1) I get _Move Error Rate_ as a percentage. What is this a percentage of?
Winning chances given up per game? per move? Percentage of total
"possible-with-best-moves" equity given up? What exactly is a 1.2% error
rate... (p.s. wouldnt it be better just to see a points per game figure
here?)
2) What is the meaning of the cube error amount?
3) You guessed it... what is the overall error rate? obviously somehow a mix
of cube and move error but they are in diffent units... percent and ?
4) What is variance reduction. Why should I believe snowie when it says my
288 game roll out is like 30000 games? What is it doing to get the "extra"
games?
All responses appreciated.
Vince Mounts wrote in message <74i7l3$1jm$1...@camel15.mindspring.com>...
Nothing. It isn't a percent of anything.
I believe that it's the average normalized gammon- and backgammon-price
adjusted total cubeless error per 100 unforced moves.
Unpacking this:
- it's a cubeless error rate. The cube doesn't enter into it in any way.
- it's gammon- and backgammon-price adjusted. Snowie uses its
match-equity table and the current cube level to calculate the gammon
and backgammon prices, and uses these rather than the money values.
- it's normalized. An error that costs you the equivalent of the value
of 100% of a plain win is an error of size 1 -- regardless of the size
of the cube or the match score.
- it's the average of this per 100 unforced moves.
Basically, if you defined the error threshold as 0.0 and then went
through and added up all your errors in a match (add up the figures
in parentheses), divided by the number of unforced moves and multiplied
by 100, you would have it.
>4) What is variance reduction. Why should I believe snowie when it says my
>288 game roll out is like 30000 games? What is it doing to get the "extra"
>games?
Take a look at the Backgammon Galore newsgroup archive, Programming
Section. There are several articles on variance reduction, although
they are somewhat technical.
If you have rolled out a complete game on 2- or 3-ply, you can
trust the number of equivalent games. If you have truncated, especially
after only a few ply, you should regard the rollout as a little less
reliable than the equivalent games says.
David Montgomery
mo...@cs.umd.edu
monty on FIBS