Figure out which data to feed into neural networks as game state and action. Also think about the interface between AI training and game logic and how to keep it efficient.
1) Transform the game log into trainable (state_before, action, reward, state_after, is_terminal) tuples.
2) Make the state as adaptable as possible to support various training techniques while still performing well.
3) Use only a single model for all agents by normalizing the game view for player=0.
4) Think about pre-computing metadata as feature vector to enhance training success.
Figure out which data to feed into neural networks as game state and action. Also think about the interface between AI training and game logic and how to keep it efficient.
1) Transform the game log into trainable (state_before, action, reward, state_after, is_terminal) tuples.
2) Make the state as adaptable as possible to support various training techniques while still performing well.
3) Use only a single model for all agents by normalizing the game view for player=0.
4) Think about pre-computing metadata as feature vector to enhance training success.