Transfer Learning in EEG-based Reinforcement Learning Brain Machine Interfaces via Q-learning Kernel Temporal...

Reinforcement learning based brain machine interfaces (RLBMIs) is an emerging technology with many possible real-time applications. Transfer learning (TL) has proved beneficial as it can improve performance of machine learning algorithms by reusing learned knowledge from similar tasks. However, its application in BMIs has mainly focused on supervised learning approaches. In this study, we investigate the effect of TL in RLBMIs to decode freewill movement related intentions using multichannel scalp electroencephalogram (EEG). We applied TL strategies to Q-learning Kernel Temporal Difference (Q-KTD), which is an algorithm to estimate the action value function, Q, by a nonlinear function approximator using kernel methods. A publicly available EEG dataset recorded while healthy adult participants conduct a key pressing task was used to decode premovement (before movement onset) and movement intention (after movement onset). Differently from most cue-based tasks, participants had freewill to choose the key being pressed, providing unique neural dynamics for decoding. TL was applied between and within subjects to decode the movement related intentions. Significant increase on success rates (p < 0.01) were observed in 96% cases. The success rate increases in each case ranged from 1.39 to 10.69%. These results support the use of TL as an effective way to improve the efficiency of RL-based neural decoder's learning.Clinical Relevance- The improved performance of the neural decoder using transfer learning provides efficient modeling strategy of RLBMIs that can assist patients with neurological disorders.