Integration of Physical Domain Knowledge and Machine Learning

Abstract

We will design an overall system architecture for agents that learn to operate in the physical world using compositional representations. We will build in advance knowledge in the forms of: the design of compositional representations, domain-independent mechanisms for reasoning withthem and learning them from data, and very fundamental general models of the underlying physical processes in our world. We will use the built-in structure and knowledge as a substrate for learning, using methods drawn from deep neural networks, Gaussian processes, and probabilistic program synthesis.We believe that by integrating structural prior knowledge with advanced learning mechanisms, we will develop the basis for constructing flexible, adaptive autonomous systems. In addition, these structures can serve as testable hypotheses about human cognition. We will experimentally investigatethe combination of built-in knowledge and knowledge learned from experience in humans, and reflect those findings in the design of our solutions. We will test our solutions in simulated and real robot systems.

Document Details

Document Type
DoD Grant Award
Publication Date
Oct 17, 2018
Source ID
N000141812847

Entities

People

  • Leslie P. Kaelbling

Organizations

  • Massachusetts Institute of Technology
  • Office of Naval Research
  • United States Navy

Tags

Fields of Study

  • Computer science

Readers

  • Agent-Based Social Robotics and Mobile-Assisted Learning in Virtual Environments.
  • Neural Network Machine Learning.
  • Theoretical Analysis.

Technology Areas

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