Neural Low-Discrepancy Sequences

Authors: Michael Etienne Van Huffel, Nathan Kirk, Makram Chahine, Daniela Rus, T. Konstantin Rusch

Overview

This repository contains the code for our preprint Neural Low-Discrepancy Sequences. It provides PyTorch implementations for model training, discrepancy-based fine‑tuning, and reproducible experiments.

Figure: Overview of the NeuroLDS architecture.

Key files & scripts

• scripts/models.py — Main model definition (NeuroLDS).
• scripts/main.py — Create/generate sequences; trains/evaluates NeuroLDS.
• scripts/utils.py — Aid/utility functions (discrepancy losses, seeding, I/O, plotting).
• scripts/hypertuning.py — Hyperparameter optimization via Optuna.
• scripts/smoke_test.sh — Sanity check; run to verify the setup works end-to-end.

Installation

# Recommended: Python 3.10–3.12 (tested on 3.11)
python3.11 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Quick check

Run the smoke test to verify that everything is set up correctly:

cd scripts
bash smoke_test.sh

License

This project is licensed under the MIT License - see the LICENSE file for details.