advantage = actual_reward - expected_reward