Core Concepts
The EEG RL-Net model significantly improves the classification accuracy of EEG motor imagery signals compared to the state-of-the-art EEG GLT-Net, by leveraging a Reinforcement Learning agent to make efficient classification decisions within a predefined time horizon.
Abstract
The paper presents the EEG RL-Net, an enhanced model for classifying electroencephalography (EEG) motor imagery (MI) signals. The key highlights are:
The EEG RL-Net combines the strengths of Graph Neural Networks (GNNs) and Reinforcement Learning (RL) to achieve superior EEG MI signal classification performance.
The GNN component extracts optimal graph features from EEG MI time point signals using pre-trained weights and an adjacency matrix with 13.39% density, obtained from the state-of-the-art EEG GLT-Net framework.
The RL component, implemented as a Dueling Deep Q-Network (Dueling DQN), learns to make sequential decisions within a predefined horizon to accurately classify EEG MI signals as quickly as possible.
Extensive experiments were conducted to evaluate the impact of varying reward settings (rright, rwrong) and maximum episode lengths (H) on the accuracy and classification speed of the EEG RL-Net.
The EEG RL-Net achieves an unprecedented average accuracy of 96.40% across 20 subjects, significantly outperforming the EEG GLT-Net's 83.95% accuracy using the optimal mg GLT adjacency matrix.
The EEG RL-Net also demonstrates superior efficiency, classifying EEG MI signals within an average of 2.91 time points (approximately 18 milliseconds) per episode, compared to the EEG GLT-Net.
Stats
The EEG RL-Net model achieves an average accuracy of 96.40% across 20 subjects on the PhysioNet EEG MI dataset.
The EEG RL-Net classifies EEG MI signals within an average of 2.91 time points (approximately 18 milliseconds) per episode.
Quotes
"The EEG RL-Net model showcases exceptional classification performance, achieving an unprecedented average accuracy of 96.40% across 20 subjects within 25 milliseconds."
"Choosing at = 0 indicates the agent's hesitance to classify due to uncertainty, leading to a decision to skip the current state with a minimal penalty until it is deemed ready to classify or the episode ends."