Neither. It's a plateau-seeking algorithm. More info in the manual under "Training".