sorin wrote:
Uberdude wrote:
sorin wrote:
Which reminded me of something John wrote sometime after AlphaGo defeated Lee Sedol, I don't remember the exact quote but along the lines of "maybe there is some simple secret to playing Go well, something that humans just overlooked so far".
Is it a secret? Play (and learn from) millions of games and you'll get really strong. Problem is humans tend to die before they can do so.
Ha-ha, good point, but that is about a method to get strong, not about a principle
I think the meaning behind the message I mentioned was that there is a simple underlying principle, which if discovered anyone can apply in realtime and play much better.
I think that humans have discovered most simple underlying principles, and some not so simple ones, as well. Efficiency is an underlying principle, which has some simple examples, but some that are not so simple.
Here is an analogy with extensive and intensive definitions. An extensive definition of
snapback, for instance, is a list of all snapback sequences of play. An intensive definition is a description that allows us to check whether a sequence of play belongs on that list or not. Humans are good at coming up with such descriptions.
Superhuman bots can give us a long list of very good plays, with some errors. Humans, especially if using such bots to check their work, will be good at refining current principles that tell us whether a play or sequence of play belongs on that list (with a few errors), and with coming up with new principles.
Note, however, that such
recognition principles are not quite enough to tell us how to play better. We also have to think of plays that will pass their tests.
There are also principles about how to do that, as well. It does not matter too much if their error rates are higher than the error rates for the recognition principles, since the plays that fail the tests can be eliminated.
Login
Register
FAQ
Search
Uberdude
