| |
BGonline.org Forums
Very impressive result from DeepMind team - Xavier's take?
Posted By: eXtreme Gammon In Response To: Very impressive result from DeepMind team - Xavier's take? (Casper Van der Tak)
Date: Monday, 11 December 2017, at 2:26 p.m.
I would be interested to see what A0 would yield for backgammon.
Their approach is very similar to what was done with BG: self-training. However, there are a few differences (at least for XG, I cannot talk about the others, but I think it is quite similar).
XG used bear off database from the beginning. XG is using human intelligence for inputs of the NN. XG NN does not consider the cube, this is done as a post process. XG uses many different NN based on the position type. Of course, Google NN is vastly more complex than mine and could avoid any preprocess work.
As for MET, this is a bit outside the process. Chess has also MET: match in 12 games will require them. The obvious example is the last game where you are losing 11 to 12: you need a win to stay alive and a draw is as bad as a loss. If A0 returns just the equity, it will not be able to use MET, to do so it would need to return the % of losing, drawing and winning. For BG it would be more complex if you do not want to rely on calculation as you would need to output winning/loosing X point where X in [1..match length]
But if could see how much better A0BG would be in unlimited that would be a great interest.
One thing I would be interested to see is how it would handle snakes, as I do not expect a stronger computer to go for a deep backgame, I expect that the snake position will almost never been reach by self-play. As a result A0BG may misplay them greatly, but they may surprise us, but it would need to be tested against a human I think.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.