Page 1 of 1
Imbalance
Posted: Sat Feb 01, 2020 9:05 am
by xr_a_y
How much to expect from the well know quadratic imbalance terms ?
I'm trying this right know, and first of all, tuning is very long to converge and yet I get nearly 0elo from it.
I was expecting a lower draw ratio with this knowledge but it doesn't seem to be the case.
Who tried it already apart from SF and defenchess ?
Re: Imbalance
Posted: Sat Feb 01, 2020 6:24 pm
by jdart
I don't use Stockfish like formulas but have quite a few special-case imbalance terms. The tuner sets these to non-zero values so it thinks they have some value.
I don't know what the ELO gain is but I think it's important in endgames. For example, you don't want to trade down from an endgame you might win to one where you have material advantage but it is drawish.
I also have a penalty for trading a minor for pawns. This is larger in the opening/middlegame.
--Jon
Re: Imbalance
Posted: Sat Feb 01, 2020 6:43 pm
by xr_a_y
I am probably finally getting something out of it
Score of minic_1.37 vs minic_1.35: 181 - 135 - 477 [0.529]
Elo difference: 20.2 +/- 15.2, LOS: 99.5 %, DrawRatio: 60.2 %
Now trying to separate MG and EG values.
Re: Imbalance
Posted: Sat Feb 01, 2020 8:59 pm
by JohnWoe
I scale down towards EG when there's no queens left. And some more when there's also no rooks.
Like this:
https://github.com/SamuraiDangyo/Sapeli ... li.c#L2298
Code: Select all
// Endgameish
if (wtm) {
if ( ! BRD->black[4]) scale *= 0.9f; // No queens
if ( ! BRD->black[3]) scale *= 0.95f; // No rooks
} else {
if ( ! BRD->white[4]) scale *= 0.9f;
if ( ! BRD->white[3]) scale *= 0.95f;
}