Compressed Sensing (CS) and parallel MRI (pMRI) have been successfully applied to accelerate MRI-data acquisition. CS requires incoherence, usually achieved by random undersampling the data, but pMRI does not. Combined, these methods allow even higher acceleration rates. However, it is unknown how the sampling pattern (SP) should be selected. It is also unknown if the SP is dependent on the reconstruction method. Here we demonstrate, using a new algorithm, that the SP can be learned from given data and reconstruction method. Our results show that the learned SP is superior to others such as Poisson disk and variable density.