Controlling learning and learning control

Friday, 2018, May 25 - 12:00
Brown University

Abstract

In order to act adaptively in dynamic environments, individuals must flexibly control their behavior based on feedback. While it may often be useful to rely on explicit models of the environment, for many real-world problems such models are insufficiently constrained to be useful. Engineers frequently address just this sort of practical challenge by using online control algorithms that are both robust and computationally frugal. Here we test whether these engineering principles can inform our understanding of human behavior, drawing on the theoretical tradition of cybernetics. In the domain of learning, we found that a popular engineering control algorithm (PID control) accurately explained participant’s predictions about a dynamic environment. In the domain of cognitive control, we found that participants parametrically adjust attention based on recent task demands, and that they track predictable task dynamics to improve performance. These experiments suggest that cybernetic control models may help to explain people’s behavior in dynamic environments.

 
Recommended Papers: 
Ritz, H., Nassar, M. R., Frank, M. J., Shenhav, A. (In Press) A control theoretic model of adaptive learning in dynamic environments. Journal of Cognitive Neuroscience. Preprint available at: https://www.biorxiv.org/content/early/2018/04/16/204271.
 
Danielmeier, C., Eichele, T., Forstmann, B. U., Tittgemeyer, M., & Ullsperger, M. (2011). Posterior medial frontal cortex activity predicts post-error adaptations in task-related visual and motor areas. Journal of Neuroscience31(5), 1780-1789.
 
Conant, R. C. & Ashby, R. W. (1970). Every good regulator of a system must be a model of that system. International Journal of Systems Science1(2), 89-97.