In these investigations, a "teaching machine" was used to train pre-school and first-grade children in a series of progressively difficult discrimination tasks, leading up to matching to sample. Such training was much more efficient than training in the final discrimination alone. The errors the subjects made were found to be a functon both of the differences between consecutive discriminations (the "size of the steps" in the program) and the length of training on each discrimination. Theoretical and practical implications of these findings are discussed.