Multi-level PEnet

Official implementation of Multi-level PEnet (MLPEnet) for parameter estimation in stochastic differential equations driven by Gaussian and non-Gaussian Levy noise.

This repository focuses on the alpha-stable OU experiment from the paper and provides the model code, experiment configs, checkpoint-based evaluation, and training/evaluation entry points.

Usage

Install dependencies:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run a small end-to-end smoke test:

bash scripts/smoke.sh

Primary paper-scale config:

• configs/alpha_ou.yaml

Checkpoint evaluation:

python scripts/reproduce.py \
  --checkpoint checkpoints/alpha_ou.pt \
  --output-dir outputs/repro \
  --fixed-grid

This command regenerates 5,000 alpha-stable OU evaluation paths, loads the released checkpoint, runs inference, and writes prediction/residual figures plus metrics. Use --force-data to regenerate an existing output dataset.

Gaussian OU is included as a sanity check. This public release focuses on the alpha-stable OU experiment.

Checkpoint-only runs write generated data, predictions, figures, and metrics under the selected --output-dir. Training runs write data, logs, checkpoints, and metrics under the selected --work-dir.

Repository Structure

• mlpenet/: model, data generation, training, and evaluation package.
• configs/: smoke and paper-scale YAML configs.
• scripts/: CLI entry points for data generation, training, evaluation, and figure generation.
• README_REPRODUCIBILITY.md: experiment and evaluation notes.
• docs/IMPLEMENTATION_NOTES.md: implementation notes for data and model conventions.
• docs/CLOUD_RUN.md: GPU-server runbook for pilot and full reproduction jobs.

Citation

If this repository is useful, please cite the associated paper:

Shuaiyu Li, Hiroto Saigo, Yang Ruan, Yuzhong Cheng.
Multi-level PEnet: A Robust Three-Stage Model for Parameter Estimation in Non-Gaussian Noise-Driven Stochastic Differential Equations.

License

MIT License. See LICENSE.