Re: [Bug-gnubg] Game Statistics

[Christopher D. Yep]

At 12:58 PM 8/10/2003 +0000, you wrote:
> On Sun, Aug 10, 2003 at 02:55:20AM -0400, Christopher D. Yep wrote
> > Within a match, gnubg gives game statistics.  However the last 3 lines of
> > the game statistics may be confusing to users (see below).  While I don't
> > consider it a "bug", perhaps we can debate a possible change.
>
> There is actually a bug. When one or more games are not properly
> "finished", e.g., no resignation is recorded, gnubg will include the
> luck from these games in the luck adjusted result.
>
> Three solutions:
>
> (a) calculate an actual result from unfinished games by using a 0-ply
>     evalution
> (b) exclude luck from games where no winner is recorded
> (c) try to calculate the actual result based on the score from the
>     following game
>
> Without much thought (b) seems to be easiest to implement.

I prefer (a) [calculating the actual result based on the equity at the 
final move of the unfinished game].  This seems like the "best" solution 
for unfinished games in both matches and money sessions.  Note that this 
also handles the case where two players play a live match (or money 
session) which is interrupted in the middle of a game (and analyzed before 
the match is resumed).  (C) will not be able to handle this case, while (b) 
ignores usable information.

Note that (a) is probably easier to implement than (c).  Also, it 
effectively solves Albert's problem with the WC match since presumably all 
unfinished games were stopped when one player was gin to win a normal game, 
gammon, or backgammon.  But additionally note that unfinished games in the 
middle of a match need not be a problem when inputting live matches since 
one can always manually add "resigns" when necessary.

> > 3. The above is from a 7 pt. match.  A 55.17% to 44.83% edge in a 7 pt.
> > match corresponds to a 68.16 point rating edge (i.e. 34.08 to
> > -34.08).
>
> Yes, Kees van Doel has suggested to change this to
>
> FIBS Rating difference:  68.2
>
> instead of the relative ones reported now.

Yes, Kees' idea is probably more intuitive for users.  On the other hand I 
still think it's better to only include the FIBS rating difference for 
match statistics and not for individual game statistics.  If player1 plays 
perfectly, while player2 makes 5.17% mwc errors in a game, gnubg will 
wrongly imply (in this game's game statistics) that the FIBS rating 
difference is 68.2, whereas in actuality it should predict an even greater 
edge for player1 in the game statistics; over a 7 pt. match player2 will 
make far more than 5.17% mwc errors, thus player1 is much more than a 
55.17% favorite over a 7 pt. match.

Chris