When playing without superko, as skill level goes up, I expect the frequency of long cycles will go up as well, just as draws are more frequent in chess among skilled players.

Now that AI has surpassed the best human players by a wide margin, I was wondering how much more common long cycles are among AI players than they are among human pros, and how far we are from the day long-cycle draws become a common result in AI play, thus emphasizing the need for superko.

Are there any statistics on this? Has anyone got a long cycle yet by pitting KataGo against itself without superko?

They're still fairly rare. It's also not a given that they'll necessarily get more common. But you can see a few, for example here:
https://katagotraining.org/sgfplayer/ra ... mes/79525/

Thanks for the link. Assuming a fair komi that doesn't allow ties (that would be 7 points plus button under area scoring, as I've seen discussed elsewhere), I think long cycles are pretty much guaranteed to get more common as skill increases, because the player that is behind in a given position is more and more likely to find them in order to avoid losing.

You're likely right. But perhaps in a given ruleset, the result of optimal play could still a be win for one particular side even given that they have to also avoid any long cycles. If that turns out to be true, it must be the case that as you converge to optimal, even if there is some intermediate period where they become more common, after that they necessarily must then get rarer again until becoming nonexistent.

So the very strongest form of that claim - if it were always the case at increasingly high levels that further skill increases made long cycles more common - would be essentially equivalent to claiming you know the result of optimal play - i.e. you've weakly solved the game. Which is a bit of a stretch of a thing to claim as "guaranteed".

And who knows? If it turns out that one side does win even if they also have to unilaterally defend against long cycles, it's also easily conceivable that in practice as you converge to optimal, long cycles don't ever become remotely close to common before the level of play reaches the point where it becomes clear which side is winning for sure. For example this certainly seems to be the case on 7x7, and looks to be increasingly likely on 9x9. Despite the fact that you can construct board positions on 7x7 or 9x9 where long cycles are forced for both players to avoid losing, it nonetheless seems that under both Chinese and Japanese-like rulesets without superko, non-integer komi values are probably outright wins rather than long cycles and it seems there is never a point where long cycles become common in practice as you converge towards optimal.

I agree on all accounts. My claim was not nuanced enough; I didn't mean it to apply all the way up to perfect play.

Superko certainly seems unnecessary on 9x9, which is a definite strong point of that board size. Of course, the frequency of long cycles increases with board size, simply because there is more room for one to appear.

Is there a simple way to search for such games in the KataGo training database?

Probably: wait until bulk SGF downloads are implemented and download the SGFs and grep for the game result "Void".



Please don't attempt to write a scraper or anything yet, in the interests of minimizing load on the site. It might take a while since bulk download is not the highest priority item, but it will be implemented eventually.

Okay, thanks.