Slashdot Mirror
Google's AlphaGo Beats Lee Se-dol In the First Match (theverge.com)
New submitter Fref writes with news from The Verge that "A huge milestone has just been reached in the field of artificial intelligence: AlphaGo, the program developed by Google's DeepMind unit, has defeated legendary Go player Lee Se-dol in the first of five historic matches being held in Seoul, South Korea. Lee resigned after about three and a half hours, with 28 minutes and 28 seconds remaining on his clock. "
Lee will face off against AlphaGo again tomorrow and on Saturday, Sunday, and Tuesday. Also at the New York Times. Science magazine says the loss may be less significant than it seems at first.
Lee will face off against AlphaGo again tomorrow and on Saturday, Sunday, and Tuesday. Also at the New York Times. Science magazine says the loss may be less significant than it seems at first.
And it will be no match for him at kickboxing.
It might he harder for Lee to beat a computer because he says that he relies heavily on reading his opponent. Unlike poker you can't just calculate odds on everything, and unlike chess there are too many permutations to plan right to the end of a game.
It really depends if he can find a way to figure the computer out without the usual cues he gets from human players.
const int one = 65536; (Silvermoon, Texture.cs)
SJW, n: "Someone I don't like, and by the way I'm a fuckwit" - AC
I respectfully disagree.
Diplomacy with it's self-references and multi-body perturbations is closer to a set of non-linear partial differential equations whose solution is extremely difficult and idiosyncratic and chaotic.
Ever had a disagreement with someone who is basing their behavior upon yours, who in turn is basing their behavior upon theirs? Now add as many as five more players, all intertwining their interactions with yours over time. We are not talking here about enumerating numerical solutions to nice set of equations, either. We are talking about inputs including revenge, boredom, capriciousness, contrariness.
Basically Chess and Go are nerfed versions of real world problems that humans have learned to deal with pretty well.
Even humans have significant problems with these cicular and self-referential domains, see R.D. Laing's book Knots, for example.
Since I doubt that most people unfamiliar with go have a way to view an sgf file here is link to the gogameguru article about the game. At the bottom there is a javascript applet you can use to play through the game.
https://gogameguru.com/alphago-defeats-lee-sedol-game-1/
Here's a talk by deepmind about this AI https://www.youtube.com/watch?...
AlphaGo did beat the European champion 5 out of 5. So Lee will have to step up his game a bit.
The European master was 2 Dan. Lee is 8 Dan (or 9, but the last one is honorary, so it doesn't count for comparing strengths). In the world of Go, that difference is rather staggering. It's like the difference between a chess ELO rating of 2100 and 2600. Someone consistently beating the lower ranked player may not have a chance against the higher rated player.
If Lee were to play the European master, he'd be expected to trounce him too. Quite thoroughly.
What counts in AlphaGo's favor here is that it has had over a year to improve. That it did win the first game says a lot more about its strength than any play against much lower ranked players.
In fact let me spell it out for people who've never paid any real attention to Poker.
At any moment, a single human player of (say) Texas Hold 'Em can see some of the state of the game, but not all of it, and, except for the final round of betting (the "river" round) there is still a random element which in most cases can be decisive.
It might seem as though a perfect player would calculate the odds that they've got the best hand, and bet accordingly. But actually that's awful because now the other players can determine from how you bet exactly what cards you've got. Instead then, a good player must "balance" their behaviour so that whatever they do their opponent doesn't learn anything valuable without paying for it. Some high end professional players have balanced play where they'll occasionally bet very strong with total air (ie they know they don't have the best hand) so that even when you suspect they have great cards you can't be sure. Seeing just one hand of this looks completely insane - but they make money every year, because they don't play just one hand, they play thousands of hands and over time this unpredictability makes them hard to beat.
I am still waiting for a computer who can recognize my bags at a conveyor belt at least as efficiently as me
I've worked with industrial vision devices in the past and trust me, you could set up a machine to recognize luggage as efficiently as a human being today if you wanted to. In fact, it will do better.
The only thing surprising about the Go event is that it did not happen like ten, or even twenty, years ago. You may be impressed, but I find this most underwhelming.
That's likely because you don't understand what it involves. Go is unlike chess in the sense that just throwing raw computing power at the problem won't help you at all; for a "small" 13x13 there are over 10^300 valid game trees to compute, and the number gets exponentially worse once the board increases in size. For reference, the estimated number of atoms in the universe is 10^130.
Google's AlphaGo engine is an actual machine-learning AI which had to be trained plays the game much like a regular person would - Myungwan Kim actually remarked that it feels like playing against a human being. Having a competitive Go engine today is a major milestone, make no mistake about it.
In game design theory, the kinds of politics you describe is normally treated as a form of luck - and as such it makes determining "who is better at this game" a meaningless question. If 3 random people are playing Risk against the best Risk player in the world (the person who understands the game the best), there's a rational argument that their best strategy is to co-operate and eliminate him first (no matter what he does or says or how he behaves). This sort of interaction effectively decouples skill from game success.
This is also why modern game design has generally abandoned games with lots of politics - they effectively all become the same game, and that game is really uninteresting after a while.
Let's not stir that bag of worms...
It requires more than skill. Pruning such a massive game tree is no minor feat - in fact, we don't know how to do it even today. All Go engines are based on some form of adaptive AI.
Again, chess is waaaaay easier in comparison. Pretty much all chess engines work the same way: they start with precomputed moves from an opening book and then move to what's an essentially brute force approach where the engine tries positions, assigns them scores and then picks the highest score available. How these positions are scored / discarded is what separates them, but the base procedure is unchanged. This is also what leads to what chess players call "computer moves" - most chess engines will favor unassuming, conservative moves yielding small positional advantages instead of, well, more "human", intelligent ones. Picking up pivotal moments from classic games (move 17 on Fischer-Byrne, for example) and feeding them to top-rated chess engines is an enlightening exercise.
This is all but impossible to perform in Go with even modest board sizes. The game complexity, given its simple rules, is just staggering.