Aline Xavier Fidêncio; Felix Grün; Christian Klaes; Ioannis Iossifidis
Non-invasive BCIs are promising assistive technologies, but real-world EEG is non-stationary—often causing performance to degrade over time. One route to robustness is to adapt the system online using feedback signals. In this work we study the limits of a closed-loop BCI framework that uses error-related potentials (ErrPs) as a reward signal for reinforcement learning.
ErrP detection is imperfect in practice, so we systematically examine how false positives and false negatives in ErrP classification affect closed-loop learning and performance. Using both synthetic and real datasets, we evaluate two contextual bandit approaches—LinUCB and NeuralUCB—that map motor-imagery time–frequency features to actions in a binary decision task.
Our results highlight that the impact of false positives vs. false negatives depends on baseline accuracy and exploration behavior, and that careful parameterization and sufficient data can partially mitigate degraded feedback quality.
Reference
In: Giuseppe Nicosia; Varun Ojha; Sven Giesselbach; M. Panos Pardalos; Renato Umeton; Emanuele La Malfa; Gabriele La Malfa (Eds.)
Machine Learning, Optimization, and Data Science, pp. 335–349, Springer Nature Switzerland, Cham, 2026.
ISBN: 978-3-032-21480-5.
Links
BibTeX
@inproceedings{fidencio2026performance,
title={Performance Boundaries for Brain-Computer Interfaces Using Error-Related Potentials and Reinforcement Learning},
author={Fid{\^e}ncio, Aline Xavier and Gr{\"u}n, Felix and Klaes, Christian and Iossifidis, Ioannis},
booktitle={Machine Learning, Optimization, and Data Science},
pages={335--349},
year={2026},
publisher={Springer Nature Switzerland}
}