A new run of KataGo released - strength comparable to ELFv2

[amnal]

In this discussion there has been mention of an OpenCL version, so I assume that doesn't require CUDA etc., right?

That is correct. It will still need an OpenCL device, normally a gpu but also supported by some cpus (although that will be relatively slow).

[splee99]

Now this is a strange game between Katago(B) and LZ230(W). Near the end of the game white played F8 so that there was a seki in ko situation.



Both sides think they are losing. Maybe Katago thought the ko was too big? Maybe LZ thought it was a bad ko to start? So at least one of the bots was wrong and finally LZ230 resigned. I hope someone can explain this because I'm not a strong go player.

[yoyoma]

splee99, it cannot be a seki because White H8 string is dead. White's only option is to start the ko by playing A5, and Black takes first. By the very end I don't see any White ko threats, so Black will win the ko.

For Black maybe doing something earlier to resolve the situation is safer -- either taking the ko (and filling it to make a temporary seki), or filling liberties of H8. But maybe Black calculated White starting the ko isn't dangerous since Black will take first.

[Bill Spight]

This is an approach ko ( https://senseis.xmp.net/?ApproachKo ). White cannot win this ko. Earlier in the game it may have been better for White to convert this to a direct ko and gain something in exchange for losing the ko. FWIW, I am unaware of any bot ever playing this strategy. I am not sure that current bots could learn it, but in any case they might be able to read the ko fight out and play it.

[splee99]

You are right it is a complex situation. However, Katago thought b was losing (w can play A5 to make it a direct ko). That's a little confusing to me. By the way, white just need to find a small ko threat to win the game. I would think it would be better than sitting there and doing nothing about the ko.

[jann]

These bots are close in strength but their weaknesses differ. Katago enjoys ladder vision and is therefore superior in positions with any ladder potential. This means in other positions like this one it is significantly weaker (as they match ~50%).

[hyperpape]

Does that follow? It seems to me that it means there must be more weaknesses overall in positions not featuring ladders, but it's still plausible that it will judge some positions better and some worse than LZ.

[jann]

Sure, just like a kyu may judge a position better than a dan sometimes. But generally we can expect it to be significantly worse since the ladder gain also seems quite significant winrate-wise.

[cyndane]

I dont think that follows at all. You are making many implicit assumptions about the frequency and severity of misjudged ladders.

[Bill Spight]

Never mind.

[jann]

You are making many implicit assumptions about the frequency and severity of misjudged ladders.

There were factual reports on this. And a few percent of actual losses on misplayed ladders suggest much more winrate loss because of search innefficiencies (misjudged HIDDEN ladders).

Bill: Are you sure B threat (g17) makes no difference?

[Bill Spight]

Bill: Are you sure B threat (g17) makes no difference?

I thought so, but after you asked, I changed my mind. But now I think I see White's reply after Black completes the threat, so that White wins by a few points. There is still a lot of play afterwards, and I don't want to post something erroneous. So I am checking the play.

OTOH, there is no question at all if White plays the ko at move 240.

[And]

katago v1.2-alpha https://github.com/leela-zero/leela-zero/issues/2448. Has anyone figured out how to run katago on windows?

[Yakago]

Have a look at https://github.com/lightvector/KataGo/issues/2 for a windows version

If you can manage to compile Lizzie by yourself, you can make it work with that. There has already been made changes that supports KataGo, but they are not yet officially released

Also, you need an NVIDIA GPU with CUDA support

[And]

it seems there is a way easier and without problems with cuda