Abstract
In the mid-1980s, the AI community was divided between building "Expert Systems" (ES) that mimic human heuristic reasoning and "Decision Analysis" (DA) models based on formal probability. This paper details a head-to-head experiment comparing these two paradigms by developing diagnostic systems for apple tree root disorders. The results demonstrate that while rule-based systems are easier to prototype, probabilistic models provide more robust reasoning, handle uncertainty more consistently, and can uncover counter-intuitive insights that improve the expert's own intuition.
Introduction: The Heuristic vs. Normative Divide
The fundamental debate addressed by Henrion and Cooley (1987) centered on whether AI should emulate human experts or follow normative mathematical theories. Proponents of Expert Systems argued that human-like "if-then" rules were more natural and computationally efficient. Conversely, proponents of Decision Analysis argued that heuristic rules were often inconsistent and fragile. To test these claims, the authors built one of each system for a real-world problem: diagnosing root disorders (Phytophthora, cold stress, and water stress) in apple trees.
Methodology: Modeling Apple Tree Disorders
The researchers collaborated with a plant pathologist to construct two distinct models:
- The ES Model: Implemented in the Knowledge Engineering Environment (KEE), this model used an inference network where links moved from evidence to diagnosis. Uncertainty was managed through seven qualitative "degrees of belief."
- The DA Model: This model utilized influence diagrams and Bayesian belief networks. Unlike the ES model, the links followed a causal direction—for instance, representing how a disease causes a specific symptom rather than how a symptom implies a disease.
The study relied on a knowledge base derived from intensive interviews with an expert pathologist. The initial influence diagram consisted of 30 nodes, while the inference net had 25.
Comparative Insights: Rules vs. Probabilities
A key finding was the difference in the "Gold Standard." In the ES approach, the expert's judgment is the ultimate benchmark; if the model disagrees, the model is considered flawed. In the DA approach, the underlying probability theory is the standard. This allows the model to reveal human cognitive biases.
Furthermore, the DA model exhibited "isotropic reasoning," meaning it could reason both from symptoms to disease (diagnosis) and from disease to symptoms (prediction). The ES model was largely "one-way," limited by the directionality of its rules.
Results and Economic Precision
The experiment highlighted several practical trade-offs:
- Encoding Efficiency: The ES model was faster to prototype because it allowed for "partial specifications" without requiring exhaustive definitions of all event combinations.
- Maintenance and Robustness: While easier to start, the ES model required more extensive "tuning" to handle edge cases. The DA model, though requiring more rigorous initial data, proved more robust.
- The "Aha!" Moment: The DA model suggested that fungicide treatment had low value because, by the time symptoms were visible, the tree was usually unsalvageable. After reviewing the model's logic, the expert realized the model's reasoning was superior to his own initial intuition.
- Economic Analysis: The DA model explicitly accounted for the costs of treatments versus the cost of tree replacement ($85), allowing for mathematically optimal recommendations based on expected net cost.
Conclusion
While the study demonstrated the power of decision-analytic models, the authors noted challenges in scalability and the relative lack of sophisticated software tools for influence diagrams at the time. They concluded that the future of AI likely lies in hybrid systems that combine the intuitive entry of heuristics with the rigorous probabilistic "backplane" of decision analysis.
Summary updated