search

Akababa / Chess-Zero

Chess reinforcement learning by AlphaZero methods.
MIT License
38 stars 14 forks source link

Known bugs #1

Closed Akababa closed 6 years ago

Akababa commented 6 years ago

I'll write the bugs I found here in case I have to rebase:

  1. Evaluate: when automatically drawing a long game, it counts as a win for one player! Really bad
  2. ChessPlayer: doesn't reset MCTS nodes in between moves. It's like keeping the hash table of moves, except it messes up MCTS node statistics in a way I don't understand
  3. Optimize: I just realized I was training everything on the starting position... ALWAYS VALIDATE INPUT DATA
  4. (minor) SL needlessly loads model
Akababa commented 6 years ago

https://github.com/Akababa/chess-alpha-zero/wiki/Known-Bugs