[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: [Bug-gnubg] Good beta release! (Suggestions)

From: Richard Anderson
Subject: RE: [Bug-gnubg] Good beta release! (Suggestions)
Date: Sun, 16 May 2004 16:09:57 -0700

> >17. The program missed several obvious hits in a back game - example 
> >attached.  Perhaps it is still using the regular game 
> analyzer in this 
> >position?  Or has it given up and is trying to get off the gammon?
> Ouch! This is one of GNU Backgammons bad sides. Back games are 
> not really good played. If it is any consolation, I can 
> assure you it has been much worse. 
> Other bot, Jellyfish and Snowie doesn't play back games really 
> good either. 
I read Snowie version 3 splits the back men with an opening 2-1 whereas
version 4 slots the five point.  The development team interpreted this to
mean that Snowie 4 plays back games better than version 3, which might have
some truth to it.  Or maybe Snowie 4 plays advanced anchor games better.

> Neural nets are hard to train in these kind of positions.

My naïve guess would be that the net should be trained using a set of
starting back game positions from expert play.  Since there are separate
nets for middle games, races and crashed positions, why not another one for
back games?  Deciding which net applies to a given position might be
difficult, though.

Perhaps what the bots are telling us is that good players don't get into the
defensive side of back games very often and that there is more to be gained
by improving the bot's play in the early moves and normal front games?

I think it is interesting that a bot can tell us what it thinks is the best
play but not why it is best.  I've read some of the "expert" interpretations
of why the bot's non-intuitive play is correct, but I am dubious about the
value of such speculation.  

I think the best way to interpret a bot's unusual play in a complex position
is to manually or with a computer do an exhaustive number of rollouts, then
categorize the critical variations and document how often each occurs.  Then
repeat for the other primary candidate moves and try to make sense of the
statistics.  Perhaps some kind of AI pattern recognition could automate some
of this process.  Reference positions could also help.

Good work, I look forward to future releases.

Richard Anderson

reply via email to

[Prev in Thread] Current Thread [Next in Thread]