Content area
One of the key obstacles to the rapid adoption of non-invasive Brain-Computer Interfaces (BCIs) for Motor Imagery (MI) is the low signal-to-noise ratio, and the substantial data requirements which can be mentally taxing for users. EEGNet, a compact Convolutional Neural Network (CNN), has long been considered the state-of-the-art (SOTA) for MI classification, demonstrating strong performance even with limited data. However, recent studies advocate for integrating Deep Reinforcement Learning (RL) to further enhance classification accuracy by dynamically optimizing feature extraction and decision-making processes. Despite this potential, practical implementations remain scarce due to challenges in stabilizing RL training and adapting it to noisy EEG data. This work addresses these limitations by extending the Shallow Convolutional Network, a SOTA model for EEG classification, with a Stochastic Policy Gradient (SPG) policy. SPG was chosen for its ability to handle continuous action spaces, making it well-suited for working with EEG signal representations. The proposed approach uses reward-driven optimization to adaptively enhance feature selection and classification performance. Using the publicly available BCI Competition IV dataset, the model performs within-subject MI classification, achieving results that are comparable to, and in some cases exceed, existing SOTA approaches.