AI can now win at Go

[littlebabyjesus]

Crafty said Kc3 after less than a minute.

Computers have solved that particular problem then.

But it took me much less than a minute. First I worked out that black couldn't break through, then I worked out that black would gobble up the pawn that takes the rook soon enough, my defences would be destroyed and I'd be totally buggered.

I'm an ok chess player, but not a properly strong one. I'd expect a stronger player to work that out very quickly - once you see that taking the rook is stupid, the only other option is to take the draw. Taking the rook is only not stupid if you can promote the pawn, which you clearly can't.

Be interested to know what kinds of things the likes of Crafty have done to improve their performance since this puzzle was formulated. How much analysis of human input was needed? As the OP says, one of the biggest challenges is to find ways to process the information and narrow down your analysis, and the go machine uses human input for that in the shape of past games. How much of that is a 'cheat' wrt evidence of intelligent thinking, or indeed 'thinking' full stop? Knowing when to stop is also a massive thing - if Crafty took a large fraction of a minute, it no doubt looked down many different lines before making its decision, many more lines than a human player would need.

 

[two sheds]

I'd sort of assumed that chess would be the more difficult to beat because Go has simpler basic rules but huge number of board positions so easier to calculate just by throwing huge processing power at it.

Rethinking, though, I suppose that just saving to disk all the known games by masters and tacking on a basic end game strategy would be enough to beat nearly all chess players.

 

[Signal 11]

I don't know a great deal about it but I don't think the changes in the engines will have made much difference compared to the advances in hardware. For chess you can use a fairly simple evaluation function and just calculate deep enough to see which is the best line, so it's more like brute force than AI.

A quick read of Computer Go - Wikipedia, the free encyclopedia suggests it's too complex to either do the evaluation well enough or calculate enough variations, so they have to use some proper AI as well.

 

[Bonfirelight]

How is it at tic tac toe?

 

[littlebabyjesus]

How is it at tic tac toe?

It will never beat me.

 

[spanglechick]

I keep reading this thread as "Al can now win Go" as in Al, short for Albert, Alan or Alasdair.

 

[littlebabyjesus]

I don't know a great deal about it but I don't think the changes in the engines will have made much difference compared to the advances in hardware. For chess you can use a fairly simple evaluation function and just calculate deep enough to see which is the best line, so it's more like brute force than AI.

A quick read of Computer Go - Wikipedia, the free encyclopedia suggests it's too complex to either do the evaluation well enough or calculate enough variations, so they have to use some proper AI as well.

Interesting reading that and about chess programming. If you can create a machine that develops its own heuristics through repeatedly playing itself or others, then you can develop a machine that can potentially give you insights.That's close to something impressive in terms of AI.

Takes some of the romance away. Tactics, intimidating, the idea of being daring, all go. Reading about the last major computer–human chess match, the human drew four and lost two. He lost the first through a massive blunder, not spotting mate next move, and lost the second taking risks to try to level the match. Computer was handicapped, though. Humans can't beat computers at chess any more. Machine wasn't machine-learning in the same way as this one, though.

 

[DotCommunist]

I doubt a computer can win monopoly from the left yet though

I sort of like the idea that we can make machines better at games than us. It provides a challenge, and as you said about kasparov it will provide a learning curve. I can imagine how furious a proper master of the game was to lose to a machine and then applying human ingenuity to see a way of outsmarting it. Who is training who
I'm still on team lee tho

 

[littlebabyjesus]

I doubt a computer can win monopoly from the left yet though

I sort of like the idea that we can make machines better at games than us. It provides a challenge, and as you said about kasparov it will provide a learning curve. I can imagine how furious a proper master of the game was to lose to a machine and then applying human ingenuity to see a way of outsmarting it. Who is training who
I'm still on team lee tho

tbh I don't doubt that an unbeatable go computer will emerge from this at some point. It took a decade from the Kasparov match, which K might have won, to computers becoming effectively unbeatable, and not just through increasing brute force - Deep Blue, which beat Kasparov, used brute force in a way that isn't used now.

I don't know what I think about it. Chess computers play dull chess, and I bet go computers play dull go. I might be wrong, of course. We don't know exactly what will come out the other end of neural network- style learning.

 

[beesonthewhatnow]

I don't know what I think about it. Chess computers play dull chess, and I bet go computers play dull go. I might be wrong, of course. We don't know exactly what will come out the other end of neural network- style learning.

The Sadness and Beauty of Watching Google’s AI Play Go

 

[littlebabyjesus]

The Sadness and Beauty of Watching Google’s AI Play Go

Hard not to like that tbh. This kind of learning-by-playing-against-yourself does seem to be a game-changer.

 

[weltweit]

Lets hope future AIs are more in the mould of Iain M Banks ideas, and less Terminator type.

 

[discokermit]

Rethinking, though, I suppose that just saving to disk all the known games by masters and tacking on a basic end game strategy would be enough to beat nearly all chess players.

no it wouldn't. it doesn't take long before your game is unique and would leave your computer on its own with it's endgame theory somewhere in the early middlegame.
they do have comprehensive opening books built in but they get through the rest of the game with brute force computing power.

 

[two sheds]

Yes probably - I was thinking that there are so many traps and it doesn't take long to be in a really poor position if your opponent has good opening technique.

 

[tommers]

How is it at tic tac toe?

The only way to win is not to play the game

 

[NoXion]

How is it at tic tac toe?

It's a trivially solvable game. Perfect play is possible for humans.

 

[JimW]

When it can beat Spatsky at the World Staring Championships we'll know AI has really arrived.

 

[Supine]

That's a good question - the puzzle was devised specifically to trip the computer up - by Bill Hartson (who presented the Chess show on the bbc in the 80s, if you remember that).

And now a regular on Gogglebox!