In this talk, I will present our recent work to apply reinforcement learning to quantum parameter estimation. In single-parameter case [1], we have shown that the control generated by our method is more generalizable than traditional methods such as GRAPE, namely the pulse sequences generated by the trained neuron network can be easily used to measure parameters having a range of values. We further extend the work to cases involving multiple parameters [2] and found that the generalizability of reinforcement learning mostly holds, which becomes much more significant for estimating an ensemble of systems with parameters varied in certain ranges. In the examples that we consider, each GRAPE run, on average, takes tens of hours on a typical CPU, while for reinforcement learning the time is only a few seconds. Therefore it quickly becomes prohibitively expensive for GRAPE to optimize controls of every parameter in the ensemble of systems, while the reinforcement learning method, generating optimal or suboptimal solutions, remains practical. Our results suggest that the usefulness of reinforcement learning, previously under-appreciated in quantum metrology, may play an important role given its generalizability, especially when massive measurements of an ensemble of systems are required.

[1] H. Xu, J. Li, L. Liu, Y. Wang, H. Yuan, and XW, npj Quantum Inf. 5, 82 (2019).
[2] H. Xu, L. Wang, H. Yuan, and XW, in preparation (2020).


Meeting Link: To request for meeting link, please write to phweb@ust.hk.