Algorithm

check if neural network has to be used in first iteration of self play or not because it is stupid (random)

Unit Tests

test MCTS
test game logic
test AlphaZero

Improvements

alte Spiele nachträglich visualisieren
evaluation also parallel
store replay buffer with less memory
profile and try to improve run time
parallelize logging
maybe different library than pathos
other NeuralNetwork (does not have to be torch)