trying out decision transformers as a Plate Heat Exchange (PHE) controller
- MATLAB script used to generate a large dataset of simulations with varying parameters and random shifts.
- Data stored in csv file
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.
Adapted the existing data collator class (DecisionTransformerGymDataCollator) to work with our specific dataset.
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.
- 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