A technique that prioritizes reusing past training samples with high advantage estimates instead of always collecting new data.