Network Structure

This page documents the ever evolving network structure of Chrysaora's brain. Chrysaora's network is originally based on that of an experimental chess engine called Giraffe. Giraffe's network structure is outlined in the thesis explaining the engine, on pages 17-21. Giraffe used 363 input features to define a board state. The network was 3 layers, where each feature group was run through one hidden layer first, then the outputs of those were run through the next hidden layer. The layers used ReLu activation functions.

The following sections document Chrysaora's network structure at each major release or change.

V 0.1.0 - Noctiluca

• 0-4: Num of white pieces (PNBRQ)
• 5-9: Num of black pieces (PNBRQ)
• 10: Side to move (1 for White, -1 for Black)
• 11-14: Castling rights (White Kingside, White Queenside, Black Kingside, Black Queenside) (1 if allowed, 0 if not)
• 15-62: White piece slots: Pawn, Knight, Bishop, Rook, Queen, King
• 63-110: Black piece slots: Pawn, Knight, Bishop, Rook, Queen, King

Piece "slots" are three indices long: existence (1 or 0), normalized x position, and normalized y position. Normalized positions are -4-4 labeled from the center of the board.

Once all of these numbers are in place, the tensor is reduced such that the max value is 1, so everything is divided by 8 (as the maximum a position can have is 8, in the number of pawns).