If the model does something undesirable or wrong, it is given the equivalent of a shock with a cattle prod. With repetition, this process reshapes the network and the model avoids the ‘bad’ areas.