PHE_transformer

trying out decision transformers as a Plate Heat Exchange (PHE) controller

LINKS:

Data Generation and Storage:

MATLAB script used to generate a large dataset of simulations with varying parameters and random shifts.
Data stored in csv file

Data Preprocessing in Python:

Python script used to load and preprocess the generated data from MATLAB.

Handle any missing or invalid data points.
Split the preprocessed data into training and validation sets and store in parquet files.

Custom Data Collator in Python:

Adapted the existing data collator class (DecisionTransformerGymDataCollator) to work with our specific dataset.

Training Pipeline in Python:

Training pipeline using the Hugging Face Trainer class.

Instantiate the Decision Transformer model and the custom data collator.
Configure the training arguments, such as the number of epochs, batch size, learning rate, and optimization settings.
Train and store checkpoints
Visualize training and evaluation loss and based on that chose the right checkpoint for testing.

Testing Integration with MATLAB Simulation:

modify the matlab script into functions for calling from python. State managed on python side, simulation (PHE state generation for each timestep) managed by matlab.
test if the transformer works on controlling a PHE