Generalizing User Models through Hybrid Hierarchical Control

Reinforcement Learning has two main challenges in the field of Human-Computer Interaction. The first challenge is generalization across tasks and environments. The second challenge is to achieve human-likeness. We propose a Hybrid Hierarchical Control framework for pointing tasks to address both challenges simultaneously. In our framework, we separate high-level decision-making from low-level motor and gaze control. This hierarchical structure promotes generalizability. By constraining the low-level control to human-like capabilities we aim to achieve human-like results. Finally, we present some applications that our framework could be used for.