|Stochastic Systems Group|
Professor Leslie Pack Kaelbling
Artificial Intelligence Laboratory
Most machine learning systems only work in highly constrained environments, with carefully engineered input representations consisting of the values of a set of state variables. Domains in which the agent interacts with a set of objects (boxes, keys, cups, etc.) cannot be effectively encoded in this way. In this talk, we investigate methods of representation and learning that allow an agent to learn to generalize over objects without getting trapped by logical intractability. As a concrete example, we would like eventually to build a robotic system capable of learning and using the information that "If object A is on object B and I move object B, then probably object A will move, too."
I'll report on a set of experiments that completely failed to work, try to explain why we think they failed, and tell you what we're working on now.
Problems with this site should be emailed to firstname.lastname@example.org