What is the best modern go game? (or one of the best)

[ermartin86]

Hi, I'm new here, so hopefully I'm not asking for too much with this question. I'm a beginner player, probably 16k in my prime on IGS. I'm looking for a game to use as a exemplary game to hopefully one day mold a computer go engine around.

Most games have at least a few moves that are considered "questionable" by some commentator or another. A game without any questionable moves, or a game with the least serious questionable moves by top players might be ideal for what I'm looking for.

Even a game where there are multiple pros on each side consulting before each move could be ideal. By modern I'm thinking within the past 5-10 years. The ideal game would also end very closely, perhaps a .5 or 1.5 win, and would go to the end without a resignation. Another thing that would be nice, for consistencies sake, would be if there were no odd or outstanding moves in the game, but of course, I'm still hoping for an astounding game.

In trying to find this game on my own, the problem I find is that the top games that pop up are really interesting, like the Lee Sedol - Hong Chang Sik - ladder game, but those games aren't necessarily great for creating an ideal program, because they might contain rare or unique situations.

Any thoughts are appreciated. Thank you so much!

-Eric

[Uberdude]

When you say mold (is that American for mould?) a go engine around, do you mean training a neural net type approach (as some recent research has been done at Edinburgh, Toronto and Deepmind) or coding an expert system approach (like ManyFaces pre monte-carlo addons or GnuGo)? If the former I would expect 50,000 professional games with small professional-level mistakes would give far superior results to just one single flawless professional game. And the later seems rather like a dead end in Go AI these days.

[John Fairbairn]

Here's a candidate for the perfect game. Shusai said it was hard to find any mistakes in Black's play, and indeed Black forced a very, very distinguished player to resign after just 80 moves without killing anything.

[Uberdude]

The OP did ask for a modern game (from the last 5-10 years) rather than over 200 years old, but perhaps finding a perfect game what with the generally shorter time limits nowadays will be rather harder. If mid-20th century is better I remember in JF's Kamakura book one of those games (maybe 4 or 5) was described as a masterpiece with practically no mistakes by Go Seigen, though Kitani may have made a few.

[Knotwilg]

Let's look at the marginal gain here. Your purpose is to make a go engine. It will have to deal with all sorts of moves by all sorts of opponents. It's already questionable that you'd be able to draw such power from one game, because the moves in such a game are dealing with a specific situation, though backed up by general knowledge. The added value of that game being truly perfect by the standards of the current professional thinking, rather than containing a few suboptimal moves, would in that respect be very marginal indeed.

[leichtloeslich]

the idea of gaining significant playing strength from looking at a single pro game is an IGS 16k idea indeed, no offense.
As Uberdude said, if anything you want a huge database to be able to extract common shapes and patterns.

those games aren't necessarily great for creating an ideal program, because they might contain rare or unique situations.

I have the feeling that you're vastly underestimating the complexity of go.
But hey, if you just want to see examples of perfect play, you can look at tsumego, local endgame sequences, or simply go on smaller boards (you can solve 3x3 go with pen and paper and many slightly larger boards have already been solved by bruteforce).
That's what I found so compelling about tsumego when I was starting out. The search space is so small that even a beginner can determine perfect play.

I still don't think this will help you in creating a decent go AI, but good luck anyway.

[EdLee]

The ideal game would also end very closely, perhaps a .5 or 1.5 win

Hi Eric,

• There are (many?) examples of pro games where one side "dominated" the game throughout (as agreed upon by qualified pros), and that the other side "had no chance at all," but the final score difference was 0.5 or 1.5.
  You are associating, incorrectly, the final score difference with: the level of the game; the level difference between the two players; the actual "closeness" of the game moves -- none of which is reflected by the final score difference.
• The opposite scenario: there are (many?) examples of nail-bitingly close games with spectacular fights and huge dead groups where neither side gave an inch, and the final score was a 50-point+ difference (or Resign); and afterwards it was still difficult for top pros to find obvious mistakes.
  ( I wonder if the famous game where Fujisawa sensei spent almost three hours on the key move to kill was such a game. )

As others have noted, some beginners have some wrong perceptions about Go: for example, the meaning of a small score difference.

Remember that two 20k players can have a 0.5 game.

To assess the actual level of a game requires qualified people to study each move.

The final score difference does not (necessarily) mean what you think it means.

Go is very hard.

[ermartin86]

Thank you for all of these responses. Definitely a lot to think about. I'll see what happens.

[ermartin86]

Oh, and Uberdude, I'm thinking of a monte-carlo based engine. Also, really any high-level game where neither side made obvious mistakes, and no resignation, would be nice.

[Uberdude]

How would you use one perfect game to help build a monte-carlo engine? The only thing I could see it being useful for would be trying to create the rudimentary local move selector for the playout logic (typical bots now use things like 3x3 shape squares with patterns like connect a one point jump against a peep or break through a keima) but I don't see how a perfect pro game is better than a pro game or games for that, and anyway adding too much complication to that move selector apparently tends to decrease the bot's strength (a combination of taking longer so fewer playouts, and reducing creativity). Or do you want to use the perfect game and play the bot as one colour in it and see if it recreates the game? There are too many points in a game where there are several perfectly good moves that that won't work.

[ermartin86]

Hi Uberdude. My goal is to create an engine that will play both sides, and through a monte carlo algorithm which I tweak, it will play through the game exactly as the "ideal" game is played. It's an ideal, and I imagine that it will be very hard to achieve. I'm not going into particulars, because I feel I may find some unique information along the way. To actually do this would be a huge achievement for me... if I could even get just the first few moves the same as the ideal game, I'd feel like I accomplished a lot. In fact, just an engine that played by the rules would be an accomplishment for me at this point. Thanks so much!

[EdLee]

Hi Eric,

if I could even get just the first few moves the same as the ideal game

"even" ? But nobody knows what the first few moves should be.
This, after a few thousand years of study.

[ermartin86]

Hi EdLee. Yes, I know. What I mean is an "ideal" game. Not truly ideal, but a high level modern game with no obvious mistakes.

[xed_over]

In fact, just an engine that played by the rules would be an accomplishment for me at this point.

start small... work you're way up.

[ermartin86]

start small... work you're way up.

Sounds good. Thanks