“…This is possible only in simulated environments, for example by providing the learning agent with full knowledge of the system's wavefunction, which supplies enough information for decision making [19, 22-24, 26, 28, 32, 36, 37]. Moreover, since in the simulation the distance to the target state or operation is known at every step of the quantum trajectory, it can be used to construct a steady reward signal to guide the learning algorithm [22][23][24]36], thereby alleviating the well-known delayed reward assignment problem. Taking RL a step closer towards quantum observability, some works do not give the agent access to the wave-function, but still use it for calculation of fidelities and expectation values in different parts of the training pipeline [20,25,27,48,49], which would require a prohibitive amount of averaging in experiment.…”