Recurrent neural networks for system identification with quadratic constraints. In this package we evaluate the use of generalized sector conditions and how they can be used in system identification.

deepsysid

The configuration was used from [deepsysid][https://github.com/AlexandraBaier/deepsysid]. In contrast to deepsysid in this work we wanted to pull out the training, simulation, saving, loading, ... functionality out of the models and therefore started a new project.

Environment variables

DATASET_DIRECTORY # The folder contains the subfolders train, test and validation RESULT_DIRECTORY # Directory in which artifacts will be stored CONFIG_DIRECTORY # Directory of configuration files, currently it requires a base.json file

Configuration

{ "epochs": "Number of training epochs", "eps": "Epsilon value for numerical stability", "dt": "Time step size", "optimizer": { "name": "Name of the optimizer (e.g., adam, sgd)", "learning_rate": "Learning rate for the optimizer" }, "nz": "Size of the latent space", "batch_size": "Number of samples per batch", "window": "Size of the time window for input sequences", "loss_function": "Loss function to be used (e.g., mse, cross_entropy)", "horizons": { "training": "Number of time steps for training horizon", "testing": "Number of time steps for testing horizon" }, "input_names": [ "List of input variable names" ], "output_names": [ "List of output variable names" ] }

Usage

For training a model you run train_model.py from the scripts directory python scripts/train_model.py --model_name <MODEL_NAME> # e.g. tanh, dzn, dznGen For validation python scripts/evaluate_model.py --model_name <MODEL_NAME> # e.g. tanh, dzn, dznGen

tikzplotlib

In utils/plot.py we use the package tikzplotlib which is outdated and does not work with most recent matplotlib package. You can simply remove tikzplotlib and its corresponding calls in plot.py, then you will only get *.png images.

mosek

For finding initial parameters and projecting the updated parameter if they are infeasible we use mosek in the cvxpy package. Other solver like cp.SCS should also work and can be configured in ConstrainedModuleConfig.sdp_opt.

torch

on newer torch versions the load_state_dict requires the weights_only=True property in src/models/base.py model.load_state_dict(torch.load(model_file_name, weights_only=True))