search

Zeta36 / connect4-alpha-zero

Connect4 reinforcement learning by AlphaGo Zero methods.
MIT License
114 stars 38 forks source link

Can u share your trained best model? #4

Open zephyrrr opened 6 years ago

zephyrrr commented 6 years ago

I have trained a lot of hours and get loss about 1.5. and can't defeat perfect online AI agent with black or white.

Zeta36 commented 6 years ago

Indeed, you cannot defeat a perfect brute force AI player, just get a draw.

zephyrrr commented 6 years ago

I use http://connect4.gamesolver.org/ to test my best model, and my model can't play the best action as web AI did.

theputernerd commented 6 years ago

A perfect agent always wins as player 1. https://en.wikipedia.org/wiki/Connect_Four

theputernerd commented 6 years ago

I would be happy to push the changes with my mcts player vs AI if you would briefly give me permissions. It provides a more definitive quality of opponent which is also incremental based on the MCTS iterations. My fork is already fairly different from this so I went back and made the minimal changes to your code.

Zeta36 commented 6 years ago

yes, @theputernerd. Please make a pull request so I can see your changes. Include also the model weights so I can check your best game :).

Regards.