Update WLD model #4373
Update WLD model #4373
Conversation
update the WLD model with about 400M positions extracted from recent LTC games after the net updates. This ensures that the 50% win rate is again at 1.0 eval. No functional change.
|
The corresponding visualization
The graphics is generated as part of the fit (https://github.com/vondele/WLD_model) |
|
Hi @vondele , I read your pull #4216 (comment) that reviewed by @ddobbelaere But increasing |
Don't see how this is a counterargument to the PR being merged. Besides, since you're not giving the position's FEN, it's basically impossible to even understand what you're trying to demonstrate. Could you be a bit more precise on what you think is wrong in the WDL/normalization update ? |
|
So what we have |
|
yes, a slight deflation. |
centipawn = 64
centipawn = 604 This comparison is not just for this pull but for all pulls that vondele have been done so far. #4216 (comment)
I want from @mcostalba @romstad @zamar @ornicar to compare this. |
|
Sorry, but I have nothing to add, nor do I understand what you are talking about. |
Thanks for reply, you were reviewer of #4216 (comment) & approved this pull, but increasing
If you read my previous comments, you will understand why this increase is wrong. |
Wrong. 0.66cp means the position has around 17% win rate.
Comparing to SF from distant history doesn't make any valid reasons. What you claming is likely more close to evaluation issue, not UCI score nornalization. |
@MinetaS This test was in June 2022 that With new |
|
So you are saying the result of 394 / 348 is 4.125? |
non-linear relation between the internal value and the reported UCI score in centipawns The result of 348 / 237 ≠ 4.93
I think perhaps this is not evaluation issue. This is due to changes in UCI score normalization. Do you think there is no difference between It's very important to know that the position is equal or clearly winning... |
|
It is linear. Before normalization One of your SFs are misevaluating. You also haven't given the position or versions you are using. |
So, it can be concluded that half of this problem is due to But I guess changing
Yes that's right. centipawn score = 3.26 is before vondele changes that is correct (before June 2020) centipawn score = 0.66 is after vondele changes that is mistake (after July 2022) |
No that can’t happen.
You still haven’t given us the position or what version you used. Nor can you say that a score is wrong. Nor can you conclude that the problem is the normalization. Better Stockfish version might see a draw while a weaker one might not |
@Disservin @Craftyawesome @mhouppin
No, White is clearly winning & |
|
Hi @LovelyChess I just enabled private replies in comments in my settings, just reply to my message and only you and I can see it. |
|
@Disservin You can see the position in my private repo now. Just don't send it to anyone. |
This comment was marked as duplicate.
This comment was marked as duplicate.
|
@LovelyChess you have now pinged 7 different people on this closed PR, please stop doing that. |
|
@vondele Thanks for reply. To clarify the issue, I calculated up to a depth of 30
centipawn score = 604
centipawn score = 64 @vondele , first centipawn is 10x bigger than second centipawn?! Do you think this is not a serious problem? |
|
That is useless information, and can't be followed up. |
update the WLD model with about 400M positions extracted from recent LTC games after the net updates. This ensures that the 50% win rate is again at 1.0 eval.
No functional change.