| |
BGonline.org Forums
True or false? Future BG bots will significantly outplay XG 2
Posted By: Tom Keith In Response To: True or false? Future BG bots will significantly outplay XG 2 (Jim Stutz)
Date: Monday, 21 September 2015, at 1:50 p.m.
Perfect Play
A time range of 50 (or even 20) years opens up the possibility that backgammon will be completely solved.
There are fewer than 18,528,584,051,601,162,496 positions in DMP backgammon. (Extending to match play increases this number by some factor, but not a huge one.) All you need is a computer with enough memory and you can calculate the exact equity of every position. Now you have a perfect-playing bot. ("Perfect play" will never lose in the long run against any other opponent.)
So your question boils down to this: How much short of perfect play is the current version of XG? I would guess perfect play beats XG much more than 51% of the time in a 15-point match.
Exploiting Play
A much scarier possibility (for XG) is an opponent programmed specifically to exploit XG's weaknesses.
Suppose a programmer had access to XG's neural net and could determine what play XG makes in every possible position. As before, you calculate a database of equities. But this time you use XG's neural net to look forward to opponent's next position instead of assuming perfect play. An "exploiting bot" based on this skewed database will maximize its equity against XG but, in exchange, will give up some equity against other bots. In particular, it will win less than 50% of the time against a "perfect bot."
It wouldn't surprise me if an exploiting bot would win 60%, 70%, or even 80% of the time against poor old XG. It's hard to say because you need to know how many positions XG botches, how badly it bungles those positions, and how easy it is to maneuver the game into a position that XG does not understand. I suspect it is pretty easy for an all-knowning bot.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.