Default Reasoning From Statistics

Fahiem BacchusAAAI 1991

Abstract

In artificial intelligence and formal epistemology, a long-standing challenge is the reconciliation of objective statistical data with subjective degrees of belief. This paper introduces a formal logic for direct inference, providing a mathematical foundation for how an agent can rationally move from population-level statistics (e.g., "75% of birds fly") to individual predictions (e.g., "I am 75% sure this specific bird flies"). By utilizing a two-sorted first-order modal logic, the framework treats subjective belief as the expected value of statistical frequencies across possible worlds, offering a robust alternative to traditional non-monotonic logics.

1. Introduction: The Dual Nature of Probability

Probability has historically been interpreted in two distinct ways: the statistical interpretation, which measures relative frequencies in the physical world, and the degree of belief interpretation, which measures an agent's internal uncertainty. While humans intuitively use statistics to inform their beliefs—such as assuming a specific doctor is likely wealthy because most doctors are—formalizing this "direct inference" has proven difficult for symbolic AI.

Traditional deductive logic is too rigid for this task; knowing that "most birds fly" does not allow for the deductive conclusion that a specific bird, Tweety, flies. Conversely, existing non-monotonic logics often require complex, ad-hoc rules to handle exceptions or fall prey to logical paradoxes. Fahiem Bacchus addresses these issues by developing a logic that represents both types of probability simultaneously.

2. A Combined Probability Logic

The core of Bacchus’s contribution is a two-sorted, first-order modal logic. This system differentiates between two types of terms:

  1. Statistical Terms ($[\alpha]_{\vec{x}}$): These represent the objective proportion of individuals in a domain that satisfy a specific formula $\alpha$.
  2. Belief Terms ($prob(\alpha)$): These represent the agent’s subjective uncertainty across a set of possible worlds.

To bridge these two domains, Bacchus introduces an Expectation Operator ($E$). Because an agent may not know the exact statistical facts of the world (i.e., the statistics may vary across different possible worlds the agent considers), the agent sets their degree of belief based on the expected value of the statistic.

2.1 The Direct Inference Principle

The Direct Inference Principle serves as the primary mechanism for belief formation. It posits that an agent's degree of belief in a proposition regarding an individual $c$ should equal the expected proportion of individuals that share $c$'s known properties. This "Type III" logic treats individuals in a vector as if they were selected independently, allowing for the calculation of marginal probabilities and providing a clear path from general population data to specific instances.

3. Key Insights and Advantages

Bacchus’s statistical approach offers several advantages over purely symbolic or qualitative default reasoning systems:

  • Natural Subset Preference: Many default reasoning systems require manual "priority" rules to handle specificity (e.g., ensuring "penguins don't fly" overrides "birds fly"). In this logic, if Tweety is known to be a penguin, the system automatically prioritizes the penguin statistic because it is a more specific reference class for the individual.
  • Resolution of the Lottery Paradox: In binary default systems, one might believe of every individual lottery ticket that it will lose, yet also believe that some ticket must win, leading to a contradiction. By assigning degrees of belief (e.g., 0.99) rather than absolute truth values, Bacchus’s system maintains consistency.
  • Flexible Contraposition: Standard logic dictates that "If A then B" implies "If not B then not A." However, in default reasoning, "Birds fly" does not necessarily imply "Non-fliers are not birds." The statistical foundation allows the logic to block counter-intuitive contrapositions when the underlying frequencies do not support them.

4. Theoretical Results

The paper establishes several critical properties of the proposed logic:

  • Consistency: The theory of direct inference is proven to be consistent, ensuring it does not produce logical contradictions provided the initial knowledge base is satisfiable.
  • Generalization of KD45: The model generalizes the KD45 modal logic, a standard for reasoning about belief. When an agent is 100% certain of a statistical fact, the logic simplifies to behave like traditional belief logic.
  • Computational Feasibility: The existence of a complete proof theory suggests that it is possible to build automated systems that calculate these degrees of belief from a structured knowledge base.

5. The Irrelevance Problem and Future Directions

A significant challenge identified is the Irrelevance Problem. If an agent knows Tweety is a "yellow bird" but only possesses statistics for "birds," the system cannot inherently determine if "yellow" is relevant to flight. Currently, the agent requires explicit statistical data stating that color does not affect flight.

Bacchus suggests two potential solutions for future investigation:

  1. Syntactic Extensions: Incorporating explicit "irrelevance" assertions into the knowledge base, similar to Reiter’s default logic.
  2. Semantic Models: Utilizing "maximal independence" or maximum entropy assumptions, where properties are assumed to be irrelevant unless evidence suggests otherwise.

6. Conclusion

By grounding default reasoning in the mathematics of statistics, Bacchus provides a rigorous framework for direct inference. This approach not only bridges the gap between objective data and subjective belief but also resolves several long-standing paradoxes in AI logic, providing a path toward more sophisticated probabilistic reasoning systems.

Summary updated

BibTeX
@inproceedings{bacchus1991default,
  author = {Fahiem Bacchus},
  title = {{Default Reasoning From Statistics}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1991},
}
Figure