Saturday, September 3, 2011

Rebel ELO Test Corrections and Results

Running tournaments on diverse test suites gave me more confidence in the correctness of my FEN and PGN parser. This time Chess for Android found bugs in the "Rebel ELO Test". Positions ELO 30, 121, 170 do not set the castling rights for either black or white properly, even though castling is the first given move (that is, the move the engine is supposed to find when used as a tactical test suite; when used as an opening suite, this move is obviously played right away).

The results of (mis)using the corrected test suite as an opening suite in a fast match between the built-in Java engine and Stockfish 2.1 on a Nexus One is shown below.                   

1 Stockfish 2.1    +421/  -4/=27 96.13%  434.5/452
2 Chess for Android  +4/-421/=27  3.87%   17.5/452

Here is one of the sparse victories of Chess for Android playing white from a very advantageous position. Nevertheless, I am happy that it found the right continuation as 1. Qc8+ (given in the suite) Kg7 2. Qf8+ Kg6 3. Qf7+ Kh6 4. Rf6+ Qxf6 5. Qxf6+ yielding a position that even the much weaker engine could win.