Utilizing Reinforcement Learning to Optimize Non-invasive BCIs for Robotic Rehabilitation
Author
Madsen, Christian Grønborg
Term
4. semester
Education
Publication year
2025
Abstract
Many brain–computer interface (BCI) systems require a calibration session for each new user. This project explores a pipeline that combines reinforcement learning (RL), Common Spatial Patterns (CSP), and a feed-forward neural network (FFNN) to adapt a pretrained model to new users without calibration. The system starts from pretrained CSP weights—spatial filters that highlight informative patterns—and optimizes them to better fit unseen data; RL guides this adaptation. In a three-class setup (hand, foot, rest), the approach reached accuracies of 45% (hand), 40% (foot), and 14% (rest). For comparison, a standard method, Linear Discriminant Analysis (LDA), was also evaluated on the same data. These results are an early step that lays the groundwork for future improvements to adaptive, calibration-free BCI pipelines.
Mange BCI-systemer kræver en kalibreringssession for hver ny bruger. Dette projekt undersøger en pipeline, der kombinerer forstærkningslæring (RL), Common Spatial Patterns (CSP) og et fremadrettet neuralt netværk (FFNN) for at tilpasse en fortrænet model til nye brugere uden kalibrering. Systemet starter med fortrænede CSP-vægte—rumlige filtre, der fremhæver informative mønstre—og optimerer dem, så de passer bedre til ukendte data; RL styrer denne tilpasning. I en tre-klasses opgave (hånd, fod, hvile) nåede tilgangen nøjagtigheder på 45 % (hånd), 40 % (fod) og 14 % (hvile). Til sammenligning blev en standardmetode, lineær diskriminantanalyse (LDA), også afprøvet på de samme data. Arbejdet er et tidligt skridt, der lægger grunden for fremtidige forbedringer af adaptive, kalibreringsfrie BCI-pipelines.
[This apstract has been rewritten with the help of AI based on the project's original abstract]
Keywords
BCI ; EEG ; Reinforcement Learning ; CSP