Training and applying the human like manager autoregressive

[LBrinkmann]

Our model needs to be able to reproduce autocorrelation of punishments of different participants by the same manager. There is (for instance in the first round) no correlate that could provide such correlation.

Redesigning the human like manager, such that it makes punishment predictions autoregressive. That means, first predicting the punishment of a random participant, then provided that knowledge predicting the punishment of a random next participant and so on.

Tasks

• Refactor evaluation f796d36
• Refactor graph model a71ad00
• Refactor api manager bac4029
• Refactor data 3df766b
• Refactor batch creation
• Refactor graph
• Rerun old training