Texel tuning Zurichess quiet like generator

[xr_a_y]

Keeping only the ones with two players > 2600 elo

I don't think there is advantage in excluding games from weaker players. You want a variety of positions. The objective is to have a decent sample of possible positions that an actual search may evaluate and these will be quite varied and many of them uneven (for example, positions the q-search will cut off due to high eval).

--Jon

Yes but I also need the outcome of the game to be "reliable" and I suspect top engines are more prone to blunder.

[fabianVDW]

I didn't have success with the dataset, sadly. STC tests went -100 Elo. My tuned values are from ZuriChess dataset which is also definitly very high class (considering every position was played out with sf)

[xr_a_y]

I didn't have success with the dataset, sadly. STC tests went -100 Elo. My tuned values are from ZuriChess dataset which is also definitly very high class (considering every position was played out with sf)

Yes what I want to do next is playing 3 fast games per position with sf to verify the outcome. Thanks for testing.

[fabianVDW]

NP, and if your super dataset with 3sf games per position is finished I am also willing to test again

BTW, on your current dataset I have very high Error with FabChess in Tuning. Error is about 0.129 IIRC which would be a residual of 0.36. Very high... On zuri dataset I have Error of 0.058 or something

[xr_a_y]

Have you fit your sigmoid constant to the new data set?

[fabianVDW]

Yeah, but k turned out to be around 0.23 which sounded way too low, that's why I set it back up to my standard (0.96). Anyway, I can also try with optimal k again, it was just so low it didn't sound right to me.

[jdart]

Don't know if it is useful but I uploaded a couple labelled data sets a few years back:

https://www.arasanchess.org/big3.epd.bz2 (about 5 million positions)

https://www.arasanchess.org/lichess-new-labeled.epd.bz2 (2.5 million positions, I think distinct from big3 - but not sure)

These are both quiet and non-quiet positions. They use the c2 EPD field to indicate the result (1.0 for White win, 0.5 for draw, 0 for Black win). There is also a c1 field which Arasan uses but you can ignore.

--Jon

[fabianVDW]

Okay, I tried again with fitted K constant.

CCRL datastet ) k: 0.2375, Error: 0.12605 Residual: 0.355, STC Tests: Cancelled Early: +6 =3 -70, -391.7 Elo

Arasan Big 3) k: 0.5326, Tuned Tempo Bonus was at 300 CP, because Non-Quiet positions => Converted to quiet positions

This conversion was actually troublesome because there were about one thousand illegal positions (where the king could be captured) which had to be sorted out first. (How would you even label these??)

Arasan Big 3 Quiet) k: 0.8796, Error: 0.081 Residual: 0.2846, STC Tests: +301 =356 -341 -13.93 +/- 21.56 Elo

Lichess Quiet( no illegal positions there) ) k: 1.0887, Error: 0.06516 Residual: 0.25526, STC Tests: +244 =261 -203 +20.14 +/- 25.61 So that seems to be a success

I will start LTC regression tests for the tuned params of Lichess Quiet against Version 1.9 and Version 1.9.2 (private version)

Fabi

[zenpawn]

Lichess Quiet( no illegal positions there) ) k: 1.0887, Error: 0.06516 Residual: 0.25526, STC Tests: +244 =261 -203 +20.14 +/- 25.61 So that seems to be a success

I will start LTC regression tests for the tuned params of Lichess Quiet against Version 1.9 and Version 1.9.2 (private version)

Is "Lichess Quiet" the second link from Jon above (perhaps with non-quiet positions removed?) or something else?

[xr_a_y]

Ok I found another bug ... sorry ...

Can I send you the new edp by mail ?