Efficiently Training Neural Networks for Imperfect Information Games by Sampling Information Sets