rooklift/nibbler

Eval graph with WDL

Opened this issue · 2 comments

carlosfmora commented

Usually, the evaluation graph only has one attribute, centipawns or percentage. What if there was the option to graph through LC0's WDL? That provides much more information for the analyst.

For example, take a look at the WDL graph of the 2024 Candidates in round 13, the game between Gukesh and Firouzja
image
image

Or the most incredible game of the Candidates, game between Nepo and Caruana in round 14, A do-or-die game.
image
image

Or a crazy position, How to understand the traditional evaluation of nearly 50% in that position, while with WDL we can see that a draw is highly unlikely
image
image
WDL in final position:
image
image

My recommendation: two graphs

  • one for WDL across an entire game,
  • and another for a specific position (circular or bar chart)

image

mooskagh commented

I think WDL is best viewed as a stacked graph. E.g. like here https://lczero.org/dev/stats/2022/ but horizontal rather than vertical.

It's a question though how to show two engines on the same axes though.

comodoro commented

Something like this, only nicer

game