Andqaser: Quantum Architecture Search Achieves 50% Accuracy Improvement with 20% Depth Reduction

Quantum computing currently faces a fundamental challenge, balancing the need for circuits with minimal depth, essential for reliable operation, and the high accuracy demanded by complex calculations such as error correction. Ioana Moflic, Alexandru Paler, and Akash Kundu, from Aalto University and the University of Helsinki, address this trade-off with a new reinforcement learning technique called QASER. This innovative approach uses carefully designed reward functions to simultaneously optimise for both low circuit depth and high accuracy, achieving significantly better results than existing methods. Benchmarking on circuits used for state preparation demonstrates that QASER consistently produces stable and more efficient designs, improving accuracy by up to 50% while reducing the number of two-qubit gates and overall circuit depth by 20%.

Quantum Circuit Design via Tensor Networks

This research explores methods for enhancing quantum architecture search, a crucial process for designing efficient quantum circuits. Finding optimal circuits for specific tasks is extremely challenging, as the possibilities are vast and evaluating their performance is computationally demanding, hindering progress in near-term quantum computing. Researchers introduced TensorRL-QAS, a reinforcement learning approach that leverages tensor networks to improve the efficiency and performance of quantum architecture search. Tensor networks provide a more compact and efficient way to represent the complex state and action spaces of quantum circuits, allowing the learning agent to explore the search space more thoroughly, addressing scalability issues and potentially leading to the discovery of better quantum circuits.

The field of quantum architecture search is rapidly evolving, with numerous related research areas. Reinforcement learning is a popular approach, but requires careful design of reward functions and efficient exploration strategies. Researchers are also investigating the use of pre-trained circuit components, unsupervised learning to discover meaningful circuit representations, and techniques for optimizing circuits based on gradient descent. Considering the limitations of specific quantum hardware is also crucial for designing practical circuits.

Reinforcement Learning Optimises Quantum Circuit Compilation

Researchers tackled the challenge of balancing circuit depth and accuracy in quantum computing by developing a novel reinforcement learning approach, termed QASER. This method centres on carefully engineered reward functions that simultaneously optimise for seemingly contradictory goals, enabling the compilation of circuits with both reduced depth and improved accuracy. Benchmarking on state preparation circuits demonstrates the effectiveness of QASER, achieving up to 50% improvement in accuracy while simultaneously reducing both two-qubit gate counts and circuit depths by 20%. The research builds upon existing quantum architecture search strategies, but moves beyond limitations encountered with methods based on matrix product states or differentiable search techniques. QASER’s reward engineering facilitates exploration of a broader architectural space, and the team addressed computational bottlenecks associated with repeated quantum simulator queries, enabling the scaling of quantum architecture search to larger qubit systems.

Accurate Quantum Chemistry with Reinforcement Learning

This work presents a novel reinforcement learning approach, termed QASER, designed to overcome the traditional trade-off between circuit depth and accuracy in quantum computation. Researchers achieved significant improvements in compiling circuits for quantum chemistry problems, ranging from 6 to 10 qubits, specifically targeting the ground state preparation of molecules like 6-LiH, 8-H₂O, and 10-H₂O. QASER consistently outperforms state-of-the-art techniques, achieving up to a 50% improvement in accuracy while simultaneously reducing the number of two-qubit gates and overall circuit depth by 20%. In a noisy scenario simulating realistic quantum hardware, QASER achieved a ground state approximation error threshold significantly better than a comparable method for the 6-LiH and 8-H₂O molecules. QASER exhibits accelerated reward accumulation during the learning process, converging to a cumulative reward faster than a comparison method, and reduces the quantum resource requirements, a crucial advantage for post-NISQ-era computation where gate errors accumulate with circuit size.

Optimizing Quantum Circuit Depth and Accuracy

Researchers have achieved a significant advance in quantum architecture search, overcoming a key limitation in the field: the trade-off between circuit depth and accuracy. They developed a novel reinforcement learning approach, called QASER, which utilizes a specifically engineered reward function to simultaneously optimize both of these critical characteristics in quantum circuits. This method demonstrably improves upon existing techniques, achieving up to 50% higher accuracy while reducing the number of two-qubit gates and overall circuit depth by as much as 20%. The team successfully applied QASER to benchmark problems in quantum chemistry, demonstrating its ability to prepare ground state energies with comparable or improved accuracy to prior methods, but with significantly more efficient circuits. QASER accelerates the convergence speed of warm-start quantum architecture search, indicating its potential for broader application.