December AI Updates

It appears downloading the AI reviewed files might not be working.
The format of the files with AI info appears to be the similar to Lizzie, meaning you can run analysis using Lizzie locally and upload the sgf that has winrates and varations to OGS and view them. Its not quite the same a Lizzie as not all varations are visible. This is especially useful for reviewing games with others. The formatting would be nice to standardize around a program and clarify which one, perhaps Lizzie. Since Lizzie, q5go, and GoReviewPartner all use different formats for their sgf’s that contain varations and winrates.

Do you have a game link where it’s not working?

And yeah standardization would be good… although really SGF is not the right format for that sort of thing

This happened in two different games, a 9x9 and 19x19. I am absolutely confident that there were no real “moves” remaining (not even dame) and the AI did not show any variations either.
passpasspass

1 Like

The issue I mentioned before is fixed, thank you.

Now I have some doubts about top three moves.
In the last game I finished (against not-supporter) there’s a nice winrate chart along the whole game. Top 3 moves are highlighted on the chart by red dots. That’s an awesome amount of information. :slightly_smiling_face:

But in all my previous games (against not-supporter) the chart is very different and confusing: it’s flat, it ends before the end of the game and there are 4 red dots.

Could it be related with the last update?

1 Like

Yeah the old fast reviews didn’t have all of the winrate information. I should probably just delete them so they are regenerated…

Re the top 3, I’ll be looking more but basically what happens is we do a very fast pass to get a hunch about how the game flow went, and then we use that to pick the top three we are going to analyze, and then do full strength reviews on those top 3 moves. After doing the full strength review on those moves the winrate will almost always change a bit, sometimes by a lot. So it’s far from perfect… but it’s not too uncommon that it’ll work out pretty well

6 Likes

I think so, since some of them are very strange: top moves on pass (as pointed by taatelikakku), top moves on first and second move (I can provide link). They seem bugged while latest ones look perfect

When I try some of the suggested moves and it shows me what to do next, it can be hard to understand because the “phantom” stones are so heavy. I think they should be even lighter - or invisible until I click them, or something.

Oh, and when I try out those variations, if there’s a ko fight, it covers up the smaller numbers, so that I don’t know where to click first initially.

3 Likes

Awesome work anoek, OGS just keeps getting better and better. :slight_smile:
Bug report: AI SGF review here errors: https://online-go.com/game/21007528
C[error enscribing AI review information at this node, review move mismatch]

1 Like

@anoek a new KataGo version is out. The new net now supports Japanese-like rules.

https://github.com/lightvector/KataGo/releases/tag/v1.3

8 Likes

There is mismatch in score in handicap in chinese rules
Score estimate thinks I won
Bot review thinks I won
Bot itself thinks I lost
Game itself thinks I lost

https://online-go.com/docs/go-rules-comparison-matrix
there is rule that “White receives an additional point of compensation for each black Handicap stone.”

So Score estimate and Bot review should be fixed.

3 Likes

Thank you for the bug report.

  1. The score estimator is as the name suggests an estimator. It always counts area + handicap. Nothing else. So you should never rely on it. The new years poll suggested that anoek is planning to improve the SE, so maybe it will get better.
  2. The bot is right
  3. AI review: @anoek KataGo is probably missing the 9 points handicap compensation. I don’t know if there is an extra parameter for it or if we have to add the handicap compensation to komi for the AI review.
  4. The final score is correct (and no bugs are known, so it should be always right)
Counting by hand gives W+6.5

white 179 area score +0.5 komi +9 handicap compensation = 188.5 points
black 182 (361-179) area = 182 points

white wins by 6.5 points

2 Likes

It doesn’t look at handicap - that is the problem. It just needs (score - 9) fix

@anoek For uploaded sgfs with odd played stones (initial position played out of order) KataGo doesn’t know the initial board position, assuming an empty board. The variations for the last move are played on too of perfect alive stones

and another variation of the same game

2 Likes

Ah yes, I believe that will be fixed pretty soon here with the next update

5 Likes

Hi @anoek, is this something that is still in the works? For handicap games it is quite hard to see from the % on each point or graph what the trend is. A toggle to territory scalars would be very interesting!

3 Likes

Thanks for the bump, it slipped off my radar, lots of fires to put out recently :sweat_smile:

4 Likes

:joy: no worries, I know the feeling. Do you manage a public Jira list or equivalent for features that we can monitor?

2 Likes

Yep:

3 Likes

Ace thanks!

2 Likes

I’ve only just noticed it’s rolled out. So slick! I love it!

1 Like