Brilliant Move from AI

illluck · Post by **illluck** » Fri Dec 30, 2016 7:32 pm

So, these past few days has seen the emergence of a number of pretty strong bots on Tygem and Wild Fox. One (presumably Korean) AI "Master" has went on an undefeated streak against strong and top pros (please refer to viewtopic.php?f=10&t=13913 for further info).

At the same time, two other Chinese AIs (as well as Zen) also started playing online. One of them (Xing Tian :b1:

) has a slightly less impressive record compared to Master (it went 1-1 against Gu Li, 1-1 against Ke Jie, and 4-1 against Park Junghwan), but is also undefeated at the 20 second byoyomi timing (which is the time setting Master uses).

This is a snapshot from its second game against Ke Jie. It is black and apparently was in a difficult position until it played a brilliant sequence that Ke Jie missed.

Solution and SGF in comment below.

:b1:

Named after a Chinese deity who waged war against the Supreme Divinity, not giving up even after he was decapitated

Edit: slight wording modification ("slightly impressive" to "slightly less impressive") and moved explanation about Xing Tian to a note.

illluck · Post by **illluck** » Fri Dec 30, 2016 7:39 pm

Solution:

SGF:

Edit: Re-uploaded solution image.

ez4u · Post by **ez4u** » Sun Jan 01, 2017 2:22 am

The upper right is an interesting position. It does not look so unusual to me. However, when I checked my database (currently just over 100,000 games) I found the following.

This corner position is quite normal of course. Out of 3,090 cases, Black continues with either 'a' or 'b' 3,006 times.

Click Here To Show Diagram Code: [go]$$Wc Search Pattern $$- - - - - - - + $$. . . . . . . | $$. . . . . . . | $$. . . a 3 . . | $$. 2 . X 4 5 . | $$. . . . . b . | $$. . . . 1 . . | $$. . . . . . . |[/go]

The other thing that happens occasionally (49 cases) is for White to get in 1 below after a Black tenuki as in the game. The position after Black jumps to 2 appears 6 times in my database.

Click Here To Show Diagram Code: [go]$$Wc White connects, Black jumps $$- - - - - - - + $$. . . . . . . | $$. . . 2 . 1 . | $$. . . . O . . | $$. X . X X O . | $$. . . . . . . | $$. . . . O . . | $$. . . . . . . |[/go]

However, Ke Jie's exchange of 1 for 2 below does not appear at all. Needless to say then, the rest of the development in the game from White's jump to 3 to the connection of 7 sets up a new situation.

Click Here To Show Diagram Code: [go]$$Wc Game continuation, setting up the tesuji. $$- - - - - - - + $$. . . . . . . | $$. . . B 1 O . | $$. . . 2 O . . | $$. X . X X O . | $$. . . 6 . . . | $$. . 3 4 O . . | $$. . 7 5 . . . |[/go]

We can go a bit further and tighten up the situation to the bare essentials. The following corner situation does not appear. I think that it makes the 1, 3, 5 combination very impressive indeed. It is certainly not anything directly "learned" from any training set.

Click Here To Show Diagram Code: [go]$$Bc Search Pattern $$- - - - + $$. . 5 . | $$X O O . | $$X O . . | $$X X O . | $$X 2 1 . | $$X O 3 . | $$O 4 . . |[/go]

Mike Novack · Post by **Mike Novack** » Sun Jan 01, 2017 7:17 am

ez4u wrote: impressive indeed. It is certainly not anything directly "learned" from any training set.

Yes, that's what training a neural net does.

You can think of a neural net as a function and training a neural net the process of getting it to implement that function. The function, given a board positions as input, return the best move. It has been trained with a finite set of data (known/assumed "best move" from a large dataset of pro games under the assumption that most of the time the pro did make the best move).

If that was ALL neural nets could do, return the function results for the points where it had been trained, they would not be all that interesting. But however hard this might be to understand, the process of training/annealing << slightly disrupting >> trepeated over and over results in the neural net being able to INTERPOLATE correctly. In other words, probably to return a correct answer for points that are in some sense "in between" the points for which explicitly trained ( it will return SOME answer; I mean the chances that it will have been the RIGHT answer to return )

John Fairbairn · Post by **John Fairbairn** » Sun Jan 01, 2017 10:56 am

Not in any way belittling the program's achievement, I have a strong memory of seeing Xing Tian's tesuji before, and (much less strongly) equate it with Go Seigen. If the latter memory is correct, it will likely be in Gengen Gokyo or Guanzipu but as a correction by Go to the traditional solution. I haven't got the time to check, but someone else may remember it.

It would in any case make sense to me that neural net trainers would use problem collections as well as games, no?

illluck · Post by **illluck** » Sun Jan 01, 2017 11:36 am

John Fairbairn wrote:Not in any way belittling the program's achievement, I have a strong memory of seeing Xing Tian's tesuji before, and (much less strongly) equate it with Go Seigen. If the latter memory is correct, it will likely be in Gengen Gokyo or Guanzipu but as a correction by Go to the traditional solution. I haven't got the time to check, but someone else may remember it.

It would in any case make sense to me that neural net trainers would use problem collections as well as games, no?

That is an interesting point! Not aware of training using problem collections (might be a bit difficult in terms of getting local problems to whole-board), but it certainly would be a possible method!

denizen · Post by **denizen** » Sun Jan 01, 2017 11:38 am

John Fairbairn wrote:Not in any way belittling the program's achievement, I have a strong memory of seeing Xing Tian's tesuji before, and (much less strongly) equate it with Go Seigen. If the latter memory is correct, it will likely be in Gengen Gokyo or Guanzipu but as a correction by Go to the traditional solution. I haven't got the time to check, but someone else may remember it.

It would in any case make sense to me that neural net trainers would use problem collections as well as games, no?

I haven't heard any suggestion that AIs are trained on problem collections.^[1] I don't think it would be of much value. You'd have to build a separate process to recognize that "aha, I can use this particular tesuji here." I don't think any such processes have been used, except as to ladders.^[2] I don't think AI developers are cranking out tons of different processes to recognize different tesuji. I think you'd also need more input to make it work than just a single problem (maybe you'd have to train a neural network to recognize the pattern in which it works by putting in a bunch of examples? not sure).

Footnotes
[1] E.g. https://www.scribd.com/doc/302719734/AlphaGo-Paper at PDF page #2 ("We trained a 13-layer policy network, which we call the SL policy network, from 30 million positions from the KGS Go Server.").
[2] http://www.cs.toronto.edu/~cmaddis/pubs/deepgo.pdf at PDF page #3 (DeepMind: "Each position s_t was preprocessed into a set of 19×19 feature planes φ(s_t), that serve as input to the neural network. The features that we use come directly from the raw representation of the game rules (stones, liberties, captures, legality, turns since). In addition, we have one simple tactical feature representing a basic common pattern in Go known as ladders; in practice this adds a small performance benefit, but the results that we report would be qualitatively similar even without these features.")

Life In 19x19

Brilliant Move from AI

Brilliant Move from AI

Re: Brilliant Move from AI

Re: Brilliant Move from AI

Re: Brilliant Move from AI

Re: Brilliant Move from AI

Re: Brilliant Move from AI

Re: Brilliant Move from AI