Warning:
This page is
provided for historical and archival purposes
only. While the seminar dates are correct, we offer no
guarantee of informational accuracy or link
validity. Contact information for the speakers, hosts and
seminar committee are certainly out of date.
Goal recognition is the task of inferring people's goals based on observing them try to achieve their goals. In software domains, goal recognition can be used to enhance automatic help systems and enable software agents to assist people with their current tasks.
Traditionally, plan and goal recognition have investigated examples containing less than one hundred plans and goals. I present three methods, drawn from work in machine learning and planning, for scaling up goal recognition to handle on the order of 100,000 candidate goals. First, I show how to automatically construct the plan library from domain primitives, a task previously performed by hand by a human expert. Second, I present compact representations and efficient algorithms for performing fast recognition on massive plan libraries. Third, I describe an unsupervised-learning algorithm that allows a goal recognizer to adapt to an individual's idiosyncratic behaviors given a sample of that person's recent behavior.
I identify a class of goals for which the goal recognition algorithm requires space and time that is logarithmic in the number of goals. I report on experiments that show my prototype implementation to be fast and show that adaptation can improve goal recognition, in terms of accuracy and coverage, by a factor of two to three.
This appears on the World Wide Web at http://www.frc.ri.cmu.edu/~mcm/seminar.html