More versatile training data

Include positions from various stages of endgame. Randomize the number of previous moves shown to the model (from 0 to 7, or in case of the currently played position from 1 to 7).

More strict loss functioin

v0.3b

Safer download process

With secure territories in training

v0.2c

Correct convolutions

Models with previous moves

Models trained on board states including all eight recent positions

New models!

Train only on positions taken from 19x19 games. In one run backbone's weights remained frozen, in the other they were unfrozen after 15K steps (roughly 75 iterations).

v0.1 knowledge distillation

Pre-trained Alpha Zero model on positions from KataGo selfplay. Ownership of local positions trained against KataGo predictions. Segmentation into local position computed from KataGo predictions. First local move taken from the actual game.