| |
BGonline.org Forums
Bot Comparison after 60/500 games.
Posted By: neilkaz In Response To: Bot Comparison after 60/500 games. (eXtreme Gammon)
Date: Thursday, 29 October 2009, at 2:44 a.m.
"The upcoming XGR+ is clearly better on cube and will improve also the checker player some. Obviously it has a cost and it is about 4 times slower.
Xavier Dufaure de Citres"
Neil adds that he suspects XGR+ would have scored about 3.5 on the total error scale. It passes the 1.267 pass that XGR thinks is a .998 take and while I haven't seen the other cube errors that XGR made, I can assure you all that during my extensive testing, XGR+ made fewer cube errors than XGR. Checker play is also slightly improved and since XGR+'s statistical C.I is smaller, there's less chance to drift .02 or so away between plays that are really similar.
My expected 3.5 also takes into account that XGR+ probably would have come up with one or two small errors not made by other settings, just do to randomness...ie just running XGR+ on these 60 games would more likely be about 3.4 or 3.45.
First of all, I have to let another cat out of the bag and say that I've seen Xavier's earlier prototypes of XG for several years. I continually stressed areas where the playing engine needed to be stronger. Well.. a few months ago, he did a serious training effort but never mentioned it to me until XG was almost ready for release! We do see the result..a very strong playing engine that is clearly faster than any other bot. Does it have a couple areas where it can further improve its strength? Well...what bot doesn't? But we can see that over the analysis of 60 high level games (Snowie playing GNU for money is certainly high level) it is more than holding it's own compared to established bots. I strongly suspect that it will hold up over the entire 500 game set.
Note that the only setting seriously competing so far with XG 4 ply and XGR is GNU 4 ply and both XG settings are much faster than GNU 4 ply. Note a couple of things:
1) XG 4 ply nomenclature is the same as GNU 3 ply and XG 4 ply is clearly well ahead of GNU 3 ply and not slower.
2) GNU 4 ply filter needs to be expanded to choose at least top 4 2 ply plays within .08 in order to be a better analysis tool, but further slowing it down.
3) Since XG suffers less odd-even play oscillations than does GNU, 4 ply choices have been filtered on every ply unlike GNU whose defaults don't look at odd plies.
4) Perhaps a GNU Roller setting can compete with GNU 4 ply or outplay it and be somewhat quicker.
Disclamer... I have no financial interest in XG.. well I get comped..but that it the extent of it.
You all can now start to see why I endorse XG so much and think that it is a must have for the serious student. The majority of my studying is now done with XG, although sometimes with a GNU match analysis also opened.
.. neilkaz ..
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.