AlphaGo vs. Lee Sedol

Social Forums Non-Philosophical Chat
1

Any other go players on ILP? I know of at least one, but I’ll let them out themself. Anyone else following the AlphaGo vs. Lee Sedol match?

A little background: Until late last year, go was a game that humans dominated computers in. Unlike chess, go playing computers could only play at the level of skilled amateurs. Go’s decision tree branched too quickly, and it was too difficult to evaluate positions, so computers couldn’t brute-force the game the way the have with chess.

Enter AlphaGo, a go playing computer that surprised the Go world late last year by beating the European go champion. That sounds more impressive than it is: go is not very popular in Europe, so the best players in Europe are not nearly the best in the world. Still, a computer beating a pro-level player was unheard of, and about 10 years earlier than expected.

That match was in October last year, and right now there’s a new competition underway between AlphaGo and Lee Sedol. Lee is one of the best go players of all time, and currently ranked among the top five in the world, so this match-up is comparable to the DeepBlue - Kasparov match in chess: human dominance in go is being challenged by a computer program.

Except this is bigger. Go has long been the holy grail of AI research, because the game is much harder for computers to grasp. The skills required have been seen as quintessentially human: flexibility, insight, intuition. And AlphaGo exhibits all these. In fact, it was born out of a system that wasn’t even built for go, but for learning to play and master Atari games. The AI at play here is getting very, very close to a general purpose AI, that learns games the way humans learn them, and can be applied to numerous problems. And AlphaGo has won the first two games.

A computer winning at chess was about breakthroughs in raw processing power; this is about breakthroughs in computer learning and problem solving. Even if Lee wins the next three games, which is looking pretty unlikely, this is a moment to take notice of. AI will move very quickly from here on out.

2

Front page of the times today…

I was talking about this with a guy a couple weeks ago… We were talking about how the last 5 moves of chess has been solved.

They used an adaptive algorithmic format for this, because it’s like 64000 iterations after just the first few moves. This decision paid off!!

3

Yeah, the beginning and end of Chess are much easier to brute force. You never have more than a couple dozen possible moves in a game of chess, and in the beginning and ending it’s often much lower. In go, you start with about 50 (because the board is symmetrical), and quickly go up to over 300 (their are 361 places on the board), and most of the game you have well over 200 available moves.

And go is harder to evaluate. Chess is pretty straightforward, because the relative worth of pieces is established, and whether a move will cost you a piece is easy to determine. Go is more ephemeral, and the worth of a particular move, and who’s winning at a given time, is frequently not a trivial question.

One of my favorite ways to think about game depth is like this: take two players A and B. Say A beats B 75% of the time they play. In that case, A is 1 ‘level’ better than B. Now ask, how many levels are there between the completely inexperienced beginner and the best player ever to play the game? Tic tac toe has maybe 2 levels, because once you understand how to play you realize immediately how to provoke a draw every time. Chess has about 14 levels. Go has something more like 40.

4

Chess programs today, missile guidance systems and DARPA robots tomorrow.

5

Yes, I used to play.

No, I’m not impressed. Bit saddened this didn’t come sooner.

6

Sure. But nuclear-proof military communication network today, cat videos and conversations about go playing computers tomorrow. Advances in technology are tools, they can be used for good or evil.

Did you expect this sooner? Prior to last October, the state of the art of computer go was at best at the skilled amateur level, and I am under the impression that even AI experts didn’t expect something like this for another 10 years. What led you to believe it should have arrived before now?

7

In Go, people will memorize the sequencing of their moves in tough matches they won, and recite it excitedly later on, to show off their finesse.

This means the moves are mnemonically represented in space and sequence simultaneously. Relative deposition of terrain is the crucial matter here, people listening to these chains visualize the board, it has a mirror neuron effect.

This is more or less what they are doing. You don’t brute force calculate, but rather focus on forcing form lines to switch their values… so the millions of calculations aren’t needed. The greats are merely playing a game of Formlessness in making purposeful patterns seem random. They are merely asserting patterns out of form lines. The calculations as a result get smaller and more certain the longer you go.

AI programmers don’t realize this. What happens if you provide a truly random jumble of lines? Cam a AI adapt? Unlikely under brute force, but a human can.

8

I think you’re underestimating how intensive it is to recognize and flexibly apply the patterns. For example AlphaGo does use millions of calculations: it has almost 2000 CPUs and almost 300 GPUs. That’s some intensive processing. But I still wouldn’t call the strategy brute force, because it’s not just trying every possibility and see which ones produce the best outcome, or even using monte carlo methods to try a whole lot of random moves to see what works. It’s actually abstracting from the millions of games in its database to intuit good moves and to evaluate who’s winning in a given situation. Processor intensive, but much more elegant than simple brute-forcing.

What do you mean by “a truly random jumble of lines”? Do you mean to say that playing randomly against AlphaGo would be a winning strategy for a human player?

9

Carleas, you overlook the corporate military industrial complex that has a stranglehold over technological innovation.

10
  1. How so?
  2. What percentage of people in the developed world do you suppose are included in either corporations, the military, or industry?
  3. What is the form of the stranglehold (e.g is it a conspiracy to suppress innovation? or is it just that those three broad classes of human activity tend to subsume most innovation?)?
  4. How is this point relevant to computer go?
11

You miss my point Carleas. And I am hardly underestimating it, I have a very strong background in classical Chinese strategic texts. The programmers are going about the dumbest way possible. You don’t need that much processing capacity.

Lets say your a cop, and your department has 5 SOPs governing tactical deployments. Just five basic rules, based on rules, to follow in placing men.

Thats your orthodoxy. When people know your orthodoxy, they adapt… you gotta adjust over time what your SOP is, but you could keep it still at five kids of deployments.

In Chinese Strategy, they use two principles, The Orthodox, and the Unorthodox. The Unorthodox is when you get all snazzy smart and do unexpected, hopefully genius things. However… after a while, this Unorthodoxy becomes the Orthodox… criminals pick up after a while you never come through the door, you go through windows, or the chimeny, ALWAYS. So when they see the police show up, instead of pointing their guns at the door, the point it at the chimney. At that point, you flipplop and turn the unorthodox orthodoxy aside, and due the unorthodox… which is following the orthodox… go through the door.

Go is govern by the exact same mentality. A master player isn’t doing a billion trillion calculations. He is doing the exact opposite… over the years, he has condensed his ideas down to a handful of principles. Its how these rules are patterned in each match is what matters… and it begins with Formlines.

Humans don’t have to do a lot of calculations, they just look at the formlines, the deposition. Its like Wittgenstein images, duck or rabbit. We aren’t doing a calculation for each spot, the calculation comes from does the image fit the economy of our needs, based on what sticks out.

You don’t need brute force calculations for this, on this level you described. I get it’s binary… I get it’s still a lot no matter what, but I also get it doesn’t have a theory of mind for us, and can’t see the board itself.

By completely random… I mean… whip out Microsoft paint, and start slapping down random shapes, like circles and squares, and then slap straight lines all over.

A human can adapt to this rather fast, if you play as close to normal as possible… just isn’t a grid anymore. Its not non-linear, but it is a confusing mess. How do we do it? Form lines, geometry. Its easier to institute Sun Tzu’s concept of formlessness as well.

We should of produced a AI for go capable of beating any human decades ago. Strategic works have been available for centuries, not hard to find, it practically spells out how to think… some guides literally read out like a computer program… I have some modders looking into Ascleciodotus for missing units for Rome: Total War… 2000 years old, but the details are that mathematical, covers everything (and came out admittedly absurd, cause he was a philosopher with no military experience).

You gotta approach the programming visually. How many Rods and Cones are in the eye? How do we see forms. How do we insert position, juxtaposition?

You can do this rather easily in binary… in a very real sense, we do. All your doing is asserting patterns, and throwing your opponents off. That sounds like it could blow a brute force computer, but it’s actually just another simple couple formulas.

12

so to condense the idea of this thread - ai is beating humans at the only thing that humans have had against AI - creativity? well, maybe ethics too.