Life In 19x19

There is a website that allows you to play Go against a deep neural network - all implemented in javascript. The website is powered by jgoboard and the ConvNetjs libraries.

Their trained downloadable neural net for Go is available, which might be of interest to some.It was trained on GoGoD games

"This neural net was trained to make moves similar to the moves made by professional players in a dataset of over 80,000 professional Go games."

https://chrisc36.github.io/deep-go/

Everything is written in javascript, runs in the browser and uses CPUs <not GPUs> - which keeps things simple.

The paper that describes all this was written by Amos Storkey (Edinburgh University) and Christopher Clark (Allen Institue for Artificial Intelligence) - presented in 2015 - on training Deep Convolutional Neural Networks to Play Go.

There is a good intro to Deep Learning on the ConvNetjs page

If you implement these "shape" moves that this NN provides into MCTS so it doesn't pick random moves but moves suggested by NN you get something similar to Alphago?

You would be 2/3 of the way, you also need the value network.

(AlphaGo also used 2 distinct policy networks, one fast (less accurate) and one slow (more accurate).)

Krama wrote:If you implement these "shape" moves that this NN provides into MCTS so it doesn't pick random moves but moves suggested by NN you get something similar to Alphago?

Nope, you would get something similar to the latest versions of Zen, Crazystone, Aya, etc which are MCTS bots which have recently added Neural Nets, or darkforest (from Facebook) which has always used neural nets.

uPWarrior wrote:You would be 2/3 of the way, you also need the value network.

It's worth adding that that missing third is the hardest and most innovative part of AlphaGo. There was this nice chart from the AlphaGo paper (Fan Hui version, it's gotten stronger since) showing the strengths using various combinations of the components:

This program plays beautiful opening, definitely better than shodan level. It is rather weak in mid-game which is to be expected.

Uberdude wrote:It's worth adding that that missing third is the hardest and most innovative part of AlphaGo.

Definitely. I think it is not only the hardest to conceptually figure out how to obtain (the RL idea was great), but also the hardest to implement without proper hardware.

On the other hand, it requires the previous building blocks to be in place, so it is nice that there is an open source effort moving forward.

Since people are probably interested in the development of a NN program that does not require much in the way of resources, we perhaps should be opening a thread to talk about Julie (which just appeared on the scene). OK, not a very strong player in the machine tournament. But it would be good if we had some information about how it might do against some of us not so strong humans. The reason is that it appears to be able to run fast on very modest hardware.

So yes it got beat by most the other programs there, but considering the minimal hardware and the minimal time it used, able to do what it did with more than two orders of magnitude less resources.

macelee wrote:This program plays beautiful opening, definitely better than shodan level. It is rather weak in mid-game which is to be expected.

I played it (with Black), won relatively easily and I estimate it at 1k-level. I think with a little bit of work (like how to pass or resign) this could become the best free-available bot. And with a little bit more work this could become the best free available 4-5d Bot.

Pippen wrote:

macelee wrote:This program plays beautiful opening, definitely better than shodan level. It is rather weak in mid-game which is to be expected.

I played it (with Black), won relatively easily and I estimate it at 1k-level. I think with a little bit of work (like how to pass or resign) this could become the best free-available bot. And with a little bit more work this could become the best free available 4-5d Bot.

Agreed. Combine it with a reasonable 'score estimator' type of code so that it knows when to resign. Should be quite easy.

Pippen wrote:

macelee wrote:This program plays beautiful opening, definitely better than shodan level. It is rather weak in mid-game which is to be expected.

I played it (with Black), won relatively easily and I estimate it at 1k-level. I think with a little bit of work (like how to pass or resign) this could become the best free-available bot. And with a little bit more work this could become the best free available 4-5d Bot.

Try to disable auto move and show analysis of the moves then play on your own against someone else and use the moves it suggests. However if it suggests a blunder type of move then you play a move you think is correct.

macelee wrote:...It is rather weak in mid-game...

Exceptionally so. I tried playing it and after about 100 to 150 moves it starts giving away groups.

Very nice. I needed an undo in a corner. At the end though, it started to make pointless moves. And at some stage claimed that suicide isn't allowed (I think it tried to suicide itself) and switched colours. That's a fine tesuji, indeed. Is there a way to score or is this no-pass go?

I started off with a common opening pattern of the fighting Chinese and it copied standard play for a while, but when I started a semeai in a corner it got confused and even after it was my 3 libs to its 2, it continued surrounding one of the outside groups (which was eyeless with about 7 libs but alive due to the semeai) whilst I proceeded to claim the rest of the board and win by loads. It also amusingly made the following local mistake (which some MCTS should fix):

Click Here To Show Diagram Code

[go]$$B
$$ . O . . |
$$ . O 3 . |
$$ . X O 2 |
$$ , X O 1 |
$$ X O X . |
$$ X O X . |
$$ . . . . |
$$ --------+[/go]

Kill all stones of the AI in sente

Well, you guys are a lot stronger than I am so I have yet to win

Luckily I am a lot more tech-savvy than I am a Go player. I noticed that the site only offered the net.js as a download, not the other pre-requisite files. So, I downloaded them in their original format and zipped it. For those wanting to play with it offline or on their own (private) webspace, here is the download link: http://millsf.home.xs4all.nl/Deepgo.zip

This is a direct, 1-on-1 download of the page. All credit goes to its creator.

To use it: unzip it to any directory and open the index.html (Chrome or Edge work fine, IE11 doesn't). I also uploaded it to a private webserver which works too.

_{I'll also ask the owner for permission, should he decide not to grant it I'll remove the link right away.}