Hi Matthew,
Thanks for sharing the project of Giraffe! I am just trying to replicate the experiments of training a NN evaluator for Giraffe. The first few iterations running looks fine. I just need to find a faster machine for more iterations (If just using i thread, a iteration will take 23 mins on my laptop).
Besides running experiments, I am just thinking to make two minor changes on Giraffe's training setting, but not sure this has been tried by you or someone else:
1) introduce a strong opponent to play against Giraffe instead of self-playing (only the positions with Giraffe's turn will be added to the training position set);
2) play to the end of game with a result instead of playing a fixed number of moves;
I know that this has been out of the target of project that tries to learn an evaluator from just self-playing, but just curious. Not sure whether this makes sense or not. I don't know how sensitive that TD-Leaf algorithm is to the quality of games (either by self-playing or playing-against-opponent).
Deep Pink: a chess engine using deep learning
Moderators: hgm, Rebel, chrisw
-
- Posts: 74
- Joined: Sat Dec 18, 2010 5:19 pm
- Location: Tianjin, China
- Full name: Chao M.