Abstract
In the history of science, progress is rarely a linear accumulation of facts; rather, it often involves the radical restructuring of existing theories when they collide with contradictory evidence. This paper, originally presented at AAAI-86, introduces STAHLp, a computational system designed to model the cognitive processes of 18th-century chemists. By integrating techniques from belief revision and assumption-based reasoning, STAHLp demonstrates how scientific discovery can be viewed as a systematic process of identifying and resolving logical inconsistencies within a theoretical framework.
Introduction
During the 1980s, artificial intelligence research into scientific discovery was largely bifurcated. One branch focused on empirical discovery—finding patterns in raw data—while another focused on belief revision—maintaining logical consistency within databases. STAHLp was developed to bridge this gap. The authors argued that scientific progress is fundamentally about fixing "broken" theories.
Earlier systems, such as the original STAHL, were capable of generating chemical models but often struggled with circular reasoning or infinite loops when faced with contradictory data. STAHLp addresses these limitations by treating scientific discovery as a problem of belief revision, specifically modeling the historical transition from the phlogiston theory of combustion to Lavoisier’s oxygen theory.
The Computational Framework
STAHLp operates as a forward-chaining production system that represents chemical substances through componential modeling. In this framework, substances are defined by their constituents (e.g., Iron = Phlogiston + Calx-of-iron). To navigate the complexities of chemical transformations, the system employs several key mechanisms:
- Reduced Lists: To prevent circular reasoning, STAHLp utilizes a tracking mechanism that remembers which substances have been removed during an inference chain. This ensures the system does not "rediscover" a component that was already accounted for in a previous step.
- Assumption-Based Tracking: Drawing on the work of de Kleer, STAHLp attaches "source tags" to every belief it generates. These tags allow the system to trace the pedigree of any conclusion back to its original premises—whether those premises were observed experimental results or theoretical assumptions.
- Unified Error Detection: The system’s most significant innovation is the reduction of various logical errors into a single physical metaphor: the unbalanced null reaction. If an inference suggests that matter has been created or destroyed (e.g., an equation resulting in
nothing -> something), the system flags a contradiction.
Hypothesis Generation and Theory Revision
When an unbalanced reaction is detected, STAHLp does not simply halt. Instead, it initiates a revision process:
- Hypothesis Generation: The system generates "effect-hypotheses" to balance the problematic equation. This might involve assuming the presence of an undetected substance or concluding that a previously accepted reaction was incorrect.
- Cost Analysis: Because multiple hypotheses can often resolve a single contradiction, STAHLp employs a computational version of Occam’s Razor. It calculates the "cost" of each potential revision, preferring the change that necessitates the fewest alterations to its existing knowledge base.
Through this process, "discovery" emerges as a byproduct of error correction. By attempting to resolve a contradiction in phlogiston-based models, the system can autonomously arrive at the necessity of an oxygen-based model.
Historical Replications and Results
The efficacy of STAHLp was tested against several landmark debates in the history of chemistry:
- The Phlogiston-Oxygen Shift: The system successfully replicated the reasoning required to move from the belief that fire is a substance released during burning (phlogiston) to the understanding that combustion involves the combination of a substance with oxygen.
- The Davy vs. Gay-Lussac Debate: STAHLp modeled the dispute over the composition of potash. It replicated the logic used by Gay-Lussac and Thenard to argue that Humphry Davy’s samples were impure, specifically containing water—a historical instance of hypothetical reasoning used to defend a theoretical stance.
- Elimination of Caloric: The system was able to determine that "caloric" (the hypothetical fluid of heat) was an unnecessary component in explaining certain chemical reactions, mirroring the eventual scientific consensus to abandon the caloric theory.
Limitations and Conclusion
While STAHLp provides a robust model for qualitative reasoning, it possesses notable limitations. It operates purely on a qualitative level (A = B + C) and lacks the capacity for stoichiometry—the quantitative ratios and weights that were historically crucial to Lavoisier’s arguments. Furthermore, its "cost function" for theory selection is a simplified heuristic; in actual scientific practice, the abandonment of a theory is often influenced by complex social, experimental, and aesthetic factors.
Despite these limitations, STAHLp remains a landmark in AI-driven philosophy of science. It demonstrates that the logic of discovery is often the logic of revision, showing how a system equipped with the ability to track its own assumptions can navigate the turbulent waters of conceptual change.
Summary updated