Umberto, Mert, Petrus, Pierre-Alexandre, Antti K., Aurelién, Petram, Fabio, Sami
Modelling Negative User Interruptions
Also if user and AI have exactly the same settings (identical tasks models) but not share the discount factor, things might go wrong.
Humans do hyperbolic discounting. The task ends (finite horizon): when is accordingly to a random variable.
Hyperbolic and exponential discounts are two common strategies. Something in the middle might work better (how to properly define it?).
User's interrupt decision can be performed by calculating the distance between the expected action and the suggested action. Quantal Response Equilibria is the game played by AI and user in order to keep playing (not intervene) or stop the game (user intervenes).
- Pierre-Alexandre to provide a longer update next week
- Antti O. pointed out this interesting paper