The objective of this research is to develop a real-time operating system for a virtual humanoid avatar that will model human behaviors such as visual tracking and other sensori-motor tasks in natural environments. This approach has become possible to test because of the development of theoretical tools in inverse reinforcement learning (IRL) that allow the acquisition of reward functions from detailed measurements of human behavior, together with technical developments in virtual environments and behavioral monitoring that allow such measurements to be obtained.
The central idea is that complex behaviors can be decomposed into sub-tasks that can be considered more or less independently. An embodied agent learns a policy for actions required by each sub-task, given the state information from sensori-motor measurements, in order to maximize total reward. The reward functions implied by human data can be computed and compared to those of an avatar model using the newly-developed IRL technique, constituting an exacting test of the system.
The broadest impact of the project would provide a formal template for further investigations of human mental function. Modular RL models of human behavior would allow realistic humanoid avatars to be used in training for emergency situations, conversation, computer games, and classroom tutoring. Monitoring behavior in patients with diseases that exhibit unusual eye movements (e.g., Tourettes, Schizophrenia, ADHD) and unusual body movement patterns (e.g., Parkinsons), should lead to new diagnostic methods. In addition the regular use of the laboratory in undergraduate courses and outreach programs promotes diversity.
