Extensions in the days of LMR?

[Mincho Georgiev]

I do 1 PLY for see negative captures as well.

P.S. to give that some context, R for see() < 0 = R-1, but I use constant R = 2. I don't think this would work beyond Rcaps = 1 with a depth-dependent reduction.

[bob]

I do 1 PLY for see negative captures as well.

P.S. to give that some context, R for see() < 0 = R-1, but I use constant R = 2. I don't think this would work beyond Rcaps = 1 with a depth-dependent reduction.

My LMR currently varies between 1 and 7 depending on remaining depth. We have been experimenting with reducing PV nodes slightly less, but nothing definitive so far beyond reducing them one ply less than other nodes.

[Mincho Georgiev]

I do 1 PLY for see negative captures as well.

P.S. to give that some context, R for see() < 0 = R-1, but I use constant R = 2. I don't think this would work beyond Rcaps = 1 with a depth-dependent reduction.

My LMR currently varies between 1 and 7 depending on remaining depth. We have been experimenting with reducing PV nodes slightly less, but nothing definitive so far beyond reducing them one ply less than other nodes.

Thanks, this just confirms my futile attempts to distinguish pv/non pv lmr. I once had the dream to get rid of the LMR at the root and pv nodes altogether, but how can I do that when the end results is way worse than LMR everywhere with the same R. Maybe I'm doing it wrong, I don't know.
Since we're on the subject, what's the overall opinion on dropping into qsearch. I've done tones of inconclusive testing on whether I should save 1 pre-leafs PLY (NO LMR at DEPTH=1 vs R=1 on DEPTH=1)

[jdart]

I recently tried this and not going into the q-search was very slightly better.

But I have fractional extensions/reductions. So the remaining depth can be a fraction of a ply.

--Jon

[cdani]

Since we're on the subject, what's the overall opinion on dropping into qsearch.

I'm currently simply pruning the move when lmr hits qsearch depths.

[Mincho Georgiev]

I recently tried this and not going into the q-search was very slightly better.

But I have fractional extensions/reductions. So the remaining depth can be a fraction of a ply.

--Jon

Thanks. Just to compare better, do you use tt at qnodes?

[bob]

I do 1 PLY for see negative captures as well.

P.S. to give that some context, R for see() < 0 = R-1, but I use constant R = 2. I don't think this would work beyond Rcaps = 1 with a depth-dependent reduction.

My LMR currently varies between 1 and 7 depending on remaining depth. We have been experimenting with reducing PV nodes slightly less, but nothing definitive so far beyond reducing them one ply less than other nodes.

Thanks, this just confirms my futile attempts to distinguish pv/non pv lmr. I once had the dream to get rid of the LMR at the root and pv nodes altogether, but how can I do that when the end results is way worse than LMR everywhere with the same R. Maybe I'm doing it wrong, I don't know.
Since we're on the subject, what's the overall opinion on dropping into qsearch. I've done tones of inconclusive testing on whether I should save 1 pre-leafs PLY (NO LMR at DEPTH=1 vs R=1 on DEPTH=1)

Dropping directly into q-search was worse for me, but I have not tested this in a year or so now to see if that remains true...

[bob]

Since we're on the subject, what's the overall opinion on dropping into qsearch.

I'm currently simply pruning the move when lmr hits qsearch depths.

That's a good point since LMP can trigger (in Crafty) well away from the horizon it is likely that many such moves never see the light of day to hit LMR and have a chance of dropping into q-search.

[jdart]

do you use tt at qnodes

.

Yes. See https://github.com/jdart1/arasan-chess/ ... search.cpp.

[lucasart]

Think of it this way: non reduction is a relative extension. You have a whole range between: +1 (extend), 0 (normal), -1 (reduce 1 ply), -2 (reduce 2 plies), etc.

For example the recapture extension is old and busted. Plus it's bogus, because it's path dependent combined with a hash table…

Instead, what people do nowadays is more like this:
* quiet moves can get aggressively reduced
* bad captures reduced by 1 ply
* good and equal captures not reduced (nor extended)

So the recaptures are a particular case of good or equal capture, and don't get reduced, so it's a relative extension compared to bad captures or quiet moves.