Active Transfer Learning (ATL)

Project Description

The ONR Active Transfer Learning program was created to foster research into new systems that interact with human experts to transfer learned knowledge from one domain of expertise to another, such as from charting warm waters to charting icy waters. This builds on prior research we've worked on in Transfer Learning, with an added component of human interactivity. Our responsibilities with regard to the program include performing original research, creating and maintaining experimentation testbeds, and assisting program management with research evaluations.

Accomplishments

We have demonstrated an extension of the ARTUE intelligent agent (see GDA) that is capable of interacting with a human to learn goal selection knowledge. This work may in the future be used to have humans teach robots to respond proactively to problems in their environment. Our experiments show that higher levels of human attentiveness produce faster performance improvement in an agent, but that any level of attention will eventually produce strong performance.