Life In 19x19

It's not often you see the game of Go in mainstream media, but I was pleasantly surprised today when I saw a deluge of articles about a new Facebook project.

[Article Link]

It seems there's more to the story than the simple fact that it's Facebook that is developing the AI. They are applying so-called "deep-learning" techniques based on the visual layout of the board. This is probably a lot closer to how the human brain approaches Go.

As you may know, the most successful algorithms in Go are based on the Monte-Carlo Tree Search (MCTS). MCTS looks at long sequences of moves and uses probability and heuristics to eliminate certain branches of the "game tree" (a diagram of all possible moves) until it chooses a limited number because it can't possibly evaluate them all. Well, this new AI is using pictures of the board to see what bad and good moves "look like." The pictures of possible moves are input into a "neural network" (a system of computers that emulates the connections in the brain) and it outputs the move that it wants to make. Depending on the success of the result, the algorithm adjusts itself over time.

This is the same technology that is used by computers to determine the content of a picture and also to parse and translate written text among other things. The theory behind this is that experts of Go can evaluate a move on a mostly subconscious level just by looking at it. It probably isn't the first time this type of algorithm has been used, but with Facebook throwing their weight behind it we might see some innovations very quickly!

Sorry if this is kind of hard to understand, but I'm a programmer and excited. I am also not an expert at neural networks, so corrections may be in order (and appreciated!).

It will be interesting to see how this develops. I actually thought IBM or Google would work on this project, not facebook.

Interesting.

People on the Computer Go mailing list are claiming it's playing as darkforest and darkfores1 on KGS: http://computer-go.org/pipermail/comput ... 08109.html. However, I'm not sure how they came to that conclusion.

Edit: originally left out the names.

a little discussion on the computer-go mainling list, and their bot doesn't seem very strong yet. Maybe high kyu to low dan level. Mostly marketing hype so far.

they think its playing as darkforest and darkfores1 on KGS

oops (hyperpape beat me to it)

Krama wrote:It will be interesting to see how this develops. I actually thought IBM or Google would work on this project, not facebook.

Google are working on a neural-net Go AI, via their DeepMind subsiduary, as are a few other researchers. See viewtopic.php?p=182825#p182825

So even if their doesn't turn out so strong, that they are doing it and making press releases that many people will read is good news for increasing exposure.

xed_over wrote:a little discussion on the computer-go mainling list, and their bot doesn't seem very strong yet. Maybe high kyu to low dan level. Mostly marketing hype so far.

I find a high kyu to low dan bot really impressive. Especially if the underlying algorithm does not rely on tried-and-true methods (Monte-Carlo). Count how many years were needed to reach low dan level with Monte-Carlo algorithms and imagine they did it again from scratch (this is probably an exageration but you get my point).

I don't know about the marketing argument, I'm not sure developping a go AI would be a cost-effective marketing plan given that media will mostly not care until the AI actually gets close to world champion level - and we have all the reasons to believe this requires an extremely large amount of effort. I would more easily imagine it as a proof-of-concept to develop new AI concepts that would later apply to other domains. Althought I guess the prestige of building "the new Deep Blue" might be a nice side effect.

Jhyn wrote:I don't know about the marketing argument, I'm not sure developping a go AI would be a cost-effective marketing plan given that media will mostly not care until the AI actually gets close to world champion level - and we have all the reasons to believe this requires an extremely large amount of effort. I would more easily imagine it as a proof-of-concept to develop new AI concepts that would later apply to other domains. Althought I guess the prestige of building "the new Deep Blue" might be a nice side effect.

I suspect the marketing is targeted at two specific groups: investors (we're not a one trick pony!) and developers (you can do cool cutting edge stuff here too!). Those former are more likely to read everything about a company, the latter is more likely to have familiarity with Go as a major outstanding AI problem. GE has been running some commercials lately that seem to be about convincing people they're a good place to work: companies seem willing to go pretty far to get top technical talent. If this pushes forward their machine learning stack as well, even better.

This is pretty cool ... if I want to build a Go AI program myself, how do I go about doing it ? Anyone has any idea ?

Tonkleton wrote:It's not often you see the game of Go in mainstream media, but I was pleasantly surprised today when I saw a deluge of articles about a new Facebook project.

[Article Link]

It seems there's more to the story than the simple fact that it's Facebook that is developing the AI. They are applying so-called "deep-learning" techniques based on the visual layout of the board. This is probably a lot closer to how the human brain approaches Go.

As you may know, the most successful algorithms in Go are based on the Monte-Carlo Tree Search (MCTS). MCTS looks at long sequences of moves and uses probability and heuristics to eliminate certain branches of the "game tree" (a diagram of all possible moves) until it chooses a limited number because it can't possibly evaluate them all. Well, this new AI is using pictures of the board to see what bad and good moves "look like." The pictures of possible moves are input into a "neural network" (a system of computers that emulates the connections in the brain) and it outputs the move that it wants to make. Depending on the success of the result, the algorithm adjusts itself over time.

This is the same technology that is used by computers to determine the content of a picture and also to parse and translate written text among other things. The theory behind this is that experts of Go can evaluate a move on a mostly subconscious level just by looking at it. It probably isn't the first time this type of algorithm has been used, but with Facebook throwing their weight behind it we might see some innovations very quickly!

Sorry if this is kind of hard to understand, but I'm a programmer and excited. I am also not an expert at neural networks, so corrections may be in order (and appreciated!).

How many times do you intend to ask that?

Javaness2 wrote:How many times do you intend to ask that?

I guess until they get an answer.

Just for amusement value of answering, here is a place you can start...

https://www.gnu.org/software/gnugo/gnugo_toc.html

They have published a paper on the bots here: http://arxiv.org/abs/1511.06410

And they are playing rated games on KGS now as darkforest and darkfores1, and they are both 1k. Right now they are just the pure DCNN versions. The author, tydsh, said he will put the DCNN+MCTS version on soon (he hasn't decided if he will use the same accounts or make new ones). There were some other DCNN bots on KGS, DCNNigo2, and it was 3k. So Facebook's bot seems 2 ranks stronger. I'm curious to see how much difference the addition of MCTS makes.

oren wrote:Just for amusement value of answering, here is a place you can start...

https://www.gnu.org/software/gnugo/gnugo_toc.html

Perhaps even better: Michi is a super-simple python implementation of the current state of the art (MCTS). It's weak, because right now there's no way to do simple and strong, and also because Python is slow. But the code will give you a good start: https://github.com/pasky/pachi

Forgive my ignorance, but what is DCNN in this context?

skydyr wrote:Forgive my ignorance, but what is DCNN in this context?

DCNN = Deep Convolutional Neural Network -- most commonly used in image recognition. So it is good at recognizing shapes/patterns on the go board.
MCTS = Monte Carlo Tree Search -- Monte Carlo part means it plays out many random games to get a feeling for how good the position is. The tree search part is reading out different variations.