Training-Free Quantum Architecture Search

Zhimin He, Maijie Deng, Shenggen Zheng, Lvzhou Li, Haozhen SituAAAI 2024 · DOI

TL;DR

This paper introduces a novel "training-free" approach to Quantum Architecture Search (QAS) that significantly speeds up the process of finding optimal quantum circuits. By using two novel proxies – the number of paths in a circuit's Directed Acyclic Graph (DAG) and its expressibility – the method avoids the computationally expensive step of training quantum circuits, achieving substantial improvements in efficiency and speed compared to existing methods.

What Problem Does This Paper Address?

Variational Quantum Algorithms (VQAs) hold great promise for near-term quantum computing, but their performance is highly dependent on the underlying quantum circuit architecture. Designing these circuits is a complex task, and while Quantum Architecture Search (QAS) aims to automate this, existing methods are notoriously time-consuming because they require training each candidate circuit to evaluate its performance. This training process, involving iterative parameter optimization, becomes a significant bottleneck, especially for larger quantum circuits. The paper addresses this by proposing a "training-free" QAS method that bypasses this costly training step altogether.

Methods

The core innovation of this paper is the development of a two-stage progressive training-free QAS (TF-QAS) framework. Instead of training circuits, TF-QAS uses two "proxies" to estimate circuit performance:

  • Path-based Proxy (Zero-Cost): Circuits are represented as Directed Acyclic Graphs (DAGs). The number of paths between the input and output nodes in the DAG is used as a proxy for circuit complexity and potential entanglement. This proxy is computationally very cheap to calculate.
  • Expressibility-based Proxy: This proxy measures how well a quantum circuit can generate a uniform distribution of states across the entire Hilbert space. It's calculated by sampling state fidelities with random parameters. While more computationally intensive than the path-based proxy, it's still orders of magnitude faster than full circuit training and is a strong indicator of performance.

The TF-QAS framework works in two stages:

  1. Initial Filtering: A large number of candidate circuits are sampled. The path-based proxy is used to quickly filter out a substantial portion of unpromising circuits.
  2. Refined Selection: The remaining circuits are then evaluated using the expressibility-based proxy to identify the high-performance ones.

This progressive approach leverages the speed of the path-based proxy for broad filtering and the accuracy of the expressibility-based proxy for fine-grained selection.

Datasets

The paper does not use traditional "datasets" for training in the machine learning sense. Instead, it evaluates its TF-QAS method on three VQE (Variational Quantum Eigensolver) tasks:

  • 6-qubit Transverse Field Ising Model (TFIM)
  • 5-qubit Heisenberg Model
  • 6-qubit Chemical Hamiltonian of BeH2 molecule

These tasks are standard benchmarks for VQE algorithms. The "search space" for circuits is generated using gatewise and layerwise pipelines, inspired by prior QAS research.

Key Insights and Ideas

  • Bypassing Training is Key: The most significant insight is that expensive circuit training can be effectively replaced by proxy metrics for circuit architecture evaluation, drastically reducing computational cost.
  • Synergy of Proxies: Combining a "zero-cost" proxy (path count) for initial broad filtering with a more informative but slightly slower proxy (expressibility) for refinement leads to a highly efficient and effective search strategy.
  • Expressibility as a Strong Predictor: The expressibility of a quantum circuit is a crucial factor for VQA performance, as it dictates the circuit's ability to reach desired states.
  • DAG Complexity as a Useful Filter: Even a simple metric like the number of paths in a circuit's DAG can provide valuable information for pruning the search space, especially when dealing with a vast number of potential architectures.

Key Results

TF-QAS demonstrates substantial improvements over state-of-the-art (SOTA) QAS methods (specifically PQAS) and random sampling (RS):

  • Sampling Efficiency: TF-QAS achieves an average sampling efficiency improvement of 5 to 57 times compared to SOTA QAS. For VQE on TFIM and Heisenberg models, it's nearly 100 times more efficient than RS.
  • Speed: The method is 6 to 17 times faster than SOTA QAS. For example, the average computational cost for VQE tasks was reduced from 8.8 hours to 1.2 hours.
  • Performance: TF-QAS successfully finds high-performance circuits that achieve ground state energies within chemical accuracy for the tested VQE tasks. In one comparison, TF-QAS achieved ground state energy with 100% probability using 35 gates, while a hardware-efficient ansatz (HEA) with 3 layers (42 gates) failed.
  • Correlation: Expressibility shows a strong correlation with ground-truth performance (Spearman correlation up to 0.85), while the path-based proxy shows a weaker but still useful correlation (up to 0.61).

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the path-based proxy has a weaker correlation with circuit performance compared to expressibility.
  • While expressibility is faster than training, it can still be time-consuming for a very large number of circuits, necessitating the two-stage approach.
  • The current work focuses on VQE tasks; extending to other VQA types would be beneficial.
  • Future Work:
    • Investigate training-free QAS that incorporates "trainability" for even larger-scale QAS.
  • Integrate the proposed proxies into existing QAS algorithms like Reinforcement Learning (RL)-based QAS and evolutionary QAS to provide feedback.
  • Explore the application of TF-QAS to other VQA paradigms beyond VQE.

Summary updated

BibTeX
@inproceedings{he2024training,
  author = {Zhimin He and Maijie Deng and Shenggen Zheng and Lvzhou Li and Haozhen Situ},
  title = {{Training-Free Quantum Architecture Search}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i11.29135},
}
Figure