A replay buffer technique that samples more frequently from experiences with larger TD errors, focusing learning on surprising or informative transitions.