Philipp and Ege shared a figure on the TRAIN information flow.
Concrete Inputs for the “Artificial Learner” (AL):
- Task performance.
- Motion and behavior predictability.
- Time dependent importance of the feedback signals.
- Cognitive load (to e.g., reduce the task or model complexity).
- Stress level.
- …
The goal of the AL is to adapt the (1) RL parameters, (2) the skill representation and (3) the feedback controller.
Concrete Inputs for the “Human Instructor” (HI):
- Task performance (the reward score).
- Human/AL entropy estimate (how well the human can explain the AL and vice versa).
https://ieeexplore.ieee.org/abstract/document/7041459 - Transparency?: https://www.frontiersin.org/articles/10.3389/fnbot.2018.00083/full
- …