferability, the scope of their analysis remains limited to a small collection of randomly generated instances. In comparison, we adopt a more comprehensive, data-driven approach, leveraging machine learning techniques to systematically identify good donor candidates on the simpler, unconstrained MaxCut problem, and transfer their parameters to the more challenging, constrained MIS problem. II-B Machine Learning for Quantum Computing Broadly, the intersection of machine learning and quantum computing can be divided into two main areas: Quantum Machine Learning (QML) and Machine Learning for Quantum Computing (ML for QC). QML focuses on harnessing quantum hardware to accelerate or enhance traditional ML tasks, offering speedups for problems such as classification, clustering, or generative modeling [ 17 , 18 , 19 , 20 ] . In contrast, ML for QC applies classical learning techniques to improve or optimize quantum algorithms and devices, tackling challenges such as parameter tuning [ 21 ] , noise mitigation [ 22 ] , or circuit