[Bug-gnubg] Luck analysis, round-off error

[Christopher D. Yep]

Windows XP, GUI version
Version 0.14-devel 1.1282 030818 (build Aug 19 2003)

After installing the above build I played a 64 pt. match, gnubg 
intermediate vs. gnubg advanced.

Here are my luckanalysis settings (unchanged throughout the experiment):

set analysis luckanalysis plies 0
set analysis luckanalysis reduced 0
set analysis luckanalysis cubeful on
set analysis luckanalysis noise 0.000
set analysis luckanalysis deterministic on

Also my MET is the Snowie MET (also unchanged throughout the experiment).

Also note: all evaluations listed below are done with cubeful checker 
evaluation and no noise (i.e. the 0-ply analysis is expert, the 2-ply 
analysis world class).

1. I then analysed the match at 0-ply checker, 0-ply cube.
Luck adjusted result: Advanced: +17.63%

If I save the .sgf file, then re-open it, the luck adjusted result is then 
reported as +17.66%.  If I then Analyse - Clear Analysis - Match and redo 
the analysis, it's back to 17.63% again.  Is this round-off error?

I repeated the above several times (including closing and reopening gnubg 
before opening the saved .sgf file).  Each time the .sgf file is open/read, 
gnubg reports +17.66%.  Each time the analysis is actually performed 
however, gnubg only reports +17.63%.

2. I then did Analyse - Clear Analysis - Match, and repeated the same 
analysis with 1-ply checker (move filter = normal), 1-ply cube (move filter 
= normal).  I repeated this experiment several times as well.

Results: every time the analysis is actually performed, gnubg reports a 
luck adjusted result of Advanced: +17.73%.  Every time the file is actually 
opened however, gnubg reports +17.75%.

3. Finally for completeness I did Analyse - Clear Analysis - Match, and 
repeated the same analysis for 2-ply checker and cube (move filter = normal 
for both).  Now every time the analysis is performed gnubg reports a luck 
adjusted result of Advanced: +33.57%, but when the file is actually opened 
gnubg reports +33.60%.  I repeated this step twice (it takes a long time 
since it's a 64 pt. match) with the same results.

Note that my luckanalysis (and MET) settings did not change throughout the 
experiment.  The luckanalysis should have been done at 0-ply even when I 
was analysing at 2-ply.  Thus all 6 of these numbers (17.63%, 17.66%, 
17.73%, 17.75%, 33.57%, 33.60%) should have been identical.  However 
the  luck adjusted result reported during the 2-ply analysis is quite 
different from the other luck adjusted results.

Are the changes between 17.63% and 17.75% due to round-off error?  There 
are 1437 moves in the file and since gnubg internally stores equities to 7 
or 8 decimal places it seems unlikely that this much error could accumulate 
in just 1437 moves.  What about the 33.57% and 33.60% figures?  I'm 
guessing that gnubg is incorrectly calculating the luck adjusted result 
using 1-ply evaluations (which are different than 0-ply evaluations and 
just by chance are different by ~16% [33% - 17%] over the course of a very 
long match).

I can send the .sgf files to anyone interested, but naturally I won't post 
them to the mailing list.

Chris