Skip to content

Ways to cheat in training

Michael Pang edited this page Dec 20, 2017 · 7 revisions

Here are the top ways to cheat during training if you don't have 5000 TPUs:

  1. Train naked network on another chess engine's output, or just write your own evaluation function.
  2. Adjudicate games during self-play based on heuristics like material
  3. Train only on positions where the network makes blunders (check with engine). Thus reducing noise during training.
  4. Instead of self-play, use engine vs. network games.
  5. Set loss weight of policy to 0.1 and replace values by your own heuristic.
  6. Train on high elo vs low elo and zero out the policy of the worse player.
  7. SCID filters: white 2000+, black 1300-, 1-0, PGN contains: checkmated