Architectures for Continuously Learning Agents
Abstract
In recent years, progress in artificial intelligence has led to increasingly ambitious efforts to build integrated intelligent agents, from robotic agents that perceive and affect the physical world, to software agents that perceive and affect their cyber-world. Despite significant progress in the component parts of these agents (e.g., progress in computer vision, language processing, etc.), a key open question remains: what kind of software architecture is needed to integrate these components into a continuously self-improving intelligent agent? This is an increasingly important question, as embedded intelligent systems in continuous operation are becoming increasingly widespread in commercial and military systems. Such continuously operating systems, with sensors and effectors that perceive and act on their environment, are exposed to a continuous stream of data that in many cases could be used for automatic self improvement, if we understood how to architect these systems appropriately. Beyond its practical importance, the question of how to architect continuously learning agents is also at the core of the scientific understanding of intelligence. We propose here an experimental and theoretical research program to study this question. If successful, this research will provide new guidance for design of continuously learning sensor-effector agents across many domains, and many embedded systems. Our goal is to produce both theoretically justified design principles and experimental demonstrations of successful continuous learning systems.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 19, 2022
- Accession Number
- AD1184973
Entities
People
- Tom M. Mitchell
Organizations
- Carnegie Mellon University