| |
BGonline.org Forums
Bots and basic holding games....(again and long)
Posted By: Ian Shaw In Response To: Bots and basic holding games....(again and long) (Bob Koca)
Date: Monday, 21 September 2009, at 1:51 p.m.
Bob,
AS you probably know, gnubg is trained to minimise errors against a set of positions rolled out cubeless for money. The current reasoning, which harks back to Tesauro's work, is that if a bot can rank chequer plays consistently correctly, then the absolute value of the position can be found by rollout.
However, this has drawbacks, since we require the bot to make decisions without being able to roll them out completely, and we know that its chequerplay is far from perfect, too.
I'm not sure how one would represent the cube info in an efficient manner, considering there are so many score-based scenarios. You would have to have positions in the database rolled out for a variety of scores, which would be a time consuming process.
Any ideas you have would be of great interest. Perhaps for money play you could have a separate net that has three sets of 5 outputs, representing the results for [win, wing g, win bg, lose g lose bg] for [centered, owned and opp-owned cube].
Douglas Zare reports that Zbot is structured for cubeful play, so he and Walter obviously came up with a scheme, but this is not likely to be part of gnubg in the near future, I fear.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.