I swear that I read that between the Fan Hui and Lee Sedol games, the AlphaGo team used professional games to train the system, in addition to (or in lieu of?) the KGS games that were originally used. Now I cannot find a reference to that, making me doubt my memory. Can anyone confirm or deny that they used pro games?

I know self-play was the overwhelming component of the system, but I'm talking about the "seed" games that the nets were trained on.

_________________
Occupy Babel!

Hello,

here Demis says that they used strong amateur games for their training: https://www.youtube.com/watch?v=yCALyQRN3hw&feature=youtu.be&t=6h9m48s

So, apparently no pro games were used.

I pointed this out at the time: the way David Silver reacts to Demis's statement and then leans over and whispers to him (plus other contradictory information) makes me wonder if that was actually 100% accurate. My suspicion is maybe it was for the Fan Hui version of AlphaGo but not the Lee Sedol version.

The answer is definitely yes, because amongst KGS 6+ Dan games, some of them must be professionals, right?
Aja Huang said in aga go Congress that he can't reveal which games he trained on, but you'll actually be surprised. Makes me wonder if they actually use even low Dan games for training. Maybe the number of games have higher efficiency than the quality of game.

I don't know what the other contradictory information is, but my assumption was that Silver's whispering prompted Hassabis' followup (after the translation) syaing that even if they had used Lee Sedol games in the training set, those games would be drowned out by the millions of other positions used and have very little effect.

Pookpooi, sure, but I meant adding a corpus that's primarily professional games.

_________________
Occupy Babel!

I believe that the training games used in the version described in the Nature paper were only from KGS.

I suggested that they might use the GoGod data set when I reviewed the draft paper, so it's possible that they added some professionals games for training purposes afterwards.

Unless we get a definitive statement from DeepMind we just don't know.

Jon