Using a Matcher to Make an Expert Consultation System Behave Intelligently

René RebohAAAI 1980

Author: Research Summary based on Rene' Reboh (1979) Affiliation: SRI International

Introduction

In the late 1970s, the field of Artificial Intelligence was dominated by the development of "Expert Systems"—programs designed to mimic the decision-making abilities of human specialists. One of the most prominent examples from this era was Prospector, a system developed at SRI International to assist geologists in the exploration of ore deposits. While Prospector was groundbreaking, it faced significant hurdles common to early AI: the difficulty of maintaining logical consistency in massive rule bases and the rigid, often frustrating nature of human-computer interaction.

In his 1979 research, Rene' Reboh introduced a "Semantic Network Matcher" designed to address these issues. By providing the system with a mechanism to understand the relationships between different geological statements, the Matcher allowed Prospector to behave with a level of "intelligence" that went beyond simple if-then logic.

The Problem: Knowledge Integrity and Rigid Interaction

By 1979, expert systems had evolved into complex webs of thousands of inference rules and probabilistic weights. This growth created two primary challenges:

  1. Knowledge Integrity: As the number of rules grew, it became nearly impossible for "knowledge engineers" to ensure that the system remained logically sound. For instance, if a system knows that "rhyolite" is a specific type of "igneous rock," it must also ensure that the probability of rhyolite being present never exceeds the probability of igneous rocks being present. Manually tracking these dependencies across thousands of rules was prone to error.
  2. Clunky Interaction: Early systems were often "black boxes" that followed a strict, pre-programmed line of questioning. If a user volunteered information out of sequence—for example, mentioning a specific mineral before the system asked about it—the system often failed to recognize how that information related to its internal goals. This led to repetitive questioning and a lack of "human-like" flow.

The Solution: The Semantic Network Matcher

The core of Reboh’s solution was the Semantic Network Matcher. This component treated every statement within the Prospector knowledge base as a set of constraints within a partitioned semantic network—a graph-based representation where nodes represent entities and arcs represent the relationships between them.

The Matcher’s primary function was to compare any two arbitrary statements (S1 and S2) and classify their relationship into one of four categories:

  • Identity: S1 and S2 are logically equivalent.
  • Restriction/Generalization: S1 is a specific case of S2 (e.g., "a rhyolite plug" is a restriction of the more general "rhyolite").
  • Disjointness: S1 and S2 are mutually exclusive (e.g., "pyrite is present" vs. "no minerals are present").
  • Overlap: The statements share common elements but do not have a subset relationship.

Applications in Intelligence

By identifying these relationships, the Matcher enabled several key improvements in how Prospector functioned:

Automated Consistency

The system could automatically update Bayesian probabilities. If a user confirmed a specific geological feature (a restriction), the Matcher ensured that the probability of the more general category was updated accordingly. This prevented logical "leaks" and ensured that the system’s internal model remained consistent without requiring the expert to write redundant rules for every possible subset.

Mixed-Initiative Dialogue

The Matcher allowed for a more natural "mixed-initiative" conversation. When a user volunteered information, the Matcher identified where that information fit within the inference network. This allowed the system to acknowledge the user's input with context-aware responses, such as "I know you doubted X, so I have reason to suspect Y." This made the user feel "listened to," significantly improving the psychological experience of the consultation.

Search-by-Content

For the engineers building the system, the Matcher provided a way to search the knowledge base using semantic descriptions rather than specific keywords. An engineer could ask the system to find all rules related to "sulfide mineralization," and the Matcher would identify relevant nodes based on their logical structure and relationships.

Limitations and Legacy

Despite its successes, the Matcher had limitations inherent to the technology of the time. Its effectiveness was entirely dependent on the quality of the underlying taxonomy; if the system was not told that a "plug" is a type of "intrusive," the Matcher could not deduce the relationship. Furthermore, the computational cost of matching statements against a large semantic network was a significant burden for 1970s hardware.

Ultimately, Reboh’s work on the Matcher demonstrated that true machine intelligence requires more than just a collection of facts; it requires an internal "self-awareness" of how those facts relate to one another. This principle remains a cornerstone of knowledge representation in AI today.

Summary updated

BibTeX
@inproceedings{reboh1980using,
  author = {René Reboh},
  title = {{Using a Matcher to Make an Expert Consultation System Behave Intelligently}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1980},
}
Figure