Published in the 1986 AAAI proceedings, this classic paper arrived at a pivotal moment in AI history known as the "Uncertainty Wars." At the time, the field was deeply fractured between proponents of classical probability and those who argued that "human-like" reasoning required entirely new formalisms, such as fuzzy logic or certainty factors. Horvitz, Heckerman, and Langlotz stepped into this fray not to advocate for a specific side, but to provide a "Rosetta Stone"—a rigorous framework for comparing and categorizing all reasoning models under a unified set of principles.
The Problem: Reasoning Under Uncertainty
In the mid-1980s, AI researchers were struggling to implement "plausible reasoning" in expert systems. While probability theory was well-established, many in the AI community viewed it as too computationally expensive or intellectually rigid for complex, real-world applications. This led to a proliferation of ad-hoc methods, such as the certainty factors used in MYCIN, and new mathematical theories like Zadeh’s fuzzy logic.
The fundamental issue was the lack of an objective metric for comparison. Proponents of different systems often talked past each other, lacking a common ground to evaluate the merits of their respective approaches. This paper sought to establish that ground by asking: What fundamental "common sense" properties must any measure of belief satisfy?
The Axiomatic Approach: Cox’s Theorem Revisited
Rather than proposing a new algorithm, the authors introduced a conceptual framework based on the work of physicist R.T. Cox. They identified seven "Fundamental Properties of Belief" that they argued are essential for any rational system of reasoning:
- Clarity: Propositions must be well-defined (the "Clairvoyant test").
- Scalar Continuity: Belief should be represented by a single real number.
- Completeness: A degree of belief can be assigned to any well-defined proposition.
- Context Dependency: Belief in one proposition can depend on the state of belief in another.
- Hypothetical Conditioning: Belief in a conjunction ($A$ and $B$) should relate to the belief in $A$ given $B$.
- Complementarity: Belief in "not $A$" must be a unique function of the belief in $A$.
- Consistency: Logically equivalent propositions must result in equal belief values.
The authors demonstrated that if a researcher accepts these seven intuitive properties, they are mathematically forced to accept the axioms of probability theory. In this view, probability is not merely one option among many; it is the necessary logical consequence of these basic desiderata.
A Taxonomy of Alternative Formalisms
By using these seven properties as a benchmark, the authors created a taxonomy to categorize alternative reasoning methods based on which property they "break" or modify:
- Generalizations: These systems weaken a property to handle specific informational gaps. For instance, Dempster-Shafer (DS) theory weakens the Completeness property. By using belief intervals (belief and plausibility), DS theory allows for cases where there is insufficient information to assign a single probability.
- Specializations: These models adopt the core properties but add restrictive assumptions, such as assuming all variables are independent to increase computational speed.
- Substitutions: These systems replace a fundamental property with a different rule. The authors noted that Fuzzy Logic primarily deviates by abandoning the Clarity property. While this allows for reasoning about vague concepts like "large tumors," it incurs a penalty in decision-making utility because the events themselves are ill-defined.
- Self-Inconsistency: Some systems fail to maintain internal logic. The authors demonstrated that the original MYCIN certainty factor model was mathematically self-inconsistent because its rules for combining evidence were non-commutative, leading to different results depending on the order in which evidence was processed.
Key Insights and Legacy
The paper’s primary contribution was the realization that most "alternatives" to probability are actually trade-offs. For example, the use of belief intervals in Dempster-Shafer theory is a direct response to a lack of completeness in the data. Similarly, the "vagueness" of fuzzy logic is a trade-off against the clarity required for strict decision-theoretic utility.
The authors concluded by distinguishing between Belief Entailment (how we update numbers) and Problem Formulation (how we decide which variables matter). While they argued that probability theory effectively solves the problem of entailment, they acknowledged that problem formulation remains a significant challenge for AI. Their work remains a cornerstone of AI theory, urging researchers to weigh the pragmatic trade-offs of their chosen formalisms against a firm foundation of logical consistency.
Summary updated