Reinforcement learning optimizes entangled gate sequences and reduces gate count in parameterized quantum circuits

The challenge of building practical quantum computers is focused on overcoming the inherent noise present in current devices, particularly the operations that link qubits. Tom R. Rieckmann, Stefan Scheel, and A. Douglas K. Plato from the Institute of Physics at the University of Rostock presented a method to significantly improve the performance of quantum circuits by intelligently optimizing the sequence of these link operations. Their work designs a more efficient circuit for preparing quantum states and demonstrates a reinforcement learning algorithm that achieves higher fidelity with fewer gates than standard approaches. This advance resolves a critical limitation of quantum computing, allows researchers to get the most out of their existing hardware, and paves the way for more complex and reliable quantum computing.

Optimization of variational quantum circuits using reinforcement learning

This work pioneers reinforcement learning algorithms to optimize quantum circuits for state preparation, addressing the limitations imposed by noise in current quantum computing devices. The researchers focused on minimizing the number of intertwined gates, which are a major source of error, while maintaining the depth of the circuit to suit systems with limited coherence time. In this work, we extend our previous approach by incorporating common single-qubit operations and consistently achieve higher state preparation fidelity with the same gate count compared to standard hardware-efficient designs. The team designed a system in which a reinforcement learning agent learns to optimize sequences of entangled gates within a parameterized universal gate set.

The agent was trained to maximize fidelity when preparing the target quantum state starting from a defined initial state. Importantly, the algorithm takes into account the specific connectivity architecture of the qubits and reflects the physical constraints of actual quantum hardware. The experiments used publicly available quantum computers from IBM, specifically the ibm_manila and ibm_quito systems, and took advantage of documented qubit connections and associated gate errors. To quantify performance, the researchers utilized fidelity, calculated as the trace of the product of the density matrices of the prepared and target states, providing an accurate measure of accuracy. This direct comparison demonstrates the effectiveness of this approach in minimizing gate count and maximizing fidelity on realistic quantum hardware, paving the way for improved performance of variational quantum algorithms. The technique is designed to be deployed in any gate-based quantum computing system, providing a versatile solution for noise reduction and performance improvement.

Optimize noisy quantum circuits with reinforcement learning

Scientists have achieved a breakthrough in optimizing quantum circuits for state preparation, demonstrating a reinforcement learning algorithm that significantly increases fidelity while minimizing the number of entanglement gates. This work addresses a critical limitation of current quantum computing devices that are susceptible to noise generated from gate entanglement and provides a path to more reliable computation on noisy intermediate-scale quantum (NISQ) systems. The researchers focused on optimizing the sequence of entangled gates in a parameterized quantum circuit so that they can limit the total number of gates needed for a given operation while respecting the connectivity architecture of the qubits. The research team developed a reinforcement learning agent that can optimize entangled gate sequences for a parameterized universal gate set, and specifically applied it to the task of quantum state preparation.

The results show that our approach consistently reaches higher state preparation fidelity compared to hardware-efficient analysis, even when using the same number of CNOT gates. This improvement is particularly important because entanglement gates are a major source of error in most experimental systems, and minimizing their use directly leads to improved computational reliability. The algorithm was tested and validated using IBM’s publicly available quantum computer, demonstrating its adaptability to real-world hardware constraints. Experiments reveal that reinforcement learning approaches can tailor gate sequences to the specific architecture of a quantum processor and effectively navigate the complexities of qubit connectivity.

By incorporating arbitrarily parameterized single-qubit gates, the team created a gate set that closely aligns with native gate sets employed in many experimental systems, such as those developed by IBM. This adjustment allows for a more effective optimization process and significantly increases the fidelity of state preparation. The team quantified accuracy using fidelity, a measure of how well the prepared state matches the target state, and demonstrated clear advantages over traditional hardware-efficient analysis. This breakthrough provides a promising new way to improve the performance of quantum computation on NISQ devices, paving the way for more complex and reliable quantum algorithms.

Optimized gate sequence improves quantum fidelity

This study demonstrates a reinforcement learning algorithm that can optimize the sequence of entanglement gates used in quantum state preparation and achieve improved fidelity compared to standard hardware-efficient approaches. By intelligently choosing gate sequences, this algorithm reduces the number of required gates while respecting the connectivity limitations of quantum hardware. The team successfully applied this technique to simulations of existing IBM quantum devices, consistently achieving higher state preparation fidelity with fewer gate counts than traditional layered circuit designs. The findings reveal that gate sequence optimization is particularly beneficial in devices with high noise levels, where minimizing the number of gates is important to maintain fidelity.

Analysis across different devices such as ibm_manila and ibm_quito shows that the algorithm adapts to different noise characteristics and achieves peak fidelity at approximately 9 to 11 CNOT gates. The authors acknowledge that there is a difference between the performance of the simulation and the real device due to the inaccuracy of the noise parameters used in the simulation, and point out that further refinement of these parameters is required for accurate modeling. Future work will focus on addressing the limitations of the simulation by incorporating more realistic noise models and validating the algorithm’s performance on real quantum hardware. The team also plans to investigate the effects of different qubit connectivity constraints and explore the possibility of extending this approach to more complex quantum algorithms.