Life In 19x19

Cassandra wrote:

CutFirstThinkLater wrote:From what I understand AlphaGo's knowledge of pro games are mainly used for reading ahead(predicting moves) and selecting candidate moves. Since those two compose a bulk of one's playing skill, I think it's likely that AlphaGo's skill will be stuck at 'human level' for quite some time, until its experience piles up enough to see the flaws in those man-made textbooks. So maybe your guess will be right after all, at least for a while.

Don't you think that this effect ("seeing" some flaws) might have happened once already, with the textbooks written by strong amateur players (KGS 6d - 9d) ?

Yeah, I heard the process was under the way (i.e. discarding games it now considered 'weak' - like the 6~7d games), but not completed (i.e. wasn't able to discard 9d level games yet) by the time of the Fan Hui matches.

Anyway, what happened with those 'KGS textbooks' seems largely irrelevent since I heard they are now mostly feeding the program 'professional made textbooks' to prepare for the Lee Sedol matches. And in this case I think the 'surpassing textbook' stage will start a lot later than with the KGS games, because the quality of the games are not only higher but much less diverse. In the case with KGS games there was a large difference in strength so the program was able to sort out the weaker ones with the hints it gained from the stronger ones. With professional games that won't happen very much; it'll have to wait until it polishes it's skills internally and that will take a lot more time. Which is why I don't think it will happen until the 'Big Game' starts.

John Tilley wrote: I decided to join Life in 19x19 today - so this is my first post!

Hi John, long time no see!

Myungwan Kim criticized alphago for following bad joseki. I wonder if Lee Sedol could prepare some hamate taking advantage of this. If you know the bot is going to play the "correct" move maybe you can take it off-book and confuse it early.

There was an interesting article posted earlier today, where Lee Sedol was interviewed after returning from the Nongshim Cup ([1]).

During dinner, after returning from Shanghai, he drank for a bit for the first time in a little while (apparently, it was Kaoliang). He'd been drinking, so maybe he was drunk during the interview (that's my own speculation), but was quoted as saying:

내가 정말 화나는 일은 패배라는 결과가 아니다. (What I'm really angry about is not the fact that I lost.)

Rather, he said that though Ke Jie made some fatal mistakes, he (Lee Sedol) made more fatal mistakes. And if he (Lee Sedol) dwelt on that fact, he was really angry at himself.

나 자신에게 만족하는 바둑을 두고 싶다. (I want to play Go that is satisfying to myself.)

Lee Sedol always says, and it means that the various mistakes that he made during the game against Ke Jie were extremely painful.

When asked whether these same kind of mistakes would happen in his upcoming match against AlphaGo, Lee Sedol replied that he'd make similar mistakes. But, he declared:

그 실수가 오늘처럼 패배로 이어지지는 않을 것이라고 단언했다. 왜냐하면 ‘알파고’의 기력은 프로급이 아니기 때문이라고 했다. (Those mistakes won't decide the game like they did today. That's because 'AlphaGo' isn't really pro level.)

In a recent public interview about AlphaGo, it was brought up that AlphaGo had gotten to a fairly strong level, and they asked if AlphaGo could play evenly with Lee Sedol. To that he replied:

“선바둑이라는 말이 곧 프로급이 아니라는 뜻이다”라며 “선바둑이란 구체적으로 말하면 내가 ‘가지고 놀 수 있다’는 소리”라고 답했다. 선바둑에게 자신이 지는 일은 거의 없다고 했다.

("Just because AlphaGo plays evenly with me doesn't mean that it's at pro level. If you really want a detailed answer, playing evenly with AlphaGo, I'd be able to mess/play around (and still win)." There's just about no way I'd (Lee Sedol would) lose in an even game.)

He later said:

“사실 판후이 2단과 둔 기보를 보면 두 점 접바둑을 둬서도 이길 수 있다. 그러나 ‘알파고’가 매월 수만 판의 바둑을 두면서 스스로 학습한다기에 선바둑으로 높여 준 것”이라며 “하지만 그것이 지금 나를 이길 수 있다는 소리는 절대 아니다”

("Actually, looking at Fan Hui 2p's game records, even if I gave him a 2 stone handicap, I could still win. But even if you say that AlphaGo is playing lots of games every month, learning by himself, to raise to the level of playing me evenly, there's still no way AlphaGo could win.")

“정상급 프로기사들의 바둑에서는 계량화할 수 없는 ‘비틀기’나 ‘흔들기’가 나오는데, ‘알파고’는 이에 대한 훈련이나 감각이 없는 듯하다”고 설명했다. 말 그대로 기계일 뿐이고, 직관과 감각이 승부에 절대적 영향을 미치는 바둑에서, 지금 수준의 ‘알파고’가 프로를 뛰어넘을 수 없다는 것이다.

따라서 알파고와의 승부에서 자신이 5-0으로 승리하는 것은 당영한 일이니 행복한 결과이지만, 4-1로 이기는 것은 이변이며, 3-2로 이기는 것은 아주 불행한 일이라고 말했다.

("In top pro games, there are unquantifiable 'twists' and 'shakes' that aren't a part of AlphaGo's training or sense," He explained. Considering that, AlphaGo is just a machine; intuition and sense definitely have an influence on who wins the game, so the level of AlphaGo right now cannot surpass pros. Accordingly, in the upcoming match against AlphaGo, it's obvious that I'll win 5-0, so that'd be a happy result; but if the result were something like 4-1 or 3-2 (in Lee Sedol's favor), I'd (Lee Sedol would) be very unhappy with the result.)

Later, near the end of the article, Lee Sedol acknowledged that it'd actually be scary to play AlphaGo in maybe 2 years from now. So 2 years down the road, if he were given the challenge to play AlphaGo, he wouldn't do it for $1 million dollars - they'd have to pay him $3 million to play in that case




[1] http://sports.media.daum.net/sports/gen ... _SPORTS=R6

Kirby wrote:In top pro games, there are unquantifiable 'twists' and 'shakes'

Too bad Lee Sedol does not explain what he means here. (Instead he already hides it while saying it: "unquantifiable".)

RobertJasiek wrote:Too bad Lee Sedol does not explain what he means here. (Instead he already hides it while saying it: "unquantifiable".)

Yeah, I dunno. Either he really can't quantify it, or he just can't quantify it while drinking Kaoliang

Also, on a related note, there's another (IMO less interesting) article [1], where Park Junghwan 9d (ranked #1 in Korea), Ke Jie 9d, and Iyama Yuta 9d all basically say that they think Lee Sedol will win 5-0 against AlphaGo this time. Though, some argue that $1 million dollars isn't enough for this match, given the significance that it would represent if Lee Sedol 9d were to somehow lose. I guess that's why Lee said he'd increase the price if the match were held in a couple of years - he and other top pros are extremely confident that the computer has no chance.

We'll find out if their confidence is founded next week

[1] http://sports.media.daum.net/sports/gen ... _SPORTS=R1

Thanks for the nice translation! But I think I'll have to make a small correction to this part:

Kirby wrote:He later said:

“사실 판후이 2단과 둔 기보를 보면 두 점 접바둑을 둬서도 이길 수 있다. 그러나 ‘알파고’가 매월 수만 판의 바둑을 두면서 스스로 학습한다기에 선바둑으로 높여 준 것”이라며 “하지만 그것이 지금 나를 이길 수 있다는 소리는 절대 아니다”

("Actually, looking at Fan Hui 2p's game records, even if I gave him a 2 stone handicap, I could still win. But even if you say that AlphaGo is playing lots of games every month, learning by himself, to raise to the level of playing me evenly, there's still no way AlphaGo could win.")
-> CORRECTION: ("Actually, looking at its game records with Fan Hui 2p, even if I gave it a 2 stone handicap, I could still win. But even if you say that AlphaGo is playing lots of games every month, learning by himself, to raise to the level of playing me evenly, there's still no way AlphaGo could win.")

I don't think a pro will ever openly thrash his fellow pro's strength like that, even for someone as fiery as Lee.

I wouldn't be so surprised if DeepMind say they're already expecting those words from Lee Sedol.

CutFirstThinkLater wrote:Thanks for the nice translation! But I think I'll have to make a small correction to this part:

Kirby wrote:He later said:

“사실 판후이 2단과 둔 기보를 보면 두 점 접바둑을 둬서도 이길 수 있다. 그러나 ‘알파고’가 매월 수만 판의 바둑을 두면서 스스로 학습한다기에 선바둑으로 높여 준 것”이라며 “하지만 그것이 지금 나를 이길 수 있다는 소리는 절대 아니다”

("Actually, looking at Fan Hui 2p's game records, even if I gave him a 2 stone handicap, I could still win. But even if you say that AlphaGo is playing lots of games every month, learning by himself, to raise to the level of playing me evenly, there's still no way AlphaGo could win.")
-> CORRECTION: ("Actually, looking at its game records with Fan Hui 2p, even if I gave it a 2 stone handicap, I could still win. But even if you say that AlphaGo is playing lots of games every month, learning by himself, to raise to the level of playing me evenly, there's still no way AlphaGo could win.")

I don't think a pro will ever openly thrash his fellow pro's strength like that, even for someone as fiery as Lee.

Thank you. Looking at it now, you are correct (판후이 2단과 둔 기보) - I totally agree.

I read this part too quickly.

I am doing these translations partially for my own study, so if you find any other mistakes, feel free to let me know.

pookpooi wrote:I wouldn't be so surprised if DeepMind say they're already expecting those words from Lee Sedol.

I think DeepMind is confident from their side, too - to be sure.

I lost a little bit of respect for them, though, when I saw the nature video clip where one of the lead programmers on the team mentioned their upcoming challenge against "Lee See-dol".

It's a small mistake, and has nothing to do with the development of their AI - but at least it gives me the idea that he might not understand the magnitude of difference between a pro and top pro if he doesn't even know how to pronounce the name of the person they are challenging.

That being said, a lot of pros don't understand deep learning that well, either, so anything could happen next week, in my opinion

RobertJasiek wrote:

Kirby wrote:In top pro games, there are unquantifiable 'twists' and 'shakes'

Too bad Lee Sedol does not explain what he means here. (Instead he already hides it while saying it: "unquantifiable".)

He means tipping the game off balance and making it complicated and chaotic when things don't go one's way.


Although I personally think the pros are being a little bit overconfident, I somewhat agree to Lee's argument here that humans have the upper hand in their "twisting and shaking". Because although at first glance machines might seem to be stronger in complicated fights due to their precise reading ability, in reality this may not be true for the following reasons:

1. At pro level, fights are usually not just about linear reading but making many strategic choices (i.e. selecting the best variation that serves the bigger strategy) along the way, thus involving a lot of evaluation. Even though AlphaGo employs a specialized network dedicated to this part, I imagine evaluation is still not its forte (yet). OTOH, since AlphaGo's evaluation skills is not based on imitating humans, I think this is the area where it most likely can surpass the pros. Whether it can do it this time, we'll soon find out.

2. The bigger problem in complicated fights I think is not evaluation, but selecting candidate moves to read ahead. In order to narrow the near infinite number of searching-trees, AlphaGo picks candidate moves to read based on 'intuition'. This 'intuition' is basically imitating human moves and using standardized data collected from playing itself. The problem is, because of this 'standardized' selecting process, it can become weak at thinking outside the box and might be prone to missing unusual but brilliant moves (an area which Lee particularly excels.).

Just my 2 cents.

We all know that the big scientific question is whether they're going to make Hikaru no Go 4 for after this, obviously.

Kirby wrote:It's a small mistake, and has nothing to do with the development of their AI - but at least it gives me the idea that he might not understand the magnitude of difference between a pro and top pro if he doesn't even know how to pronounce the name of the person they are challenging.

It's an embarrassing and unprofessional mistake, but it may also be a product of reading more about a subject than you talk about it. I don't know how to pronounce Lee's name either. I wouldn't pronounce it 'See-dol', but I don't know how 'dol' is supposed to sound.

Related: in graduate school, I was good friends with a Serbian grad student. After knowing him for a year, he and his girlfriend got married, and she joined him in the States. I constantly mispronounced her name ('Yelen-a' instead of 'Ye-lena'). I could hear the difference, I could say it right when being careful, but in conversation, it came out wrong. I finally had to stop myself, and apologize for getting it wrong so much.

tl;dr: mouth-words are hard.

Communication is fraught with peril.

Kirby wrote:Yeah, I dunno. Either he really can't quantify it, or he just can't quantify it while drinking Kaoliang

Did he yell, "I love you, man!" halfway through the interview? That would be a tell.