Re: OSR [Was: Re: [Bug-gnubg] call for cycles (unix/libux only for now)]

[Joseph Heled]

Joern Thyssen wrote:
> 
> On Wed, Jul 31, 2002 at 01:18:09PM +1200, Joseph Heled wrote
> >
> > BTW, are you aware of the work I done about race doubling?
> 
> Not really. Can you elaborate?

When cube > 1, My cube code estimates the equity by  'a * live + (1-a) * dead'
For contact positions, the best empirical value I found is a = 0.7

For race, I developed another model

  a = M * P * I

  M = a magic scaler (0.0606)
  P = average number of rolls to race end.
      (Take the smaller value computed for the two players.)
  I = redouble incentive

P is obtained from a modified pip count. Wastage is accounted for by adding 2.5
points for every checker on the 1 point, 2 for checkers on the 2 point etc., and
the total is divided by 8.16

I is 'G/L', where G (gain) is (equity after redouble - equity at current cube)
when he wins, and L is same when he loses. I tried to capture the "eagerness" of
the cube holder to redouble.

So, the longer the race or the holder incentive, the more weight you give to the
live value.

How did I arrive at this? I performed cubefull rollouts on many positions, for
each determining it's "best" alpha ('a'). Then I tried to find a formula that
will fit this data. I tried several other factors, but none correlated with the
computed alpha's.

You can see the code under analyze/danalyze.cc (positionFdt)

I would love to hear comments!!

-Joseph

  I 

> 
> > Unfortuanetly, this is probably not applicable since it is
> > incorporated into my doubling algorithm, not GNU's
> 
> Is the code in "sagnubg"? Maybe the concept can be moved (properly
> re-implemented).
> 
> Jørn
> 
> --
> Joern Thyssen, PhD
> Vendsysselgade 3, 3., DK-9000 Aalborg, Denmark
> +45 9813 2791 (private) / +45 2077 2689 (mobile) / +45 9633 7036 (work)