A CNN identified by reinforcement learning-based optimization framework for EEG-based state evaluation.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Electroencephalogram (EEG) data is crucial for assessing human psychophysiological states.
• Deep learning, particularly Convolutional Neural Networks (CNNs), shows promise in EEG analysis.
• Optimizing CNN hyperparameters is vital for model performance but time-consuming.

Purpose of the Study:

• To develop an automated framework for efficient CNN architecture search for EEG data.
• To reduce the time and labor required for expert hyperparameter tuning.
• To enhance the performance of CNN models in EEG-based classification and prediction tasks.

Main Methods:

• A step-by-step framework combining Reinforcement Learning (RL) and Particle Swarm Optimization (PSO).
• Deep Q-Network (DQN) in RL determines convolutional layer depth and connections.
• PSO fine-tunes the number and size of convolution kernels.
• The framework is applied to EEG-based sleep stage classification and driver drowsiness evaluation.

Main Results:

• The proposed framework identified high-performance CNN models outperforming state-of-the-art methods.
• Achieved high overall accuracy and improved root mean squared error in both tasks.
• Demonstrated the efficiency of the step-by-step search strategy in narrowing the search space.

Conclusions:

• The RL-based framework efficiently searches for optimal CNN architectures for EEG analysis.
• This approach significantly saves researchers' time and effort.
• The framework holds potential for broad application in various classification and prediction tasks beyond EEG analysis.