There has in several scientific articles related to designing gates for superconducting qubits, been proposed to use reinforcement learning to design pulses with high fidelity and short duration.
Some of the articles:
https://arxiv.org/abs/1902.08418
https://arxiv.org/abs/2311.03684
A simulated example:
https://pennylane.ai/qml/demos/tutorial_rl_pulse/
My question is whether optimizing the parameters for micro-wave pulses even is a reinforcement learning problem, or if it is a more general black-box optimization problem. Reinforcement learning models predict actions based on a state of a system. In the framework proposed in the article, it aims to solve the problem of generating pulse parameters one segment at a time, where actions are pulse parameters, and states are quantum states. I wonder if this problem is really the kind you would use reinforcement learning for. You generally want to design a pulse working for an arbitrary state.
I would really appreciate any opinions on the subject!
