Go Programs and counting...

[jts]

That does not solve the problem because in general the humans (except at the very highest levels of strength) cannot determine dead or alive either. Have you never seen a really difficult life and death problem? (you are asked to determine the status of a group, alive unconditionally or live with ko or seki or dead).

But in the human case each player can be expected to act as his own advocate, and only agree that a group is dead if he sees no conceivable way for it to live. (Or vice-versa.)

[Mike Novack]

But in the human case each player can be expected to act as his own advocate, and only agree that a group is dead if he sees no conceivable way for it to live. (Or vice-versa.)

But we are discussing the time rquired to come to a correct count. Of course two 10 kyu human players might come to an agreement about whether a group is alive or dead but is their answer correct? (assuming that was a 2 kyu life and death problem).

This all began with curiosity about how different computer programs might come up with a significantly different count for the same board position. Not completely different from the situation with humans (those two 10 kyu players counting their game vs a 2 dan player doing it for them)

[jts]

That's fair enough. It's hard to tell what exactly the original poster meant, which perhaps has muddied the waters.

If I asked "Is it hard for a computer to do your taxes?" you might be inclined to respond "no, not at all - there are dozens of cheap programs that will do that for you and never make a mistake." But then I might be disappointed to find out that the computer can't gather my receipts for me and figure out whether my partner is my dependent or figure out which parts of my income are taxable and which aren't.

I think Karakalis and Li Kao and I all see "counting" as a distinct skill, for bots as for humans. For humans, counting can be hard (both in terms of accuracy and bias), and it can be unpleasant, and it can take some time, and it can require practice. But for computers counting, and even the endgame skills that are a prerequisite for an accurate count, are relatively easy. Everything else in Go, and in particular life and death, is much harder for them. However, just as with humans, when they've screwed up a L&D problem their count is likely to be spectacularly off. But if I play a lost game to the end and then realize I had misjudged a L&D problem, I wouldn't say "Argh! Damn, I suck at counting." I would blame my counting if I lost a game by 0.5 because I backed down from the final ko - that's the sort of mistake even mediocre programs never make, because their counting is awesome.

But in your interpretation, "counting" is the same as "being good at Go" - after all, if a computer can count the board after all of its possible candidate moves and never err, it should never lose. And yes, in this conception counting is quite hard!

[Sneegurd]

The original poster meant: "Why do different Go programs calculate different results?". OK, now I understand that finding correct results requires correct (whatever that is - shortest winning path?) continuation of the game - I'd say that is the dark side of go. Not as easy as 0!=1.