Minigo: A minimalist Go engine modeled after AlphaGo Zero, built on MuGo This is a pure Python implementation of a neural-network based Go AI, using TensorFlow. While inspired by DeepMind's AlphaGo algorithm, this project is not a DeepMind project nor is it affiliated with the official AlphaGo project. This is NOT an official version of AlphaGo. Repeat, this is not the official AlphaGo program by DeepMind. This is an independent effort by Go enthusiasts to replicate the results of the AlphaGo Zero paper ("Mastering the Game of Go without Human Knowledge," Nature), with some resources generously made available by Google.
They explained it very clearly what exactly is MiniGo on their GitHub page.
On their website CloudyGo they did about 7 training runs and v10 got 23 million games, v12 24 million games and v13 which is the latest one and the one from which I got all the positions in this article, played 23 million games.
Amazing, isn’t it? If we sum up all the training runs that they did, we could end up getting about 98 million games (includes 28 million games of other training runs which I do not described above).
About 98 million games on 20 blocks, truly fascinating, Leela project is still on 11 million self play games. Indeed it is hard to estimate how strong MiniGo is without official tests, that’s why this article is not exactly about that.
In this article, the point is to cover some of the most interesting things that I find about MiniGo that makes me watch it’s games more and more. The first time when I watched the games, I thought the games are easier to understand than any other AI, it was just completely different style even though it had san-san invasions into play. I thought I am exaggerating, that’s why I sent it to some students and their response was that the games were really easier to understand comparing to other AI. It would be nice to see if the readers of this article would check some games and agree with the statement.
Tesuji moves and sequences which looks truly beautiful were always my passion.
So, below I am gonna share some situations in games of MiniGo which I rechecked a minutes ago, however, do not forget that these games are match games and they are not really played at too high visits, although the amount of visits that they used is probably enough to crush almost all professionals.
I find it pretty hard to comment anything on the situations below, so for the reader, this article would be mostly just watching beautiful moves and ideas, which I believe would not be boring at all, especially since Go is an Art.
As you can see, MiniGo found some extraordinary moves as AlphaGo too.
It is wonderful that it can play flexible without worrying that some of it’s stones can die or get hurt.
Don’t dare to play a normal connection move, attach full power instead.
When you want to grab some points and minimize your opponent territory, attachment is a good idea, however I feel like MiniGo got addicted to it, attachments never end.
These small things like this move always make impression to me. I can barely explain it, but you can feel how calm this move is, how mature in a way.
You see these two openings up to 25/21 move? We can call it MiniGo opening, it loves it.
Ever wonder how to effectively escape or make sabaki – > Attach
Such a beautiful move, never give your opponent easy time settling his group!
Attachment is good technique to enclose roughly territory as well.
You remember when your teacher said “Do not place your stones close to thickness” How can I ever teach someone anything, after this? :- )
Oh, there is a little open space over there, let’s invade!
Yes, this slow move, not the keima, thick.
If I say that MiniGo has Takemiya Masaki style, I wouldn’t be wrong I guess.
These kind of strange and ugly looking moves, can sometimes be the best local moves, do not miss them just because of the looking of the shape, stay sharp!
Do you remember the old technique about reading 1-2-3 – > directly 1-3, well MiniGo successfully learned it.
If you think MiniGo is not creative enough … check this out
You thought you have cash over there and you counted the points..? It is time for MiniGo to do some magic then!
Bonus one more fascinating invasion..
I also really like the simple moves that MiniGo plays, it makes you even wonder sometimes “Can I win by playing such moves?”
MiniGo also dies ( or sacrifices) as a friend says “AI also sucks, just at higher level, so that we cannot notice easily”
In memory of Kitani Minoru it also played his joseki!
White invades, what is your first intuition? Would you attack or strengthen yourself first?
Finally, the cute scenario of all zero bots, they do not quite know ladder, same applies to MiniGo. About 23 million self play games, strong enough to crush top professional, but it does not understand ladder, awkward !
I hope you enjoyed this article, subscribe and stay tuned for more !Follow me in social media: