Towards Creating Better Interactive Agents: Leveraging Both Implicit and Explicit Human Feedback

My work aims to create interactive agents that are more effectively able to help people. The way in which people want to be helped can vary based on a number of factors, such as person, task, or time. Thus, an important capability of interactive agents is to be able to tailor their behavior based on a person's preferences throughout an interaction. Typically, interactive agents can learn a person's preferences from explicit feedback, such as evaluative (good versus bad) feedback, corrections, or demonstrations. However, there are downsides to relying only on explicit feedback. Therefore, it would be advantageous if interactive agents could also adapt to a person's preferences based on feedback provided implicitly. Implicit human feedback can include information such as eye gaze, facial reactions, or a person's own choice of actions in a task. This line of research investigates reasoning about both implicit and explicit human feedback together during an interaction. For example, we propose reasoning about implicit human feedback in order to proactively solicit explicit feedback. This could allow an interactive agent to proactively tailor its behavior to the preferences of the person with whom they are interacting.