Umberto, Fabio, Antti O., Pedram, Pierre-Alexandre, Petrus, Antti K., Aurélien, Mert
Kashyap - Pedram - Mert - ...
Modelling Made Easy
|Fabio||Human Interaction in Bayesian Optimization: Look-Ahead|
Fabio: Human Interaction in Bayesian Optimization
user learn AI behavior and provide feedback accordingly. Each preference is expressed according to GonzalezICLM2017. In practice we have a matrix of preferences. There are then 2 possible directions: (i) keep using BO and comparing the error between the original GP and the GP that incorporates the model user and (ii) BO-look-ahead acquisition. The latter is one step ahead, based on the Copeland score resulting from the matrix of preferences. Currently there is some instability in the GP classification.
Kashyap - Pedram - Mert: Modelling Made Easy
Practical project: teacher-user model, based on some loss function. Given some issues, the problem has been rethought focusing on interpretability and trust.
From a general and abstract standpoint, the model has 2 agents: the modeler and the assistant. There a 3 spaces: observations (available for assistant), perception (available for the assistant) and model space (available for both assistant and modeler). The assistant will pick a model and the modeler will decide if it is helpful or not.
- Antti O. to show the lineup next week
- Antti K. to provide a longer update next week or 2 weeks