Physics-Guided Learning for Sample Efficient Spatiotemporal Decision Making

Abstract

The recent breakthroughs in Artificial Intelligence (AI) and Machine Learning (ML) demonstrate new promises for decision making in complex spatiotemporal environments. Unfortunately, a key challenge to deploy ML in real-world is the large amount of labeled training data required to train complex deep learning models. We propose to develop a sample efficient physics-guided machine learning framework to learn from spatiotemporal data such as trajectories and videos. Our framework will learn complex spatiotemporal dynamics in various combat and non-combat scenarios, including autonomous vehicle tracking and navigation, multi-agent team behavior modeling, and atmospheric turbulence simulation. Our framework will drastically reduces learning sample complexity by infusing common sense physics knowledge into data-driven machine learning systems. Specifically, our scientific approach consists of three major components: (1) physics-guided data sampling: performing active learning guided by physical intuitions and querying data in the most informative spatiotemporal regions (2) physics-guided learning objective: incorporating physical constraints into the optimization objectives to guide the learning algorithms. and (3) physics-guided model architecture: designing novel neural network models that are guaranteed to preserve physical characteristics. Scientific Merit This proposal aims to advance the fundamental knowledge of spatiotemporal reasoning in environments with complex dynamics. The proposed research will lead to novel machine learning methods that can (1) significantly reduce the operational cost of labeling large amount of training data in the field (2) accurately predict the behavior of enemies or autonomous vehicles using rich spatiotemporal data collected from night vision sensors, weapons sights and navigational devices, and (3) tremendously expand our capabilities to model diverse unknown environment dynamics and heterogeneous behavior patterns. Broader Impact Algorithms and methodologies developed in this project will have a major impact on the theory and practice of large-scale spatiotemporal data analytics by integrating new concepts and tools from physics and machine learning. It will further contribute to our burgeoning understanding of spatiotemporal decision making, dynamical systems, large-scale optimization and properties of deep neural networks. Additionally, the PI plans to continue her outreach activities to facilitate dissemination of research findings to the scientific communities, including giving tutorials, organizing workshops, and teaching a graduate course on physics-guided learning. The project will also train a new generation of interdisciplinary scientists, especially those from underrepresented groups, who can break domain boundaries between physical sciences and machine learning, and contribute to the urging needs of homeland security and global resilience.

Document Details

Document Type
DoD Grant Award
Publication Date
Oct 22, 2020
Source ID
W911NF2010334

Entities

People

  • Qi Yu

Organizations

  • Army Contracting Command
  • Defense Advanced Research Projects Agency
  • University of California, San Diego

Tags

Fields of Study

  • Computer science

Readers

  • Neural Network Machine Learning.
  • Systems Analysis and Design

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks
  • Autonomy
  • Autonomy - Autonomous System Control