[Bug-gnubg] Creating GnuBg's own MET

[Ian Shaw]

Neil Kazaross wrote on GammOnLine:
"And speaking of ME's I think it time I use this new GNU build to roll out a ME 
table for a 7 point match. I will need some help from GNU programmers like 
yourself to do this project."

This sounds like a job for another co-operative venture, now that we've 
finished rolling out Joseph's database for now. We would have to start at low 
scores and build up the table to include higher scores. It would be a slow 
start, but as ME values are established for each score it opens up more of the 
higher match scores.

I'd be prepared to co-ordinate the effort, but I think we need to discuss the 
parameters and method beforehand.

How many rollouts are required to get statistical significance? Does it vary 
according the score and maximum possible cube level?
As I understand variance, at DMP the variance is low because the results are 
always +1 or -1. At 7-away 7-away the variance is higher because the possible 
results are [7 4 2 1 -1 -2 -4 -7].
Therefore more rollouts would be needed at longer trials. 

Which ply should be used?
There is the usual trade-off of accuracy vs. speed. I would be in favour of 
doing 2-ply, to get GnuBg playing at it's strongest. I don't think 3-ply or 
above is feasible at today's PC speeds.

How exactly does the Rollout As Initial Position setting work? 
IIRC, it ignores doubles, but doesn't determine who wins the opening roll. We 
would need to rollout the position with each side on roll.

My current thinking about a procedure would be:

DMP is obviously 50% each. 
Create an MET with only this value in it.

Start with 2-away Post Crawford, where trailer has free drop.

Create a script which -  
Rolls out the initial position a set number of times with Player 0 on roll.
Saves results to a text file.
Rolls out the initial position a set number of times with Player 1 on roll.
Append results to the text file.
Combine the results to get the overall ME for the score.
Create a new entry the MET.

Repeat for 3-away Post Crawford and 2-away Crawford.
etc.

Participants could roll out a score, then send the results in to be combined 
into the new MET.

Any comments?

--Ian Shaw