On the suitability of incremental learning for regression tasks in exoskeleton control

In recent times, a new generation of modern exoskeleton robots has come into existence, that aims to utilize machine learning to learn the specific needs and preferences of its users. A simple way to facilitate a personalization of an exoskeleton to the end user is to make use of incremental algorithms that keep learning throughout their deployment. However, it is not clear, if any standard algorithms are fast enough to keep pace with sudden change points in the data stream, like for example the change in movement pattern from a normal walk to going up the stairs. In this paper, we study how well common incremental regression algorithms are suited to predict such an ongoing data stream. We use both, theoretical benchmarks and real world human movement data, to evaluate how fast an algorithm reacts to change points in the data, and how well it is able to remember reoccurring patterns. The results show that a simple KNN algorithm outperforms all other more sophisticated models.