Saturday, March 17, 2012

Murphys law

As I now do some code fine tuning in iCE I run long test series to see that the changes does not make the engine weaker. I started to see iCE losing some games because of illegal moves.

I was absolutely dumbfounded by that. The move generator of iCE passes all perft test, I use a 64 bit hash signature to make undetected collisions very unlikely and as an additional safety belt I verify each move from the hash again for legality. So illegal moves should just be not possible.

After some investigation I found that is was always the same illegal move in the same position and it actually came from the new opening book of iCE. So when I assembled the book I generated an illegal move and put it in the book.


In this position the move to play was specified as 5. bxa6. The iCE book creator translated that move into Pawn from b2 to a6. So when it looked what pawn actually captures on a6 it picked the wrong one. The reason was an uninitialized bitboard and this was a quick fix.

But it still did not explain why iCE actually used that move and did not detect it as illegal. After debugging I found that the move validator had also a small bug (a bracket was closed at the wrong spot) that just effected white pawn captures and so the move went ok through validation.

I fixed that bug too and corrected the books but it showed "whenever something can go wrong it will".

No comments:

Post a Comment