Re: [Bug-gnubg] new contact net

[Joern Thyssen]

On Tue, Jun 10, 2003 at 09:21:01AM +0000, Joern Thyssen wrote
> On Tue, Jun 10, 2003 at 04:56:41AM -0400, address@hidden wrote
> > In a message dated 10/06/2003 09:53:00 GMT Daylight Time, address@hidden 
> > writes:
> > 
> > 
> > > I estimate that the luck analysis will double the time need
> > > to play a match, 
> > 
> > Interesting. Do you know this for sure or is that a guess?
> > I ask because I'm always looking for ways to speed up analysis and don't 
> > usually have interest in luck,  just errors.
> 
> A 0-ply luck analysis cost 21 0-ply evaluations, which is the same as a
> 0-ply chequerplay decision requires in average (assuming an average of
> 21 legal moves per roll). The cube analysis costs one 0-ply evaluation.

Sorry, this is not correct :-(

A 0-ply luck analysis corresponds roughly to 21 0-ply chequerplay
analysis. Of course there will be overlap between different rolls, e.g.,
42 and 51. 

So a 0-ply luck nalysis will probably cost in the order 5-21 times a
0-ply chequerplay analysis, so it may not be worthwhile for Joseph to
calculate the luck adjusted result for each of the 500,000 matches.

> 
> totally different. Assuming you, on average, analyse two moves on 2-ply
> for each roll, you'll have close to 900 0-ply evaluations per roll for
> the chequerplay analysis. The cube analysis is another 441 0-ply
> evaluations, so the 21 evaluatiosn for the luck analysis is nothing
> compared to 900-1350 needed for the cube and chequerplay analysis.
> 
> Conclusion: for 0-ply analysis the luck analysis costs a factor of 2,
> but for the more common 2-ply analysis it's less than 5%.

I don't think the 5% changes much due to evaluation caching.

Jørn