Meta-learning digitized-counterdiabatic quantum optimization

Abstract

The use of variational quantum algorithms for optimization tasks has emerged as a crucial application for the current noisy intermediate-scale quantum computers. However, these algorithms face significant difficulties in finding suitable ansatz and appropriate initial parameters. In this paper, we employ meta-learning using recurrent neural networks to address these issues for the recently proposed digitized-counterdiabatic quantum approximate optimization algorithm. By combining meta-learning and counterdiabaticity, we find suitable variational parameters and reduce the number of optimization iterations required.
We demonstrate the effectiveness of our approach by applying it to the MaxCut problem and the Sherrington-Kirkpatrick (SK) model. Our method offers a short-depth circuit ansatz with optimal initial parameters, thus improving the performance of the state-of-the-art QAOA.

Shortcuts to adiabaticity: theoretical framework, relations between different methods, and versatile approximations

“…The performance of CD-QAOA depends on the choice of initial parameters and classical optimizers. Reinforcement learning was used for optimization [98], and a meta-learning technique using recurrent neural networks was considered for initial parameter setting [101]. As another development, the power of single-layer QAOA [102] and improvement by higher-order counterdiabatic terms [103] were confirmed, and the theory was extended to photonic systems with continuous variables [104].…”

Physics-informed neural networks for an optimal counterdiabatic quantum computation

“…However, practical implementation often requires speeding up adiabatic evolution, leading to various approaches like shortcuts to adiabaticity (STA) [34,35], including fast-forward methods, invariant-based inverse engineering, and counterdiabatic (CD) protocols, which is the main focus of this study. Readers interested in more detailed developments on quantum optimization processes for CD protocols and direct applications are directed to works such as [36,37] (as well as references therein).…”

Enhancing Quantum Approximate Optimization with CNN-CVaR Integration

“…Notably, the trade-off between circuit depth and performance, constrained by coherence times in existing and near-term quantum processors, poses a significant challenge [19,20,21,22]. Additionally, the complexity introduced by the QAOA ansatz, with its increasing number of variational parameters, presents challenges for classical optimizers [23,24,25]. Previous studies have shown that the optimal QAOA parameters exhibit specific patterns [26,27], leading to depth-sequential strategies and machine learning-based methods for parameter initialization [10,23,28,29,30].…”

“…Additionally, the complexity introduced by the QAOA ansatz, with its increasing number of variational parameters, presents challenges for classical optimizers [23,24,25]. Previous studies have shown that the optimal QAOA parameters exhibit specific patterns [26,27], leading to depth-sequential strategies and machine learning-based methods for parameter initialization [10,23,28,29,30]. For instance, Alam et al adopted a regression model to predict high-depth parameters from low-depth ones [26], while Amosy et al applied iterative neural networks for parameter initialization, selecting the most promising parameter sets from cluster centers for the given problem [31].…”