In the context of this thread and its amateur students of the game, it is irrelevant whether this paper is right, wrong or just flawed.
Centuries of experience have taught us that the best way for humans to be good at something is to spend a long time at it so as to build up a huge database in the brain - one's personal reference library, or one's intuition - whatever you want to call it.
Experience has also taught us that, while time spent may be the most important tactor, there are other factors at play. We know, for example, that it is good to start young. But, as any parents will know, starting does not imply finishing. Think of all those trombones and violins languishing in the loft. Or adults' rowing machines. Psychological and social factors come into play, too. But if a child can stick at a subject, it is likely that starting early has some influence on the way they build up their intuition. With still malleable brains, it seems that whatever they learn can be stored in the most efficient way for that subject. But the way something is learned can have a major effect on how something later is learned - even something similar. The effect of that can be seen best with language learning. Everyone learns their own native language to a very high level. No-one, essentially, can learn a foreign language to that level as an adult.
No doubt natural talents are also prominent factors. For example, in go, it may be that people like Sumire can memorise evaluated chunks of (say) 10x10 points, whereas most of struggle with 4x4 chunks. In the same way that most people can hold 6 or7 items in their frontal short-term memory but a lucky few can do 8 or 9. Possible the size of chunks memorised in go depends on the age you started.
The precise mechanisms or numbers are not relevant here. The important point is that amateurs have clearly not spent enough hours yet to become strong at the game. What they need to become stronger is simply more time at the board. For some, the bling of AI has provided that incentive. They may end up stronger, but it is the effect of the extra hours spent that is giving the improvement, not any fanciful notion of understanding how AI "thinks". If free lollipops for every extra hour spent at the go board were enough incentive for someone, their strength improvement would happen just as inexorably.
Pros are in a different situation. Their internal go databases (intuition) are already huge and probably close to saturation point. But even then we know from people like Shibano Toramaru that they still spend a lot of time looking very quickly at games to note what is new and so potentially worth storing in the brain. That method has worked for them in the past, so why not try some more?
There are, of course, pros who also try to think about theory. But, there, there may not be a single right answer. We saw this with Shin Fuseki. Go Seigen thought the way to achieve improvement was to achieve more speed in the opening. Kitani thought the key was the integration of joseki and fuseki. Both ideas together sound remarkably like whatever it is that characterises AI go - though probably there are other concepts.
One concept that appears to be relatively new in modern human go is the willingness of Korean (especially) and Chinese players to gamble with the odds. Wang Xi wrote a seminal paper on this, pointing out that the Japanese insistence on playing souba go may be what is holding them back internationally. Players of the Japanese school still prefer to choose a safe line that seems to offers a small lead that translates into a 51% chance of victory, whereas Koreans and Chinese have found more success with choosing risky lines that may offer a 75% chance of victory. The much faster time limits of Korea and and Chinese go favoured that style of play, which is now ensconced, and it seems to have given them an advantage in the AI era.
Top pros seem to be able to play games with around 90% of their moves approved by AI. The very best, Sin Chin-seo, are scoring about 96%. The difference between the second-tier pros, Sin and AI seems (according to what is being said in the Far East) to be down to deeper reading. But this does not mean tsumego-type reading. Rather, it seems to be the ability to make EVALUATIONS of positions at deeper and wider levels - encompassing bigger areas of the board. 19x19 chunks instead of 10x10 or whatever. Covering a wider area of the board like this still mainly demands building up one's intuition as the prime way forward (not theoretical understandings) but because of the size of the chunks they are now tackling (i.e. whole boards) concepts such as Go's 'speed' and Kitani's 'integration' are being seen in a new light. Because of the need for speed, which in a sense comes about because you feel the need to be in two or more places on the board at once, more stress is now being put on the initiative (not sente in the usual western sense), and the value of miai (as per Shuei, of course) is likewise now enhanced. My sense of the new approach to miai is that, while you can't be in two places at once, you can at least guarantee to be in one, as per the traditional view. But nowadays you can try to leave aji in the other, and with a bit of luck you can go back there later and so achieve an effect close to being in two places at once. We can be sure the top players are not giving all their secrets way, but these seem to be the things they are talking about from what I read.
But, interesting as all that is for go fans, amateurs who want to improve will probably get much more mileage out of studying go the traditional way and worrying about the bling only once they have mastered that.
[Incidentally, it seems that we can see the emergence of Japanese economic strength a few decades ago as an application of the souba theory (and this is where the term comes from). Traditional evaluation methods based on that, such as candlestick theory, worked well in that environment. But with globalisation (bigger boards!), the risk-loving methods of the West that offer much bigger rewards have come to the fore, and now Japan is going bnackwards, in economics as well as in go.]
how has NN go engines changed way the top people play
-
John Fairbairn
- Oza
- Posts: 3724
- Joined: Wed Apr 21, 2010 3:09 am
- Has thanked: 20 times
- Been thanked: 4672 times
-
xela
- Lives in gote
- Posts: 652
- Joined: Sun Feb 09, 2014 4:46 am
- Rank: Australian 3 dan
- GD Posts: 200
- Location: Adelaide, South Australia
- Has thanked: 219 times
- Been thanked: 281 times
Re: how has NN go engines changed way the top people play
It sounds like you've looked into it more deeply than I have. I'd be interested to hear more about your conclusions some time.RobertJasiek wrote:I have also paid close attention to the timing of forcing moves: it differs greatly. In some environments, forcing occurs early. In other environments, it occurs late. In yet other environments, there is a long period during which it can occur, but there are exceptional moments when temporarily something elsewhere is more urgent (such as preventing a big cut in a joseki) when the forcing option is interrupted.
-
RobertJasiek
- Judan
- Posts: 6272
- Joined: Tue Apr 27, 2010 8:54 pm
- GD Posts: 0
- Been thanked: 797 times
- Contact:
Re: how has NN go engines changed way the top people play
If simply this were true, I'd be 100 dan.John Fairbairn wrote:it is the effect of the extra hours spent that is giving the improvement
Evidence? What means "approved"?Top pros seem to be able to play games with around 90% of their moves approved by AI. The very best, Sin Chin-seo, are scoring about 96%.
You try to construct a competition between non-AI and AI learning methods but all methods are useful.amateurs who want to improve will probably get much more mileage out of studying go the traditional way and worrying about the bling only once they have mastered that.
The more relevant question is about go theory (other than theorems): only traditional versus also AI. Traditional theory has a limited scope. Also using AI-born theory expands the scope and provides solutions where the former was wrong.
-
xela
- Lives in gote
- Posts: 652
- Joined: Sun Feb 09, 2014 4:46 am
- Rank: Australian 3 dan
- GD Posts: 200
- Location: Adelaide, South Australia
- Has thanked: 219 times
- Been thanked: 281 times
Re: how has NN go engines changed way the top people play
Yes, I was underwhelmed by that paper. There are some more fundamental problems. Correlation does not imply causation! They haven't commented on the most important confounding variable: the passage of time. Yes, their data shows that people play "better" go in 2020 than in 2014, for some measure of "better". But they can't separate "had access to Leela Zero on their own computer" versus "had an extra two years to study AlphaGo games" versus "spent more time learning go by other means".RobertJasiek wrote:Here are my comments on the peer-reviewed paper Human Learning from Artificial Intelligence: Evidence from Human Go Players’ Decisions after AlphaGo by Minkyu Shin, Jin Kim and Minkyung Kim in Proceedings of the Annual Meeting of the Cognitive Science Society...
The pseudomathematical notation on page 1797 is irritating. It's just an opaque way of saying "measure the average winrate drop". Dressing it up in symbols this way makes it harder to read without adding value. (I've seen equally bad things done in music theory papers! It's a way of trying to improve your chances of publication by impressing reviewers with your "cleverness",)
What I'd like to see is:
- Use of historical data as a control. By these metrics, did humans get any better at go between, say, 1950 and 2000?
- Critical examination of the metrics used and consideration of alternatives. When Leela Zero first came on the scene, there was plenty of discussion in this very forum about how to interpret "winrate". And whatever it's measuring, I don't think it's measuring on a linear scale, so taking a simple average isn't good mathematics. Other metrics that occur to me are frequency of a human choosing the AI-recommended move, frequency of "large mistakes" (winrate drop over a certain threshold), and comparisons of outputs from a neural network trained on a database of human games versus the networks trained on AI self-play games.
I do think AI is less of a black box than it's painted out to be. It's possible to have a "conversation" with the AI if you allow the possibility of non-verbal communication, and I believe that humans can and do learn from AI. And I think there's potential to slice and dice datasets in different ways to show evidence of this. But the paper at hand is only a small first step.
-
xela
- Lives in gote
- Posts: 652
- Joined: Sun Feb 09, 2014 4:46 am
- Rank: Australian 3 dan
- GD Posts: 200
- Location: Adelaide, South Australia
- Has thanked: 219 times
- Been thanked: 281 times
Re: how has NN go engines changed way the top people play
Yes, but beware of confirmation bias! It's not enough to ask "What do the successful people do?" The real question is "What do they do differently from others?" For example, the many people you can see on IGS or Fox who've played thousands of games but are still well inside the kyu ranks. They've also built up a huge database, but it's sometimes a case of "garbage in, garbage out". Access to pro game records -- or access to AI -- lets you put something else in.John Fairbairn wrote:Centuries of experience have taught us that the best way for humans to be good at something is to spend a long time at it so as to build up a huge database in the brain - one's personal reference library, or one's intuition - whatever you want to call it.
Now this is interesting. Hard to pin down in a few words, but it starts to make sense of the chaos that is modern high-level go.John Fairbairn wrote:Because of the need for speed, which in a sense comes about because you feel the need to be in two or more places on the board at once, more stress is now being put on the initiative (not sente in the usual western sense), and the value of miai (as per Shuei, of course) is likewise now enhanced. My sense of the new approach to miai is that, while you can't be in two places at once, you can at least guarantee to be in one, as per the traditional view. But nowadays you can try to leave aji in the other, and with a bit of luck you can go back there later and so achieve an effect close to being in two places at once.
-
xela
- Lives in gote
- Posts: 652
- Joined: Sun Feb 09, 2014 4:46 am
- Rank: Australian 3 dan
- GD Posts: 200
- Location: Adelaide, South Australia
- Has thanked: 219 times
- Been thanked: 281 times
Re: how has NN go engines changed way the top people play
So here's an example that popped up for me today. I'm belatedly working through Relentless. Here's one of the variations from game 5:
In the game, black played 'a' instead of the marked stone. But if black does play the marked stone, where should white reply? The book recommends a shoulder hit at 'b'. But KataGo goes off in a very different direction:
In the game, black played 'a' instead of the marked stone. But if black does play the marked stone, where should white reply? The book recommends a shoulder hit at 'b'. But KataGo goes off in a very different direction:
). I mean it to illustrate "AI style" (or possibly "contemporary human style after the influence of AI"), not necessarily best play. To my amateur eye, you toss in 
and 
to create some aji for future fun, play a forcing move at 
(makes sense: after black has played the marked stone, you don't want to force from the other side), then start a new fight at the top. And depending on black's reply to 
, you might continue the fight at the top, or come back and move out from -
xela
- Lives in gote
- Posts: 652
- Joined: Sun Feb 09, 2014 4:46 am
- Rank: Australian 3 dan
- GD Posts: 200
- Location: Adelaide, South Australia
- Has thanked: 219 times
- Been thanked: 281 times
Re: how has NN go engines changed way the top people play
Yes, I took my time over this. But here is something I noticed today.xela wrote:... when I review games with KataGo. It seems to play forcing moves earlier than was recommended a few years ago. I've always been taught to avoid pushing the opponent around unless I can see a clear benefit. Leave things open, because you don't know what options you might need later. But it's a fine line between leaving it open versus waiting too late and losing the chance entirely. Except that KataGo seems to be suggesting it's not such a fine line: just do it as soon as the chance comes up (usually, not always).
...
Sorry I'm too lazy to pull out specific examples from pro games today. I might come back here next time I notice something.
This is a game between Hashimoto (white) and Takagawa in the 13th Old Meijin tournament. At move 72, after chasing the centre group for a while, white comes back and plays the forcing exchanges a for b and then c for d. I'm looking at this and thinking: how do you learn this sort of timing? White could have played those moves much earlier; why is now the right time for it? But KataGo says no, actually, the pro 9 dans did not have a perfect sense of timing. White should have exchanged a for b back on move 37, as soon as this shape appeared on the board.
Full game for anyone who's interested:
- Attachments
-
- early_forcing.sgf
- (1.41 KiB) Downloaded 393 times
-
lightvector
- Lives in sente
- Posts: 759
- Joined: Sat Jun 19, 2010 10:11 pm
- Rank: maybe 2d
- GD Posts: 0
- Has thanked: 114 times
- Been thanked: 916 times
Re: how has NN go engines changed way the top people play
I looked at the example and based on my understanding of how the search and neural net interact I suspect you shouldn't read too much into this case, unless you can actually find some variations where failing to make the exchange earlier costs something? Sometimes you can actually find some variations that do show that you have to exchange now. Sometimes there isn't one, where the exchange will always be makeable safely later.
Assuming I didn't miss actual variations that punish missing the exchange (don't trust me entirely, I might have!) then in this case, I would guess that the neural net probably rates the position as very slightly better for white with a stone at "A" and probably the net "understands" that "A" is sente-ish but it's not quite as salient to the neural net if the exchange is not made. So the search on average is returning very slightly higher evaluations for making the exchange earlier, rather than delaying it and having a higher proportion of positions in the tree without the exchange. However, if in all variations the rest of the fight never becomes so forcing as to prevent making it later, then it's at least as good if not better to delay, and the higher evaluations of the neural net for making the exchange is in some sense "fake", it's a small correlated error in the judgment of the relative values of the positions with and without the exchange in positions as being different when objectively they are equal.
Bots will usually refrain from making significantly bad exchanges unless horizon-effecting is happening, but for exchanges that will *probably* only be played in one way (i.e. the optionality of the other choices isn't really of much value), and/or where it's very unlikely that correct play will result in a ko in the intervening time for which the move would be a good threat, for those exchanges there's little to stop them from being preferred. The cost to playing them a bit earlier is very very small or zero, and so some for some fraction of them correlated evaluation noise will have the right sign to outweigh that ~zero cost and give the bot a clear preference to play earlier... and precisely because the cost is very small it doesn't appreciably cost much if any strength in practice for the bot to doing so.
Assuming I didn't miss actual variations that punish missing the exchange (don't trust me entirely, I might have!) then in this case, I would guess that the neural net probably rates the position as very slightly better for white with a stone at "A" and probably the net "understands" that "A" is sente-ish but it's not quite as salient to the neural net if the exchange is not made. So the search on average is returning very slightly higher evaluations for making the exchange earlier, rather than delaying it and having a higher proportion of positions in the tree without the exchange. However, if in all variations the rest of the fight never becomes so forcing as to prevent making it later, then it's at least as good if not better to delay, and the higher evaluations of the neural net for making the exchange is in some sense "fake", it's a small correlated error in the judgment of the relative values of the positions with and without the exchange in positions as being different when objectively they are equal.
Bots will usually refrain from making significantly bad exchanges unless horizon-effecting is happening, but for exchanges that will *probably* only be played in one way (i.e. the optionality of the other choices isn't really of much value), and/or where it's very unlikely that correct play will result in a ko in the intervening time for which the move would be a good threat, for those exchanges there's little to stop them from being preferred. The cost to playing them a bit earlier is very very small or zero, and so some for some fraction of them correlated evaluation noise will have the right sign to outweigh that ~zero cost and give the bot a clear preference to play earlier... and precisely because the cost is very small it doesn't appreciably cost much if any strength in practice for the bot to doing so.
-
kvasir
- Lives in sente
- Posts: 1040
- Joined: Sat Jul 28, 2012 12:29 am
- Rank: panda 5 dan
- GD Posts: 0
- IGS: kvasir
- Has thanked: 25 times
- Been thanked: 187 times
Re: how has NN go engines changed way the top people play
These forcing moves are of the kind which white will be making anyway at some point since they are good endgame. The thing that can change in this position is that black could give up on the cutting stones. From what I learned it is important to get these kind of forcing moves before something changes in the position and they stop becoming forcing.
The forcing moves are also important in the middle game for making eyes. That is the reason I see for why they were played when they were played. Even without this reason it is important to get your forcing moves but for awhile it looked like it was really just endgame. It seems rather farfetched that white could miss out on these as endgames.
In short it feels natural to make the forcing moves before testing if black will give up on the cutting stones. However, it is a bit of a question of what you prefer. Would you like to capture the cutting stones or the corner? I'd rather not have to capture the corner if black can make two equal moves elsewhere (but how likely is that?) and I wouldn't mind capturing the cutting stones. Basically, making or not making the forcing moves both feel natural to me.
We can enjoy having new insight into the minute details of old games with not much effort. For example if you have black play a one space jump on move 55 (those are never wrong) then the issue of when to play the forcing moves appears to come to a head. In this case black didn't defend the cutting stones and we have the typical situation that white would want to make the forcing move before ever capturing the cutting stones. In some other situation capturing the cutting stones immediately is suggested by KataGo as coming before the forcing moves, but with some patience and faith in old Go theory it has so far always updated to the two being about the same.
There is also another reason for delaying forcing moves. It is more important for us that are not the top players
It is that you shouldn't be making unnecessary moves. Those moves are a sign of not having known what to do. It is expected that top players, who appear to know what to do, are so meticulous about making the moves that need to be made and not all the other moves. It shows that they work hard to always play better.
The forcing moves are also important in the middle game for making eyes. That is the reason I see for why they were played when they were played. Even without this reason it is important to get your forcing moves but for awhile it looked like it was really just endgame. It seems rather farfetched that white could miss out on these as endgames.
In short it feels natural to make the forcing moves before testing if black will give up on the cutting stones. However, it is a bit of a question of what you prefer. Would you like to capture the cutting stones or the corner? I'd rather not have to capture the corner if black can make two equal moves elsewhere (but how likely is that?) and I wouldn't mind capturing the cutting stones. Basically, making or not making the forcing moves both feel natural to me.
We can enjoy having new insight into the minute details of old games with not much effort. For example if you have black play a one space jump on move 55 (those are never wrong) then the issue of when to play the forcing moves appears to come to a head. In this case black didn't defend the cutting stones and we have the typical situation that white would want to make the forcing move before ever capturing the cutting stones. In some other situation capturing the cutting stones immediately is suggested by KataGo as coming before the forcing moves, but with some patience and faith in old Go theory it has so far always updated to the two being about the same.
There is also another reason for delaying forcing moves. It is more important for us that are not the top players
It is that you shouldn't be making unnecessary moves. Those moves are a sign of not having known what to do. It is expected that top players, who appear to know what to do, are so meticulous about making the moves that need to be made and not all the other moves. It shows that they work hard to always play better.-
dust
- Lives with ko
- Posts: 161
- Joined: Sun Feb 28, 2016 4:01 am
- GD Posts: 0
- Has thanked: 138 times
- Been thanked: 23 times
Re: how has NN go engines changed way the top people play
It's the increased use of probes that I find most interesting.
There seem to be at least 2 main types:
- the early probe invasion which seems to ask or try to force the opponent's hand in committing to claiming an area, and may also leave possible aji
- the contact probe, which I'd speculate may often be previously researched and potentially opens up a sequence to new variations, or offers an opportunity of a fight or exchange
There's probably a lot of other uses of such moves as well that pass completely above my comprehension.
There seem to be at least 2 main types:
- the early probe invasion which seems to ask or try to force the opponent's hand in committing to claiming an area, and may also leave possible aji
- the contact probe, which I'd speculate may often be previously researched and potentially opens up a sequence to new variations, or offers an opportunity of a fight or exchange
There's probably a lot of other uses of such moves as well that pass completely above my comprehension.