Public code using AI for 9x9

General Go Discussion
1

Hello!

I have read a lot about Leela Zero, Crazy Stone Deep Learning, Facebook codes, etc, I think all of them for 19x19.
Is there anything similar for 9x9 ? (I know minusgo in OGS is a developing codeof this kind, but is not publicy available, unfortunately :frowning:

1 Like
2

Iā€™m curious why itā€™d be any different?

3

How to make minusGo (paraphrasing Alderi from memory):

  1. Go to config.h
  2. change: BOARD_SIZE = 9
  3. compile
  4. train/play
4

Different? I am interested in the code itself, to improve my gameā€¦ and I play 9x9 gamesā€¦

5

I am afraid I am not too good with computersā€¦

Go to config.h (where?)

Sorry if this is a very silly question.

6
2 Likes
7

Most of the recent and highly successful efforts in developing strong go AI have focused solely on the 19x19 even game (reasonable komi + no handicap stones). These recent AI methods require a significant amount of computationally intensive training that tailors the bot to the particular choice of parameters of the game (board size, komi, etc). However, this makes these bots quite inflexible to be used on different board sizes (or even different komi, if very strong play is desired).

In principle, with the open-source bots, one could download the code and retrain the bot for a different board size, komi, etc. However, that would require a very significant amount of computational effort (especially if working alone, starting from scratch), and at least some technical understanding to get all of the software/hardware properly configured.

I think the original poster is looking for existing efforts where people have already developed, trained, and packaged strong bots for the 9x9 game. Unfortunately, I am not aware of any. However, I would not be surprised if they do already exist. A quick search seems to show that at least some efforts/experimentation in this area (which OP mentioned that he was already aware of):

3 Likes
8

Yes - the training is different, but the code is the same.

When you said ā€œcodeā€ I thought you meant you are interested in the code :smiley:

1 Like
9

Sorry if I did not explain accurately.
What I was looking for was just a program to analyze my own games (as mark5000 suggested in a recent post).

I have neither the time or the knowledge to develop an AI researchā€¦ maybe if I were 20 years younger :wink:

I have read people can get Leela Zero together with a file ā€œwith the weights of the neural networkā€ and run in an interfaceā€¦ (I apologize if I do not use accurate words again) and I was wondering if there were something similar for 9x9ā€¦

At present I used the Monte Carlo version of Leela (I think is Leela 11) but it does not play wery well (even me can notice it!!). On the other hand minusgo plays beautifulā€¦

I know minusgo is available to play with in OGS (thanks to Alderi!!) but the system does not allow fork the games from a position I am interested in (It seems that is not possible if difference in playing strengh is wider than 9d even for unranked games). So I am afraid I need to reach 1d in order to analyze my games with minusgo!!! Bad luck!

Thanks everybody for your answers

3 Likes