Co-adaptation improves performance in a dynamic human-machine interface

Abstract

Despite the growing prevalence of learning algorithms in daily life, methods for analysis and synthesis of how these systems interact with people are limited. We studied optimization-based algorithms that co-adapt with people in the presence of dynamic machines, finding limitations on current theory that motivated us to conduct an experiment where human subjects interact through a dynamic interface with a machine that has complex dynamics. Experimental results provided evidence of co-adaptation and a trade-off between performance and the "effort" of the human and interface, defined as the norm of their output signals. We developed a parsimonious model of the human adaptation strategy observed in our experiments and conducted a simulation study using this model. Our computational results matched the empirical results, suggesting our human subjects adapted to minimize a combination of error and effort. These results demonstrate how co-adaptation between humans and intelligent interfaces shapes behavior and performance, and introduces a modeling framework that can be used in future work to systematically design interaction outcomes.