forked from Zeta36/chess-alpha-zero
-
Notifications
You must be signed in to change notification settings - Fork 14
Preventing draws by repetition
Michael Pang edited this page Dec 17, 2017
·
7 revisions
So there is the problem of every game drawing by repetition...
I don't know much about ML but I guess this is similar to the situation in generative adversarial networks where both sides are performing "well enough" that no training is happening, so maybe research there for a solution.
- Same model on both sides (duh)
- I flipped all the data so white and black data points are isomorphic
- Not enough exploration or noise?
- Not enough training data
- Bug in MCTS??
- Train white and black separately with a color plane
- Augment training data with SL
- Adjudicate the drawn matches
- Keep training (fake it til you make it)
- Introduce some artificial noise, ex. idk??