lantonov
commented
7 years ago Completed the first tests of the NN (67, 128, 64, 10) with actual chess positions. The results are encouraging. With 10925 positions and using the common logistic activation function, it starts from 3545/10925 (32.45%) and reaches to 10678/10925 (97.74%). With the ramp function, the same as used by Moha in his experiments, starting from 3545/10925 (32.45%) it reaches 10493/10925 (96.05%). The progress depends upon the settings of hyperparameters (eta and lambda), especially in later stages (above 94%) which take also the most time. Before this, I made a helper program to extract positions from TCEC games (these 10925 positions are from TCEC 9 Superfinal) and their evaluations in separate files. The program ignores positions from the opening, for which there are no evaluations. Another helper program converted these FEN positions to format convenient for input into NN (67 nodes). A third helper program converted evaluations into categorical variables (10 categories). This was necessary because I wanted to use exactly the same NN which was used for the MNIST data (which are categorical). A preprocessor program converted the input arrays into explicit numpy arrays exactly the same as those used for the MNIST data. All programs (helpers and NN), I wrote in Python at which I am a little better than at assembly.
Ipmanchess
tthsqe12
I made a simple plain vanilla NN in Python with adjustable layers. This is its take on the MNIST data set with 30 hidden nodes:
With optimisation of hyperparameters and regularisation (L2) I can make it reach 99%. To start a serious work, first I need to make some helper programs to extract and work with FENs and evaluations, translate them in the necessary input format and create a database to be used for network training and testing. Also, these discussions, including the closed issues will be moved to Wiki in respective topics to be easier to find them for reference.