I have now debugged Valkyria with the test positions, 5 bugs was found
directly in the move generation preventing some tesujis to be played.
2 bugs was also found for moves deeper in the search tree. Some bugs
affected more than one position.
With the new version Valkyria does:
1000 35 correct
10000 41 correct
100k 46 correct
out of 50 test positions.
These results are the strongest we have seen sofar. This is because
Valkyria has very heavy playouts. On my 1.4GHz Pentium M 100k takes
about 90 seconds, which means that in real games with a clock it will
not be able to search as deep as other competitors.
Valkyria prunes bad tactical moves with respect to eyshape heavily,
and this was the reason I had so many bugs initially. Some of the
pruning bugs was such that it was correct at least 99% of the time and
would have been very hard to catch without these tests. So once again
thanks to Yamato for the hard work creating the test suit. It confirms
that the strength of Valkyria is tactical fighting at the expense of
positional judgment.
In two of the remaining positions, 127 and 134 I think Valkyria cannot
find the correct move because of a strong bias in the AMAF
implementation. So maybe I can benefit even more from looking closely
at those positions.
Note that the bugs corrected were real bugs. I would have have changed
the move generation also in positions where these moves were not the
correct moves.
-Magnus
--
Magnus Persson
Berlin, Germany
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
