Meta-Learning State Space Models

This repository includes source code for training and evaluating meta-learning for system identification and control via neural state-space models, proposed in IFAC World Congress paper:

@article{chakrabarty2023meta,
  title={Meta-learning of neural state-space models using data from similar systems},
  author={Chakrabarty, Ankush and Wichern, Gordon and Laughman, Christopher R},
  journal={IFAC-PapersOnLine},
  volume={56},
  number={2},
  pages={1490--1495},
  year={2023},
  publisher={Elsevier}
}

Table of contents

1. Installation
2. How to run
3. Included datasets
4. Pre-trained model weights
5. Contributing
6. Copyright and license

Installation

Clone this repo and create an anaconda environment. Then in that new environment, pull the repository and install the dependencies:

git clone https://github.com/merlresearch/meta-learning-state-space
cd meta-learning-state-space && pip install -r requirements.txt

How To Run

1. Activate the correct environment.
2. Run v1_train_metaLearn_fomaml.py with first_order=True for first-order MAML (FO-MAML), and first_order=False for classical MAML
3. (Optional) Run v1_train_supervised.py for supervised learning with limited dataset.
4. (Optional) Run v1_train_universal.py for supervised learning with all tasks dataset (same data as MAML).
5. Run v1_test_metaLearn_fomaml.py to compute MAML/FO-MAML performance metrics on arbitrary test data, and v1_test_benchmark_fomaml.py to compute MAML/FO-MAML performance on nonlinear Bouc-Wen benchmark.
6. Base-learner network architecture can be changed inside networks directory. Default base-learner is KoopmanNetMixed.
7. All relevant data is in the data directory. Pre-trained weights are in the saved_weights directory. More information on these directories is provided below.

Included datasets

This directory contains the data used for training and testing the meta-learning models. The specific files include:

• train_data/BoucWen.MetaLearn.TrainTestDataset.Simple.mat: This contains the dataset used for meta-training. The dataset $\mathfrak D_{\sf train}$ contains data from $N_{\sf train}=100$ Bouc-Wen systems with 7 parameters ($m_L$, $c_L$, $k_L$, $\alpha$, $\beta$, $\gamma$, $\delta$), where for each system, the parameters are selected via uniform sampling within $\pm$50% of the nominal parameter values. Each system is simulated for $T=12$s at $750$Hz sampling. The excitation $u=120\sin(2\pi t)$ for all systems. The script v1_test_metaLearn_fomaml.py is a sanity check to ascertain the quality of the meta-inference on a system within $\mathfrak{D}_{\sf train}$. This dataset was generated by MATLAB R2024a.

• test_data/BoucWen.BenchmarkDataset.mat: This is the benchmark dataset proposed in https://www.nonlinearbenchmark.org/benchmarks/bouc-wen. We have ensured no system in the meta-training set has these benchmark parameters. The first 20% of that 40960-sample noisy benchmark dataset is assumed available as $\mathfrak D_{\sf test}$. The final 80% of the output trajectory is to be predicted by adaptation, which is implemented in v1_test_benchmark_fomaml.py. This data is available at: https://data.4tu.nl/articles/_/12967592.

Pre-trained model weights

This directory contains pre-trained weights for the meta-learning models. The specific files include:

• benchmark/maml_v1_benchmark.pt: Weights obtained after meta-inference or adaptation phase of MAML meta-model on the Bouc-Wen benchmark data.
• benchmark/fomaml_v1_benchmark.pt: Weights obtained after meta-inference or adaptation phase of FO-MAML meta-model on the Bouc-Wen benchmark data.
• benchmark/reptile_v1_benchmark.pt: Weights obtained after meta-inference or adaptation phase of Reptile meta-model on the Bouc-Wen benchmark data.
• metatest/maml_v1_test.pt: Weights obtained after meta-inference or adaptation phase of MAML model on the meta-testing (randomly varying parameters, not benchmark) systems.
• metatest/fomaml_v1_test.pt: Weights obtained after meta-inference or adaptation phase of FO-MAML model on the meta-testing (randomly varying parameters, not benchmark) systems.
• maml_v1_final.pt: Weights obtained during meta-training the MAML model on the meta-training systems.
• fomaml_v1_final.pt: Weights obtained during meta-training the FO-MAML model on the meta-training systems.
• competitors/reptile_metatrain.pt: Weights for the Reptile meta-model trained on on the meta-training systems. This model is trained in the v1_train_metaLearn_reptile.py script.
• competitors/universal_learned_model.pt: Weights for the universal model trained on all meta-training systems. This model is trained in the v1_train_universal.py script.
• competitors/supervised_learned_model.??percent.pt: Weights for the supervised model trained greedily on the benchmark system. This model is trained in the v1_train_supervised.py script and the percentage of benchmark system data seen by the model is described in the title of the weight e.g. 20 or 80.

Contributing

See CONTRIBUTING.md for our policy on contributions.

Copyright and license

Released under AGPL-3.0-or-later license, as found in the LICENSE.md file.

All files, except as noted below:

Copyright (C) 2024 Mitsubishi Electric Research Laboratories (MERL)

SPDX-License-Identifier: AGPL-3.0-or-later

The following files:

• maml/base_learner.py
• maml/maml.py
• maml/utils.py

was taken without modification from https://github.com/learnables/learn2learn/ (license included in LICENSES/MIT.md)

Copyright (c) 2019 Debajyoti Datta, Ian Bunner, Praateek Mahajan, Sebastien Arnold

SPDX-License-Identifier: MIT

The dataset:

• data/test_data/BoucWen.BenchmarkDataset.mat

was taken without modification from https://data.4tu.nl/articles/_/12967592 (license included in LICENSES/CC-BY-SA-4.0.md)

Copyright (c) 2020 by Jean-Philippe Noël, Maarten Schoukens

SPDX-License-Identifier: CC-BY-SA-4.0