“The loss function for AlphaZero has two main components: one that trains the policy network to match the MCTS search distribution, and another that trains the value network to predict the final game outcome.”