A neural network-based chess evaluation library/program

Demo showing the interactive gameplay when the model files (provided in Releases) are in the active directory.

In the "Releases" section, a set of example models has been provided. You should place them in the same flat directory as game.py. Then, to start the interactive gameplay, you can run

python3 game.py

The game will expect you to enter your moves in algebraic notation, with the starting square and the ending square. To exit, you can type Ctrl-C.

Currently, training can only be performed by giving the model a dataset containing games in .pgn format. In the REPL, you can call

>>> from train import *
>>> train_iteratively_on_games_and_export(, )
pgn_file>, <epochs>)

Where le> represents whichever dataset you are using and hs> is the number of iterations over that dataset you want to perform.

This model was based on the ideas proposed in the ConvChess paper with modifications.

The model takes the board input as 12 8x8 grids (represented by a 12x8x8 tensor), with a bitboard-style binary encoding for where each piece is located. The output is given through 2 tensors: the first gives a probability for each square being the starting place of the moved piece, and the second gives the probability for each square being the ending square of a given piece. To yield the best move, these probabilities are multiplied.

The selection of parameters and layers was Ad-hoc but very loosely based on ConvChess. The design was strongly influenced by the paper's design.

Right now, there are 4 layers:

1. Convolution + Linear + Leaky Relu (separated by channel)
2. Convolution + Linear + Leaky Relu (interacts between all 12 channels)
3. Convolution + Linear + Leaky Relu (reduces channels to 1 channel)
4. Convolution + Linear + Leaky Relu (acts on the one remaining channel)

Then, the result is flattened and run through a softmax layer to yield probabilities.

The data used for training was from LumbrasGigaBase, specifically a small subset of the games from 0001-1899. I used a small data set due to limited time and computing resources, but in theory the model should work (and be more accurate) given a larger dataset.