Researchers Cut Quantum Circuit Gate Count with Reinforcement Learning

A new method optimises quantum circuits for near-term applications. Khoa Dang Tao and colleagues report that the approach delivers shorter, more effective circuits suitable for noisy intermediate-scale quantum (NISQ) devices. The technique addresses operational limitations caused by decoherence and hardware noise through key improvements to existing optimisation processes. It provides a flexible set of tools which builds upon, rather than duplicates, the functionality of established deterministic techniques. This enables greater efficiency in learning-based circuit reduction.

Deterministic simplification guides quantum circuit gate reduction via reinforced learning

Scientists at Hanyang University, alongside collaborators at Korea 4Research Institute and Korea 3Hanyang Institute, achieved a 44.9 per cent reduction in gates within complex quantum circuits. This level of optimisation exceeds previous methods because it focuses computational effort only where genuine benefit is delivered. Standard reinforcement learning agents repeatedly relearned simplifications previously addressed by conventional techniques, making such gains unattainable until now.

The team’s framework embeds a deterministic Commutation-and-Reduction (CR) algorithm into the training environment, enabling concentration on non-trivial improvements instead of rediscovering basic reductions. Gate counts decreased by an average of one hundred and sixteen across twenty-qubit circuits utilising the CNOT+Pauli basis. Earlier methods like CR alone managed thirty-two gate removals, while standard reinforcement learning achieved eighty.

Applying this approach to larger Clifford’T circuits, five times bigger than those used for initial training, still yielded improvements over conventional techniques; it outperformed Qiskit’s transpiler with between eighteen percent and twenty-two percent reduction in gate counts across various system sizes. Despite these sharp optimisation results, evaluation was conducted on relatively small circuits and does not yet demonstrate scalability towards the thousands of qubits required for practical quantum computation.

Limited generalisation currently constrains scalable quantum circuit design

Ever-shorter circuits are essential for stable quantum computation as they mitigate disruptive noise affecting today’s devices. The team’s new reinforcement learning framework offers a promising route to this goal by intelligently combining machine learning with established optimisation techniques. However, current demonstrations only extend generalisation to circuits five times larger than those used during training; this limitation raises concerns about scalability when tackling genuinely complex problems demanding thousands of qubits, a key hurdle for practical applications.

Automatically simplifying circuit diagrams through cancellation of redundant steps represents genuine progress in embedding established optimisation techniques into the machine learning process. A method for optimising quantum circuits has been successfully demonstrated at Hanyang University via integration of machine learning and existing optimisation methods. Their framework embeds a deterministic Commutation-and-Reduction algorithm directly into reinforcement learning’s training environment; it simplifies circuit diagrams by cancelling redundancies allowing the agent to focus on complex improvements rather than relearning basic reductions already handled by conventional approaches.

The research demonstrates that combining reinforcement learning with a deterministic simplification technique reduces the number of gates required in quantum circuits. This matters because shorter circuits are less susceptible to errors arising from current limitations in quantum hardware, improving computational stability. Across tests using Clifford+T circuits, this approach removed between eighteen percent and twenty-two percent more gates compared to standard optimisation methods like Qiskit’s transpiler. The authors note evaluation was limited to relatively small circuits and further work is needed to confirm scalability towards larger systems.