Romann Weber, RPI Graduate Student

Romann Weber, RPI Graduate Student


Many complex tasks can be performed using guidance from visual information alone, and a great deal of research has been concerned with determining the strategies that map optical information to actions—or, at a minimum, determining the sources of optical information being used.  The standard approach to this problem is grounded in traditional methods of null-hypothesis statistical testing, which are of limited value when the hypothesis space is too complex to be known a priori.

 My work employs the fundamentally different approach of using data-mining techniques to directly analyze information-action patterns embedded in subject data during the execution of behavior guided by otherwise unknown strategies.  Within this new framework, extracted patterns that relate optical information to subject actions become candidate control strategies, which can be validated and implemented in computational models.

 In this talk, I will briefly review the problem's motivation and the results I obtained by applying these methods to discrete behavior, specifically binary actions, in the analysis of a car-braking experiment.  In the main part of the talk, I will discuss how I have extended these methods to more general cases and surmounted the considerable challenges inherent in analyzing behavior executed in continuous time.

 Again using a braking example, I will describe two new approaches—the intermittent-control model and the control-margin model—and the results they produce.  I will also outline planned extensions to the continuous-time approach that extend its generality and applicability to the analysis of other behaviors.