Re: [Bug-gnubg] Re: Adaptive doubling

[Joern Thyssen]

On Mon, Sep 29, 2003 at 06:09:41AM +1200, address@hidden wrote
> 
> The real problem is that there is no known method of adaptive doubling 
> thayt works. I had a failed try with mgnutest some timer back. Also, when I 
> asked two FIBS "Experts" (whose names I can't recal) if doubling late weak 
> opponents makes sense, one said yes and the other said early makes sense. 

(1)  you want to double late to avoid playing for many points
     (reduce the variance)
(2a) you want to double early because your opponents chequer play errors
     are multiplied with the value of the cube
(2b) you want to double early because your opponent will not be able to
     use the cube efficiently (this should shift your doubles toward the
     dead cube double point, as you'll often effectively be handing over
     a dead cube)

The big question is how 1, 2a, and 2b should be weighted and calculated.

> 
> What you have to realize is that changing the met table is just a part of 
> the puzzle. It give you the equity on the Drop part. On the Accept path you 
> need to figure out the expected gain from op errors until the end of the 
> game and add them up to the Accept equity. 

Yes, this is the tough part.

A simple minded approach would be to use the previously calculated error
rate per move and multiply this with the expected number of moves left
in the current game. I think the error rates should be split into
different classes (bearoff, race, contact) for this to work. The error
rates should be much lower for bearoff and race (although you may see
some weak players playing very bad in gammon-save races).

There are so many sources for errors and unknowns in this, that I
personally doubt it'll be possible to get anything useful out of it.

Jørn

pgpiNvGPL7DX4.pgp
Description: PGP signature