Mysterious Tartrate Conquers All At Go

Rubyflame writes "As noted on the Sensei's Library resource for the ancient Chinese boardgame Go, Tartrate, a very strong and mysterious Go player, has recently returned to the Kiseido Go Server (KGS) after a long absence. The game records can be found here. Tartrate first appeared in March, and has yet to be defeated - his identity is unknown." This intriguing story is a little reminiscent of Bobby Fischer's online chess appearances - the Go players on KGS even log their Tartrate number: "tartrate has a tartrate number of 0. If you have played a game with tartrate, your tartrate number is 1. If you have played a game with someone whose number is 1, your number is 2, and so on."

Comments

Good.

[LordJezo]

Pretty hardcore. Anyone here ever play him or see him?

heh

[Deternal]

Interesting to see a news about this actually.

Saw one of his games in june - didn't quite get the fuss then since I was a beginner, but I definitely get it today :)

Deep accurate reading seems easy for this guy.

Re:sgf

[Deternal]

you can download his games as SGF files and view them from the KGS archives if you want :)

Is that you, Sai?

[Firehawke]

At least he isn't going by the name "Sai".. that'd be downright scary. Still amusing that life mirrors art..

Tartrate is a jerk...

[clambake]

He uses cheat codes!

Re:Tartrate is a jerk...

[Alkaiser]

Go used to be so much more fun before they invented the Aimbot.

AI?

[torpor]

How feasible is it that its an AI being used to play tartrates games, anyone know?

I've seen some amazing Go games in my life (while I lived in Tokyo) and I know that the Go mojo is not something you're going to just up and code without being really, really good yourself ... but it is interesting that since we know nothing much about tartrate himself, the first thing that came to my mind is 'someone is running a good Go bot' ...

Not to detract from his skills, mind. I'm just interested if any of those who have played him could not have been defeated by some of the various Go-playing algorithms which are floating around out there. Some of them are too good.

Re:AI?

[Anonymous Coward]

He's beaten a 6-dan professional.

Re:AI?

[fstrauss]

Go has a rating system which briefly is explained as follows:

Begginers start at about 30kuy, as you get better your kuy rating decreases. 1kuy is better than 2kuy. Better that 1kuy is 1dan, dans count upwards to about 7dan. Better than that you start with pro ratings which are not easy to come by.

AI is far from beating pros at Go
The best go playing software is rated about 12kuy.

In otherwords, there are people in my local go club who would beat the best go playing ai :)

Re:AI?

[Anonymous Coward]

Why would you assume it is a bot, not knowing anything about the situation except that the person is very good? Considering how good AI is for Go, it's a bit like an anonymous writer releasing an incredible novel online, and you saying, "the first thing that comes to mind is somebody ran a good writing bot."

I don't think so. Humans ares till way, way, way better at plenty of things, Go and writing among them.

Re:AI?

[sgf]

'Too good'? The best go programs are still very weak compared to humans. These programs can be defeated by not particularly strong amateurs (unless Tokyo is keeping something secret from the rest of the world).

The relatively simple search techniques used in chess can't be applied to go, as the number of possible moves makes the space too big, so it may stay like this for some time (although some novel ideas are being tried). Tools like online joseki dictionaries could be useful (at least for an amateur), but that wouldn't help him with any of his reading.

My guess is a pro. They're just scary.

Re:AI?

[Infinite93]

I would be very suprised if it was an AI. Last time I looked (I will admit it has been while), the best AI out there was comparable to an average amateur player. There was even a $1M prize until Y2000 (donor died) for any AI that could beat a Taiwanese professional player. Of course, it could be a very clever way of performing a kind of 'turing test' for someone's next generation AI.

Re:AI?

[torpor]

Why would you assume it is a bot, not knowing anything about the situation except that the person is very good?

I'm not assuming anything - I'm just curious as to whether it could have been an AI, and as you can see other /.'ers have pitched in with the idea that it is not a very feasible idea, which is as interesting to me as the fact of tartrates success.

I have a lot of respect for good Go players, and tartrate is one of them. Just because I ask if it could be an AI doesn't mean that I think he is ...

Re:AI?

[psavo]

On the article page there was a notion that he did lose some games, but they were in a simultaneous play situation (eg. one master against several younglings at a time). His single-play account is still unbeat.

AI's don't lose in that kind of situations unless there's severe CPU resource starvation. Anyhow, if it's some AI, then it almost has to be some undergraduate/researcher with some new algorithm, so they probably have quite enough CPU power at hand.

Re:AI?

[cgenman]

Anyhow, if it's some AI, then it almost has to be some undergraduate/researcher with some new algorithm, so they probably have quite enough CPU power at hand.

I KNEW those SETI @ Home guys were up to something!

Re:AI?

[edbarrett]

'someone is running a good Go bot'

Is he a mighty Go bot?

Is he a mighty vehicle?

Am I an idiot?

Re:AI?

[eoyount]

The word is kyu, not kuy.

Re:AI?

[xenocide2]

If its a computer program, then it won't be detracting from anything. Go is a notoriously difficult game to write an AI for that beats people. If this is an algorithm, the author has a cunning way of proving its strength, and can expect a long and wealthy life.

Re:AI?

[spottedkangaroo]

More than that. gnugo supposedly plays around 10kyu, but I (18kyu) can beat it by exploiting it's weakneses. Because it always plays the same.

Tartrate beats all sorts of opponents and never loses. I don't think we have computers that good yet. "They say it will be 100 years ..." -- Hikaru No Go.

Re:AI?

[rodentia]

An average amateur player with a year or two of experience can beat any go AI.

The 15-12k rating of Many Faces and others is highly suspect. A few games against the machine and you can see how to beat it. Keep many open positions and don't pursue local conflicts. It is very easy to maintain sente against any of the programs. Against anyone with knowledge of the machines' style, it rates closer to 24-20 kyu.

Re:AI?

[scrytch]

> The best go playing software is rated about 12kuy.

Manyfaces is 6 kyu according to Nihon Kiin, and Handtalk is 3 kyu. Yet Handtalk gets handily beaten by amateurs with a 25 stone handicap. I suspect it's that the programs are a good deal more rigid than any decent human player and their idiosyncracies are well known. Kasparov would almost certainly beat the pants off any of the Deep * series if he played against them over and over (yes, Deep Fritz plays at grandmaster, but chess _is_ an easier problem)

Re:AI?

[curtisk]

>>Am I an idiot?

For quoting Go-Bots cartoon, yes! :D

Cheap Transformers knock-offs! hurmmmmph!

Re:AI?

[SamSim]

Out of interest, has anyone tried not directly coding Go AI but breeding genetic algorithms instead? Given the current shortfalls of Go AI it seems an attractive alternative to me.

Re:AI?

[Discoflamingo13]

What's really sad is how some people will assume that you're a troll, just because they're too young to remember how much Go-bots sucked when compared to Transformers. No as much as the "power of living rock" guys, but still stucked quite a bit.

And just in case somebody thinks this is off-topic, remember that Japanese robots have two eyes and are considered "living shapes" for the purpose of territory captured.

Re:AI?

[Anm]

I've seen some amazing Go games in my life (while I lived in Tokyo) and I know that the Go mojo is not something you're going to just up and code without being really, really good yourself

Not necessarily. Pick of a copy of Blondie24: Playing at the Edge of AI (ISBN: 058-3743638-9346720). It details how a couple of grad student wrote a genetically design neural network to play very good checkers online. Not only did the programmer not know how to play good checkers, but they were very careful to not design hints into the system.

Now, checkers is a lot simpler than go, but the possibility that it could be done is not impossible. The size of the board the number of possible and number of moves per turn would grow the problem significantly, but the students in the book worked off a single PII 400 throughout their entire project. The design detailed in the book would be very easy to distribute (the neural net evaluate each possible board position at the next turn, multiple machines could evaluate multiple boards in parallel).

Anm

Re:AI?

[pontifier]

In other news: The Earth Simulator has been renamed the Go Simulator for unknown reasons...

Re:AI?

[Anonymous Coward]

Not like anyone will read this, but Go-Bots came out before Transformers. Transformers just had better marketing.

Re:AI?

[AndyMouse GoHard]

No offense but I would assume all AI tricks currently know are being tried. These problems tend to obsess programmers so chances are very good that alot of things are being tried.

Bill

Re:AI?

[Anonymous Coward]

My point was that it makes no sense at all for "bot" to be the first thought in this situation, jsut as it would make no sense if an incredible novel were published anonymously for one's first thought to be "bot."

Re:AI?

[cgenman]

Right, but a sufficiently large search space will not be reached in a human time frame. You are talking about the number of open spaces left on a board with 361 possibilities for the first move, 129,960 possibilities for the second move, 46 million possibilities for the third move, etc. By move ten you have 33 septillion possibilities, and you still have no easy way to gauge whether or not the move was a good one. By the time each player has made 15 moves, or a solid start to a game, your search space is 20 google, and that's assuming no taking has taken place. Yes, you can do things with rotation and identical board positions, but 5 Google is still a lot for a 486. If you want the whole possibilities, trimmed down to only the best moves a computer can make at any given point, that's still a space of... little TI-92 programming here... 4*10^384 possible trees.

A Neural Net of Go would suffer from a similar problem of scale... You have 19x19 locations, with 3 possibilities at each. That's a learning space of 17 sexquinquagintillion, or 17 octovigintilliard for our british friends. We could again divide by 4 for each possible rotation, but 4*10^171 is still pretty big. Assuming each board discovered and knew the one "best" next move to make, the storage required would be enormous. If there are 10^81 atoms in the universe (a high estimate), and one were to further ludicrously assume that there are as many universes as there are atoms in this one, each QUARK in every atom would have to store a bit of data to have a pre-stored list of the best next moves. That's not including how much space would be required to store all of the associated failure rates with the other positions, let alone a system capable of reading it all. Learning computers require a lot more RAM than a straight programmed one. And unlike image recognition, blurry go boards just won't cut it.

Genetic programs and neural nets are great at some things, but certain problems don't play out so easily. Any program that will play well at Go will have to have some extremely high-level thinking, of which we are not capable of producing or breeding today. Otherwise, the sample space just falls apart.

Re:AI?

[Anm]

Ugh... I can't believe people modded you up on that BS. Do you realize you arguments apply against human understanding of Go, let alone an AI's understanding of checkers (both proven to work)? Stop thinking anyone is going to consider and memoirze every possible move. That is why procedural programming does not work.

Instead, you start with a random sampling and just see how well they perform. Obviously they start bad, but the mutation process pushes the better perfromers forward.

The neural net is just a convient evaluation function that is both extremely flexible and is fairly straight forward to serialize in a genome.

Additionally, the feed forward neural nets used in the book don't store anything about future board states, let alone "a pre-stored list of the best next moves". The neural net just computes a single number, which an higher architecture inteprets as how good a board position is. Compare the evaluation numbers of a few board positions, and board after them in an A* fashion as time permits. You should be able to compare a few tens-of thousands of _significant_ board possibilities on a single processor (number out of my ass) in a minute of turn consideration. Memory is only needed to store which turn lead to what evaluation number (and maybe last calculate board position to compute possible following turns). And as I said before, the architecture is VERY easy to parallelize.

The problem here is the neural nets don't encounter a variety of end-game scenerios, and thus their performance in the end-game is critical. The book details one game where the AI that trained for 6 months lost in a series of very stupid end-game mistakes.

Anm

Re:AI?

[cgenman]

I believe my arguments apply to any understanding of how to program a machine to program itself to understand GO. Obviously a machine will not be able to understand every position. The quality of a position in go can be radically different based upon a single stone. And how is that quality to be quantified and judged, without interjecting a more than superficial understanding of the game? Through evolution? You're going to be spending a lot of time evolving there, to some equally ludicrous number.

I'm sorry, but living near MIT, and having worked with neural nets, there is a point where each fresh face puts their finger up and says either "Neural nets will solve everything" or "evolutionary programming will solve everything." Sorry, they have limitations.

And is it any surprise that my comment was modded interesting? Slashdot moderators always fall for really big numbers.

Re:AI?

[Anonymous Coward]

The correct spelling of 10^100 is googol (Surprise, google is not a real word)

Re:AI?

[tunah]

By the time each player has made 15 moves, or a solid start to a game, your search space is 20 google, and that's assuming no taking has taken place.

That's googol... a word I'm sure is disappearing from the language as we speak ;)

Re:AI?

[Anm]

I'm sorry, but living near MIT...

Wow, you must get a lot of job offers with that on your resume!!

And how is that quality to be quantified and judged, without interjecting a more than superficial understanding of the game? Through evolution? You're going to be spending a lot of time evolving there, to some equally ludicrous number.

The "understanding" of the game is broken into two parts. First, the legal sequence of board states, and thus turn-based nature of the game is written into the state selection mechanism (hence before any neural net evaluation) of the A* algorithm. The more qualitative state of a given board are what the neural net "understands", albeit not in any easily extractable form. The layered nature of a feed-forward neural net explicitly allows for "radically different [evaluations] based upon a single stone". If it didn't, the whole architecture would not work since the board positions compared often only differ by one or two squares. It's worth noting that the structure of the neural net has a lot of influence on the possible "concepts" in the neural net. In the blondie24 example, they explicitly wired nodes for every 5x5, 4x4, 3x3, and 2x2 square to allow the system to recognize spatial concepts such as corners and adjacency.

As for time evolving: Yes, it does take a while. Blondie24 took six months (24/7 on a single CPU). But in six month, it evolved the skill to occassionally beat master checkers player. It's a lot of time, but it is definitely possible. Go would take even more CPU time, but machines are cheap.

"Neural nets will solve everything" or "evolutionary programming will solve everything." Sorry, they have limitations.

I never said either, and don't believe it either. I just made reference to an existing successful example in a strongly related field. But I do believe similar problems (exceeding wide yet finite search space with definitive win conditions) problems are well matched to the GA+NN solution. Genetic algorthims simulate searching the state space and maximize the use of partial solutions, while neural nets provide quick state evaluation function that is capable of generalizing qualities of between states.

And is it any surprise that my comment was modded interesting? Slashdot moderators always fall for really big numbers.

Touche.

Re:AI?

[Anonymous Coward]

Your sig is a soul gouging mutilation of Simpsons quotation.

Re:AI?

[Anonymous Coward]

Not like anyone will read this, but Go-Bots came out before Transformers. Transformers just had better marketing.

And better animation and better characters

Re:AI?

[MilenCent]

Worse, they fell for incorrectly spelled big numbers. The word for the one-followed-by-100-zeros number is googol.

I'm not trying to cut ya down though, I would have made the same mistake myself a few short months ago. I agree with you more than your derisive sparring partner, go is an AI problem of a far greater magnitude than chess. There are many reasons for this, but the number of possible moves is definitely one of them.

Re:AI?

[millette]

you mean with genetic programming or with neural networks ?

Nothing a quick search engine couldn't solve...

Re:AI?

[Anonymous Coward]

Instead, you start with a random sampling...
OK so far...

and just see how well they perform.
ooooh, that's the hard part.

Sai?

[fstrauss]

This reminds me of Sai playing online via Hikaru in Hikaru No Go.

Re:Sai?

[Your_Mom]

Thank goodness.
I'm glad I wasn't the only one who had that first pop to mind when I read this.

I, for one...

[MyDixieWrecked]

Welcome our new Go overlord.

Tartrate number == Shuusaku number

[cthulhubob]

Before Tartrate got so famous (infamous?) amateur go players were keeping track of how many games they were away from having played Honinbo Shuusaku of the Edo period.

My Shuusaku number is 5 -- Shusaku (0) - Iwasaki Kenzo (1) - Honinbo Shusai (2) - Iwamoto Kaoru (3) - James Kerwin (4) - Ethan Baldridge (5).

One of the coolest games on the KGS archives is Tartrate vs. Redrose (Irina Shikshina, a Russian woman who is a 1st dan Korean professional). Tartrate was black and played his first move on tengen (the center of the board), which is an unusual opening. There were two ENORMOUS ko fights, and everybody thought Redrose had won after the first one was over. Check it out, it's a great game.

If anybody wants a Shuusaku number of 6 and/or a Tartrate number of 3, my username is ethanb on both KGS and DGS (kgs.kiseido.com, and www.dragongoserver.net).

Re:Tartrate number == Shuusaku number

[funkhauser]

Yes! I was on KGS when the redrose vs. tartrate was going on. There was a big ruckus in the English room about it, so I started watching about 30 moves in. What a game! Of course, it only added to the mystique of tartrate. :) I'm not on KGS very often, but my user name is tetsujin23. I'm 17k? but I think I'm closer to 15k. Perhaps we should play sometime.

Japanese, not Chinese

[kenthu]

I've always thought go was Japanese, not Chinese. m-w.com seems to agree with me:

a Japanese game played between two players who alternately place black and white stones on a board checkered by 19 vertical lines and 19 horizontal lines in an attempt to enclose the larger area on the board

Re:Japanese, not Chinese

[kenthu]

On the other hand, another page says go got started in China:

One of the oldest strategy game in existence is the game called GO. It came to existence over 3000 years ago in China where it was given the name, "Wei-chi".

Eh. Never know who you can trust on the internet.

Re:Japanese, not Chinese

[cthulhubob]

Go was brought to the west via contact with Japan - that's why it's called "Go" here. The game is known as "Igo" in Japan, "Wei qi" in China, and "Paduk" in Korea. The technical terms used in the west are also all Japanese terms (most amateur go players in the US will know what "miai", "hane", "tengen", "joseki", and "aji" mean, for example), even though China and Korea have their own equivalents.

Evidence shows that go was originally brought to Japan via Buddhist monks from China though. Evidence of go in China predates written records, so it's not certain whether it originated there or was brought from elsewhere.

Re:Japanese, not Chinese

[Deternal]

It is true that Go is often times mistaken for a Japaneese game.

It is also likely that it wouldn't have proliferated as far into the west as it have today if it wasn't for the Japaneese interest in the game.

The game is between 2000 and 4000 years old and stems from China. The first written sources on the games history stem from about 500 bc wherein among others Konfutse wrote about the game.

Konfutse did not believe the game helped anything, whereas the Taoists believed that it was a means to contemplation and selfunderstanding.

In the T'ang dynasty (618-906) the game is recognized as something which should belong to common knowledge.

About the year 700 the game comes to Japan. Where it later would be deemed as part of necessary training for samurais. Thru Go, they thought, warriors could practice tactical and strategic training which could be used in real life battles.

On a more domestic (for me) note, the first Go club in Denmark was established in 1970 :)

Re:Japanese, not Chinese

[BeProf]

Another reason that Go is known as a Japanese game is the fact that the standard rule set we use today was developed in Japan during the Tokugawa period.

I'm not certain, but I believe that what was added in Japan was the handicapping system and the half-point komi to prevent ties.
In Korea, for instance, I know that the work 'baduk' is used to refer to Japanese-rules Go, but another word is used to refer to 'old' Go.
I don't know about China.

Re:Japanese, not Chinese

[stonecypher]

Try not looking for it under our bastardization of its Japanese name - Wei Qi, Padhooq, et cetera. The game has a history lost in time, much like that of Chess, but it's probably a Chinese invention for attempting to provide strategy lessons for military generals.

A hint: m-w.com is not the authority on Asian culture.

Tartrates' Identity Revealed! That's right...

[Braintrust]

... you guessed it!

Frank Stallone.

Re:Tartrates' Identity Revealed! That's right...

[SuperMo0]

Noooo! I heard from someone that it was Anna Nicole Smith. >_>

I'm unimpressed - and I'd like to qualify that.

[TerryAtWork]

This guy is nothing but some 9 Dan pro slumming. No big deal. The REAL big deal would be if it was a program. That would be UBER-REVOLUTIONARY as programs famously suck at Go.

To cover my ass though - a 9 Dan pro is God Almighty at Go. I will never beat one. I saw a 9 Dan pro play a 6 Dan amateur on a Go server. He spotted the amateur 9 stones and was behind all the way to the end where he pulled ahead and just beat the guy by 3 stones. He knew all along what he was doing. It was slick as hell.

Here's the kicker though - while he was doing this he was also playing another guy at the same time. That's right - he was playing two games at ther same time and he STILL beat a 6 Dan amateur with a 9 stone handicap. Amazing.

Re:I'm unimpressed - and I'd like to qualify that.

[atheist666]

Pros are strong, even at the lower levels. And the computer programs kind of suck.

I think there was one exhibition where Janice Kim, then 1-dan Korean pro, spotted GnuGo (3.2 or thereabouts) something like 25 stones. That's 25 stone handicap. She pulled off the win in the end.

Anyhow, Tartrate is probably Jie Li.

Nominations

[scrytch]

Mysterious Tartrate Conquers All At Go

I nominate this article title for "Most Surreal Slashdot Title Ever".

six

[Scrameustache]

Kevin Bacon's Tartrate number is, of course, 6.

: )

Pros are ALL God Almighty

[cthulhubob]

You should attend a workshop taught by Yang Yilun. He's a 7-dan Chinese pro who teaches in the U.S. Usually the workshops run all weekend for about $200-250. He is an excellent teacher and has written several books, including one coming out soon from Slate and Shell.

The most impressive thing I've ever seen is at the one workshop I've been to by him. He took all of the students (eighteen), divided us up into pairs (so they can discuss moves with one another), and played us all simultaneously. Then after beating us all (even the pair composed of Keith Arnold, 5 dan and Eagles Song, 4 dan) we cleared off the boards, then he sat down with the first pair, replayed their game from memory, and commented on what they could have done better. Then he replayed the second game from memory... and kept going all the way around the circle.

He's got another workshop coming up in June, I believe. It's in New Jersey. I'm definitely making the trek.

Re:Pros are ALL God Almighty

[Anonymous Coward]

I wonder if he plays online.

BUUWAHAHAHAHAAAAHAA!!!

[Anonymous Coward]

(not funny) ...wanker

A good book about Go

[Laplace]

Read "The Master of Go" by Yasunari Kawabata. It's an amazing novel about not only the culture and play of Go, but also the rise of Western ideals after the fall of Imperial Japan. It won the Nobel Prize for Literature. Read it; you'll like it.

Re:A good book about Go

[Anonymous Coward]

Read "The Master of Go" by Yasunari Kawabata. It's an amazing novel about not only the culture and play of Go, but also the rise of Western ideals after the fall of Imperial Japan. It won the Nobel Prize for Literature. Read it; you'll like it.

While it is an excellent book, it did not win the Nobel Prize. Kawabata did win the prize, but had many other brilliant works besides this one.

CthulhuBob.

[Anonymous Coward]

Off the "Go" topic. I just had to let you know that your handle is cool. Because I play a character in Call of Cthulhu named Bob Cthulhuslayer. That is all.

Re:sgf

[millette]

yes, you can download all his kgs games there.