Utilizing Reinforcement Learning to Optimize Non-invasive BCIs for Robotic Rehabilitation
Author
Term
4. semester
Education
Publication year
2025
Submitted on
2025-06-04
Abstract
This project implements a novel RL and CSP FFNN pipeline to try and create a BCI system that can adapt to new users without a calibration session before use. Specifically, the system can optimize pretrained CSP weights so that they fit to new, unknown data, thereby enhancing and adapting the spatial filters. The results showed that it achieved accuracies of 45\% for hands-, 40\% for feet-, and 14\% for rest class. For comparison, an LDA classifier was used on the same data. This project lays out the groundwork for future work and research using this pipeline.
Keywords
BCI ; EEG ; Reinforcement Learning ; CSP
Documents