Influencing User Learning in Co-Adaptive Myoelectric Interfaces 

As we aim to design more generalizable and customizable user-machine interfaces, we need to be able to couple user learning with machine learning. I propose using online adaptive machine learning methods, that are derived from game theory, to shape how the user learns in closed-loop interfaces. 

Parameterizing User Performance in Adaptive Human-Machine Interfaces

With a novel adaptive decoder, we parameterized user performance of a myoelectric-controlled trajectory tracking task. Subjects were asked to track a pseudo-random 2D continuous task using only their muscle activity (recorded through surface electromyography). We varied multiple decoder parameters to characterize the effect of adaptive decoder parameters on user performance, with the goal of more robust adaptive decoder designs. 

Modeling User-Machine Co-adaptation using Game Theory

 

Neural interfaces face limitations in controllability and long-term viability, hindering the widespread adoption of these technologies. In neural interfaces, both the user and the algorithm (decoder) work in conjunction to control an external system (e.g a cursor on the computer screen, or a robotic limb).  Smarter algorithms that adapt in concert with the user to help the user learn to control this external system present a possible avenue to constructing more robust neural interfaces. These co-adaptive interfaces leverage both user and decoder adaptation but form a two-learner system–the user and the machine are both jointly learning to control an external device. We can leverage existing frameworks from game theory to consider the interactions of two-learner systems. In neural interfaces, game theory provides theoretical tools to model user-machine co-adaptation and analyze interactions based on different computer adaptation parameters.