Artificial agents (including chatbots, robots, game-playing agents, …) can make use of interactions with human experts to acquire new skills in a flexible way. A variety of available algorithms allow human experts to interactively teach agents optimal or near-optimal policies in dynamic tasks. Depending on what the algorithm allows, the teaching signal can be one or a combination of evaluative feedback, corrections, demonstrations, advice, etc. (Li et al., 2019, Celemin et al., 2019, Ravichandar et al. 2020).
Existing research on human-interactive agent learning has focused on designing efficient agent learners, but largely ignores factors pertaining to the human teachers that can have a direct or indirect impact on the outcomes of the agent’s learning process. We propose that modeling inter-individual differences in teaching signals should go hand in hand with designing efficient algorithms on the agent side, as it could help agents explicitly reason about, adapt, and possibly influence different types of teachers.
This project is aimed at identifying how features of an individual’s teaching signal, including the type, timing, accuracy, or frequency of human input, can affect the performance of the agent as well as the teaching experience of the human. Through a series of studies involving human participants, we propose to investigate teaching signal variability in interactions between human teachers and state-of-the-art human-interactive machine learning algorithms, in typical reinforcement learning benchmark tasks. The output of this research will be a collection of models capturing inter-individual differences between teachers that can explain different learning outcomes on the agent side. Such models may unlock new possibilities for designing learning agents that are more efficient, more flexible, and more human-centered.
Researchers: Kim Baraka, Daniel Preciado Vanegas