Building Agents from Shared Ontologies through Apprenticeship Multistrategy Learning

AAAI 1998

TL;DR

This paper introduces the Disciple Learning Agent Shell, a system designed to build intelligent agents by importing existing knowledge from shared ontologies and teaching agents through apprenticeship. It tackles the challenge of creating adaptable agents by integrating symbolic learning, knowledge acquisition, and inductive probabilistic reasoning.

What Problem Does This Paper Address?

Developing intelligent agents that can learn and adapt efficiently is a significant challenge. Traditional approaches often require extensive manual knowledge engineering or rely on purely data-driven methods that may struggle with generalization and explainability. This paper addresses the need for a more robust and flexible agent development environment by proposing a system that leverages pre-existing knowledge (ontologies) and learns from expert guidance (apprenticeship) in a multistrategy learning framework. The goal is to create agents that can understand and solve problems by importing domain knowledge and learning from examples, rather than starting from scratch.

Methods

The core of the proposed system is the Disciple Learning Agent Shell. Its architecture involves:

  • Ontology Importation: Agents import concepts from shared ontologies using the Open Knowledge Base Connectivity (OKBC) protocol. This provides a foundational understanding of the application domain.
  • Apprenticeship Learning: Domain experts teach agents how to solve specific problems by providing examples and their solutions. The agent then generalizes these examples into problem-solving rules.
  • Multistrategy Learning: The system integrates various learning approaches. Notably, it incorporates:
  • Analogical Reasoning: Agents can hypothesize explanations for new problems by drawing analogies to previously understood problem-solving episodes.
  • Inductive Probability Theory (Cohen, 1989): This approach is used to refine learned rules. Instead of traditional probabilistic measures, it uses a notion of "provability" and a systematic process of experimentation to falsify rules. This helps in identifying conditions under which rules fail and provides a measure of inductive support for the rule.
  • Inductive Probabilities for Rule Selection: The system uses inductive probabilities to guide rule evaluation and selection during problem-solving. The probability of a rule instance is based on its inductive support and the knowledge of its falsifying conditions.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Leveraging Existing Knowledge is Crucial: Importing concepts from shared ontologies significantly reduces the burden of knowledge acquisition, allowing agents to build upon existing, structured knowledge.
  • Apprenticeship as a Powerful Teaching Mechanism: Learning from expert-provided examples and explanations is an effective way to imbue agents with domain-specific problem-solving skills.
  • Inductive Probability Offers a Non-Numeric Alternative: Cohen's inductive probability theory provides a way to reason about uncertainty and rule reliability without requiring explicit numerical probabilities, which can be difficult to elicit or maintain.
  • Synergy of Symbolic and Probabilistic Reasoning: The integration of symbolic learning (apprenticeship, analogy) with inductive probabilistic reasoning creates a more powerful and adaptable agent development environment.

Key Results

The paper describes the development of the Disciple Learning Agent Shell and its architectural components. It provides an example of how the system can generalize numerical information to symbolic concepts and discusses the theoretical underpinnings of its inductive probability approach. While specific quantitative results are not presented, the paper outlines the potential for creating efficient agent development environments through the proposed integration of techniques.

Limitations and Future Work

  • Limitations: The paper does not present empirical evaluations or quantitative results to demonstrate the effectiveness of the Disciple Learning Agent Shell. The focus is primarily on the architecture and the theoretical integration of different learning paradigms.
  • Future Work: The authors suggest that using existing ontologies and synergistically integrating symbolic multistrategy learning, knowledge acquisition, and inductive probabilistic reasoning can form the basis of an efficient agent development environment. Further research would likely involve implementing and evaluating this integrated system on various benchmark problems.

Summary updated

BibTeX
@inproceedings{1998building,
  title = {{Building Agents from Shared Ontologies through Apprenticeship Multistrategy Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1998},
}
Figure