Extended Learning on SOAR

Abstract

The major goal of this project was to develop the science and technology for building autonomous knowledge-rich learning agents - computational systems that have significant competence for performing tasks without human intervention, but also have the ability to learn new tasks and improve almost any aspect of their behavior through learning. Our plan was to concentrate on building agents with a variety of architectural learning mechanisms, including reinforcement (to capture statistical regularities), episodic (to capture experiences) and semantic learning (to capture facts). We also planned to study their integration, including integration with Soar's chunking mechanism (which captures procedural knowledge). The result of this project has been to develop initial versions of episodic and semantic memory, while at the same time creating a robust implementation of reinforcement learning in Soar. By the end of this project, we had implementations of all three learning mechanisms, and each of them integrated with Soar's chunking mechanism. In follow on work, we are refining episodic and semantic memory and creating a complete integration of these learning mechanisms.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Mar 14, 2006
Accession Number
ADA445048

Entities

People

  • John E. Laird

Organizations

  • University of Michigan

Tags

Communities of Interest

  • Autonomy
  • Human Systems

DTIC Thesaurus Topics

  • Accuracy
  • Algorithms
  • Arrays (Data Structures)
  • Artificial Intelligence
  • Artificial Intelligence Computing
  • Artificial Intelligence Software
  • Automatic
  • Coding
  • Computational Complexity
  • Computations
  • Dynamics
  • Intelligent Agents
  • Language
  • Ratings
  • Reinforcement Learning
  • Specifications
  • Test And Evaluation

Fields of Study

  • Computer science

Readers

  • Artificial Intelligence

Technology Areas

  • AI & ML