A technique to find good weight values in a neural network by trying different weights, and seeing if the change contributes positively to prediction quality.