Authors: Douglas B. Lenat (Stanford U. and Teknowledge), Albert Clarkson (ESL), Garo Kircmidjian (ESL)
Abstract
This paper, published in the early 1980s, describes a pioneering attempt to apply expert systems to the high-stakes domain of military intelligence. The researchers developed a prototype system designed to assist analysts in Indications and Warning (I&W) analysis—the task of predicting armed conflicts by sifting through hundreds of daily intelligence reports. By utilizing a "Dual Blackboard" architecture and frame-based representations, the system identifies critical missing data and provides a clear line of reasoning for its predictions.
Introduction
In the early 1980s, military intelligence analysts faced a significant "data deluge." Processing hundreds of reports daily from media, reconnaissance, and human operatives to predict potential conflicts was a daunting task. This challenge was compounded by high personnel turnover, the need for rigorous justification of predictions, and the inherent incompleteness of intelligence data. To address these issues, the authors developed a prototype expert system capable of monitoring military activity and predicting outbreaks of war within a simulated environment.
System Architecture and Methods
The system was built using Interlisp-D on Xerox "Dolphin" machines and integrated several key AI architectures of the era:
- Blackboard Architecture: A shared data space where information at various levels—from raw reports to high-level strategic states—was posted. The X-axis represented time, allowing the system to project predictions into the future.
- Frame-Based Representation: All entities, including military units and intelligence reports, were represented as "frames" (data structures with attributes and values). This allowed the system to understand complex hierarchies, such as identifying a specific tank as a type of tracked vehicle.
- Dual Triggering (Strong vs. Weak Conditions): This novel contribution involved running two simultaneous simulations. Every rule had an
IfStrongcondition (strict, high-certainty) and anIfWeakcondition (loose, speculative). By comparing these "Dual Worlds," the system could pinpoint exactly which facts were needed to turn a speculative warning into a high-certainty prediction.
Experimental Setup
Due to the sensitive nature of real-world intelligence, the system was tested using a fictitious scenario involving two countries: "Upper X" and "Lower X." The knowledge base consisted of the distilled wisdom of professional I&W analysts, encoded into 60 condition/action rules and 170 object frames.
Key Insights and Results
The system demonstrated several significant capabilities:
- Temporal Reasoning: It used "inverse rules" to infer past events (e.g., deducing that troops must have mobilized days prior based on current rail traffic) and forward-chaining to predict future movements.
- High Performance: Even on 1980s hardware, the system processed 30 reports per minute, outpacing human analysts' ability to follow the reasoning in real-time.
- Interactive Visualization: A dual-screen setup provided an abstract blackboard view alongside a color map display that updated as reports were processed, aiding analyst intuition.
Professional analysts validated the prototype, finding its "stylized English" explanations particularly useful for justifying predictions to superiors.
Conclusion and Future Work
While the system was a successful small-scale prototype, the authors noted the need for a broader rule base for operational use. Future research directions included "Scenario Generation," where the system would proactively simulate potential cause-and-effect chains to help analysts identify early warning signs of undesirable events before they occur.
Summary updated