| |
BGonline.org Forums
Leon, a new very strong backgammon program!?
Posted By: GeorgeMJ23
Date: Thursday, 1 September 2011, at 6:48 p.m.
Yesterday i run into a site that it was referring to a new program for playing tavli. Tavli is the Greek game that it contains Portes(Backgammon without doubling cube and without backgammon), Fevga and Plakoto variants.
Its author says it's a program that used training of artificial neural networks so it's very very strong, for both 3 AI's for Portes, Plakoto and Fevga. Here is the official page: http://www.leon.tournavitis.de/ The program is free for only 7 days and after that only if you buy it will work.
Anyway my point is not any of the above, but the claim of the author that after 10000 games against latest GNUBG, it did beat GNUBG with a bit more than 0.04 points per game(i guess since Leon plays only Portes and not Backgammon that these were single games at 1). If my math don't betray me this means something like a 5200-4800 score for Lean against GNUBG. Is this significant? OF COURSE this score means nothing if luck factor doesn't get extracted from the games. any opinions on this. I find it amazing if there is another (new) program that got close to GNUBG.
So my question is(because i have many years to involve in Backgammon bots comparisons so i forgot about the "numbers" etc. It would be easy to look up on my notes but it's easier to ask here :-) ) is this 5200-4800 score conclusive? Or the luck factor can be so significant, even after these 10000 games, that it's totally inconclusive?
I have to notice that the play of Leon in Plakoto and Fevga is mediocre and a pretty well, respectively and this is because of some huge errors it makes from time to time(and not so rarely) especially in Plakoto. E.g in a game of 50 moves in Plakoto it would play the 45 of them for example, in a very very good level indeed, but it will play some kind of moves(that i brought to notice in its author) extremely badly!
I played 2 Portes(Backgammon) games with it and let ExtremeGammon 2 to analyze and it rated Leon as Supernatural with only a tiny teeny equity loss. In fact one of Leon's moves that XG2+ rated as second best, XG2++ rated it as the best. I'm very puzzled by all these and how such a new try, i.e a new "backgammon" program, can achieve such a good results.
But unfortunately the program sucks in terms of analyzing abilities and features as you can just play against it and not even analyze moves, etc. Not even to speak about having a rollout ability. There is no export feature also and this is the crucial thing that makes finding its strength almost impossible.
Its author of course is not new in the area and he is the author of another extremely strong Reversi(Othello) program, Herakles.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.