[Bug-gnubg] GNU errors in the Match Stats

[Albert Silver]

I had a funny idea and thought I'd share it. It is expected that when GNU
analyzes itself at the same depth, it will find no error, and thus it has a
perfect zero error evaluation. Thus if a player somehow actually played
*better* in 5 moves, and as well in the others, he would be rated as having
erred 5 times. The thing is that if a move of its is debunked it is not
reflected in the match statistics. It's as if GNU were saying it had meant
to play that move all along, which is false. 

My idea is that if a move of it at 2-ply were debunked by a greater depth
analysis such as 4-ply, or a rollout, the difference of its 2-ply choice
compared to the 4-ply's choice would be reflected in the error stats. I
think that the move's error rate should be added in favor of the player's
FIBS rating and not deducted from its own. The reason is that the
theoretical 2050 FIBS rating it is adjudged is quite valid, and its moves
and results reflect this, so this would make it possible to achieve a higher
rating than its own 2050.

What do you think?

                                        Albert