Dealing with Changing Contexts in Myoelectric Control

Myoelectric prostheses approximate the motion and flexibility of biological limbs, especially when compared to their mechanical counter-parts. Machine learning enhances the functionality of these devices; however, in an everchanging environment, the effectiveness of conventional approaches is impeded. We introduce Partition Tree Learning—a method for learning predictions in an ongoing fashion whilst being able to identify and adapt to new contexts automatically. We compare the performance of PTL to that of a stochastic gradient descent learner on a stream of data from a participant actuating a myoelectrically controlled robot arm. In a consistent context both learners’ predictions are comparable. After a context switch, PTL is able to adapt to the change and outperform the gradient descent learner. These preliminary results indicate that PTL may effectively deal with change in real-world prosthetic use, lending its ability to learn over varying situations to the constantly changing environment of powered prosthetics.