Re: [Bug-gnubg] GnuBG faring poorly in back games

[Rod Roark]

I don't know what "Tesauro pub_eval" means, but I can see
how this might be a chicken-and-egg problem.

My impression is that gnubg is overvaluing positions that
have a large number of opposing checkers in its home board.
For playing rollouts, how about using a customized version
of gnubg that applies an adjustment factor to its computed
MWC values that depends on the number of such checkers?

For example let's take Øystein's cue and say that a position
with all 15 opposing men back has MWC of only 40% of what is
computed.  And let's say that B is the number of such
checkers, that MWC is the computed (and overvalued) MWC, and
that MWC2 is the revised MWC that we really want to use as
our evaluation.  We might compute something like this:

  XB = (B > 3) ? B - 3 : 0;
  MWC2 = MWC * (1 - 0.6 * XB / 12);

This is probably too simple because we'd also want to
consider how the back men are spread among their points.
But does this make sense as a general concept?

-- Rod

On Friday 10 June 2005 03:04 pm, Joseph Heled wrote:
> On 6/11/05, Rod Roark <address@hidden> wrote:
> > On Friday 10 June 2005 05:50 am, Øystein O Johansen wrote:
> 
> > 
> > I assume the problem lies with the weights file.  How would
> > one go about making a better one?
> > 
> 
> The biggest problem I see is where to bootstrap from. For the more
> typical style of play, Tesauro pub_eval plays surprisingly well -
> where I define well as "rollouts based on play are better than
> evaluations" - which means you can incrementally improve your
> evaluations, which is basically what I did starting from gnubg 0.0.
> 
> Now, can you write a (simple?) playing strategy which plays this
> position (both sides) that will not contain too many horrible/obvious
> mistakes - i.e such a strategy that we can get meaningful rollouts
> from?
> 
> -Joseph
> 
> > -- Rod