Skip to content
Michael Pang edited this page Dec 22, 2017 · 22 revisions

Welcome to the chess-alpha-zero wiki!

What I'm doing on this fork:

  • Input: Positions are 12x8x8 binary feature planes, for last 8 positions. 5 auxiliary 8x8 constant planes for 4 castling flags and 50-move rule for total of 101x8x8. Side to move is always on the bottom of the board.
  • No history, just 18x8x8 input. Simple and reduces overfitting (in theory)
  • Adjudicating games by material to speed up training.
  • Training on the material value and messing around with loss weights
  • Multithreaded MCTS: CPU/lock bound in Python, and just slow even with multiple workers running. However I finally maxed out my GPU usage so "good enough" for now.

TODO:

  • Implement MCTS in C++
  • Variable regularization....

Goals:

  • Get a model that beats the materialistic MCTS agent