It is currently Thu Mar 28, 2024 6:30 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 28 posts ]  Go to page Previous  1, 2
Author Message
Offline
 Post subject: Re: Guess KataGo's next move
Post #21 Posted: Wed Dec 07, 2022 1:45 pm 
Honinbo

Posts: 9545
Liked others: 1600
Was liked: 1711
KGS: Kirby
Tygem: 커비라고해
Mike Novack wrote:
But we aren't so much being asked to imagine the move as to recognize what might be good about it after having seen it.


In a game, you have to imagine the move. Maybe by analyzing these kinds of moves after seeing them, we get better at that. But you must come up with your own moves during a game.

Quote:

And perhaps not so much AI being better at understanding the center as in understanding the whole board. Because I think whole board is how we should look at it.


Of course the whole board is important. But humans are already relatively good at understanding corners and sides. That's what most joseki is based off of, for example.

_________________
be immersed

Top
 Profile  
 
Offline
 Post subject: Re: Guess KataGo's next move
Post #22 Posted: Thu Dec 08, 2022 3:17 pm 
Lives in sente

Posts: 1037
Liked others: 0
Was liked: 180
Kirby wrote:
Mike Novack wrote:
But we aren't so much being asked to imagine the move as to recognize what might be good about it after having seen it.

In a game, you have to imagine the move. Maybe by analyzing these kinds of moves after seeing them, we get better at that. But you must come up with your own moves during a game.
[quote]

One step at a time, please. I agree with you that the ULTIMATE goal is to be able to imagine wonderful move. However that would be impossible if EVEN being shown the move we cannot see why it is wonderful.

In other words, SOMETIMES when the AI or a stronger player shows us what the right move was, we can immediately recognize, yes, that was the right move. But this isn't one of those cases, is it. In other words, being shown the move I can see some of the possibilities but it still isn't obvious. Or am I wrong about that? Is it just because I am a weaker player?

Top
 Profile  
 
Offline
 Post subject: Re: Guess KataGo's next move
Post #23 Posted: Thu Dec 08, 2022 3:33 pm 
Lives in gote

Posts: 470
Liked others: 62
Was liked: 278
Rank: UK 2d Dec15
KGS: mathmo 4d
IGS: mathmo 4d
Gérard TAILLE wrote:
The exchange :b1: :w2: may be good for black for a fight in the center but it is not a good exchange for a fight in in bottom of the board due to the possiblity to play


I think this is incorrect.

Because B doesn't mind responding at B17, which is the whole point of this sequence. The question is whether B17 is bigger or A10 since they are close to miai here. B9 isn't sente and nor does it connect without ko. Even after B9, the B2 hane connect has a similarly large threat as before.

This is perhaps showing where human "creativity" was a blind spot to skill compared to the AI's preference to settle things earlier, just taking the bigger points.

Top
 Profile  
 
Offline
 Post subject: Re: Guess KataGo's next move
Post #24 Posted: Thu Dec 08, 2022 10:03 pm 
Honinbo

Posts: 9545
Liked others: 1600
Was liked: 1711
KGS: Kirby
Tygem: 커비라고해
Mike Novack wrote:
Kirby wrote:
Mike Novack wrote:
But we aren't so much being asked to imagine the move as to recognize what might be good about it after having seen it.

In a game, you have to imagine the move. Maybe by analyzing these kinds of moves after seeing them, we get better at that. But you must come up with your own moves during a game.
Quote:

One step at a time, please. I agree with you that the ULTIMATE goal is to be able to imagine wonderful move. However that would be impossible if EVEN being shown the move we cannot see why it is wonderful.

In other words, SOMETIMES when the AI or a stronger player shows us what the right move was, we can immediately recognize, yes, that was the right move. But this isn't one of those cases, is it. In other words, being shown the move I can see some of the possibilities but it still isn't obvious. Or am I wrong about that? Is it just because I am a weaker player?


I don't disagree. It's not clear to me, either.

_________________
be immersed

Top
 Profile  
 
Offline
 Post subject: Re: Guess KataGo's next move
Post #25 Posted: Fri Dec 09, 2022 2:28 am 
Oza

Posts: 3647
Liked others: 20
Was liked: 4626
I looked at the starting position on KaTrain.

The peep Black 1 did not appear in even the top ten of the candidate moves (most of which clustered on the top right side).

When I played Black 1, KaTrain reported a 0.3 point loss.

Black 1 may not even be the true best move with other versions of KataGo. Is there something equivalent to the horizon effect of computer chess? Or a randomising element, or whatever?

Either way, it is clear that we can't make any sense of why Black 1 was chosen. In contrast, we can make perfect sense of why it might be questionable (it forces White to connect but then Black faces horrible ijime on the left side and so has to defend there, losing sente). Furthermore, we can make perfect sense of why KaTrain chose a move in the upper right (keep away from thickness, etc).

There may be some universe in which Black 1 truly is better than any other move, but until we know where it is and can get there, it seems much better to stick in our own human universe.


This post by John Fairbairn was liked by: Elom0
Top
 Profile  
 
Offline
 Post subject: Re: Guess KataGo's next move
Post #26 Posted: Fri Dec 09, 2022 11:00 am 
Honinbo

Posts: 9545
Liked others: 1600
Was liked: 1711
KGS: Kirby
Tygem: 커비라고해
I feel like it is more beneficial to try to learn from the flow of sequences we see from AI, rather than single moves in a static board position. If we see the flow and how KataGo uses the move, it becomes something easier to interpret and assign meaning to. Looking at a single hard-to-predict move on a static board position doesn't seem that fruitful.

_________________
be immersed

Top
 Profile  
 
Offline
 Post subject: Re: Guess KataGo's next move
Post #27 Posted: Fri Dec 09, 2022 1:11 pm 
Oza

Posts: 3647
Liked others: 20
Was liked: 4626
Quote:
I feel like it is more beneficial to try to learn from the flow of sequences we see from AI, rather than single moves in a static board position. If we see the flow and how KataGo uses the move, it becomes something easier to interpret and assign meaning to.


Although I agree with that completely, I don't think it is a satisfactory answer because it lacks the time element. I'll explain my thinking in two parts.

Part 1: Why I believe flow is crucial. Scottish country dancing often touted as a superior recreation because it provides great physical and mental stimulation. At my age it is the physical stimulation I notice most: wondering the next day how I am going to hobble downstairs to breakfast with aching feet - we dance on the balls of our feet for four hours almost non-stop at big balls. But the we are all also aware of the mental stimulation. It can be a real challenge to remember all the moves in over 20 dances per evening, and much of the chatter between dances is "how does the next dance start again?" and stuff like that. But at a recent ball, graced by many young dancers from Cambridge University, it occurred to me that I know dancers from both Oxford and Cambridge (the elite), but I have not come across dancers from any other universities in England (there are plenty in Scotland, of course). During the break, I asked a Cambridge man about this and had my own jokey 'explanation' ready: the Oxbridge people are toffs who come from stately homes and they need to practise for the big balls they hold when not shooting grouse. I was, however, a bit taken aback by the chap's answer: "You have to be very intelligent to do Scottish country dancing." It was not the words that surprised me so much as the very serious tome in which he said them.

Thinking about that on the drive home, I tried to pin down the elements that I thought made SCD 'intellectual'. There is the challenging element I have already mentioned, remembering many complex patterns, but at another extreme there is the ability to guide beginners or the less experienced seamlessly through a live dance (mostly non-verbally). It is my view that the 'glue' between these two major elements is understanding the flow of a dance.

In dancing (and this applies to other forms such as ballet), we get enormous external help, primarily from the music. A teacher will often ramble on and on with words and when they sense they've lost their audience (as almost invariably happens), they wrap it all up with "JUst listen to the music!" And 9 times out of 10, that's enough. The other thing that teachers say ad nauseam, but absolutely correctly, is "don't look at the floor; it can't help you." If you look at the other dances (and smile) you can both get and give help, and of course it boost the social aspect of the evening.

Although Garry Kasparov wrote a book about how life imitates chess, I prefer to take the opposite stance and have a habit of thinking how go can imitate life. I therefore pondered whether these two adages from SCD could help with playing go. I concluded that the advice not to look at the floor (a very, very common habit) is akin to ignoring the overall position in go. Being obsessed with josekis, or the kill, or forcing moves, or atari, atari, atari. All bad habits.

But "listening to the music?" I was inclined to think that that equated to achieving an intuitive flow by copying the masters. As per standard pro advice, play over lots and lots of master games so that your hands respond as instinctively in go as your feet in dancing. In SCD the footwork can be very intricate and so you need to practise steps carefully, to get good technique. Just like practising tesujis. Once you got that down pat, so that you don't need to think about your technique any move, then you can work on the flow (in SCD this broadly means getting the right combination of long and short steps so that you end up in the right place on the right beat without crashing into anyone else). It can take quite a long time to master this, and many never do. But some (the Sumire's of the dancing world) just seem born with the knack.

The music helps by proving a beat. What is the go equivalent? I had a strong sense that there may even be the equivalent of a beat in go, but I couldn't quite get a firm grasp on that idea, so I left it in abeyance.

My mind, however, was thoroughly sensitised so that, a few days, later when I was finishing off a draft of my latest book, I was able to follow the thread deeper into the Labyrinth. This was Part 2: the time element.

Because this book, the Yi Mo (Brush, Ink and Go, of 1662), showcases quite a lot of old Chinese players I was adding a section on their biographical details. In quite a few cases, the later "Meijin" Xu Xingyou (the one who followed and probably equalled Huang Longshi) had heavily criticised the way these players played. And, in his ultra-famous Jianshantang Yipu, he explained why. I added his comments to the biographies. This translation is part of my Museum of Go Theory project, and so any comments by the likes of him on theory are grist to my mill

Simplifying greatly here, Xu saw the old players of Ming times as too obsessed with local plays. There was no dancing round the board. A game between, say, Lin Fuqing and Guo Bailing would typically fill one quarter of the board, drift to a second quarter, which would be filled up and then a third quarter would be targeted. Guo Bailing, however, was at least aware there was more to the dance than that: he occasionally remembered not to look at the floor. But early Qing times we can see more players sensing the potential of the bigger board. There were players who were prototypes of players who see today. Yong Haoru had a reputation for being good at the end game and he wrote about the importance of following theory. He wasn't very strong, however. At the other extreme, we have Jiang Junfu who was noted for daring innovations and for taking no notice of the old manuals. He wasn't so strong either. Those who were strongest seem to be those who steered a middle course.

But even they were lambasted by Xu Xingyou. He noted their attempts to bring the sides into play, but he felt there was too much reliance on the 8-3 points. Previous players had gradually moved up from 6-3 to 8-3 via 7-3 plays, but Xu was dismissive. However, even he was not quite at the cutting edge, and players after him (such as Fan Xiping and Shi Xiangxia) were putting more emphasis on the 9-3 points.

My point is that this process of moving from the corners bounded by the 6-3 points, to the 6-3 as a side play, thence to the 7-3, 8-3 and 9-3 points was a process of improved understanding that took two or three centuries, even though the area on the go board was so limited (i.e. changing the flow of the game from the corners to the the third line on the sides, remembering that Chinese games started with four corner stones). It was a hard grind even for the greatest of players.

We see something similar with Shin Fuseki. Overall, we can say it probably led to improved play, not least because it opened players' eyes to move they would never have thought of before. These new moves likewise altered the flow of the game (Go Seigen's emphasis on speed, for example). But, ultimately, nobody ever really understood what was going on (with the exception of Go, maybe) and attempts to understand the merits of tengen were explicitly given up as "too hard."


My musings, allied to what happened in history, lead me to believe that trying to understand AI moves individually will fail in the same way as tengen research failed. Trying to understand the flow of AI moves will also fail, but is likely to have beneficial spin-offs. However, these will take a long time to be absorbed, and it is unlikely that amateurs will have the time or skill to absorb the lessons properly. I further believe that amateurs would be better off just dancing to the music, i.e. just playing over many master (or AI) games. If they can brush up their tesujis so that they really feel the beat in the flow of moves, so much the better. But the physical work of putting in the hours must not be neglected either.

I don't see how guessing a bot's next move contributes to that. It can be fun, of course, but it just seems to equate to doing a daily Sam Lloyd chess problem in a newspaper and vainly expecting that to make you a stronger chess player.

Top
 Profile  
 
Offline
 Post subject: Re: Guess KataGo's next move
Post #28 Posted: Fri Dec 09, 2022 1:25 pm 
Gosei

Posts: 1590
Liked others: 886
Was liked: 527
Rank: AGA 3k Fox 3d
GD Posts: 61
KGS: dfan
John Fairbairn wrote:
I looked at the starting position on KaTrain.

The peep Black 1 did not appear in even the top ten of the candidate moves (most of which clustered on the top right side).

When I played Black 1, KaTrain reported a 0.3 point loss.

I used a recent strong network for multiple thousand visits. It took a little while for that to become the favorite move. It is possible that you used a different network or fewer visits.

Quote:
Black 1 may not even be the true best move with other versions of KataGo.

I don't know exactly what you mean by "true best move", but yes, other networks very well might choose different moves.

Quote:
Is there something equivalent to the horizon effect of computer chess?

Sure, some moves' evaluations change a lot when the tree is explored further.

Quote:
Either way, it is clear that we can't make any sense of why Black 1 was chosen.

I thought people here did a pretty decent job, but I may not be strong enough to judge.

Quote:
There may be some universe in which Black 1 truly is better than any other move, but until we know where it is and can get there, it seems much better to stick in our own human universe.

Please feel free to stick in the human universe! I thought the move was interesting and that it might be possible to learn something from the fact that a very strong player chose it.

Quote:
[...] I don't see how guessing a bot's next move contributes to that. It can be fun, of course, but it just seems to equate to doing a daily Sam Lloyd chess problem in a newspaper and vainly expecting that to make you a stronger chess player.

My main intention was to create fun, of course, and I hope I succeeded. I also thought that thinking about what could lead a strong player (whether human or computer) to make such a move could be an interesting exercise that could potentially create some insight. I very specifically said "KataGo's next move" rather than "the best move" because I was hoping to avoid this very sort of debate. I apologize if I was not clear enough that I wasn't trying to make any sort of claim for it to be "the" best move.

Anyway, I like Sam Loyd's (one L, please!) chess problems too, so that shows you where I am coming from.

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 28 posts ]  Go to page Previous  1, 2

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users and 1 guest


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group