Extracting state data from game memory (such as boss coordinates, health points, and animation frame numbers) rather than using raw pixel input can improve training efficiency by one to two orders of magnitude.