Using sparse human demonstrations to reshape a critic's value estimates, guiding autonomous exploration toward expert-preferred behaviors.