[bobfromhuddle]

Heuristics. Really complex heuristics that defy simple codification: the feeling that a particular group of stones just isn't quite safe yet, the feeling that there is weakness in a structure on the other side of the board that can be exploited, the feeling that this corner is too hot right now, so you should definitely extend instead of the hane.

If you're not already a go player, and want to get started, come sign up on online-go.com and ping me, I'm pathogenix.

[eli_gottlieb]

>Heuristics. Really complex heuristics that defy simple codification:

The whole point of a heuristic is that it's a simple rule that works reasonably well, one might even say admissibly as they do in undergrad AI classes, for dealing with an unsolvably complex problem.

Saying "humans use complex heuristics" amounts to just saying, "Humans use some algorithm I don't know."

>the feeling that a particular group of stones just isn't quite safe yet, the feeling that there is weakness in a structure on the other side of the board that can be exploited, the feeling that this corner is too hot right now, so you should definitely extend instead of the hane.

This mostly just sounds like probabilistic, bounded-rational prediction and evaluation of positions, which is what we currently think human cognition is anyway, but hey.

[white-flame]

It's not that humans only _use_ heuristics, it's that humans _create_ heuristics, and seem to be able to optimize the speed of the heuristic with training and use. They're also introspectable to some level, and can be combined with rational observation and feedback.

[Retric]

One possibility is humans actually have more training time in go to fine tune things than computers so far. 6 hours a day * 300 days * 5-20 years is a lot of training, but multiply that by millions of people who don't all come up with great models.

The idea being people with better heuristics end up as better players adding. So, more people effectivly adds more training time backing up the best models.

PS: This also means each player is using a different algorithm while playing.

[tel]

Sure, but there are a lot of them, they involve several layers of metaheuristics to balance different approaches, and nobody has been able to automatically grow something which matches the same success rate as highly trained human heuristic layering.

Which maybe takes some fun out of it, but this does seem to be an area where humans consistently out-perform AI: when local optimizations have to be balanced across many medium to medium-large optimization criteria as well. Similar things happen in language, at least metaphorically.