Selecting a small subset of real samples from a large dataset that best represent the original data distribution.