Realizing a deep reinforcement learning agent for real-time quantum feedback.

Area of Science:

• Quantum technologies
• Artificial intelligence
• Control systems

Background:

• Precise real-time control is crucial for quantum technologies, requiring faster-than-coherence-time operations.
• Model-free reinforcement learning (RL) offers a path to discover control strategies without system models.
• Implementing real-time, feedback-driven RL for quantum systems has remained a significant challenge.

Purpose of the Study:

• To implement a reinforcement learning agent for real-time control of a single qubit.
• To demonstrate the agent's capability in efficiently initializing a superconducting qubit.
• To overcome the challenge of developing and training RL agents for low-latency feedback systems.

Main Methods:

• Developed a sub-microsecond-latency neural network agent on a field-programmable gate array (FPGA).
• Utilized model-free reinforcement learning for training the agent.
• Employed measurement-based feedback for agent training and qubit initialization.

Main Results:

• Successfully implemented a real-time reinforcement learning agent for single-qubit control.
• Demonstrated efficient initialization of a superconducting qubit using the RL agent.
• Achieved sub-microsecond latency in the neural network feedback loop.

Conclusions:

• This work presents a viable approach for real-time, model-free reinforcement learning in quantum control.
• The implemented FPGA-based agent facilitates efficient superconducting qubit initialization.
• This represents a significant step towards integrating RL for controlling quantum devices and other low-latency feedback systems.