A method uses an artificial neural network (ANN) to automatically produce a magnetic resonance (MR) pulse sequence. A first MR signal corresponding to a first tissue and a second MR signal corresponding to a second tissue are identified. An RF pulse to be applied to the first and second tissues is selected. Based on at least the first MR signal, the second MR signal, and the RF pulse, an updated first MR signal and an updated second MR signal are determined. A difference is computed between the updated first MR signal and the updated second MR signal. The difference is added to an accumulated difference. The RF pulse selecting, updated first and second MR signal determination, difference computation and adding are repeated. The ANN is controlled to use reinforcement learning to select the MR imaging pulse sequence based, at least in part, on the accumulated difference.