Request: AI review chart with score difference beside win rate

I am a huge fan of the AI review that were recently introduced. So much so, that the AI review is one of the main reasons I support the site.

Having said that, I frequently have a strange feeling of missing out something important, when I look at the AI review charts. I play a lot of handicap games, and the chart frequently looks like this:
AI review - lost
Or, like this:
AI review - won
The charts suggest that both games were boring. In the first, white had no chance, and in the second, black must have made a huge blunder. But the reality is, that both games were close and exciting. In handicap games the win rate swings from 0 to 100% very fast - hence the flat charts.

I was thinking on an alternative representation of chances. Instead of displaying win rates, OGS could display the score difference instead of win rate. kata-bot comments this information in Malkovich:
Winrate 0.01% ScoreLead -72.3 ScoreStdev 24.1

Using score lead and score deviation information an exciting new type of chart could be show. Something like this:

(The black line shows the expected score, the gray area shows the uncertainty of the score estimation.)

On a related note: it would be great to show the score value of individual moves in the AI review, instead of change in win rate.


Yup, would be wonderfull.

As per discussion in December AI Updates this is something that Anoek is planning to implement, just might take a while still.


Agreed - this recently got bumped up again to be on @anoek’s radar. Nice sample graph! Win rate is great for competitive matches between AI and machine, but territory + stdev is really evocative for human games and handicap!


As a kyu player I find it difficult to learn from AI. To me the most useful method is to open a game in sabaki/lizzie and start interacting with the AI: make variations and see how my bad ideas get instantly refuted. If it fixates on one thing then just add stones so it can think about another question instead.
Players may consider an important group dead but keep playing anyway. AI sees how to make it live and therefore both winrate and variations get stuck for rest of the game. Territory estimate is an ok workaround as that number keeps changing regardless of who has already won.
But interaction is still the best.
Realtime analysis could be done for “free” if katago was running in the web browser. Nowadays we even have webGL that could enable neural network evaluation on GPU, overcoming any performance issues. But still, a brand new javascript katago implementation is way too much work to implement.


This topic was automatically closed 91 days after the last reply. New replies are no longer allowed.