Reproducibility File: PSO-MLP Model for Intelligent Assessment of Student Learning

1. Updates
- Provide full environment setup for replicating experiments.
- Include command-line examples for preprocessing, training, evaluation, and visualization.
- Incorporate support for multimodal physiological signal processing.

2. Environment (Python 3.9)
We recommend using Anaconda to create a dedicated environment for this project:

# Install Anaconda
wget https://repo.anaconda.com/archive/Anaconda3-2022.05-Linux-x86_64.sh
bash Anaconda3-2022.05-Linux-x86_64.sh

# Create and activate the environment
conda create -n psomlp_env python=3.9
conda activate psomlp_env

# Install dependencies
conda install pytorch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 -c pytorch
pip install numpy==1.23.5
pip install pandas==1.5.2
pip install scikit-learn==1.1.3
pip install matplotlib==3.6.2
pip install seaborn==0.12.1
pip install scipy==1.9.3

3. Dataset: DEAP
Download the DEAP dataset from:  
https://paperswithcode.com/dataset/deap  
(Original DOI: 10.1109/T-AFFC.2011.15)

Place extracted files under:  
`./data/deap/`

4. Preprocessing

# Run preprocessing (EEG filtering, normalization, interpolation, segmentation)
python preprocess.py \
--data_dir ./data/deap/ \
--output_dir ./data/processed/ \
--window_size 1 \
--normalize zscore \
--fill_missing linear \
--eeg_bandpass 4 45

5. Feature Extraction

# Extract features for each modality (EEG, GSR, HR, facial)
python extract_features.py \
--input_dir ./data/processed/ \
--output_dir ./features/ \
--modalities eeg gsr hr facial \
--segment_size 5 \
--overlap 0.5

6. GAN-Based Augmentation (Optional)

# Generate balanced samples for underrepresented emotion classes
python augment_gan.py \
--input_dir ./features/ \
--output_dir ./features_augmented/ \
--labels_file ./data/labels.csv

7. Model Training: PSO-MLP

python train_pso_mlp.py \
--features_dir ./features_augmented/ \
--num_particles 30 \
--max_iterations 50 \
--hidden_layers 2 \
--dropout_rate 0.14 \
--batch_size 95 \
--activation relu \
--optimizer pso \
--save_model ./checkpoints/pso_mlp_model.pth

8. Evaluation

python evaluate_model.py \
--model_path ./checkpoints/pso_mlp_model.pth \
--test_data ./features/test_set.npy \
--metrics accuracy precision recall f1 confusion_matrix

9. Visualization

python visualize.py \
--metrics_file ./results/metrics.json \
--confusion_file ./results/confusion_matrix.npy

10. Notes
- Label mapping: 0 = Exhausted, 1 = Negative, 2 = Active.
- All preprocessing and training configs can be adjusted in `config.yaml`.
- The PSO-MLP architecture used 14 input neurons, 2 hidden layers with ReLU, and was optimized using inertia weight ω = 0.5, acceleration constants c₁ = c₂ = 1.
- The code supports reproducible runs using `random_seed` flag.