- A blog explain how mcts works
- AlphaZero connect four, there are someting interesting ideas, like use Q to instead of Z to be training target.
- Some notices about mcts
- tensorflow AGZ implement(NOT an official version of AlphaGo)
- A renju program follow AlphaZero. Here is a chinese doc about it. I think this implemention has some bugs. For example the author compute final z by mistake, but when do backup he adds a negative sign, so maybe the result is ok.
- This project build with bazel==2.0.0
- python requirements
-
Training about 24 hours, each move with 100 mcts simulations:
-
Training about 48 hours, each move with 400 mcts simulations:
test