Hello forum,

I hope everybody is going well.
I am coming back to go after a very long hyatus and I saw that AI has been a big disruption in recent years.
I was KGS 8 kyu, and I struggled to improve mainly due to the fact I didn't have the opportunity to review my game often.
While not as beneficial as a real teacher for a beginner like me (you still need to interpret an AI proposal) I believe it would still be a useful tool while I am re-starting Go.

Sorry for the personal introduction, now comes my question :
What AI do you recommend to analyze my games and how does it work (I guess we just need to upload the game file into the soft)?

Thank you very much

Hi BP, welcome.

Unlike AlphaGo (Zero), current consumer AI makes mistakes with ladders, sekis and LD. Therefore, never simply trust AI but verify its suggested moves or analysis by your own tactical reading.

Ignore naive influence maps because they represent proximity very much more than influence. (I have not evaluated KataGo's points judgements but would be surprised if they could meet human territorial positional judgement. Also note that positional judgement has many more aspects than territory.)

Do not use AI to study endgame because AI can (often) play (very) suboptimal endgame if only the win is maintained.

Probabilties mean almost nothing as absolute values. Relative changes of probabilities might give hints though. Huge changes can indicate mistakes.

A list of top (probability) move candidates need not include all good moves. Therefore, do not exclude moves just because they are not suggested by AI.

The top-most probabilities of the first move are more meaningful than those of alternative first move candidates or later moves. Therefore, trust alternatives and later moves much less.

AI hardly gives strategy or reasons. You need to invent those yourself. AI top move may be chosen for the sake of complex implied strategy but such might not always be the best for (especially weak) human players. Instead of AI suggestions, find and respect "keep things simple" alternatives. Simple moves might lose a fraction of a point (or, for kyu moves, a few points) but who cares? Simple is good for learning but AI suggestions might oppose this.

AI can point out (the biggest) mistakes of opening and middle game.

AI can also point out mistakes (typically oversights) of tactical reading but you often need to notice by yourself that it is about tactical reading and verify by doing it.

AI is a source for 3-3 under 4-4 josekis, which - unlike outdated human josekis - do not unnecessarily strengthen outward thickness too much. However, AI evolves such josekis. Therefore, especially learn the simple ones and do not expect the complex variations to persist.

Use AI for play.

However you use AI, do not use it exclusively. There is very much human teachers or books teach that AI does not (or not explicitly) teach.

E.g., AI used for analysis in between games cannot replace one's own analysis during one's own games.

Look for significant blunders. A significant drop in win rate is a good indicator of where you messed up. Another place is where AI consistently recommends a move as blue spot, but both you and your opponent continue to ignore the move. This can give you a sense of where the hotspot is during the game.

AI is also very good at openings, and finding big moves for points or for unsettling opponents. Look out for those.

Other subtle recommended moves with a few point different in percentage is really meaningless.

You can download Lizzie: https://github.com/featurecat/lizzie/releases

Once you open the program, you can choose between the engines "LeelaZero" and "Katago". I suggest Katago, which has been trained for ladders. For a kyu player, it will play good moves at every stage of the game, even with low playouts (e.g. 100 playouts). When you review a game, explore variations, check alternative responses.

OC, everybody is different, but at your level I would suggest using AI in three different ways:

1)To find blunders. You can use the game analysis option for that, or just walk through the game looking for moves in the game that lose a large winrate percentage to the AI's top choice. BTW, such moves may have very few playouts, which means that their evaluations are not reliable. Always check by making the move yourself and seeing how the AI evaluates its top reply. That will give you a better comparison with its top choice. Sometimes the difference will be dramatic.

2)To find good plays. At your level you have started to develop certain habits of play. Sometimes these are bad habits, because you are playing against weak opponents who do not know how to reply to those plays, or make them themselves. Even if these habitual plays are not blunders, you can learn better plays from the AI.

3)To explore the game. You can do this for fun by making the AI's choices or by making other plays and seeing how the AI responds. Also, some AIs will report long variations. These can be useful to see what it has been thinking about, but do not trust all the plays, especially towards the end, as they are based upon fewer and fewer playouts. Instead, you can walk through the variations yourself, either making an AI choice or a different play that looks interesting.

It has rightly been pointed out that today's AI bots can make mistakes, even blunders, with ladders, semeai, and life and death. You can use the AI to check itself. For instance, you may think that a group needs to protect itself, but the AI plays somewhere else. There are any number of reasons it might be doing that, but you can try and see if you can kill the group. Often you can't. Maybe you can make a ko. If you can kill it, maybe the AI wants to take a sente elsewhere and then come back to save the group. Or maybe it is willing to sacrifice the group. On maybe it has made a mistake. Whatever the case, you can learn something by trying out moves.

As for which bot to recommend, you can learn from any of them. I would suggest KataGo. It is much less likely to make a mistake involving ladders, and it makes score estimates, which can be easier for us humans to understand than winrates.

Good luck!

Edit: The Elf team in its commentaries on the GoGoD database, did not trust any play with fewer than 1500 playouts. I wouldn't either. If you make the play yourself you should get more playouts than that.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Good advice but ...

I for one would, if I would not have a nice beefy Nvidia card in my digital Wú Qīngyuán. And I did it in fact when I only had a lowly substitute aka laptop at hand for reviewing. While there will be much more errors at lower playouts, I would still put more trust on the AI&me than purely on myself. There is a lot to be learned at less than 1500 playouts too.

I agree with Gomoto. With my slow laptop, I need 30 seconds to get 1500 playouts. For 300 moves, this represents theoretically 2.5 hours, but thanks to tree reuse it amounts to maybe 1.5 hours, which is more time than I want to spend on a game review. My guess is that Katago with 100 playouts is at least mid-dan level, which is strong enough for a kyu player like me. Most of the time, the AI will point out several elementary mistakes : shape problems, invading too deeply, not thinking about playing at the border of two moyos, not seeing a big point, failing to protect a weak group, not seeing the vital point of the opponent's shape, forgetting to play a peep... Occasionally I give Katago more playouts if I really want to remember a sequence like a joseki, but most of the time it's not necessary.

Everybody's different.

I remember being with some players, one of whom asked Takemiya how long he took to review games of other pros. He replied, 15 minutes. We were all astonished that he took such little time, and he said that he was only looking for ideas. That's one of the ways I suggested using AI.

Suppose that you take 30 minutes to review a game. At 30 seconds per move you can generate 60 plays. To my mind, that's way too many plays to focus on in review. A handful is more like it. (Edit: Maybe only one.) Yeah, waiting 30 seconds or more for the program to make a move can be boring, but you can think about the position while you wait. When I used to kibitz pro games in real time I might have to wait 5 or 10 minutes for the next play. IMO, it was time well spent.

And what about online services? I just tried out Deep Leela at https://www.deepleela.com . It uses Leela11, which is far from a top bot these days. In the opening it usually seems to take around 5-6 seconds to produce 2-3 plays with more than 1500 playouts, among others with fewer playouts. Often the top choice has more than 10k playouts. There are services that use top bots, I understand.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

1 playout on Leela11 is not equivalent to 1 playout on Leelazero. The latter takes about 100 times longer. I manually ran a match between Leela11 with roughly 20000 playouts at each move, vs. Katago with 100 playouts. Katago won a semeai early in the game, but Leela11 didn't count liberties correctly and thought she was leading.

For most moves in my own games, I prefer analyzing at low playouts. What does it matter that AI can read very deeply to find another move is slightly better than mine? I won't be able to find that move in the future, until my reading ability suddenly improves tenfold. So as long as my move is one of the AI's first moves it looks at, and not that bad of a winrate loss, I can trust that my move made some sense. As long as I had an idea behind my move, and it's apparently sensible, it's good enough for me. I certainly still make tons of moves AI doesn't look much at and lose a lot of winrate - those are the ones that I let AI play a little bit with, while I analyze.

My point was about online speed, not accuracy. Are you saying that if it usually takes Leela11 5-6 seconds to return 2-3 moves with more than 1500 playouts, it would take LeelaZero 500-600 seconds (8 min. 20 sec. - 10 min.) to do the same?

Edit: The Elf team did not trus plays with fewer than 1500 playouts by last year's Elf, which is still among the top bots, I guess.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

My slow laptop (CPU only) can make 1500 visits in 30 seconds with Leelazero, and more than 100000 (didn't count exactly) with Leela 0.11.

(If a program reads a variation ... , this counts as 10 visits if I understand correctly).

What I was calling playouts is the number reported with the winrate of by Elf. You get a smaller number with because it is only one of the replies to , and a smaller number with because it is only one of the replies to . I assume that the number you get by hovering the mouse over a prospective play by Deep Leela means the same thing, but Deep Leela calls it visits. In any event in its commentary ELf does not report any of its prospective plays with fewer than 1500 visits or playouts, even though I am sure that it builds a tree with lower numbers in its nodes.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

For instance, here is the 4th Oza title match between Shibano and Iyama.



At move 91, after 30 seconds, Katago has visited K6 1.3k times, P7 979 times and O5 214 times. The total number of visits is 2554.



After 30 seconds, Leela 0.11 has visited O5 126924 times, N7 1618 times, etc. The total number of visits is about 129000.



So for that move, a visit on Katago takes 50 more times than a visit on Leela 0.11. This number may differ for other moves.

Thank you very much to all the people who participated in this topic.
I already learned a lot about how to leverage AI while reviewing my games.
I really appreciate that you took your time to develop your point of view.

I hope this topic can be useful for anyone who wants to start using AI to help them strengthen their game.
For this purpose, in addition to all the comments about how to review a game, the AI you recommend would be :

(please feel free to propose if you have a better interface for these engines).
- LeelaZero (with Lizzie interface)
https://github.com/featurecat/lizzie/releases
- Leela11 (is it an upgraded version fo Leela zero ?)
https://sjeng.org/leela.html
- KataGo (with Lizzie interface)
https://github.com/featurecat/lizzie/releases

I found this website which uses KataGo :
https://ai-sensei.com/
But you have to pay for every review.
I would prefer to buy a license one time and have unlimited access to it.
Or does it exist an open version? (I guess on Ai sensei what we are paying is not really the AI itself but the sever that are used to run it)

Thank you all for your help

Leela 11 is an AI which completely different from Leelazero, despite the name. Leela 11 is less strong but some people still prefer it because it plays more human moves.

Instead of ai-sensei, you can try www.zbaduk.com

What about zbaduk's reviews? Do they provide a review for you or can you play around with positions, making plays to compare results?

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

In some cases though, bots can make very weird statements at low playouts and they need a couple of thousand to get beyond some bizarre line of thought.

But otherwise I agree and:
- use the bot's intuition (aka "policy") to improve mine
- look at the underlying sequence to check my reading

The thread dedicated to zbaduk is here: https://lifein19x19.com/viewtopic.php?f=9&t=16563

You can try it for free, but need to register. In order to use the full capacity of the processor, you need to pay membership fees (5€/week, 15€/3 months, 35€/year or 55€/2 years).

It lets you play out variations, just like Lizzie.