In MCTS I have searched and found that the mcts itself has not very succesful results but when we combine it with deep reinforcement learning we get perfect results like Google have done in Alpha Go Zero. My question is how to combine mcts with deep reinforcement learning like Alpha Go Zero?
Pure MCTS does not scale to very large-state spaces efficiently by itself. In Alpha Go, the search skeleton still exists but most computationally heavy operations are off-loaded to neural networks. For instance, while MCTS needs to perform rollouts to simulate games, AlphaGo Zero does not, but just uses a trained neural net to evaluate states.
So can you recommend a source to understand the theory behind alpha go zero?
This is a great curriculum to understand & master all the concepts for AlphaGo Zero.
http://www.depthfirstlearning.com/2018/AlphaGoZero
Is it also possible to apply the algorithms to the 2 player board games which have a chance factor like in the backgommon? Especially the MCTS algorithm. I mean how to apply the strategy used in alpha go zero to the games I mentioned above?
Well, even though the mentioned chance factor will induce high variance in game outcomes, I think it is still worth trying self-play based approach. The chance factor will increase the state-space complexity, but with enough number of games, it might still learn. You might also want to check the distributional RL for such domains. https://mtomassoli.github.io/2017/12/08/distributional_rl/
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling