If AlphaGo Zero had trained solely on AlphaGo games, I would think there was a significant chance that it had learned and was exploiting an AlphaGo weakness. But AlphaGo Zero had never seen an AlphaGo game. It learned Go exclusively by playing against itself. So it seems to me exceeding unlikely that it happened to evolve a strategy that was effective against AlphaGo and not against human players. I've never played a chess grandmaster. But I've played and lost consistently to players who play and lose consistently to masters, who play and lose consistently to grandmasters. It's conceivable, since who-beats-who is nontransitive, that I could consistently beat chess grandmasters. But it's so wildly unlikely as not to be worth thinking about. AlphaGo beat the best human Go player decisively in a match. It's possible that the twelfth best human Go player would trounce Alpha Go. But I know which way I'd bet. Andy On Thu, Oct 19, 2017 at 8:55 AM, James Propp <jamespropp@gmail.com> wrote:

Just because AlphaGo Zero can beat its predecessor, which in turn can beat expert humans, does it follow that AlphaGo Zero can beat expert humans?

https://www.quantamagazine.org/artificial-intelligence-learns-to-learn-entir...

It seems possible to me that AlphaGo Zero could be exploiting weaknesses in AlphaGo that humans might not share.

It's not obvious to me that transitivity applies in this kind of situation. I was surprised that Kevin Hartnett didn't address this question. (Or am I missing something?)

Jim Propp _______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun

-- Andy.Latto@pobox.com