Updated: May 3
The latest version of the Wasp chess engine brings a strength increase of about 50 Elo points and the Neural Network used for position evaluation is now embedded into the Wasp binary file.
Logo by Wilhelm Hudetz (Austria).
Below is information from John Stanback, who is the author of the Wasp chess engine.
"The Neural Network structure is as follows:
- 1824 Inputs (2x2x7x64 for piece/square combinations plus 32 for material difference).
- Only one hidden layer with 640 nodes, using "leaky ReLU" activation.
- Three outputs, using sigmoid activation. If either side has any non-pawn pieces the evaluation is interpolated from the first two outputs. For king and pawn endgames the third output is used.
The NN was trained on a dataset of about 130M positions taken from games between Wasp and various engines of roughly similar strength (including previous versions of Wasp). The target value used for training is based 50% on game result and 50% on Wasp's score from a 10K node search. The net was trained using a total of about 150 billion samples which took about 56 hours using 24 threads. I continue to add positions to the training dataset and hope that this will give some improvement for future versions.
A few small changes were made to the search. These include modifications to the criteria used for static null move pruning, constants used for updating the move history counters, and late-move reduction criteria."
Source: Frank's Chess Page