AI plays go

http://www.bbc.com/news/technology-35420579

Yet another hurdle passed.

The post that was quoted here has been removed

It took the AI programmers a long time to encapsulate strategic thinking in a finite set of unambiguous instructions, but once they cracked the chess problem, it took relatively little time to crack the semantic problem (viz. the Jeopardy tv game show). It's only a matter of time now, and I suspect not ten years, before the AI community cracks go, too.

Well, there's still sumo wrestling...

The post that was quoted here has been removed

But it is a harbinger of things to come. Computers continue to get faster and software more powerful, it seems only a matter of time till a comp is 10th dan.

It looked like in the video the games were blitz. It may have had a different outcome if they were at normal go time scale which could be hours long.

I was in tech school in Denver and got very ill and was in hospital at an army hospital and a general there was in the bed next to me and he taught me to play Go. A really nice guy. Didn't seem like a general in attitude, just another soldier. He didn't condescend to me, a lowly airman, at all.

One game I experimented with is 3D go. I found if you had a real 3D Go board, 19X19X19, you need at least 150 mm between boards if you want to be able to get your fingers in between boards to place the stones.

That would be 19 stacked boards about 3 meters high, not realistic.

So the only way to realize such a game would be with computers where the resultant cube could be manipulated to any angle so you could view the board from any perspective and then just use a 3D co-ordinate system to make moves, maybe something like this: looking at the boards, left to right, A to S (19 lines) call that X direction, then 1-19 for Y then revert to letters for Z, so the lower lefthand corner as convention, would be A1A and upper right hand corner S19S and so forth.

The big question is whether you could have the same kind of territories as 2D Go. Like a combination of moves on each board to encircle a sphere shape or a cube and what would constitute the ability to take a piece, totally surrounding all lines leading away in three dimensions.

I think it would be a mind boggling game!

If Go is anything like Tae Kwon Do, you can't get 10th Dan until you're dead.

Originally posted by moonbus
It took the AI programmers a long time to encapsulate strategic thinking in a finite set of unambiguous instructions, but once they cracked the chess problem, it took relatively little time to crack the semantic problem (viz. the Jeopardy tv game show). It's only a matter of time now, and I suspect not ten years, before the AI community cracks go, too.

Well, there's still sumo wrestling...

Unfortunately for the Japanese, Hawaiians and Blacks seem to be better at Sumo than the Japanese are.

"So the only way to realize such a game would be with computers where the resultant cube could be manipulated to any angle so you could view the board from any perspective and then just use a 3D co-ordinate system to make moves, maybe something like this: looking at the boards, left to right, A to S (19 lines) call that X direction, then 1-19 for Y then revert to letters for Z, so the lower lefthand corner as convention, would be A1A and upper right hand corner S19S and so forth.

The big question is whether you could have the same kind of territories as 2D Go. Like a combination of moves on each board to encircle a sphere shape or a cube and what would constitute the ability to take a piece, totally surrounding all lines leading away in three dimensions. I think it would be a mind boggling game!"

Probably only a Romulan could master it.

Originally posted by moonbus
"So the only way to realize such a game would be with computers where the resultant cube could be manipulated to any angle so you could view the board from any perspective and then just use a 3D co-ordinate system to make moves, maybe something like this: looking at the boards, left to right, A to S (19 lines) call that X direction, then 1-19 for Y then reve ...[text shortened]... mensions. I think it would be a mind boggling game!"

Probably only a Romulan could master it.

My idea would be to allow a move to be made on each level so the game would be about the same length as a normal 2D version. But now you have lines going up one level and down one level besides the left and right and in and out of the 2 D board. I think you could enclosed volumes that way.

The post that was quoted here has been removed

From the BBC story linked above:

"DeepMind now intends to pit AlphaGo against Lee Sedol - the world's top Go player - in Seoul in March."

The post that was quoted here has been removed

What did it feel like playing a 9? Did he give you stones? Were you overwhelmed? What was the score at the end?
Was the time control fast, like blitz? What were you doing in Tokyo to meet such a star at Go? It would seem to me like me playing tennis with Djokovic or Federer🙂

Originally posted by moonbus
It took the AI programmers a long time to encapsulate strategic thinking in a finite set of unambiguous instructions, but once they cracked the chess problem, it took relatively little time to crack the semantic problem (viz. the Jeopardy tv game show). It's only a matter of time now, and I suspect not ten years, before the AI community cracks go, too.

Well, there's still sumo wrestling...

Don't try Sumo wrestling with a robot. The interesting thing about this is that they seem to have used a neural network for the position evaluator. So they seem to have a normal alpha-beta search with a neural net to assign a "how good" number to the leaf nodes. This makes me wonder as I know little about neural networks, but imagine them to take some time to reach a decision. What one needs with an alpha-beta search is a very fast final evaluation as huge numbers of leaf nodes need to be evaluated. They talk about a three stage process (learn from human position is stage 1, play itself is stage 2, then stage 3 it's competing) what they have done is use a neural network to learn what a good position is and then the neural net to train some sort of Bayesian filter to make a quick decision for the alpha-beta search.

The post that was quoted here has been removed

What is your level? I think I went all the way to ten Q🙂

How many stones did you get? He could have given me all 9 and it wouldn't have made the slightest difference🙂

Where did you learn the game? Were you in a city that had enough players for a club?

The only club I saw was an informal one at Bell Labs, when I was a field service engineer for Varian ion implanters, my field for 20 years. When Bell was open in it's full glory, there was a noon time go club and chess club upstairs from the cafeteria, playing both games with clocks.

Were you using a clock for your games? How many other folks were playing this 9?