A variance reduction technique in RL that subtracts the average reward of a batch from individual rewards to reduce noise.