Frequency-domain observations on how humans learn to control an unknown dynamic system

This paper presents results from an experiment that is designed to explore the approaches that humans use to learn to control an unknown linear time-invariant dynamic system. In this experiment, 10 subjects interacted with an unknown dynamic system 40 times over a 2-week period. We use subsystem identification to model the control strategies that the subjects employ on each of their 40 trials. In particular, we estimate feedback and feedforward controllers used by each subject on each trial. The controllers identified on the 40th trial suggest that the subjects learned to use the inverse plant dynamics in feedforward. Moreover, the identified feedforward controllers converge to the approximate inverse dynamics in fewer trials (i.e., more quickly) at middle frequencies than at low and high frequencies.