Accelerated VQE: Parameter Recycling for Similar Recurring Problem Instances

Tobias Rohe, Maximilian Balthasar Mansky, Michael Kölle, Jonas Stein, Leo Sünkel, Claudia Linnhoff‐Popien

Abstract: Training the Variational Quantum Eigensolver (VQE) is a task that requires substantial compute. We propose the use of concepts from transfer learning to considerably reduce the training time when solving similar problem instances. We demonstrate that its utilisation leads to accelerated convergence and provides a similar quality of results compared to circuits with parameters initialised around zero. Further, transfer learning works better when the distance between the source-solution is close to that of the target-solution. Based on these findings, we present an accelerated VQE approach tested on the MaxCut problem with a problem size of 12 nodes solved with two different circuits utilised. We compare our results against a random baseline and non transfer learning trained circuits. Our experiments demonstrate that transfer learning can reduce training time by around 93% in post-training, relative to identical circuits without the use of transfer learning. The accelerated VQE approach beats the standard approach by seven, respectively nine percentage points in terms of solution quality, if the early-stopping is considered. In settings where time-to-solution or computational costs are critical, this approach provides a significant advantage, having an improved trade-off between training effort and solution quality.

Introduction

The Variational Quantum Eigensolver (VQE) has emerged as a cornerstone of the Noisy Intermediate-Scale Quantum (NISQ) era. As a hybrid algorithm, it leverages both quantum and classical resources to find the ground state energy of a Hamiltonian. However, VQE faces significant hurdles: it is computationally expensive to train, often plagued by slow convergence and the presence of "barren plateaus"—regions in the optimization landscape where gradients vanish, stalling progress.

In many industrial applications, such as logistics or network optimization, practitioners frequently encounter recurring problem instances that share structural similarities. Standard VQE implementations treat each new instance as an independent task, typically initializing parameters near zero. This research explores whether "recycling" optimized parameters from a previously solved similar problem—a form of Transfer Learning (TL)—can provide a "warm-start" to bypass the most intensive phases of training.

Methodology

Dataset Generation and Workflow

To evaluate the effectiveness of parameter recycling, the researchers utilized the MaxCut problem, a classic combinatorial challenge. The dataset was derived from the California road network, from which 700 pairs of 12-node graphs were sampled. Each pair consisted of a "source" graph and a "target" graph.

The experimental workflow: (a) generating source and target graphs from the California Street Network and (b) the quantum experiment process for solving target instances using recycled parameters.

The similarity between graphs was quantified using the Hamming Distance (HD), measuring the number of bit flips required to transform the optimal solution of the source graph into that of the target. This metric allowed the team to test the "Proximity Principle": the hypothesis that TL is most effective when the source and target problems are mathematically close.

Circuit Architectures

Two distinct circuit architectures were tested:

  1. Base-circuit: A standard variational form.
  2. Hadamard-circuit (Had-circuit): Includes an initial layer of Hadamard gates to create a global superposition of all possible solutions, acting as a safety net to avoid poor initial partitions.

Key Results

Accelerated Convergence

The primary finding of the study is a dramatic reduction in training effort. By utilizing recycled parameters ($\theta_{pre}$), the VQE reached high-quality solutions significantly faster than standard initialization. In "Accelerated VQE" tests capped at only 50 iterations, the TL approach achieved a 93% reduction in training time compared to the 1,000 iterations required for full convergence from scratch.

Comparison of cost function convergence for Base-circuit (left) and Had-circuit (right). The orange lines represent the accelerated "warm-start" provided by transfer learning, showing immediate lower energy states compared to standard initialization (blue).

Solution Quality and Winning Rate

When forced to stop early (the "sprint" scenario), the TL approach outperformed standard initialization by 7 to 9 percentage points in approximation ratio. The TL method proved superior in approximately 56-58% of all tested instances. Notably, the Had-circuit demonstrated greater robustness, as the initial superposition prevented the algorithm from starting at the "worst-case" energy levels often seen in the base circuit.

Discussion and Conclusion

This research demonstrates that parameter recycling is a highly effective strategy for recurring optimization problems. By breaking the habit of "zero-start" initialization, researchers can significantly lower the computational barrier for VQE.

However, limitations remain. The study focused on 12-qubit systems; as systems scale to hundreds of qubits, the complexity of the parameter space may introduce new challenges, such as trapping the optimizer in local minima if the source and target problems are too dissimilar. Future work will involve testing these methods on larger, more complex combinatorial problems like the Capacitated Vehicle Routing Problem (CVRP) to further validate the scalability of the transfer learning approach in quantum optimization.

Summary updated

Abstract

Training the Variational Quantum Eigensolver (VQE) is a task that requires substantial compute. We propose the use of concepts from transfer learning to considerably reduce the training time when solving similar problem instances. We demonstrate that its utilisation leads to accelerated convergence and provides a similar quality of results compared to circuits with parameters initialised around zero. Further, transfer learning works better when the distance between the source-solution is close to that of the target-solution. Based on these findings, we present an accelerated VQE approach tested on the MaxCut problem with a problem size of 12 nodes solved with two different circuits utilised. We compare our results against a random baseline and non transfer learning trained circuits. Our experiments demonstrate that transfer learning can reduce training time by around 93\% in post-training, relative to identical circuits without the use of transfer learning. The accelerated VQE approach beats the standard approach by seven, respectively nine percentage points in terms of solution quality, if the early-stopping is considered. In settings where time-to-solution or computational costs are critical, this approach provides a significant advantage, having an improved trade-off between training effort and solution quality.
BibTeX
@inproceedings{rohe2025accelerated,
  author = {Tobias Rohe and Maximilian Balthasar Mansky and Michael Kölle and Jonas Stein and Leo Sünkel and Claudia Linnhoff‐Popien},
  title = {{Accelerated VQE: Parameter Recycling for Similar Recurring Problem Instances}},
  booktitle = {I4CS},
  year = {2025},
  doi = {10.1007/978-3-031-94263-1\_5},
}
Figure