Excellent. Thanks for running these tests. Convincing empirical evidence of the non-linear manner in which some chess engine features combine. Indeed the sum is much greater than the parts.lithander wrote: ↑Mon Mar 08, 2021 9:21 pm So just adding QSearch to a material-only engine was +50 ELO, just adding PSTs provided +200 ELO despite the horizon effect. Finally the combination of PSTs + QSearch was worth +500 ELO in self play. Almost double of what the sum of it's parts would suggest.
Quiescence Search doesn't improve strength
Moderators: hgm, Rebel, chrisw
-
- Posts: 434
- Joined: Thu Apr 26, 2012 1:51 am
- Location: Oak Park, IL, USA
- Full name: Erik Madsen
Re: Quiescence Search doesn't improve strength
My C# chess engine: https://www.madchess.net
-
- Posts: 12540
- Joined: Wed Mar 08, 2006 8:57 pm
- Location: Redmond, WA USA
Re: Quiescence Search doesn't improve strength
I am not surprised.
I guess that it boils down to branching factor.
A really good eval means a much better branching factor.
Even in pure alpha-beta, that can mean a huge speedup.
I guess that it boils down to branching factor.
A really good eval means a much better branching factor.
Even in pure alpha-beta, that can mean a huge speedup.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
-
- Posts: 2487
- Joined: Tue Aug 30, 2016 8:19 pm
- Full name: Rasmus Althoff
Re: Quiescence Search doesn't improve strength
Actually, we had found the opposite to be true in this discussion because a more accurate eval gives more different evals and thus fewer beta cut-offs on the first try. What is at work here however is that while NPS decreases and BF / time to depth increase, the quality of these calculations increases considerably.Dann Corbit wrote: ↑Tue Mar 09, 2021 5:26 amI guess that it boils down to branching factor.
A really good eval means a much better branching factor.
Rasmus Althoff
https://www.ct800.net
https://www.ct800.net
-
- Posts: 10282
- Joined: Thu Mar 09, 2006 12:37 am
- Location: Tel-Aviv Israel
Re: Quiescence Search doesn't improve strength
1)The effect get larget until it get smaller.mvanthoor wrote: ↑Mon Mar 08, 2021 11:35 pmCoollithander wrote: ↑Mon Mar 08, 2021 9:21 pm So just adding QSearch to a material-only engine was +50 ELO, just adding PSTs provided +200 ELO despite the horizon effect. Finally the combination of PSTs + QSearch was worth +500 ELO in self play. Almost double of what the sum of it's parts would suggest.
It seems, when the engine thinks a bit deeper, the effect gets larger: in Rustic, I tested this effect PST+QSearch = +600 Elo over an engine that is material only without QSearch. That suggests that, in very simple engines without other evaluation terms, this effect gets larger when the engine can reach larger depths.
If you search deep enough to solve the game without PST+Qsearch then the effect in a direct match is 0 elo.
2)Elo difference is dependent also on the order of moves and I guess that there is a significant difference in the strength of only material evaluation based only on the move generator.