Re: [Bug-gnubg] Analysis of race rollouts

[Joseph Heled]

David Montgomery wrote:
> > The power (and limitation) of the OSR is that it is one sided, 
> > i.e. plays each 
> > side independently, so that it has no notion if it needs to make 
> > desperation 
> > plays or not. Even if it did, I doubt it is possible to find 
> > those moves cheaply.
> >
> > -Joseph
>
>
> It's true that one-sided racing databases that just have the mean 
> rolls to bear off can't make desperation plays.
>
> But one-sided racing databases with the probability distribution
> of rolls to bear off *can and do* make desperation plays, if you
> choose plays by convolving the two distributions.

But I am computing the probability distribution by playing out each side 
seperatly, ignoring desperation. After that it is too late.

-Joseph

> I believe that it will be very rare for a rollout done in this
> manner to differ noticably from the exact cpw from a two-sided 
> database.
>
> If a normal approximation is used, you can still get desperation
> plays.  Basically the trailer chooses high-variance plays, even 
> those which might have a slightly higher mean rolls to bear off.  
> This must be less accurate than using the actual single-sided 
> distribution, but for positions far from the end of the race, it 
> probably doesn't give up much.
>
> David