Multi-agent Learning

Learning and adaptation is crucial to intelligence, and as our world becomes increasingly populated with myriad agents and robots, designing adaptive, intelligent algorithms will be increasingly important. In multi-agent settings, no-regret methods allow us to avoid making restrictive opponent assumptions and instead focus on the agent's own capabilities using regret as the basic performance metric. We have developed learning algorithms that are able to efficiently consider large sets of reactive strategies, while guaranteeing no-regret and safety value.

Collaborators: Leslie Kaelbling, MIT.

Cognitive AI

We have been working on developing a system we call "Jean". Jean is an architecture that models developmental learning in an autonomous agent platform. Jean acts and learns in simulated 3D environemnts, building up a richer understanding of the world her over time as she interacts with her environment. Image schematic representations, and learning algorithms based on these representations, form the core of Jean's capabilities.

Collaborators: Paul Cohen, Clayton T. Morrison, USC ISI; Robert St. Amant, NCSU.

Mobile Ad-Hoc Networks

We have applied reinforcement learning techniques to problems of routing and movement in mobile ad-hoc networks, and showed that nodes that learn to cooperate to accomplish simple targeting and relaying tasks.

Collaborators: Tracey Ho, Caltech; Leslie Kaelbling, MIT.

Interactive Games

Games and the Internet offer a new means of conducting AI research. Millions of humans are available online, as long as they are entertained, and could serve as teachers for learning agents. Simulation technologies in games have developed to a point where games can serve as a realistic test-bed for AI techniques. We are developing WubbleWorld, an online game in which children can teach a cute learning agent to do varioius tasks in a 3D environment. We are also developing ISIS, a 3D real-time strategy game that serves as a military battlefield simulation environment.

Collaborators: Wesley Kerr, Daniel Hewlett, Shane Hoversten, USC ISI; Travis Ho, NUS.

Natural Language

We are developing an aligned sentence-scene corpus that will allow language learning algorithms to benefit from scene information represented in an image schematic language. This corpus is currently being gathered within the WubbleWorld game mentioned above.

Collaborators: Tim Oates, UMBC; Paul Cohen, USC ISI.