The University of Arizona

Home     About     Colloquia     People     Research     Undergraduate Program     Courses     Get Involved     Support     Contact Us

SISTA Colloquium

CategoryColloquium
DateMonday, December 5, 2011
Time12:00 pm
Concludes1:15 pm
LocationGould-Simpson 906
DetailsLight refreshments in Gould-Simpson 9th floor atrium before talk.
SpeakerPaul Ruvolo, Ph.D. Student
School/Dept.Computer Science & Engineering
AffiliationUniversity of California, San Diego

Inverse Optimal Control for the Computational Analysis and Synthesis of Social Behavior

Algorithms from the field of stochastic optimal control provide principled approaches to programming robots that can effectively negotiate the unpredictable worlds of daily life. These approaches take as input a description of a task that the designer would like the robot to solve as well as a reward function defining the goal of the robot, and in turn compute a behavioral policy that maximizes the robot\\\\\\\'s expected reward. Recently, there has been a surge in development within the field of machine learning of new algorithms for a related problem called \\\"inverse optimal control\\\". In this setting, the algorithm takes as input a task description along with examples of a decision making agent (i.e. a robot or a human) solving the task in an optimal fashion with respect to some unspecified reward function. Given this input, the algorithm then infers the reward function (or goal) that is most consistent with the observed behavior of the agent. While the approach has been successfully applied to the problem of programming robots to perform new tasks from human demonstrations, these tools also have an enormous unrealized potential for behavioral scientists to understand human decisions in new ways.

In this talk I will discuss my work in both extending the computational field of inverse optimal control as well as applying algorithms from the field to the task of understanding human behavior. Specifically, I will talk about two studies conducted in collaboration with a team of developmental psychologists at the University of Miami. In the first study I used inverse optimal control algorithms to study the emergence of social behavior in early infancy. The patterns of timing of infant smiles were used as a signature to compute the most likely reward function of infants engaged in face-to-face interaction with their mothers during their first four months of life. Or more colloquially we asked, \\\"what are these infants trying to do?\\\" In the second part of the talk I will discuss a class of control problems that can be used to model a wide-range of mechanical and biological sensorimotor tasks. I will show my work for deriving methods both for inverse optimal control and goal-based imitation for this class of problems. I will conclude by showing some preliminary results of applying this new framework to the analysis of a database of mother-infant interaction (with dense motion capture on both mother and infant) that we are currently collecting in collaboration with the University of Miami.