Combination of Reinforcement and Deep Learning for EEG Channel Optimization on Brain-Machine Interface Systems

Brain-Machine Interface (BMI) has enabled a new paradigm for controlling external devices using brain signals. There are multiple types of brain signals, yet a non-invasive electroencephalogram (EEG) is widely used because of its high-temporal resolution and portability. Recent advancements in BMI systems are achieved by accurate decoding of Deep Neural Networks. Several research works show that CNN based classification rate of EEG signals is significant compared to other techniques. In general, EEG is measured on different locations on a subject's scalp. These locations ('channels') provide different information of brain activation. However, several works indicate that some channels on the system contain redundant information. Moreover, large number of channels affect the training time since more computation is required. In this paper, we propose an optimal channel selection method based on a combination of reinforcement learning and deep learning. We use Q-learning algorithm with experience replay for the agent to learn the best combination, while deep learning provides a reward for the agent to evaluate its performance. The environment is translated to fit with a reinforcement learning property. The results show a potential improvement on a classification accuracy while the number of channels is significantly reduced.