Life In 19x19

lightvector wrote:For example, if you adjust the komi to make the winrate 50% instead of 97%, ...

97 ==> 50 is a bit extreme, isn't it?

I would like to turn briefly to win-rates BELOW 50%.
This is based on my experience with IH120, which is probably not completely transferable, since this is not a position from a real game.

For displayed win-rates > 35% (maybe even > 30%), I think that you can be sure that KataGo's favourites are all reasonable moves.

For win-rates < 20% (perhaps < 25%), there is a risk that KataGo will turn to desperation measures, but which do not really make sense.
In this case, it might be advisable to successively adjust the komi.

With IH120, leaving the range of 35% - 65% (maybe even 30% - 70%) does NOT automatically mean that the game is already decided!
Robert, so in the future, be more relaxed about differences of only a few percent.

Yep I think your experience is transferable, but probably to midgame positions with extreme and deep tactics, some of which end in huge winner-take-all swings. For opening and for "common" game positions, the komi and/or current winrate doesn't make nearly as much of a difference, to the degree that in practice I find it usually not worth spending the time fiddling with.

For example, KataGo plays still pretty reasonably in 3 or 4-stone handicap games, in which the initial winrate for white < 1%, and well enough to win against amateurs and usually severely challenge professional players (although 4 stones, and even often 3 stones is too much to overcome versus strong pros). Mostly good moves that gradually close the gap as the opponent makes mistakes or plays inefficiently, rather than mostly "desperation moves" that can be refuted and all just fail. In some ways, even too respectful of the opponent sometimes.

lightvector wrote:If you're studying theoretic openings, then sure, go for the fairest initial komi. That should be 7.0 with area scoring

This is my current objective indeed.

if you were trying to find the theoretical sharpest and least-forgiving ways to play.

Maybe, but there are limits to my capability of understanding the most complex AI fights...

***

Many thanks for your, as always, very useful clarifications!

lightvector wrote:Yep I think your experience is transferable, but probably to midgame positions with extreme and deep tactics, some of which end in huge winner-take-all swings.

In games between Karl's 40b-IH120 (with Black) vs. my 60b-IH120 (with White), White likes to enter a "normal" middle game (i.e. without Main Semeai), if she gets the chance to do so.
There exist several moments for playing my Guzumi (40b likes these a lot), which provide White with the above mentioned option.

Apparently, the "extreme and deep tactics" of the Main Semeai tend to be too extreme and too deep for Katago, too.

For opening and for "common" game positions, the komi and/or current winrate doesn't make nearly as much of a difference, to the degree that in practice I find it usually not worth spending the time fiddling with.

That is what I also wanted to emphasize.
I think that over time Robert will also become more "insensitive" to these parameters.

For example, KataGo plays still pretty reasonably in 3 or 4-stone handicap games, in which the initial winrate for white < 1%, and well enough to win against amateurs and usually severely challenge professional players (although 4 stones, and even often 3 stones is too much to overcome versus strong pros). Mostly good moves that gradually close the gap as the opponent makes mistakes or plays inefficiently, rather than mostly "desperation moves" that can be refuted and all just fail. In some ways, even too respectful of the opponent sometimes.

With the "desperation moves" in IH120, there are no longer any meaningful possibilities to turn the tide. In addition, the board is already very full of stones.
This should absolutely not be the case in the handicap games you described.

And yes, KataGo is very very smart in accumulating small advandages!

Katago TensorRT + LizzieYZY, filling of the physical 64GB RAM:

Event at ~2m visits of each of the two top moves: they swap their order of percentages and scores.

Event at RAM 91.9%,
swap file becomes relevant and is used up to 84.5%,
drop to RAM 77%,
increase CPU up to ~34% regular use (~16% is normal of 8C/16T),
drop to GPU ~84% (94~97% is normal),
VRAM 767 MB (about this is normal).

Event at RAM 99.8%,
the top three moves have 19.6m / 20.6m / 9.0m visits, respectively,
the top two moves have 48.0% or 48.2%, respectively; each has the score -0.2,
the third move has 48.8% and the score -0.1 (White to move),
LizzieYZY process is killed automatically (I do not know about the fate of Katago's process),
drop to GPU 0%,
increase RAM 99.9%,
0 MB RAM available,
Windows GUI very slow,
Windows does not free RAM to less than 99.9%,
restart Windows is still possible,
afterwards Windows is normal.

On a desktiop, one might use the reset button. If you plan to fill the RAM on a notebook, first inform yourself how to reset or cold-shutdown it. If you think that you can run Katago over night and see the result in the morning, you might need to activate permanent logging to an autosaved file.


EDIT

If you're using LizzieYZY, there should be a setting where you can configure it to pause analysis if the total number of visits hits a certain number. If you use a value that will be large enough for your purpose but that won't exhaust all your RAM, then it should be fine for you to leave it unattended overnight.

Nice, I will try this!

During the opening and middle game, Katago TensorRT on my RTX 4070 usually has 2700 ~ 7200 visits/s but occasionally up to some 15,000.

In a local L+D problem during the middle game, it is up to some 120,000 visits/s. The GPU has its usual 92~100% but there is up to 65% use of the CPU (8C/16T) for several seconds at up to 88°C so CPU and system fans occasionally become fast and loud so that less flat fan curves for lower RPMs might be better.

RobertJasiek wrote:Now, I have refined everything and put it on one webpage:

https://home.snafu.de/jasiek/AI_Computer.html

THANK YOU ROBERT!!!

I finally upgraded my gpu to a more modern model (4060 in my case). I had no idea what to do. Your extremely detailed instructions worked great!