[Welcome to Sensei's Library!]

StartingPoints
ReferenceSection
About


Paths
Speculation

Referenced by
ComputerGoProgram...
ComputerGoProgram...

 

Neural Networks and Go
Path: Speculation   · Prev: SomePhilosophicalQuestionsAboutComputersAndGo   · Next: DanielDennettAppliedToGo
    Keywords: Software

Discussion

Does someone know if neural networks have been used with Go ?

AGiss: (By "neural networks", I mean the computer's tool to solve problems, that has merely nothing to do with the things within everybody's brain!) While neural networks can be a big faillure with Go, we can't know if the NNs is well suited to Go or not unless we try it.

As I think I'm not a genius or even someone who has good ideas before others, I think some other people have already tried to use that concept. But I have no reference of such research. Does someone has a clue or URL ?

My idea would be to use the 19x19 intersections as 19x19 input (values 1, 0, -1) and 19x19 outputs (that's huge) values -1 or 1 for (bad move or illegal move) / (good move)

A good move is a move that at pro has already done on the board.

I don't have a big pro database to try such a project!

That would take a lot of time/CPU/memory to make that, but if it really works, that would make an incredibly fast Go program. -- AGiss

WinHonte: from the Jellyfish people uses neural networks in part of the game engine. As I understand it they haven't tried to teach one big neural network to play the game as described above. Rather, they have a conventional architecture for a Go-playing program, except that where most programs have a hand-made database of patterns for suggesting local moves, WinHonte uses a neural net trained from professional games.

Kungfu (3k* IGS):

I have some questions and comments about this. First of all, from what that cached page says, the NeuroGo program was a *success*. It seems that after 4500 games against itself, it was able to beat Manyfaces at a reasonably strong playing level. So the failures of NeuroGo have really only been attributed to the computational resources available to it and not (like the other programs) entirely on the expert knowledge they have hardwired into them.

1. What are the resources required to play at the 5 dan amateur level on a 9x9 board?

2. Are the computational resources necessary to play as 1 dan on a 19x19 board greater or lesser than those required for a Many Faces / GNU Go style program?

gonzo: Part of the problem is that neither GNU Go nor Many Faces can play at a 1 dan amateur level. Many Faces claims to be able to play around 8 kyu at its strongest level. If anything can play at a 1 dan level with reasonable (polynomial) resources, I am sure someone would come up with the computing power!

3. What is the limit of strength for a given network "size"?

4. More web resources on this!!

Comment: It then seems obvious from what we know of computer Go that a hybrid approach needs to be used. A large if not complete database of things like shapes, joseki, whatever, needs to be connected to such a neural network.

Experiments should be done! If it needs one day to come up with a pro or high level amateur move using distributed computing, then so be it. Let's do it. At least we'll only be a few hundred years behind the times of the human players ;)

Links to related Papers

SAS: There's an article called [ext] The Integration of A Priori Knowledge into a Go Playing Neural Network. (Also available [ext] here. There's also a [ext] PostScript version.)

SAS: Another paper on the subject is [ext] Co-Evolving a Go-Playing Neural Network.

ilan: In my opinion these papers exactly show the weakness of the neural network approach. They tackle the classical artificial intelligence questions (the ones based on making machines do difficult things that people usually consider a representation of "intelligence") instead of following the scientific approach of first solving simple and well understood questions. In particular, a good first problem for a neural network would be to solve integer arithmetic (addition and multiplication). Note that this is in fact necessary to play go (you need to know how to count and how to add to play go)! Of course, the reason for all of this is that you don't get grants or promotions if you study addition and multiplication because these don't require "intelligence".



Path: Speculation   · Prev: SomePhilosophicalQuestionsAboutComputersAndGo   · Next: DanielDennettAppliedToGo
This is a copy of the living page "Neural Networks and Go" at Sensei's Library.
(OC) 2004 the Authors, published under the OpenContent License V1.0.