The Quantum Approximate Optimization Algorithm (QAOA) and its variants are leading algorithms to help solve challenging combinatorial optimization problems. QAOA maps these problems to Ising Hamiltonians such that their ground state is the desired solution. Candidate solutions are then drawn from a variational quantum circuit. The performance of QAOA hinges on a set of variational angles. However, finding good angles is an instance-specific and computationally expensive non-convex classical optimization task. Here, we learn the mapping from the problem instance to high-quality QAOA angles. Our learned models replace the per-instance optimization, that we call the solver, with a single forward pass that is several orders of magnitude faster than the solver with a near identical quality. Our models further generalize to unseen topologies such as Barabási-Albert and 2D grid graphs. Experiments on quantum hardware confirm that QAOA samples with angles produced by our open-source models are indistinguishable from solver angles. Crucially, models sensitive to the topology of the problem instance increase the quality of the QAOA angles. Our work amortizes a major bottleneck in QAOA by an up-front data generation and training cost. Importantly, this cost can be amortized over time in settings where many similar problems are optimized on a regular basis, as is often the case in applications for industry.
AI Summary
The paper asks whether instance-specific QAOA angles can be predicted directly from weighted problem graphs, replacing classical optimization without subsequent refinement. It treats angle selection as supervised graph-level regression, comparing seven architectures, including a global-statistics MLP, graph networks, and transformers. Separate models are trained for depths $p=1,\ldots,5$ across a dataset of 12,191 MaxCut instances spanning random-regular, Erdős–Rényi, and weighted heavy-hex graphs with up to 144 vertices. Per-instance weight-scale normalization reduces variation in the $\gamma$ targets.
Training labels come from a depth-one grid search and COBYLA optimization, followed by depth interpolation; energies above depth one use truncated Pauli propagation. On held-out instances, inference is reported to be several orders of magnitude faster than this solver, excluding upfront data-generation and training costs. The best GNN achieves mean solver-relative approximation ratios of 99.946% at $p=3$ and 99.692% at $p=5$; respectively, 86.6% and 52.0% of instances reach at least 99.9% of solver performance. Tests cover unseen grid and Barabási–Albert topologies and Erdős–Rényi graphs with 150–200 vertices, although generalization depends on architecture and the GNN deteriorates substantially on Barabási–Albert graphs. Hardware tests on 630 heavy-hex instances, using 8,192 shots per circuit at $p\leq3$, find similar upper-tail CVaR performance for predicted and solver angles.
Stated limitations include distribution shifts, mismatch between angle-MSE training and energy optimization, and approximate labels that constrain achievable quality. Exact-simulation validation of Pauli propagation is limited to 20-vertex instances at $p=2$; some approximate evaluations yield ratios above one. Evaluation is restricted to MaxCut.
Some signs of AI-generated textIncrementalMissing comparisonOverclaims
Central claim may not hold: The hardware results establish agreement of averaged upper-tail CVaR statistics, not the abstract's stronger claim of indistinguishable sample distributions.
Near-solver agreement for the true QAOA objective is not independently established throughout the claimed range because most comparisons share a truncated evaluator, with exact validation limited to p=2 on 20-node graphs.
The appendix incorrectly states that GNN's worst-instance Reach across seeds at p=5 is 93.41%, whereas its detailed table reports 86.35%.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
The Quantum Approximate Optimization Algorithm (QAOA) and its variants are leading algorithms to help solve challenging combinatorial optimization problems. QAOA maps these problems to Ising Hamiltonians such that their ground state is the desired solution. Candidate solutions are then drawn from a variational quantum circuit. The performance of QAOA hinges on a set of variational angles. However, finding good angles is an instance-specific and computationally expensive non-convex classical optimization task. Here, we learn the mapping from the problem instance to high-quality QAOA angles. Our learned models replace the per-instance optimization, that we call the solver, with a single forward pass that is several orders of magnitude faster than the solver with a near identical quality. Our models further generalize to unseen topologies such as Barabási-Albert and 2D grid graphs. Experiments on quantum hardware confirm that QAOA samples with angles produced by our open-source models are indistinguishable from solver angles. Crucially, models sensitive to the topology of the problem instance increase the quality of the QAOA angles. Our work amortizes a major bottleneck in QAOA by an up-front data generation and training cost. Importantly, this cost can be amortized over time in settings where many similar problems are optimized on a regular basis, as is often the case in applications for industry.
The paper asks whether instance-specific QAOA angles can be predicted directly from weighted problem graphs, replacing classical optimization without subsequent refinement. It treats angle selection as supervised graph-level regression, comparing seven architectures, including a global-statistics MLP, graph networks, and transformers. Separate models are trained for depths p=1,…,5 across a dataset of 12,191 MaxCut instances spanning random-regular, Erdős–Rényi, and weighted heavy-hex graphs with up to 144 vertices. Per-instance weight-scale normalization reduces variation in the γ targets.
Training labels come from a depth-one grid search and COBYLA optimization, followed by depth interpolation; energies above depth one use truncated Pauli propagation. On held-out instances, inference is reported to be several orders of magnitude faster than this solver, excluding upfront data-generation and training costs. The best GNN achieves mean solver-relative approximation ratios of 99.946% at p=3 and 99.692% at p=5; respectively, 86.6% and 52.0% of instances reach at least 99.9% of solver performance. Tests cover unseen grid and Barabási–Albert topologies and Erdős–Rényi graphs with 150–200 vertices, although generalization depends on architecture and the GNN deteriorates substantially on Barabási–Albert graphs. Hardware tests on 630 heavy-hex instances, using 8,192 shots per circuit at p≤3, find similar upper-tail CVaR performance for predicted and solver angles.
Stated limitations include distribution shifts, mismatch between angle-MSE training and energy optimization, and approximate labels that constrain achievable quality. Exact-simulation validation of Pauli propagation is limited to 20-vertex instances at p=2; some approximate evaluations yield ratios above one. Evaluation is restricted to MaxCut.
The emergent practical applicability of the Quantum Approximate Optimization Algorithm (QAOA) for approximate combinatorial optimization is a subject of considerable interest. One of the primary limitations of QAOA is the task of finding a set of good parameters, which is usually done using a variational optimization loop. Parameter transfer, or parameter concentration, is a phenomenon where QAOA angles trained on problem instances that are self-similar tend to perform well for other problem instances from that similar class. This suggests a potentially highly efficient and scalable nonvariational learning method for QAOA angle finding. In this work, we systematically study QAOA parameter transferability from small problem sizes (16 and 27 decision variables) onto large problem instances (up to 156 qubits) for heavy-hex graph Ising models with geometrically local higher-order terms using the Julia based QAOA simulation tool to perform classical angle finding for up to 49 QAOA layers ( p ) . Parameter transfer of the fixed angles is validated using a combination of full state vector, projected entangled pair states, matrix product state, and LOWESA numerical simulations. We find that the QAOA parameter transfer from single instances applied to other (unseen) problem instances does not in general provide monotonically improving performance as a function of p —there are many cases where the performance temporarily decreases as a function of p —but despite this the transferred angles have a general trend of improved expectation value as the QAOA depth increases, in many cases converging close to the true ground-state energy of the 100 + qubit instances. We also sample the hardware-compatible Ising models using the ensemble of transfer-learned QAOA parameters on several superconducting qubit IBM quantum processors with 127, 133, and 156 qubits. We find continuous solution quality improvement of the hardware-compatible QAOA circuits run on the IBM noisy intermediate-scale quantum processors up to p = 5 on , up to p = 9 on , and up to p = 10 on .
The quantum approximate optimization algorithm (QAOA) is a powerful heuristic that seeks to solve combinatorial optimization problems using quantum hardware and classical optimization in tandem. Various methods exist to train the parameterized quantum circuits that serve as an ansatz in QAOA. However, which method works best to identify optimal angles for a given problem instance remains poorly understood, especially at utility-scale, i.e., 100 qubits or more. In this work, we address this challenge through utility-scale benchmarks from which we distill operational guidance for QAOA practitioners. First, we investigate approximation techniques, such as matrix product states and Pauli propagation, to find optimal angles. Second, we train QAOA on small-scale representative problems and transfer the angles to larger ones. We then validate the results on quantum hardware for utility-scale problem instances that can be meaningfully executed. In this way, we identify insights for QAOA angle setting strategies that work best for problems at the utility scale, including as a function of resource cost for the search. Crucially, the operational implications we draw from our benchmarks will help quantum optimization practitioners execute QAOA end-to-end pipelines efficiently on current and future hardware.
Abstract As combinatorial optimization is one of the main quantum computing applications, many methods based on parameterized quantum circuits are being developed. In general, a set of parameters are being tweaked to optimize a cost function out of the quantum circuit output. One of these algorithms, the Quantum Approximate Optimization Algorithm stands out as a promising approach to tackling combinatorial problems. However, finding the appropriate parameters is a difficult task. Although QAOA exhibits concentration properties, they can depend on instances characteristics that may not be easy to identify, but may nonetheless offer useful information to find good parameters. In this work, we study unsupervised Machine Learning approaches for setting these parameters without optimization. We perform clustering with the angle values but also instances encodings (using instance features or the output of a variational graph autoencoder), and compare different approaches. These angle-finding strategies can be used to reduce calls to quantum circuits when leveraging QAOA as a subroutine. We showcase them within Recursive-QAOA up to depth 3 where the number of QAOA parameters used per iteration is limited to 3, achieving a median approximation ratio of 0.94 for MaxCut over 200 Erdős-Rényi graphs. We obtain similar performances to the case where we extensively optimize the angles, hence saving numerous circuit calls.
The quantum approximate optimization algorithm (QAOA) generates an approximate solution to combinatorial optimization problems using a variational ansatz circuit defined by parameterized layers of quantum evolution. In theory, the approximation improves with increasing ansatz depth but gate noise and circuit complexity undermine performance in practice. Here, we investigate a multi-angle ansatz for QAOA that reduces circuit depth and improves the approximation ratio by increasing the number of classical parameters. Even though the number of parameters increases, our results indicate that good parameters can be found in polynomial time for a test dataset we consider. This new ansatz gives a 33% increase in the approximation ratio for an infinite family of MaxCut instances over QAOA. The optimal performance is lower bounded by the conventional ansatz, and we present empirical results for graphs on eight vertices that one layer of the multi-angle anstaz is comparable to three layers of the traditional ansatz on MaxCut problems. Similarly, multi-angle QAOA yields a higher approximation ratio than QAOA at the same depth on a collection of MaxCut instances on fifty and one-hundred vertex graphs. Many of the optimized parameters are found to be zero, so their associated gates can be removed from the circuit, further decreasing the circuit depth. These results indicate that multi-angle QAOA requires shallower circuits to solve problems than QAOA, making it more viable for near-term intermediate-scale quantum devices.
The main object of study of this work is the Quantum Approximate Optimization Algorithm (QAOA) —the best-known heuristic quantum algorithm for addressing combinatorial optimization problems. Despite its popularity, QAOA poses major challenges. Namely, its performance remains poorly understood, especially in the low depth regime most relevant to near-term quantum computing. Besides, the algorithm is of variational nature, relying on several hyperparameters to be classically trained. In this work, we contribute to addressing these issues by developing techniques to understand the performance of QAOA at low depth, and deducing from these pretrained hyperparameters, thereby eliminating or considerably reducing the need for a classical optimization loop. We first consider applying QAOA to constraint satisfaction problems with the goal of maximizing the algorithm’s approximation ratio (fraction of satisfied constraints). As a particular example, we consider QAOA applied to MAXCUT on non-regular graphs, showing how analytic techniques provide good guess hyperparameters in this setting. Next, we analyze the performance of QAOA on a well-studied constraint satisfaction problem: random k-SAT. In this case, the performance is measured by the probability of finding an assignment satisfying all constraints (“success probability”). To estimate this quantity, we develop new techniques, mapping quantum expectation values of the QAOA circuits to classical spin models. We further show how these techniques can be generalized to a larger family of random constraint satisfaction problems. We conclude the work by two feasibility studies inspired by real-world use cases encountered in collaborations with industrial partners: peptide sampling, and solving flow problems. These problems are naturally phrased by encoding combinatorial variables in qudits, a relatively unexplored setting in the existing literature. We present partial analytic and numerical results for optimal QAOA hyperparameters and the scaling of its success probability in this framework, underlining potential specific difficulties when using QAOA with qudits.
Quantum Approximate Optimization Algorithm (QAOA) enables solving combinatorial optimization problems on quantum computers by optimizing variational parameters for quantum circuits. We investigate a simplified approach that combines linear parameterization with parameter transferring, reducing the parameter space to just 4 dimensions regardless of the number of layers. This simplification draws inspiration from quantum annealing schedules providing both theoretical grounding and practical advantages. We compare this combined approach with standard QAOA and other parameter setting strategies such as INTERP and FOURIER, which require computationally demanding incremental layer-by-layer optimization. Notably, previously known methods like INTERP and FOURIER yield parameters that can be well fitted by linear functions, which supports our linearization strategy. Our analysis reveals that for the random Ising model, cost landscapes in this reduced parameter space demonstrate consistent structural patterns across different problem instances. Our experiments extend from classical simulation to actual quantum hardware implementation on IBM's Eagle processor, demonstrating the approach's viability on current NISQ devices. Furthermore, the numerical results indicate that parameter transferability primarily depends on the energy scale of problem instances, with normalization techniques improving transfer quality. Most of our numerical experiments are conducted on the random Ising model, while problem-dependence is also investigated across other models. A key advantage of parameter transferring is the complete elimination of instance-specific classical optimization overhead, as pre-trained parameters can be directly applied to other problem instances, reducing classical optimization costs by orders of magnitude for deeper circuits.