Introduction
In the late 1960s and early 1970s, the field of Artificial Intelligence was dominated by the pursuit of "General Problem Solvers"—the belief that a small set of powerful, domain-independent logical mechanisms could solve any problem if given sufficient computational resources. However, these systems frequently succumbed to "combinatorial explosions" when applied to complex, real-world scenarios. In his seminal 1977 paper, Edward A. Felgenbaum proposed a radical shift in perspective: the "power" of an intelligent system does not reside in its inference engine, but in the specific, high-quality knowledge it possesses. This paper marked the birth of Knowledge Engineering.
The Generality vs. Power Trade-off
Felgenbaum’s central argument addresses the fundamental trade-off between general-purpose methods and specialized performance. He posits that for AI to be practically useful in professional fields like medicine or chemistry, it must be knowledge-rich, even if it remains methods-poor. While general logic is elegant, it is often too weak to navigate the massive search spaces of specialized domains. To achieve expert-level performance, a system requires the "private" heuristic knowledge—the rules of thumb and "good guesses"—that human experts use to prune search spaces and reach conclusions efficiently.
The Methodology of Knowledge Engineering
The paper defines "Knowledge Engineering" as the process of bringing the tools of AI to bear on difficult informational problems. This involves a collaborative effort between a "knowledge engineer" and a domain expert. The engineer acts as an architect, extracting, representing, and "polishing" the expert's knowledge into a machine-readable format.
Technical Foundations
Several key technical approaches underpin this methodology:
- Rule-Based Systems (Production Rules): Knowledge is represented in modular
IF...THEN...formats. This modularity allows for incremental improvements; rules can be added or modified without a total system overhaul. - Heuristic Search: Instead of exhaustive calculation, systems use "rules of good judgment" to navigate complex data.
- Generation-and-Test: A paradigm where the system generates plausible candidates (e.g., molecular structures) and uses domain knowledge to discard those that do not fit the observed data.
- Explanation Subsystems: Felgenbaum emphasizes that for a system to be trusted by professionals, it must be able to explain its "line-of-reasoning." A "black box" is unacceptable in fields like medicine; the AI must speak "Doctor-ese."
Case Studies in Expertise
The validity of the knowledge-based approach is demonstrated through several landmark systems:
DENDRAL and META-DENDRAL
DENDRAL was designed to infer chemical structures from mass spectrometry data. It became the first AI system to outperform human experts in a complex scientific task, resulting in over 25 publications in major chemistry journals. Its successor, META-DENDRAL, went a step further by "discovering" new rules of mass spectrometry previously unknown to human experts, which were subsequently validated and published in scientific journals.
MYCIN
Focusing on infectious disease diagnosis and therapy, MYCIN utilized approximately 500 rules to provide clinical advice. In blind evaluations, a panel of expert judges found MYCIN’s prescriptions to be "as-good-as" or better than those of human specialists in 90% of cases.
PUFF
A specialized system for diagnosing pulmonary function disorders, PUFF demonstrated the efficiency of the knowledge engineering process. Using 55 rules extracted from 100 clinical cases, it achieved between 90% and 100% agreement with human experts when tested against a separate set of 150 cases.
Key Insights: The Knowledge Acquisition Bottleneck
One of Felgenbaum’s most enduring observations is the Knowledge Acquisition Bottleneck. He noted that the primary challenge in AI is not the coding of the inference engine, but the "elicitation" of knowledge. Experts often "know more than they are aware of knowing," making the extraction of their internal heuristics a difficult and iterative process.
Furthermore, Felgenbaum envisioned a shift in the "What-to-How" spectrum. He predicted a future where humans move away from traditional programming (telling a computer how to do something) and toward knowledge-based interaction (telling a computer what needs to be done).
Limitations and Challenges
Despite the success of these early systems, Felgenbaum identified several hurdles that remain relevant to the history of the field:
- Hardware Constraints: The "heavy" knowledge-based systems of the time were limited by the available computer hardware.
- The "Export" Problem: Moving systems from the laboratory to the field was hindered by poor software packaging and user interface design.
- Lack of Cumulation: A lack of shared software libraries meant researchers often "reinvented the wheel" for every new project.
- Human Capital: There was a severe shortage of trained knowledge engineers capable of bridging the gap between AI theory and domain-specific expertise.
Conclusion
"The Art of Artificial Intelligence" served as the manifesto for the Expert Systems movement. By asserting that "knowledge is power," Felgenbaum shifted the focus of the field from abstract reasoning to the practical application of human expertise. This paradigm dominated AI research and industry for two decades, laying the groundwork for the sophisticated, data-driven systems of the modern era.
Summary updated