Life In 19x19

Ke Jie 9p lost yesterday to Huang Yunsong 5p (http://www.go4go.net/go/games/sgfview/56666) and Aja Huang made the following interesting comment on facebook with a link to an article on the game http://new.eweiqi.com/portal.php?mod=view&aid=27246.

Aja Huang wrote:Ke Jie 9p might receive wrong information about AlphaGo's opening.

文章中提到"卫冕冠军柯洁对黄云嵩一局，执黑的柯洁序盘就不走寻常路，第11手不挡右上角对手的点角，脱先挂左上，据说这是网上流传的AlphaGO自我对弈棋谱视频的着法。"

AlphaGo在自戰譜中並沒有脫先掛左上。AlphaGo總是先拐擋33那一子完成定式後，再去掛角。在我們第2局解說的視頻中，樊麾老師也沒有顯示棋譜的手數。柯傑九段應該是接受到錯誤的訊息了。

From the garbled machine translations I get the impression Ke Jie's approach of 11 at the top left was a bad move (and white 12 hanging connection a good one) and gave him a bad game, but that he may have played this move based on rumours that AlphaGo likes this move, but then Aja says this rumour was incorrect information. Can anyone fluent in Chinese shed some light on this please?

Click Here To Show Diagram Code

[go]$$Bm11 Ke Jie (black) vs Huang Yunsong
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . 2 . |
$$ | . . . . . . . . . . . . . O . X O . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 1 . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

Whatever the specifics here, it is clear that some pros are trying out new moves based on AlphaGo's play. That tells me that they probably think AlphaGo is the strongest player around, because it has long been normal for pros to ape moves by those they regard as their betters. You can call it following fashion. It's more likely a desperate keenness to try and get an edge.

The curious thing about this process is that they clearly do it without understanding the new moves. It's clearly a suck-it-and see approach.

I mentioned recently that I saw evidence, when transcribing 1920s games for the GoGoD database, that the Shin Fuseki of the 1930s had been prefigured in the 1920s. One element of that was a probe contact play facing the centre. I found them very strange but there was a rash of them, accompanied by an unusually high number of early contact plays of other types. The following game shows several examples. It is not that each such move is strange, but the high frequency of them in this game and others of the same period (the type of fuseki shown here was also of unusually high frequency).



The relevant contact plays are moves 23, 28, 59, 70 (especially) and 86.

In those days the typical commentary was not done by a go journalist but by both players. Both White and Black condemned one of their moves as "very bad" (59 and 70). It seems odd that each should play such a "very bad" move in a two-day game where time was not a factor.

It seems the urge to experiment or explore was stronger than normal go sense. That may be what is also going on at the moment.

I'm not sure if Ke Jie play this because of AlphaGo's influence, may be he's trying his own new sequence?

Just because they are inspired by another player doesn't necessarily mean they think that the other player is stronger than they are.

Just because they lost one game doesn't mean they don't understand what they're playing.

I find that some people often make pretentious comments based on many logically flawed assumptions.

But what I can say is two things:

1. I just rewatched Ke Jie's commentary for the Lee Sedol-AlphaGo games again, recently, and in one of the games, Ke Jie clearly said that even if it's proven that AlphaGo's moves are better, professionals won't necessarily play them because they have to understand why the move works themselves.

So if he decides to try out an "AlphaGo move", it would still have been based on a lot of research, until he feels he understands it.

2. I don't remember when exactly, but I remember watching a short interview a month or two ago (so well after the AlphaGo matches, with enough time for the pros to do thorough analyses and learn more about AlphaGo) where the interviewer said that Shi Yue claims he has a 50% chance against AlphaGo v18, and asked Chang Hao what he thought, to which Chang Hao answered he thinks Shi Yue is probably correct.

Of course some pros probably think that AlphaGo is the strongest player now. But certainly not all the top pros think that way. Whether it's because they still don't understand AI or how bots work well enough, clearly some still think they have a good chance against AlphaGo after the Lee Sedol matches.

idontgetit wrote: Of course some pros probably think that AlphaGo is the strongest player now. But certainly not all the top pros think that way. Whether it's because they still don't understand AI or how bots work well enough, clearly some still think they have a good chance against AlphaGo after the Lee Sedol matches.

Some pros may think that way, but I'm fairly confident now that they're mistaken - they probably don't understand Deep Learning. It's simply the same attitude that Lee Sedol had after seeing the Fan Hui games. He assessed AlphaGo's strength based on a handful of game records, and suddenly had confidence. And we can see how that worked out

Uberdude, I think your understanding is right on. It appears that Ke Jie thought was a move in one of the self-playing games by AlphaGo. Interestingly the Chinese article says 局后黄云嵩说曾经看到过黑11的下法，但自己觉得走到12白棋局面还可以(After the game Huang Yunsong said he had seen this move before and believed was a good enough response.). So seems to me there must be some AlphaGo self-playing game video out there that has this move and Ke Jie, maybe also Huang Yunsong had seen it.

The article went on to say Ke Jie, Huang Yunsong and Lian Xiao had long discussions about move 11 after the game, and Ke Jie joked "maybe one should't imitate blindly".

In general, I think that Ke Jie and other pros that copy AlphaGo are taking the wrong approach. It's good to be inspired by ideas that AlphaGo gives, but fundamentally, AlphaGo plays the game in a different way than humans do.

One of the examples I see so often in pro games is the bad exchange here:

Click Here To Show Diagram Code

[go]$$c
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . a . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . B . . . |
$$ | . . . , . . . . . , . . X . X , O . . |
$$ | . . . . . . . . . . . . . X O O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Prior to AlphaGo, many pros would simply say that this move is bad. Besides being aji keshi and a waste of a ko threat, it greatly reduces the threat of a play at 'a' later by black. Locally, this exchange is bad.

I still believe that, locally, the exchange is bad.

More insight into this position can be found from Fan Hui's presentation during the European Go Congress: From this presentation, Fan Hui explains to us that, not only does AlphaGo think that the exchange was good, AlphaGo thought that white's move was bad:

Click Here To Show Diagram Code

[go]$$c
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . W . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . X . X , O . . |
$$ | . . . . . . . . . . . . . X O O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

So what does this mean? Is AlphaGo correct? Is this old joseki flawed? I believe the answer is no.

Following Fan Hui's presentation further, we see that AlphaGo had a plan for its idea of an "ideal" fuseki:

Click Here To Show Diagram Code

[go]$$c
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . O . . . . . . . . . . . X . . |
$$ | . O X O O X O X . , . . . X . X . O . |
$$ | . X X . X O . X . . . . . . . . . . . |
$$ | . . X . X . . . . . . . . . . . O . . |
$$ | . . . O . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . O . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . O . . . . . . . . . . . . . . X . . |
$$ | . X X X . X . . . . . . . . . . X O . |
$$ | . X O O . O O . . . . . . . . . X O . |
$$ | . X X O . . . . . , . . X . X X O . . |
$$ | . X O . . O . X . . . . . X O O O . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Now, if AlphaGo is correct, it very well may be that this board is even for both sides. But what about the bottom right corner?

Click Here To Show Diagram Code

[go]$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . X O . |
$$ | . . . . . . . . . . . . . . . . X O . |
$$ | . . . , . . . . . , . . X . X X O C C |
$$ | . . . . . . . . . . . . . X O O O C C |
$$ | . . . . . . . . . . . . . . . . C C C |
$$ | . . . . . . . . . . . . . . . . C C C |
$$ ---------------------------------------[/go]

White has something like 10 points of territory, but I would totally bet that black's influence is worth more than that.

So what has happened here?

AlphaGo sees its ideal way of how the fuseki will play out, and is willing to play locally inferior plays in order to get the globally ideal result.

This isn't something that we only see in the opening. As we saw even from Game 1 against Lee Sedol, AlphaGo is winning to play point losing moves locally, as long as it increases its confidence that it will be ahead at the end of the game.

Humans don't play this way. In the late middle game, for example, humans are not able to see the end board position. They can't iterate over all possibilities. As a result, pros count point values and aim to maximize profit around the board. Concepts like sente, reverse-sente, and gote are taken into account and even factored into calculation.

But in general, since humans cannot see the final board position, you don't see the point-losing moves that AlphaGo plays. It would be too risky for a human to do that - maybe their point-losing move could cost them the game.

--

So from this, I conclude that humans can be inspired by AlphaGo, but we must remember that AlphaGo doesn't play the same way as humans do. AlphaGo is willing to play point-losing moves in the opening, middle game, and endgame, provided that it is able to increase its confidence in success. Sometimes AlphaGo will play aji-keshi to simplify the board and ensure that it gets a decent result.

But what we've seen from pros is some sort of copying of AlphaGo's moves. While this is interesting, I don't think it's a good idea: It's not the particular moves that AlphaGo plays that makes it powerful; it's it's powerful global evaluation ability, which can read a globally ideal sequence, and is willing to play locally inferior moves in order to have more confidence in its success.

Kirby wrote:AlphaGo is winning to play point losing moves locally, as long as it increases its confidence that it will be ahead at the end of the game.

Humans don't play this way.

Speak for yourself.

since humans cannot see the final board position, you don't see the point-losing moves that AlphaGo plays. It would be too risky for a human to do that - maybe their point-losing move could cost them the game.

I think that you are conflating two things: 1) Whole board play vs. local play; 2) Plays that increase the so-called "probability" of winning, as per relatively weak playouts. Humans -- some of us, anyway --, will definitely take a local loss to take a whole board gain. We do not, as a rule, take a local loss to increase our chance of winning by stupid play -- unless we are desperate or wish to win by trickery.

I disagree, Bill. The first example I gave is clearly an example of where AlphaGo evaluated the entire fuseki, but played a locally bad play.

Humans do utilize local heuristics that AlphaGo might not use. Joseki is a good example. Pros analyze what exchanges are good or bad, because it's a way of simplifying the game. They can't read out as far, so they use these heuristics. AlphaGo doesn't play that way. AlphaGo is willing to play aji-keshi and other moves that pros wouldn't think of, because it has confidence in its readout.

Take a few minutes to listen to Fan Hui's talk. I think it's clear from his discussion that AlphaGo's method of analysis is different from that of humans.

Bill Spight wrote:Humans -- some of us, anyway --, will definitely take a local loss to take a whole board gain.

Also, this may be the case, but perhaps humans don't always do this. Otherwise, it's not consistent with all the pros we see stealing these AlphaGo moves in situations where they are not called for.

According to wolfking, even Ke Jie said he shouldn't have blindly imitated the move from this thread.

Clearly, he didn't do the same type of analysis as AlphaGo did, and was interested in just the move itself.

And FWIW, Hajin Lee also described a difference between AlphaGo's play vs. professional play. She also indicated that human pros often aim to maximize their profit (make sure they get good exchanges, etc.), whereas AlphaGo could have confidence in just being a 1/2 point ahead.

I guess the only point I am really trying to make is this: It is not useful to copy individual moves that AlphaGo makes. Maybe you can be inspired and come up with some new ideas, but AlphaGo wasn't designed to invent new joseki - it was designed to win games, and the moves it plays are specific to the evaluation for that particular game.

Analyzing joseki, good and bad exchanges, etc., are still tasks for humans to undertake, since they're tools that we use for our way of playing the game.

Uberdude wrote: From the garbled machine translations I get the impression Ke Jie's approach of 11 at the top left was a bad move (and white 12 hanging connection a good one) and gave him a bad game, but that he may have played this move based on rumours that AlphaGo likes this move, but then Aja says this rumour was incorrect information. Can anyone fluent in Chinese shed some light on this please?

$$Bm11 Ke Jie (black) vs Huang Yunsong
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . 2 . |
$$ | . . . . . . . . . . . . . O . X O . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 1 . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+

Click Here To Show Diagram Code

[go]$$Bm11 Ke Jie (black) vs Huang Yunsong
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . 2 . |
$$ | . . . . . . . . . . . . . O . X O . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 1 . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

To me, it's not so obvious that is a bad move. Even less obvious is how white somehow punishes black and leads to a superior result for white. I think the game is still playable for both here. certainly isn't as crazy a move as the moves John was referring to.

Monkey wrote: To me, it's not so obvious that is a bad move.

And it was not obvious to Ke Jie either The article did not say if they reached a definitive conclusion about this . But even if it was unfavorable for black, I would imagine it was not "obvious" judging by a) Ke Jie played it, and b) they had a lengthy discussion afterwards.

Thanks captain english.