From viewtopic.php?p=232131#p232131

Why am I interested in AI? Over the years conventional go thinking has developed. This move is bad and that move is good, etc. Sometimes I agree and sometimes I don't but I am nowhere near pro strength so I simply have to take the word of strong players. Now there is a way for me to make a reasonable judgment as to whether these conventional ideas have merit or if they do not, at least some of them.

AI has clearly shown that some ideas conventionally held by strong players are not 100% correct. I find that fascinating.

Regarding a) it seems the strong bots can actually make some of the same blunders from blindspots as humans. Here's a human amateur dan game I was reviewing with Lizzie recently. When I saw white play the attacking/checking extension of 1 I thought "Oh, not c2?" Whether this was necessary for life or just a big reverse sente I didn't know without reading. So black played there and then came out.

Click Here To Show Diagram Code

[go]$$Wc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . O . . . . . |
$$ | . . . O . . . . . . . . X O . O O O . |
$$ | . . . , . . . . . X . . . . X X X O . |
$$ | . . O . . . . . . . . . X . . . X X . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . 7 9 . . . . . . . . . . . . . . |
$$ | . . 1 6 . . . . . . . . . . . . . . . |
$$ | . . . 8 . 0 . . . . . . . . . . X . . |
$$ | . . X , . . . . . , . . . . . , . . . |
$$ | . X X O . . . . . . . . . . . . . . . |
$$ | X O X O . O O . . . . . . . . . . . . |
$$ | . O X X X X O . . . . . . . . . . . . |
$$ | . O O X X O O X . . . . . . . . . . . |
$$ | . . O X O . O . . . . . . . . . . . . |
$$ | . . O X X O . . . , . . . . . , X . . |
$$ | . . O O X . O . . O . X . . X . . . . |
$$ | 5 3 2 X . . . . . . . . . . . . . . . |
$$ | . 4 . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

White reduced on the right, and black made a placement inside the group. Boom! With the cut of 6 white was dead as short of liberties to make the 2nd eye if a-b (so white played b and made a ko to save some of the group).

Click Here To Show Diagram Code

[go]$$Wc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . O . . . . . |
$$ | . . . O . . . . . . . . X O . O O O . |
$$ | . . . , . . . . . X . . . . X X X O . |
$$ | . . O . . . . . . . . . X . . . X X . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . O O . . . . . . . . . . . . . . |
$$ | . . O X . . . . . . . . . . . . . . . |
$$ | . . . X . X . . . . . . . . . . X . . |
$$ | . . X , . . . . . , . . . . . 1 2 . . |
$$ | . X X O . . . . . . . . . . . . . . . |
$$ | X O X O . O O . . . . . . . . 3 . . . |
$$ | b O X X X X O . . . . . . . . . . . . |
$$ | . O O X X O O X . . . . . . . . . . . |
$$ | . a O X O . O . . . . . . . . . . . . |
$$ | 5 4 O X X O . . . , . . . . . , X . . |
$$ | . 6 O O X . O . . O . X . . X . . . . |
$$ | O O X X . . . . . . . . . . . . . . . |
$$ | . X . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

So what does LeelaElf (the Facebook Elf OpenGo network converted to LeelaZero format) think in this tsumego? It doesn't see the cut, and just wants to play a ko (the only other moves it considers a small amount are on right side). I suspect this is because it's not seeing that the white group has no external liberties and is just looking at the local corner shape (in which the atari for ko is the best move if the white group has an external liberty). I've seen other instances of bots playing local shapes like this ignoring the lack of liberties on large outside chains. However, once I play the black cut and evaluate again, Elf rapidly recognises that white is bad with a 30+ swing in win rate. Also, just to make sure Elf wasn't thinking as black it didn't like white making the ko 1st line connection, I added an exchange there so white was totally dead with the cut and it still wanted to play the ko mistake.



LeelaZero (#145) has the same blindspot (after 3k playouts). Maybe it or Elf would find the cut with more playouts?

How about regular Leela 0.11? It starts off liking the a5 atari, but b3 is a distant 3rd/2nd choice at least considered with <1% of playouts. Doesn't find it after 0.6m sims.

Would poor old maligned GnuGo get it right? I don't have it to check, but think it might.

How about humans? Well, here is Frank Jansen former 6d also missing the shortage of liberties to give the British player a valuable win in the Pandanet league this season (in fact he also missed that capturing the outside stone effects the life of the outside group so I think it becomes a seki instead of dead, but he resigned):



P.S. In the first game the white player was suspected of using a bot in another game, so does him missing the shortage of liberties suggest he was doing so here? Yes (a little bit), if we take it as given that human mid-high dans won't make that mistake. However, Frank is a nice counterexample that they do (he would have avoided this situation by playing hanging not solid connection some moves earlier). I am also aware that missing shortage of liberty situations is a common cause of my mistakes in tsumego problems.

AI can also fail. Not only AI but also a) weaker players and b) mathematical go theory have shown that strong players can make mistakes or have flawed theory. Anyway, what gave you the idea that strong players would be 100% correct...? :)

Maybe you mean more general how to approach the game, but that early 3-3 invasions are okay/good seems a pretty obvious way in which humans have changed their thinking based on the bots. Maybe you dismiss this as mindless copying? I expect there are other more subtle changes in thinking that pros in Asia have made that we don't know about in English. The 3-3 invasions related idea that walls without eyes can be attacked isn't new, but does I think have new emphasis these days. Or there's Lee Sedol's comments after losing to AlphaGo that he should rely more on reading with an open mind than his now apparently restricted and faulty intuition (of course a wise person could have told him this before bots but they really hammered the point home).

As for me personally, I've happily adopted the "[when you don't know what to do] attach on top of your opponent's stones" idea from bots (e.g. see my recent European league games in my study journal, also did a few in the British Challengers' league that I've not recorded yet). It has mixed results in terms of winning, but always makes for fun and interesting games so I count that as a positive contribution.

What I mean when I say current AI contributes nothing to how humans should think is that it does not explain anything. It only gives raw data. We have to do the inferring (i.e. guesswork). We may occasionally be inferring something useful but not necessarily for the right reasons and not necessarily everything there is to infer.

Nor do we ever know if we are starting at the right place - my suspicion is that all the discussion here about joseki moves is a case of starting off at the wrong place. As the wise Irishman said to the tourist lost in the countryside and asking how to get to Dublin: If you want to go there I wouldn't start from here.

There is also the aspect that what works for an AI that doesn't make blunders might not be best for us - a certain style of safety-first play may be better for humans.

I'm not saying the inferring process is not fun, of course.

Interesting, and may hold true from before Alphago— maybe a teacher could say, 'this cutting move here is good for you, because you are good at reading and can handle this situation, but as for you, you are good at whole-board judgement, play this move which leads to a positional game', or reverse it to train a students weaknesses?

_________________
On Go proverbs:
"A fine Gotation is a diamond in the hand of a dan of wit and a pebble in the hand of a kyu" —Joseph Raux misquoted.

I get your sentiment here, John, but isn't it similar to one of the major benefits someone would get from, say, GoGoD? There are thousands of uncommented games, but users can still get benefit from seeing where pros played, even if they don't always have the explanation to back it up. The dataset may not yet be as large with some AI programs, but presumably, there could be similar benefit achieved to getting new ideas from moves, couldn't there?

I agree with the point that "what works for an AI that doesn't make blunders might not be the best for us", but I guess that could be applied to pro play, too(?)... What works for a pro may be hard for me to practically apply in my own games since they are reading a lot of variations that I don't understand. At least with some AI programs, you can see some of these variations (i.e. you have more than just the move the pro played).

I'm generally in agreement with you about AIs not being able to explain the strategy behind the moves, but I don't see a lot of difference between this and a collection of professional games that have no commentary, so I'm a bit surprised.

_________________
be immersed

It's even worse with pro games without commentary, because you can ask the AI "what if I play here instead?", and it will tell you what could happens, so you can kinda use some sort of (twisted) Socratic method with the bot

By contrast with top chess engines, and as Uberdude's example indicates, the main strength of top go bots lies with what, in humans, we call intuition. OC, the bots don't explain their play, but human intuition is hard to explain, as well. Also, humans are very good at coming up with post hoc explanations for their decisions. So we should not be too impressed by human explanations for intuitive plays.

The good news, I think, is that humans are very good at imitation. What is important is not so much surface imitation, as in, AlphaGo plays the 3-3, so I will, too, but what is called programmatic imitation, which is more thoughtful. Children can benefit from imitating strong go bots, and I think that the results will be apparent 10 years from now.

One place I think that humans are superior to current go bots is in the application of logic and experience. We can see that, I think, in Uberdude's spidey sense that White C-02 might have been a better move in the example than the checking extension.

Humans who develop their intuition by imitating bots and also apply logic and experience will be truly awesome by current standards. In addition, they will be able to come up with explanations for their intuitive plays.

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

Visualize whirled peas.

Everything with love. Stay safe.

Elf would have played c2 instead of the c12 extension, the only move it says (5% preference), so wouldn't have got itself into the situation where it could have blundered the life and death (even thought it's not aware it's a blunder! I suppose it thinks even a ko is bad).

Good for Elf! I stand corrected.

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

Visualize whirled peas.

Everything with love. Stay safe.

Your points about a human-play database vs AI games are possibly true for most people, but in my case I almost always have some sort of commentary in the source. Even the mini ones in the yearbooks can be insightful. I can't pretend I actually play the whole game over while studying the contents, but I usually run my eye over the commentary and do pick up quite a bit that way.

And of course many other people have books and magazines with commentaries and so like to have the sgf game for convenience while studying. Joseki books also often reference a game and it is nice to have the full record. There are other uses. The raw historical record is invaluable when researching players, and although xxZero has now dispensed with the need for machines to consult a large human database, in the early days of AI GoGoD was used by several researchers. Giants standing on the shoulders of toddlers in this case, of course, yet still...

But, truth to tell, to some degree I'm also honouring a promise to T Mark to get to 100,000 games (98,000 at present). Don't forget, too, that those 100,000 games will also contain AI games.

Let me use my thinking about the early 3-3 invasion as a small case study

Unlike a lot of beginners, I was never attracted to the early 3-3 invasion. This feeling was bolstered by discovering that pros seldom played it early. Later I learned some techniques to minimize the effectiveness of the coming wall before making the invasion.

Click Here To Show Diagram Code

[go]$$W Joseki
$$ ------------------
$$ | . . . . . . . . .
$$ | . . O O X . . . .
$$ | . . O X . X . . .
$$ | . . O X . . . . .
$$ | . O X . . . . . .
$$ | . O X . . . . . .
$$ | . . X . . . . . .
$$ | . . . . . . . . .[/go]

I accepted that this shape was the result of joseki, but felt that it was advantageous for Black. (Else why would pros put the invasion off?)

So I was shocked when AlphaGo started playing the early 3-3 invasion. Black often avoided the diagrammed result, either to take sente or to take the corner. But when Black did play as in that diagram, White deviated.

Click Here To Show Diagram Code

[go]$$W Joseki
$$ ------------------
$$ | . . . . . . . . .
$$ | . . . . . . . . .
$$ | . . O X . . . . .
$$ | . . O X . . . . .
$$ | . O X . . . . . .
$$ | . O X . . . . . .
$$ | . 1 X . . . . . .
$$ | . . 2 . . . . . .
$$ | . . . . . . . . .[/go]

Instead of playing the hane-tsugi White crawled along the second line. I think we have to say that this is joseki, not the hane-tsugi. Having seen it, I have to ask, Why didn't I think of that?

As a human, I can offer an explanation, too. OC, it is post hoc, but so what? Both the crawl and the hane-tsugi bolster the Black wall, but the hane-tsugi gives Black eye-shape, so that the wall is almost invulnerable. After the crawl, the Black wall is still strong, but it is potentially attackable. A pro could give a better explanation, I'm sure, but the fact that AlphaGo does not offer one itself is not such a loss.

Even so, I have to question my intuition, because Black still looks better to me in the second diagram. (By better, I mean more than 14 pts. better; Black has played one more stone than White.)

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

Visualize whirled peas.

Everything with love. Stay safe.

Where did I say that I had that idea?

Is this too hard to understand? "Why am I interested in AI? Over the years conventional go thinking has developed. This move is bad and that move is good, etc. Sometimes I agree and sometimes I don't but I am nowhere near pro strength so I simply have to take the word of strong players."

But this is not unique to AI. We do it all the time with human moves (2nd-line probes against a shimari were rare until little more tan a century ago, 5-5 was new, 10-10 was new, josekis go in and out of fashion...). The benefit of using human games as a basis is that we either have or can get a human explanation we have a chance of understanding, however imperfectly. AI has just widened the pool (by pissing in it, in my view )

One potential danger of post hoc explanation is that the explanation might have bias. Explanations are often a bit subjective (e.g. "black is good because he gets sufficient influence here"). When you know the "answer" beforehand, you can come up with reasons to support the quality of the move that you might otherwise not buy into, simply because you know that there has to be some reason the AI played that way.

I'd argue that post hoc explanation can still be valuable to make sense of a position, but I can't say that it's very scientific.

_________________
be immersed

Please note my previous statement that human explanations about their decisions, particularly about intuitive decisions, are typically post hoc. This is clear in experiments involving hypnosis and split brain research.

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

Visualize whirled peas.

Everything with love. Stay safe.

This is an example of what I was talking about:

I learned by AI:

How to haengma

How to surround

How to kill

How to live

How to joseki

How to fuseki

How to honte

How to tesuji

How to tenuki

How to surround (intentionally mentioned a 2nd time)

How to count

How to decide

How to compare

How to connect

How to split

How to read

How not to pray and play

(Never forget AI has a nice intuition, but it is also quite competent at reading)

And sometimes I win nowadays ... dont care much about explanations ... i just play the moves I learned by AI