The Disciple Integrated Shell and Methodology for Rapid Development of Knowledge-Based Agents

AAAI 1999

TL;DR

This paper introduces Disciple-LAS, an integrated shell and methodology designed to accelerate the development of knowledge-based agents. It empowers domain experts to build these agents with minimal assistance from knowledge engineers by leveraging an apprenticeship learning approach and a task-reduction problem-solving strategy.

What Problem Does This Paper Address?

Developing sophisticated knowledge-based agents (KBAs) is a notoriously time-consuming and resource-intensive process. It typically requires extensive collaboration between domain experts and knowledge engineers, creating bottlenecks in knowledge acquisition and system development. The paper addresses the challenge of making KBA development more accessible and efficient, particularly by enabling domain experts to play a more direct role in building agents that can solve complex problems.

Methods

Disciple-LAS employs a novel methodology centered around an "apprenticeship learning" paradigm. The core components include:

  • Ontology Management: A hierarchical representation of concepts, objects, features, and tasks, built using frames and adhering to the OKBC protocol. This allows for structured knowledge representation and reuse.
  • Task Reduction Problem Solving: Agents solve problems by breaking down a high-level task into a series of simpler, elementary tasks. This is achieved through a knowledge base of "task reduction rules."
  • Cooperative Learning: Domain experts teach the agent by demonstrating solutions to problems. The agent uses this demonstration to learn new task reduction rules. This process involves:
  • Cooperative Problem Solver: Guides the expert through the task reduction process.
  • Rule Learner: Generalizes expert demonstrations into task reduction rules.
  • Explanation Generator & Hint Editor: Facilitate the agent's understanding of why a particular reduction is correct, allowing the expert to provide feedback and hints.
  • Rule Refiner: Generalizes or specializes existing rules based on expert feedback.
  • Knowledge Import/Export: Modules to import knowledge from external sources (e.g., via OKBC or KIF) and integrate it into the agent's ontology.

Datasets

Not discussed in the paper. The paper focuses on the methodology and shell itself, illustrating its application with two example agents (a planning agent and a critiquing agent) rather than evaluating performance on specific datasets.

Key Insights and Ideas

  • Empowering Domain Experts: The most significant insight is the shift in KBA development from knowledge engineer-centric to domain expert-centric, significantly reducing the need for specialized AI expertise during the building phase.
  • Apprenticeship Learning as a Paradigm: The paper demonstrates the effectiveness of an apprenticeship model where the expert acts as a teacher, guiding the agent through problem-solving and rule acquisition. This mirrors how humans learn complex skills.
  • Task Reduction as a Universal Strategy: The paper shows that task reduction can be a powerful and generalizable problem-solving strategy applicable to diverse domains, from planning to military strategy analysis.
  • Synergistic Expert-Agent Interaction: The Disciple-LAS system fosters a collaborative environment where the expert and agent complement each other's strengths, leading to more efficient and accurate knowledge acquisition.

Key Results

The paper doesn't present quantitative results in terms of accuracy or performance metrics. Instead, it highlights the applicability and potential of Disciple-LAS by showcasing its use in building two distinct agents:

  • Planning Agent: Capable of generating plans for repairing or bypassing damage to bridges or roads.
  • Critiquing Agent: Analyzes military courses of action (COAs) against principles of war and army operations.

The key qualitative result is the demonstration that Disciple-LAS can facilitate the rapid development of practical, knowledge-based agents by domain experts with limited AI assistance.

Limitations and Future Work

  • Limitations: The paper acknowledges that Disciple-LAS is a shell and methodology, implying that the quality of the developed agent is highly dependent on the domain expert's knowledge and teaching ability. The "intelligence" of the modules is described as a "certain degree," suggesting room for improvement in their autonomy.
  • Future Work: The paper suggests that Disciple-LAS contains many general modules that can be part of any Disciple agent, but also allows for customization. This implies a future direction of refining these general modules and exploring further customization options to achieve an optimal balance between reusability and domain specificity. The implementation in JAVA and Common LISP also suggests portability as a key feature for future adoption.

Summary updated

BibTeX
@inproceedings{1999disciple,
  title = {{The Disciple Integrated Shell and Methodology for Rapid Development of Knowledge-Based Agents}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1999},
}
Figure