Message-Passing Monte Carlo: Generating low-discrepancy point sets via graph neural networks.

Area of Science:

• Computational Geometry
• Machine Learning
• Numerical Analysis

Background:

• Discrepancy measures the uniformity of point set distributions.
• Low-discrepancy point sets are crucial for efficient space-filling in diverse scientific and engineering fields.
• Existing methods for generating low-discrepancy points have limitations.

Purpose of the Study:

• Introduce a novel machine learning approach for generating low-discrepancy point sets.
• Develop a new class of low-discrepancy points named Message-Passing Monte Carlo (MPMC) points.
• Extend the framework for generating custom-made points emphasizing specific dimensional uniformity.

Main Methods:

• Leverage Geometric Deep Learning and graph neural networks.
• Employ a machine learning model inspired by the geometric nature of point set generation.
• Develop an extension for higher-dimensional applications.

Main Results:

• MPMC points demonstrate state-of-the-art performance, significantly outperforming previous methods.
• Empirically shown to be optimal or near-optimal in low dimensions for small point sets.
• Achieved superior uniformity and space-filling capabilities.

Conclusions:

• The proposed MPMC method offers a powerful new tool for generating high-quality low-discrepancy point sets.
• MPMC points provide a flexible and efficient solution for applications requiring uniform point distributions.
• This machine learning approach advances the field of computational geometry and numerical methods.