A training technique that constrains model behavior by penalizing divergence between policies under different conditions or representations.