Accelerated VQE: Parameter Recycling for Similar Recurring Problem Instances

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

This paper, recently published by researchers at LMU Munich and Aqarios GmbH, explores a practical shortcut for one of the most popular algorithms in the NISQ (Noisy Intermediate-Scale Quantum) era: the Variational Quantum Eigensolver (VQE).

TL;DR

By "recycling" optimized circuit parameters from previous, similar problem instances—a technique known as Transfer Learning—researchers reduced VQE training time by 93%. This approach significantly outperforms standard initialization methods in time-sensitive scenarios, providing a "warm-start" that bypasses the most computationally expensive parts of quantum optimization.

What Problem Does This Paper Address?

The Variational Quantum Eigensolver (VQE) is a hybrid algorithm used to solve optimization problems by toggling between a quantum processor (which evaluates a cost function) and a classical optimizer (which updates circuit parameters). However, VQE is notoriously "expensive" to train. It often suffers from slow convergence and "barren plateaus"—flat regions in the optimization landscape where the algorithm gets stuck.

In industries like logistics (e.g., vehicle routing), companies often solve the same type of problem repeatedly with only minor variations. Standard VQE treats every new instance as a total mystery, starting parameters from scratch (usually near zero). The authors ask: Can we use what we learned from yesterday’s problem to solve today’s problem faster?

Methods

The researchers applied Transfer Learning (TL) to the MaxCut problem (partitioning a graph to maximize the edges between sets). Their workflow involved:

  1. Pre-training: Solving a "source" graph instance to find optimal parameters ($\theta_{pre}$).
  2. Parameter Recycling: Using those parameters as the starting point for a "target" graph instance.
  3. Circuit Variants: They tested two architectures: a "base-circuit" and a "had-circuit" (which includes an initial layer of Hadamard gates to create a superposition of all possible solutions).
  4. Accelerated VQE: A "sprint" version of the algorithm capped at only 50 iterations (compared to the standard 1,000) to test how well the recycled parameters perform under strict time constraints.

Datasets

The authors generated a custom dataset derived from California’s road network (originally containing 1.9 million nodes).

  • They sampled 700 pairs of 12-node graphs.
  • Each pair consisted of a "source" and a "target" graph.
  • The pairs were categorized by Hamming Distance (HD), which measures how many "bit flips" are needed to turn the optimal solution of the source graph into the optimal solution of the target graph. This allowed them to quantify exactly how "similarity" affects performance.

Key Insights and Ideas

  • The Proximity Principle: Transfer learning isn't a magic wand; its effectiveness is directly tied to the Hamming Distance between solutions. If the source and target problems are mathematically "close" (HD 0–3), the recycled parameters provide a massive advantage.
  • Superposition as a Safety Net: The had-circuit (using Hadamard gates) proved more robust. By starting in a state of global superposition, it avoids the "worst-case" starting point of the base circuit, where all nodes are initially lumped into a single partition.
  • Breaking the "Zero-Start" Habit: Standard practice is to initialize parameters near zero. The paper shows that for recurring problems, this is the least efficient path. Recycled parameters start the optimization at a much lower "energy" level, essentially placing the algorithm on the home stretch before it even begins.

Key Results

  • 93% Time Reduction: In the "Accelerated VQE" tests, the algorithm reached high-quality solutions in just 50 iterations, representing a 93% reduction in training effort compared to full convergence.
  • Superior Solution Quality: When forced to stop early (at 50 iterations), the Transfer Learning approach beat the standard "start-from-zero" approach by 7 to 9 percentage points in solution quality (approximation ratio).
  • Winning Rate: The TL approach outperformed standard training in approximately 56-58% of all tested instances in the accelerated configuration.

Limitations and Future Work

  • The Scalability Question: The study focused on 12-qubit systems. While promising, the authors acknowledge that further research is needed to see if these benefits hold as we scale to hundreds of qubits, where the "parameter space" becomes exponentially more complex.
  • Local Minima Traps: There is a risk that if a source and target are too different, the recycled parameters might trap the optimizer in a local minimum, preventing it from finding the true global optimum of the new problem.
  • Expanding the Problem Set: The authors suggest moving beyond MaxCut to more complex combinatorial challenges like the Capacitated Vehicle Routing Problem (CVRP) and exploring weighted graphs.

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},
}

Loading related content...

Figure