[gnugo-devel] 2 sample game teasers

[david doshay]

On Saturday, January 10, 2004, at 05:13  PM, Gunnar Farneback wrote:

> David wrote:
> > We save all game records and look-ahead branch points (but not the
> > lookahead paths) along with the evaluation of the board value at the
> > end of each lookahead. We will be happy to share all records and our
> > code, but would prefer to make a few more improvements first. If you
> > really want to see it sooner we can do that too. We also will be
> > gathering better statistics.
>
> Some game record would be a good teaser.

Here are 2 sample games.

To explain the log file:

Two evaluations of the board are made, OURs and THEIRs (where Gnu Go 
would play as black if we were to pass). In the Dec 28 games we then do 
4 lookahead paths: workers 0, 1, and 2 do the 3 best from our and 
worker 4 does the lookahead if we play the best of their. For this set 
the lookahead depth is 16. At the end of the lookahead Gnu Go's 
influence function is called, and this is the number reported along 
side the move value of the original move. We then normalize both the 
influence and original move values, multiply the normalized influence 
by 0.75 and add to the normalized move value to get the score. We 
select the move with the highest score. It is clear in the log file 
when we choose the "their" move by the array that is populated by the 
"our" values around that point.

If that is not clear then feel free to ask further questions.

For this run of games the scores were:
Game 1: W+19.5
Game 2: B+42.5
Game 3: W+4.5
Game 4: W+49.5
Game 5: W+40.5
Game 6: W+8.5
Game 7: W+49.5
Game 8: W+2.5
Game 9: B+7.5
Game 10: W+42.5

here is game 10

dec-28010.sgf
Description: Binary data

dec-28game10.txt
Description: RTF file

The Jan 11 games were played, as per a comment from Gunnar, with Gnu Go 
(black) at level 15 and our MPI-Gnu Go
with lookahead working at level 10. In these games workers 0 thru 3 are 
doing 16 move lookahead on the top "our" moves, and workers 4 and 5 are 
doing lookahead on the top "their" moves.

The scores from this run were:
Game 1: W+33.5
Game 2: W+40.5
Game 3: B+45.5
Game 4: W+50.5
Game 5: W+1.5
Game 6: B+28.5
Game 7: B+13.5
Game 8: W+31.5
Game 9: W+11.5
Game 10: B+17.5

I chose to send game 5 only because the score was so close. When I 
looked at it I thought it full of strange moves.

jan-11005.sgf
Description: Binary data

jan-11game5.txt
Description: RTF file

As I said earlier, these are all quite preliminary, but perhaps of some 
interest.

Cheers,
David