Machine Learning and Friends Lunch
A Relational Representation for Procedural Task Knowledge
Steve Hart
UMass
Abstract
Generalized task knowledge can be decomposed into declarative and
procedural components. The declarative structure captures abstract
knowledge about the task; e.g. to pick up an object, we must first
find the object, reach to it, and then grasp it. The procedural
structure captures knowledge about how to instantiate the abstract
policy in a particular setting; e.g. we must use our left hand to pick
up the object and use an enveloping grasp.
In this talk I discuss a methodology for learning procedural
knowledge. More specifically, I will discuss how Relational Dependency
Networks can be used to learn the effect of sensorimotor features on
the predicated quality of task success. These relationships can then
be used to choose actions that will most likely produce a desired
behavior. I will give an example where a task expert for picking up
objects is learned from experience with a humanoid robot.