Adaptive Classification for Brain-Machine Interface with Reinforcement Learning

Abstract

Brain machine interface (BMI) is an interface that uses brain activity to interact with computer-based devices. We introduce a BMI system using electroencephalography (EEG) and the reinforcement learning method, in which event-related potential (ERP) represents a reward reflecting failure or success of BMI operations. In experiments, the P300 speller task was conducted with adding the evaluation process where subjects counted the number of times the speller estimated a wrong character. Results showed that ERPs were evoked in the subjects observing wrong output. Those were estimated by using a support vector machine (SVM) which classified data into two categories. The overall accuracy of classification was approximately 58%. Also, a simulation using the reinforcement learning method was conducted. The result indicated that discriminant accuracy of SVM may improve with the learning process in a way that optimizes the constituent parameters.