On Monday, July 27, 2015 at 12:32:44 AM UTC-4, greedygammon wrote:
> So how can 2 ply to 3 ply to 4 ply jump around like that double the error
> rate? Since 2 4 and rollout all agree it is less than .050 I have to assume
> 3 ply must be wrong no? So 2 ply can sometimes make a better move than 3 ply?
Our friend Tim Chow from rec.games.backgammon says...
Well, anything can *sometimes* occur---the nets aren't perfect.
Regarding GNU 1.0 2-ply versus 3-ply, there is some conflicting evidence. The Depreli study suggests that 3-ply is usually slightly stronger than 2-ply:
http://www.extremegammon.com/studies.aspx
On the other hand, recently on the GNUBG mailing list, Lucas reported that 2-ply was significantly better than 3-ply in a head-to-head competition. See the following post and its followups:
http://lists.gnu.org/archive/html/bug-g ... 00006.html
The idiosyncrasies of how GNUBG was trained has led to what people have called an "odd/even effect," meaning that 3-ply isn't as much stronger than 2-ply as you might expect it to be. That was version 0.x and version 1.x has supposedly mitigated the odd/even effect, but as I said, the evidence is not conclusive.
> I've been doing 3 ply analysis which takes about twice as long as 2 ply.
> It is worth it?
Hard to say. I personally wouldn't bother. However, I think that a truncated rollout is generally worth it for complicated decisions. See here for some decent truncated rollout settings:
http://www.bgonline.org/forums/webbbs_c ... =134413%FD
---
Tim Chow