https://github.com/leela-zero/leela-zero/issues/2240
Apparently they wrote some kind of C program to convert 19x19 weights to other board sizes - don't ask me how. You have to reconfigure and rebuild LZ to support a boardsize other than 19, but I've verified that if you do that and give it the size-21 weights file, it does work and it's strong enough to totally clobber me.

I suspect this probably won't answer questions like whether opening play on larger boards is similar to 19x19 with the common 3-3 invasions. If I had to guess I'd say most of the typical opening play would carry over from the smaller net it's based on. But who knows - I didn't expect such a thing to be possible.

Originally I held the position: the bigger the board the lesser the impact of AI (because the horizon effect increases bigger and bigger). So for a 25x25 board I'd predict top humans still can match top AI while not being able to match it on 19x19. Unfortunately it's not quite testable since no human will take the pain to play serious 39x39 games. But maybe a bot tournament could test it for it would mean: my theory is false if two bots with a 90-10 ratio on 19x19 would have the same ratio on 25x25. I am keen to see that test.

That seems very unlikely. I partially agree (if this is what you say) that today's bots rely more on reading than humans. This can be seen by the relatively large strength difference between a raw net and a full search. But this doesn't necessarily mean they would be weaker on larger boards.

IMO the deciding factor is that smaller boards see the strength difference manifest in closer winrates (1d has more chances against 2d on 13x13 than 19x19). The value of one stone may be similar, but the same strength difference does more work on bigger boards (more moves). I think if a pro takes one stone from a bot on 19, he would need 2 or more on larger boards, or see smaller winrate.

(I recently played a bit with my old idea of strength measurement with error distribution, and I think board size increase would reduce (push deeper to negative) the pointwise mean, and increase absolute (but decrease relative) variance.)

See, I claim it the other way around. First one would need to establish a reliable ratio. Let's say you play AGMaster vs. AGZero on 19x19 10.000 times and the result is: AGMaster 2.000, AGZ 8.000, i.e. 20%: 80%. Then I claim that if you play them on bigger board sizes this gap will strictly converge to 50% : 50% eventually. I even suspect one could prove that purely theoretically for it seems obvious: the bigger the board the higher the complexity the higher the mistake rate the more entropy/randomness.

IMO when two players play the distributions of their error totals look something like this. Doubling the move count with similar size mistakes means the mean of the total is doubled, and variance is doubled as well. But double variance means only sqrt(2) times deviation (the width of the bell shapes), so the deviation-wise distance between the same players (' means) increase.

The converted weights, the corresponding leelaz engines, and a version of Lizzie that supports all these odd board sizes, are also available as a package from https://pan.baidu.com/s/1tutkE8PfuY-Jzi9hBn50dQ
can someone download and upload to another file sharing service?

I think bots will be more better than humans on bigger board sizes, at least if they are trained on them rather than converting weights. Given the huge combinatorial complexity of Go, humans are actually remarkably good at it IMO. This is thanks to our pattern matching , analytical, logical thinking etc skills plus the fact we aren't working on it in isolation, but have the collective wisdom of thousands of years of other go players to help us. That knowledge is based on the 19x19 board [yeah Tibet 17] so things like relative importance of corners, 3rd vs 4th line, positional judgement etc is deeply linked to the board size. It is likely this current human understanding is less appropriate on bigger boards. How easily can we adjust it? Something similar probably happens when converting weights. On the other hand a new LZ started from scratch on 25x25 is likely to reach the same superhuman level after a year of training or whatnot as it did on 19x19 as it didn't rely on that 19x19 collective human wisdom.

alreadydone (https://github.com) responded to my request:
https://drive.google.com/file/d/1goxpaw ... sp=sharing
Thank you very much, alreadydone!

Some computer go programmers believe 9x9 is actually the last frontier.

http://computer-go.org/pipermail/comput ... 10708.html

I suggest to read the whole discussion