The integrated AI Review feature for OGS

Announcements OGS Announcements
102

I have one more thing to note. Was looking at another top three moves for a game:

If you look at some variations that involve playing on the same point multiple times it seems to show the later move first, so for example at move 35, if I play (one of) the suggested move(s) of A, then the sequences involving the two stone edge squeeze tesuji involve sacrificing two stones and throwing in at the same points.

I just find it confusing that if I’m trying to play through the suggestions that it would show move number 11 where move number 3 should be, and move number 12 where move number 7 should be, and then once moves 11 and 12 are played it then shows the numbers 3 and 7 afterward.

I’m hoping this isn’t coming across as negative, I’m enjoying the feature so far :slight_smile:

103

Ah yes, that is a bit confusing. I’ll have to ponder on how to best display situation

2 Likes
104

I just ran my first analysis and then downloaded the SGF, but the AI info isn’t included. Is there a way to download the analysis so I can use it off-line on my PC with Sabaki or other SGF editor?

3 Likes
105

Not yet but that should definitely be a feature, I’ll add it to the todo list.

7 Likes
106

I have a suggestion for spreading this around. Would it be a good idea to shoot a banner up on OGS at some point in order to show the new perks for the AI implementation, thus encouraging more site supporters? You could just use the “Support OGS” page for the link. A great deal of users don’t even use the forums, so this should help spread the news, thus increasing support on OGS. However, I hear news of PayPal acting up, so it would be a good idea to resolve that first. :grinning:

1 Like
107

I just played my first OGS game in probably a year and was blown away by the top-3-moves thing at the end. Wow! This is a fantastic reviewing tool. Many thanks to the OGS folks for adding this.

6 Likes
108

Great to have you back Monocle! Glad you’re enjoying some of the new features :heart:

3 Likes
109

I think will be a fun and useful feature, but I can’t figure out what I’m looking at. What do the letters mean? What do the colors mean? Why is there a 40% / 60% win estimate with zero stones on the board?

With full analysis available, let’s say I’m interested in answering the question “What was the losing move in this game?” How would I do that?

Edit: This is the 60/40% I’m referring to. My opponent in this game destroyed me, so it’s not taking into account what actually happened. It can’t be basing it on rank, because I’m the lower ranked player. Is it trying to tell me that it thinks that white in general is a 60/40 favorite in go?

AI
AI.png939×528 63.1 KB

3 Likes
110

We may get documentation on this, but until then:

They are LeelaZero’s move choices in order of preference, starting with A.

The bolder the color, the more variations LZ considered starting from that move. Sometimes the most searched move is not “A” because it changed opinions late in the search.

Because LZ thinks the komi is too damn high.

Look for the last significant drop in win rate (>10%).

5 Likes
111

Oh, the graph. I suppose the spikes are particularly interesting moves / blunders. Thanks for the explanations.

The komi thing is interesting. I hadn’t considered that AI might eventually cause a shift in komi. I guess it really likes taking white.

1 Like
112

Does the AI take custom komi into account? Does it matter to the AI if the games are under Chinese rules or Japanese rules?

113

Afaik, the network is trained specifically on 7.5 komi Tromp-Taylor, and doesn’t take settings as input.

2 Likes
114

Hang on, is there something else going on here?

Leela doesn’t think that the open board is 60/40, she thinks it’s 53.5/46.5

22
222054×1536 1020 KB

3 Likes
115

My lizzie analysis gives precisely the same win rate at the start of the game as GreenAsJade’s, not the 40% / 60% that we are seeing on OGS. I have the CPU version, but perhaps with higher playouts this changes?? Probably not though. :thinking:

116

There’s something else weird going on.

This what I got when I analysed a game the first time, when the feature went live:

39
392562×1431 811 KB

Here’s one I did today:

24
242506×1508 863 KB

“Ahah” (I thought) maybe the estimate being shown is after the first move - that would make sense for this screen shot. Q16 +1.1 pp - adds up!

Unfortunately this doesn’t explain the earlier poster’s screenshot.

117

is the komi 6.5 in yours and greenasjades games?

119

The komi is different in both your screenshots. I imagine that makes a difference. I’d prefer and feel more comfortable with a komi of 6.5 than 0.5

120

Oh my goodness you are correct!

I didn’t notice that about the first one.

I’m amazed if it makes a difference to what Leela thinks, but I agree we should be comparing apples to apples. I’ll look around some more!

Note that the second one I posted and the original poster’s picture do seem to be both komi 6.5, but do show the difference anyhow…

1 Like
121

So part of the issue is probably the Monte-Carlo system:
Not every game gets the same playouts from move 0, and so the evaluation may be different.

Also: a lot of it also has to do with what weights the reviewing AI is using (not al LZs are the same)

So it makes sense that there can be quite a bit of variation with the move 0 inputs, and some supporters get heavier networks, and more playouts due to the fact that they are better paying, which further reflects upon the evaluation.

122

Are you using the default network.gz? The networks that came out after Lizzie 0.6 was released have much sharper value heads, because they are more confident in predicting game outcomes. On my machine, it is 40/60.