“The training of AlphaGo is an off-policy method because it uses a replay buffer of past games to train the current policy.”