Classification of Movement-Related Potentials for Brain-Computer Interface: A Reinforcement Training Approach

摘要

This paper presents a new data driven approach to enhance linear classifier design, where the classifier obtained through theoretical model, is optimized through reinforcement training, to fit the data better as well as to improve its generalization ability. Applied to motor imagery experiment data in EEG based Brain-computer interface (BCI) applications, this method achieved a rather lower mean squared error of 0.59, and by which our group got a second place in the BCI competition III (dataset IVc).