Re: [Bug-gnubg] GNU notes

[Jim Segrave]

On Mon 16 Sep 2002 (10:33 -0300), Albert Silver wrote:
> 
> a) I have been having quite a bit of fun playing against GNU (WC++ as
> the ego would allow no less) and then analyzing my match with it. I have
> also saved it in the Jellyfish format so that I could also analyze it in
> Snowie 3 (S4 is on the way) and compare notes. There are times when
> Snowie earmarks one of GNU's move as a mistake. Sometimes it is dead
> wrong as rollouts show, which is Snowie's problem, and sometimes it is
> GNU that is wrong. Is there some place to send such positions so that
> furture versions of GNU might avoid repeating the same mistakes?

I know that some work is being done on looking at increasing the
strength of gnubg. I'm not sure who would be the best for contacting
on this (my expertise on neural nets is about the same level as my
backgammon play - "student" would be a kind expression)

> b) I like to keep a folder of interesting positions, whether they be
> problems for me, or for GNU. The problem is that many of these positions
> originate in games, and I don't want to save the entire game or match
> just for one position. Not just for reasons of space (hardly a problem),
> but because accessing the position means going through the entire game
> or match until I find the right moment. There should be a way to save
> just the position at any given moment.

The match ID and position ID should be sufficient (so far they have
been for me). It encodes the length of the match, current score,
current cube level and ownership, board position, whose roll it is
and, if the dice have been rolled, what the roll was. It should not be
necessary to step through a game, much less a match to reach a
position.

Where match ID and postion IDs aren't available, you can start a match
of the appropriate length, then use the edit button to set the chequer
position and Games-Set Cube/Player/Dice to set the rest of the
situation. This allows you to take examples from published matches.

> c) I would also like to be able to save cube decision problems. Right
> now I can't exactly. I can start a match, set up a position, and save it
> with no dice played, but when I open it the dice are always played even
> if both players are human. Even if it is because the dice are set
> automatically, it should only roll them when I ask it, as I might easily
> want to consider turning the cube.

If I set up a match with a position etc, but don't set the dice, then
I can use hint and eval to get gnubg's estimate of what the correct
cube decision is. I just tried setting up an arbitrary position in a
match where I owned the cube and it was my turn. I tried hint and
evaluate to look at cube decisions. I then pressed edit, moved some
chequers around, unclicked edit and got new hints and evaluaations.
> 
> > 
> > > 3)                   The statistics do not take into account changes
> in
> > > evaluation on moves made after a match analysis. On a few occasions
> I
> > > have asked GNU to roll out a move to confirm its criticism and have
> seen
> > > it change the evaluation. Soemtimes for the better, and sometimes to
> > > tell me I am a bigger bonehead than it had initially thought. The
> > > statistics window (and the Game Record where the punctuation
> commentary
> > > is) do not update to take these changes into account. They should of
> > > course IMHO.
> > 
> > Yes - I've noticed this. Of course running Analyze Game again is very
> > quick and does update things, but this is sub-optimal. But I see Joern
> > has already flagged it as a mis-feature.
> 
> Won't that just give the same results? After all, the changes are made
> by examining only a few select moves at a greater depth (3-ply or
> rollout for example). If I analyze the whole match at 2-ply, won't it
> just show the same results?

No, if you've done rollouts or higher ply analysis of some
possibilities on a particular move, they are preserved. I just was
playing with this last night. I had a move flagged as doubtful, but I
disagreed (alright, I wanted to make my results look better), so
instead of 2 ply, I did a 4 ply analysis of my move. It leapt to the
top of the move list, but was still flagged as doubtful. Running
Analyze game removeed this blot on my record and still showed the 4
ply analysis of my move. The sad part is I then went and did a 4 ply
of gnubg's originally preferred move and gnubg had it's revenge by
marking my move not merely doubtful, but bad. Running analyze game
again had a '?' showing, still, a bit of hubris is good for the soul. 

[snip]

Your three styles can be made into 3 small command files by
simply discarding every line except the 'set appearance board' from
your .gnubgautorc files. Then doing:

  load commands "Boards/Red and Green"
or
  lo co "Boards/Red and Green"
or 
  File->Open->Commands and selecting Boards/Red and Green

will change the board appearance, leaving everything else the same.

I just tried changing through files like this from your three boards
and my preferred one during a game and it worked just fine.

There is a bug in the current release - the line for setting the board
appearance works in .gnubgautorc, but the set board appearance line is
missing a space:

.........shape=0.44chequers0=#FFFF00;...
                   
needs a space in front of the 'chequers0='. I managed some ghastly
colour combinations while experimenting, since I could no longer
change the colour of player 0's chequers.


-- 
Jim Segrave           address@hidden