MLQM: Machine Learning Approach for Accelerating Optimal Qubit Mapping

Wenjie Sun, Xiaoyu Li, Lianhui Yu, Zhigang Wang, Geng Chen, Guowu Yang · 2024

Read MLQM: Machine Learning Approach for Accelerating Optimal Qubit Mapping by Wenjie Sun, Xiaoyu Li, Lianhui Yu, Zhigang Wang, Geng Chen, Guowu Yang (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-03249.txt.

Introduction

Wenjie Sun and co-authors use machine learning to speed up optimal qubit layout for quantum circuits, in work from 2024. Mapping a logical circuit onto real hardware is slow when done by exact solvers. Redundant search steps eat most of that time.

Their method, MLQM, prunes the global search space using prior knowledge and learned cues. A data scheme grows the training set by rearranging gates and qubits, fixing sample scarcity. Local search is pruned too, through a solver setting that adapts on the fly. Together the cuts raise solution speed while keeping layouts optimal. Faster mapping matters as circuits outgrow hand tuning.

Catalog record

CREATORWenjie Sun, Xiaoyu Li, Lianhui Yu, Zhigang Wang, Geng Chen, Guowu Yang
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE52 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-03249

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

MLQM: Machine Learning Approach for Accelerating Optimal Qubit Mapping Wenjie Sun, Xiaoyu Li, Lianhui Yu, Zhigang Wang, Geng Chen, Guowu Yang Abstract Quantum circuit mapping is a critical process in quantum computing that involves adapting logical quantum circuits to adhere to hardware constraints, thereby generating physically executable quantum circuits. Current quantum circuit mapping techniques, such as solver-based methods, often encounter challenges related to slow solving speeds due to factors like redundant search iterations. Regarding this issue, we propose a machine learning approach for accelerating optimal qubit mapping (MLQM). First, the method proposes a global search space pruning scheme based on prior knowledge and machine learning, which in turn improves the solution efficiency. Second, to mitigate the scarcity of effective samples in the learning task above, MLQM incorporates a data augmentation and refinement scheme. This scheme enhances the size and diversity of the quantum circuit dataset by exploiting gate assignment and qubit rearrangement. Finally, MLQM also further improves the solution efficiency by pruning the local search space, which is achieved through an adaptive dynamic adjustment mechanism of the solver variables. Compared to state-of-the-art qubit mapping approaches, MLQM achieves optimal qubit mapping with an average solving speed-up ratio of 1.79 and demonstrates an average advantage of 22% in terms of space complexity. † † highlights: MLQM introduces a machine learning approach with novel data augmentation, significantly reducing the quantum circuit mapping search space. An adaptive hyperparameter mechanism dynamically transforms search trends into constraints, enhancing mapping efficiency. Experiments show MLQM achieves 1.79× solving time speedup and 22% memory reduction, advancing quantum circuit optimization. 1 Introduction Quantum computing has the potential to solve complex problems that are intractable for classical computers [ 1 , 2 , 3 , 3 , 4 ] , and its technical research and industrialization process have achieved initial successes in the NISQ era [ 5 , 6 , 7 ] . Among the various technological approaches to implementing quantum computing [ 8 , 9 , 6 ] , superconducting quantum circuits have emerged as one of the most prevalent due to their advantages in designability and scalability [ 10 ] . To ensure the correct execution of quantum programs on superconducting quantum computers, logic synthesis and qubit mapping are two fundamental processes. Logic synthesis involves decomposing high-level gates in quantum programs into single-qubit or two-qubit gates supported by the quantum hardware. Qubit mapping, also known as layout synthesis, is employed to insert swap gates into quantum circuits to satisfy the hardware connectivity constraints of the quantum computer. The evaluation metrics of qubit mapping include circuit depth, number of swap gates, fidelity, etc., and the optimal qubit mapping refers to a mapping in which one or more metrics above are optimal. In previous research, methods for addressing the qubit mapping problem can be broadly categorized into two classes: heuristic methods and solver-based methods. Heuristic methods [ 11 , 12 , 13 , 14 , 15 , 16 , 17 ] utilize meta-heuristic algorithms to determine quantum circuit mapping schemes, with a notable example being the SABRE method proposed by Gushu Li et al. [ 12 ] . While heuristic methods generally exhibit excellent execution speeds, the solutions obtained are often far from optimal and lack sufficient interpretability [ 18 ] , leading to limited practicality. Solver-based approaches to the qubit mapping problem involve transforming the problem into a satisfiability modulo theories (SMT) problem, encoding the objective as constraints, and employing solvers to search for solutions that satisfy the constraints [ 19 , 20 , 21 , 22 ] . Robert Wille first proposed utilizing solver-based Pseudo Boolean Optimization to achieve the optimal number of swap gates in 2014 [ 22 ] . In 2020, Bochen Tan et al. treated layout synthesis as a mathematical optimization problem and employed time-space variable encoding to obtain solutions with either optimal depth or optimal swap gate count [ 21 ] . They also proposed a coarse-grained time model called TB-OLSQ, which reduced time consumption, but the optimality of the circuit depth could not be guaranteed. One year later, Bochen Tan et al. introduced a solver-based approach based on gate absorption methods, using solver constraints to find swap gates absorbable by \text{SU}(4) gates [ 20 ] . Qubit mapping approaches combined solver methods and heuristics were also proposed to address the problem. Abtin Molavi et al. proposed a qubit mapping method considering the problem as a MaxSAT problem [ 23 ] , and local relaxation and cyclic circuit relaxation techniques were applied to reduce time consumption. Tsou-An Wu et al. proposed a divide-and-conquer method, using greedy algorithms to partition the quantum circuit into circuit blocks [ 19 ] , applying solvers within the blocks to find feasible mappings, and employing heuristic methods to find swap gate insertion methods between circuit blocks. Although the time consumption was reduced in heuristic-solver combined approaches, the heuristic part made the methods difficult to find depth-optimal solutions. In summary, as the most promising and practically applied method for quantum computers, solver-based quantum qubit mapping approaches have made substantial progress. However, while obtaining better solutions, the increased search time of solver-based methods has become a pressing issue. The state-of-the-art solver-based method, OLSQ2 [ 24 ] , proposes an iterative approach to solve for the optimal depth and swap gate count, reducing the solution time by simplifying the variable counts and employing a more efficient method of encoding variables and constraints. Nevertheless, there is still room for improvement in this method. In OLSQ2, each global search targeting all candidate solutions consumes significant time, but only the two searches closest to the optimal solution provide practical benefits. Based on the above observations, it is efficient and reasonable to improve the solution efficiency by pruning the solver’s search space. In this work, we propose a machine learning-based approach for accelerating optimal qubit mapping (MLQM) using quantum circuit feature datasets. By constructing quantum circuit feature datasets based on quantum circuits and employing data augmentation and refinement techniques, MLQM aims to provide prior knowledge about the circuit depth and swap gate count using machine learning before the solver’s optimization process and conducts the global search space pruning. The scheme allows the global search over all candidate mapping solutions to start from an initial state closer to the optimal circuit depth and swap gate count. Furthermore, MLQM optimizes the local search process within the solver for constraint verification of candidate mapping schemes, ultimately achieving a comprehensive improvement in the efficiency of quantum circuit mapping. The main contributions of this work are as follows: 1. MLQM innovatively introduces a concise and efficient machine learning approach to the quantum circuit mapping problem, providing high-quality prior knowledge to the solver. With a tailored data augmentation method, MLQM significantly reduces the global search space and accelerates the solving process. 2. MLQM proposes an adaptive adjustment mechanism for solver hyperparameters, which converts search trends into dynamic constraints, reduces the local search space and substantially enhances mapping efficiency. 3. Extensive comparative experiments on widely used quantum computer architectures and circuits demonstrate that MLQM achieves an average speedup ratio of 1.79 in solving time and an average reduction of 22% in memory footprint compared to state-of-the-art methods while ensuring optimal mapping quality. These results

Read the complete text