Change centipawn fallback to account for sharper WDL with high WDLCalibrationElo #2075
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
With WDL sharpening at 3600 Elo (most commonly used value e.g. in TCEC both for playing and for kibitzing), the old centipawn calibration is off by about a factor 2 compared to Stockfish and generally takes over too quickly around +2.00 while it should only take over around +4.00 since up to there, `WDL_mu` behaves well enough. With lower calibration Elo (e.g. for analysis of human games / openings), the takeover point is significantly later due to lower Q from broader WDL, so this change doesn't affect anything. Doesn't yet fix the jumpy eval behavior in draws with very low W or L but substantial L resp. W remaining.
initial oversight: in a +1 position we want to display +128, that shouldn't change
borg323
approved these changes
Oct 12, 2024
borg323
pushed a commit
that referenced
this pull request
Oct 20, 2024
…ibrationElo (#2075) * half eval fallback formula With WDL sharpening at 3600 Elo (most commonly used value e.g. in TCEC both for playing and for kibitzing), the old centipawn calibration is off by about a factor 2 compared to Stockfish and generally takes over too quickly around +2.00 while it should only take over around +4.00 since up to there, `WDL_mu` behaves well enough. With lower calibration Elo (e.g. for analysis of human games / openings), the takeover point is significantly later due to lower Q from broader WDL, so this change doesn't affect anything. Doesn't yet fix the jumpy eval behavior in draws with very low W or L but substantial L resp. W remaining. * changed factor to +128 convention initial oversight: in a +1 position we want to display +128, that shouldn't change
uwuplant
pushed a commit
to uwuplant/lc0
that referenced
this pull request
Nov 29, 2024
…ibrationElo (LeelaChessZero#2075) * half eval fallback formula With WDL sharpening at 3600 Elo (most commonly used value e.g. in TCEC both for playing and for kibitzing), the old centipawn calibration is off by about a factor 2 compared to Stockfish and generally takes over too quickly around +2.00 while it should only take over around +4.00 since up to there, `WDL_mu` behaves well enough. With lower calibration Elo (e.g. for analysis of human games / openings), the takeover point is significantly later due to lower Q from broader WDL, so this change doesn't affect anything. Doesn't yet fix the jumpy eval behavior in draws with very low W or L but substantial L resp. W remaining. * changed factor to +128 convention initial oversight: in a +1 position we want to display +128, that shouldn't change
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The previously used
centipawnformula #1193 is also used as fallback forWDL_muintroduced in #1791 mostly to reproduce the extreme centipawn values in clearly decisive positions users are accustomed to. The formula however was calibrated with the raw (unsharpened) WDL NN output, which means that together with WDL sharpening due to highWDLCalibrationEloit climbs too quickly, which also means it takes over faster than intended.WDL_muwithout the fallback usually produces meaningful evals up to around+4.5, which is roughly where the fallback formula should take over. However, in the described scenario withWDLCalibrationElo: 3600this already happens below +2, causing a discrepancy in behavior between Lc0 and SF eval roughly like this between Lc0 and SF in the range between +2 and +5 for SF where Leela regularly shows about double the eval of SF.The fallback formula is therefore updated in two ways: 50% reduction in the scaling, while slightly increasing the constant to still meet +128 at
wl=1.0.