lichess.org
Donate

Evaluating Sharpness using LC0's WDL

In the section titled Calculation of the Score, the formulas are not showing up for me. I'm not sure if it's a browser issue of mine or if something went wrong on your end.
Thanks for pointing this out, it's fixed now.
Some interesting implication for cheat detection perhaps? If a player can find accurate moves with high "sharpness", that seems like a stronger indicator of cheating than if they same happens with low sharpness.
I don't know much about cheat detection but it's certainly more complicated than just looking at one number. I would leave the cheat detection to the chess websites since they have all the necessary data.
@jk_182 said in #3:
> Thanks for pointing this out, it's fixed now.

The formula shows up now, but I think there is a typo. In your previous article*, you add the log terms, but in the formula for this article you are multiplying them. I think adding them makes more sense because in that case sharpness is infinite only when W/D/L = 0.5/0/0.5. If you multiply the terms, then you get infinite sharpness any time either W=0.5 or L=0.5. That would include the can't-lose situation 0.5/0.5/0 and the can't-win situation 0/0.5/0.5, which doesn't really make sense to me.

* www.chess-journal.com/evaluatingSharpness2.html
what would be interesting is a bot that always chooses the sharpest move (whilst trying to minimize the centerpawn loss) - I guess it would make a lot of sacks and create interesting games
You should take a look at the LC0 WDL contempt feature (lczero.org/blog/2023/07/the-lc0-v0.30.0-wdl-rescale/contempt-implementation/). With this, you can basically tell Leela that she's playing against a lower rated opponent and here play will be sharper.
There are a couple of bots on Lichess with this feature for odds play:
@LeelaQueenForKnight
@LeelaKnightOdds
Also take a look at GM Matthew Sadler's videos about this feature: www.youtube.com/watch?v=u9i71vdm_Ew&list=PLxhvNMc95Uo9N6W48TUAyIwkjK0w1wHKj&index=10