Characterization of Model-Based Reinforcement Learning for Dynamical Decoupling

Dynamical decoupling pulse sequences are frequently employed in nuclear magnetic resonance, quantum information and sensing to decouple quantum systems from environmental noise, with developments that have informed sequence design across an array of experimental platforms. We algorithmically search for variations on dynamical decoupling pulse sequences by allowing a neural network based agent to iteratively select pulses from a library of experimentally feasible pulses, with the goal of optimizing the quantum state fidelity. The pulses are applied to simulations of dipolar coupled spin systems, with varying interaction strength, magnetic disorder, and pulse control errors. We employ both model-based (the Dreamer model series, for example) and standard RL techniques (Proximal Policy Optimization, Deep Q Networks) and compare their performance, informing future work in algorithmic pulse sequence discovery.