2 ms·
The current AlphaZero system has little in common with deep reinforcement learning; the NN-guided MTCS is effective but not something that can be applied to gen
by PeterisP 9y ago
The current AlphaZero system has little in common with deep reinforcement learning; the NN-guided MTCS is effective but not something that can be applied to general reinforcement learning tasks. And the first AlphaGo system was highly reliant on standard supervised learning (training the value network and policy network on grandmaster games).
If anything, AlphaGo, AlphaGo Zero and AlphaZero are illustrations of how "pure" deep reinforcement learning is insufficient, that the non-RL parts have an enormous impact.