Hacker News new | ask | show | jobs
by awwducks 3761 days ago
Too early to tell, IMO. The next game will be the bigger game since Lee Sedol has a far better idea what he's up against. If the next game goes like this one did, I would be more inclined to agree with you for the remaining 3.
1 comments
imglorp 3761 days ago
He has the human advantages of adaptability and intuition, the better to try a new strategy next game.
link
krastanov 3761 days ago
Adaptability is not all that human (plenty of machines learn from their mistakes and adapt to new settings). Intuition is so poorly defined that depending on what you mean machines easily have it (heuristics, Bayesian inference, etc) or it is just sufficiently vague of a notion that it does not matter.
link
arcanus 3761 days ago
> plenty of machines learn from their mistakes and adapt to new settings

Curious what you see as examples of this.

> Intuition is so poorly defined that depending on what you mean machines easily have it (heuristics, Bayesian inference, etc)

As a working scientist and a bayesian practitioner, I'm sceptical algorithms have intuition. From my perspective, almost all models that one codifies are extremely brittle and will produce catastrophic failures (or just nonsense) unless the user possesses enough expert knowledge or intuition to a-priori know not to use the model in this regime.

However, I agree with the spirit of the text... go is a well-defined game and adaptability and intuition will be highly limited. For instance, the human can't just turn the board over, or unplug the game!

link
imglorp 3761 days ago
I guess I was referring to strategy specifically. The tactics are probably well in hand for both human and AI.

For the AI, the result of the first match will result in one more game entered in the its database. If it's like chess history, it's probably slanted a little towards that player's history in particular.

But the human player is well aware of the machine studying his strategic history and weighting it. If he's well studied like the chess guys are (is that how go players study?) he could employ a strategy he thinks would be surprising to the AI, or even plan to switch strategies in the middle. If one knows they are playing a pattern matcher, you can try to lead it to a local minimum and then leave it there.

Just speculating :-)

link