Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / GM Hammer video on AlphaZero and some answers ...
- - By Chris Whittington (**) [fr] Date 2017-12-13 15:56 Upvotes 3
This is a long and slow video on AlphaZero, Hammer is admittedly a bit clueless on how it works but the cluelessness is interesting and revealing ....

He observed
a) incredible patience of AZ, I have some ideas on this, more later.
b) the tendency of AZ to play repeat moves once and then not go for the repeat again and play a different move. Hammer kept asking why (he has a chat window open when live). He argued if AZ finds the "repeat move" best, then why doesn't it find it best on second repeat? OK, answer is that AZ neural net is inputted i) a lack of progress count (non pawn moves) and ii) a repetition count, so AZ knows not just board position, but a bit of move history in effect, when a position is trending to a draw, the network output will reflect.

[vid=youtube]
https://m.youtube.com/watch?v=RqU9Q211A68%5b/vid%5d
Parent - By udokaiser (**) Date 2018-01-07 10:32
Thanks for the video. Although many speculations and 1h30 chess talk it's interesting.
Up Topic The Rybka Lounge / Computer Chess / GM Hammer video on AlphaZero and some answers ...

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill