Re: [Bug-gnubg] luck analysis suggestion

[O.M. Riordan]

> I'm not sure that I agree that the move selection should be done at the
> same level as the final evaluation. First, it would make the luck
> analysis much slower, and, second, I don't see why it's necesary.
> 
> Perhaps I'm overlooking something!

For variance reduction to work well, we need a good estimate of
luck= (a) equity after rolling - (b) equity before rolling,
i.e., of (a) for each roll.
The equity after rolling is the equity after playing the best move, so if
you choose a move on 0-ply and then find the equity after that move on
2-ply you don't get a very good (2-ply standard) estimate of the equity
after rolling.

It will be slower, but I think using the same level to select and evaluate
the move is the consistent thing; presently it is accurately measuring a
rough guess (0-ply move) of the equity, which doesn't really make sense.

In practice, my guess is 1-ply evaluation of move and equity (perhaps with
pruning) is likely to be most useful - this is faster than playing on
2-ply, as it roughly corresponds to analysing the previous move played on
2-ply.

> You forgot to attach the file!

Sorry!

> I did a few trials, and the fix above seems to fix this as well.

Yes - it's fine in the example I forgot to attach as well.

The analysis is fixed for the first move of the match, and probably also
of a game within a match (not so easy to tell: the equity estimate after
the previous move is just a value in the MET, which doesn't have to agree
with gnubg's n-ply estimate of the equity at the start of the next game). 

Oliver