IanO wrote: ↑Thu Jun 11, 2020 12:28 am
Another possibility is to use stricter endgame counts which subsume the 50-move rule when training.
As an example, consider the Thai chess variant, Makruk. It has a host of different counts, which depend on the amount of material left.* Does anyone know if there is a project to apply the LC0 framework to Makruk?
* These counts only come into effect when the last pawn is traded off or promoted, so a NN might instead learn to preserve the last pawns as a general endgame technique.
Be aware that it is " Leela Chess Zero" project! ( No human idea except chess rules or pure Neural Network learning from scratch).
They have tons of idea to tweet!
It would probably happen only in " Leela Chess One". project.
The current preferred "zero" method of addressing this issue is MLH (Moves-Left Head). The idea being to learn to predict how many moves remain until the end of the game from a position, and giving a bonus for moves which lead to shorter-game positions when ahead (and a bonus for longer-game positions when behind)
This appears to work pretty well - it doesn't fully solve the problem, but it certainly helps a great deal. It'll debut in TCEC (non-bonus) for the first time this SuFi, it's already been used in CCCC some.