An Experiment in Direct Knowledge Acquisition

Peter W. MullarkeyAAAI 1990

Introduction

In the field of oil exploration, monitoring sensor data from boreholes is a high-stakes task requiring years of specialized training. Historically, capturing this expertise in artificial intelligence systems was hindered by the "knowledge acquisition bottleneck"—the difficulty of translating human intuition into machine-readable rules via a middleman "knowledge engineer." This paper presents the Logging Quality Monitoring System (LQMS), a knowledge-based system for diagnosing anomalies in oil-well logging data that bypasses this bottleneck through a "direct knowledge acquisition" approach.

The OSR Framework and Graphical Representation

The core of LQMS is the Observation-Situation-Relation (OSR) network. This framework moves away from traditional rule-based logic in favor of a graphical representation that aligns more closely with how domain experts conceptualize problems. In an OSR network, Situations (hypothesized causes or tool failures) are connected to Observations (specific patterns in sensor data) via Relations such as AND, OR, and COMB (a weighted average).

The Observation-Situation-Relation (OSR) Editor interface, showing the hierarchical structure of diagnostic logic.

This representation allows experts to approach diagnostics from two perspectives: "symptomatically" (starting from an anomaly to find a cause) or via "failure models" (predicting data patterns from known tool issues). The OSR networks provided a "sweet spot" of representation—expressive enough for complex diagnostics yet intuitive enough for engineers without computer science backgrounds to manipulate directly.

System Architecture and the Expert Workbench

LQMS utilizes a two-tiered architecture consisting of a Signal-to-Symbol Module, which monitors raw sensor streams for anomalies in real-time, and an Analysis Module, which uses OSR networks to generate causal explanations for those anomalies.

To enable direct acquisition, the system includes a graphical workbench. This editor allows experts to build and refine networks using a "box and link" interface. A pivotal feature of this workbench was an integrated edit-tracking and versioning system. By providing a "safety net" where mistakes could be easily undone, the system reduced the psychological barrier for "computer-naive" experts, fostering a sense of ownership and encouraging continuous refinement of the knowledge base.

Field Validation and Performance

The system was developed and tested using real-world commercial logging data from five global sites: Houston, New Orleans, Cairo, Midland, and Paris. This global deployment allowed the system to incorporate asynchronous expertise, integrating knowledge about how tools behave in diverse geological environments (e.g., Egyptian vs. Texan soil).

By the conclusion of the study, the knowledge base had expanded to include models for nine different tools, comprising over 280 networks and 680 relations. The system achieved an estimated 85% accuracy relative to experienced field engineers. Crucially, the jump from early performance (~60%) to expert-level accuracy was achieved primarily through domain experts refining the networks themselves, rather than through programming changes.

Conclusion

The LQMS experiment demonstrates that the knowledge acquisition bottleneck can be mitigated by providing domain experts with intuitive, graphical tools that allow for direct manipulation of the knowledge base. By empowering experts to become the primary builders of the AI, the resulting system not only achieved high diagnostic accuracy but also gained the trust and "part-ownership" of the users it was designed to assist. The prototype's success led to its transition from research into commercial engineering phases.

Summary updated

BibTeX
@inproceedings{mullarkey1990experiment,
  author = {Peter W. Mullarkey},
  title = {{An Experiment in Direct Knowledge Acquisition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1990},
}
Figure