Pawn value

[MJZ1977]

Small tune of PawnValue parameters -4 / -7 with "closedpos.epd" opening book.

STC:
LLR: 2.96 (-2.94,2.94) {-0.50,1.50}
Total: 58776 W: 11787 L: 11511 D: 35478
Ptnml(0-2): 975, 6876, 13443, 7086, 1008
https://tests.stockfishchess.org/tests/view/5eb5aa712326444a3b6d3e33

LTC:
LLR: 2.98 (-2.94,2.94) {0.25,1.75}
Total: 137544 W: 19687 L: 19115 D: 98742
Ptnml(0-2): 988, 13219, 39901, 13561, 1103
https://tests.stockfishchess.org/tests/view/5eb67a392326444a3b6d3e9a

Non regression STC with "noob_3moves.epd" opening book
LLR: 2.94 (-2.94,2.94) {-1.50,0.50}
Total: 98168 W: 18545 L: 18499 D: 61124
Ptnml(0-2): 1647, 11396, 22951, 11444, 1646
https://tests.stockfishchess.org/tests/view/5eb7e489e0300e8e8c896203

Bench 4696646

[vondele]

@MJZ1977 as this is (AFAIK) the first test that passes with the closedpos book. I think we should do a non-regression test on our standard noob_3moves book, as you suggested in #2646 (comment)
I would expect it to pass.

(it is indeed pretty similar to https://tests.stockfishchess.org/tests/view/5eafbcbfa460c0e4b8b9ba2e which changed (-8, -8) and was slightly Elo positive on the standard book).

[Vizvezdenec]

I don't really like this tbh.
I'm now trying to do some test on this idea (pawns worth less in closed positions is a general motto of it) - I think that we can introduce book-arbitrary and more broad heuristic that would be useful in any position.
https://tests.stockfishchess.org/tests/view/5eb7c06de0300e8e8c8961e9
https://tests.stockfishchess.org/tests/view/5eb7245f2326444a3b6d3f12
At least I hope that we wouldn't have this committed before my tests finish :)

[MJZ1977]

This is my original idea:

• we test with closed book and find new optimum,
• test new optimum with standard book,
• if we see that the optimum is not the same, we try something more complex.

We are now at step 2.

[vondele]

@Vizvezdenec I agree we should see if there is something more generic, and I already planned to wait for your tests to finish. I also see that @MJZ1977 tries something more generic. At least this test shows there is potential, and that's always good.

[MJZ1977]

Non regression with "standard book" seems not going to pass. For me it is one of first prooves that show the importance of opening book on key parameters and the need of use of different opening book to improve the understanding of these parameters.
https://tests.stockfishchess.org/tests/view/5eb7e489e0300e8e8c896203

[ghost]

I think a better way is to tune connected pawns (which are abundant in closed positions) formula in pawns.cpp , not changing the base value.

[adentong]

The non regression actually passed. What happens now?

[MJZ1977]

I updated the PR with non regression test. If Vizvez or another developer finds a better solution, we can compare it to this commit.

[vondele]

@Vizvezdenec I don't see any more running tests on this topic, and since the test passes non-regression on the standard book, and improves on the closed book I plan to go ahead and merge later tonight.

[silversolver1]

So just hypothetically, if (or after) this does get merged, what if someone was able to then get a +4/+7 PawnValue to pass both STC and LTC with the normal opening book? would the resulting patch be merged as well (resulting in overall no change)?

it might be worth doing at least two more tests then: this patch as the base branch vs current master as the test branch and seeing whether that would pass STC and LTC with standard bounds and the normal opening book

edit: although now that i think about it, this could really be said for almost any tests that pass using a different opening book

[Vizvezdenec]

Yeah I guess there is nothing more to be done there... at least for now :)

[vondele]

@silversolver1 note that applying this patch passed non-regression against the the standard book. That means the revert of this patch won't pass as an Elo gainer.

[xoto10]

I was trying to see if we could [make] the pawn values variable, and set them just once, before the search, but a local test suggests it's very broken somehow: master...xoto10:pawnvalvar1
I don't think this will be any use for this PR, but if anyone can point out my errors it might improve my c++ skills :)

[vondele]

Thanks!