Causal and Plausible Reasoning in Expert Systems

Gerald LiuAAAI 1986

In the mid-1980s, the field of artificial intelligence faced a significant divide in the development of expert systems. "Shallow" systems utilized numerical likelihoods, such as Bayesian networks or Certainty Factors, to handle uncertainty but often lacked the structural depth to explain the underlying logic of their conclusions. Conversely, "deep" systems employed symbolic logic to provide clear explanations but struggled to quantify the strength of their beliefs. Gerald Shao-Hung Liu’s 1986 paper, "Causal and Plausible Reasoning in Expert Systems," presented a seminal attempt to bridge this gap by proposing a unified framework that integrates the nature of an inference (causality) with its strength (plausibility).

The Dichotomy of Expert Systems

The paper addresses the inherent "superficiality" of purely numerical approaches. In standard Bayesian models, different types of evidence—such as a "supportive" fact versus a "necessary" condition—are often treated identically if their conditional probabilities are equivalent. Liu argues that this lack of qualitative distinction limits the reasoning capabilities of AI. He critiques existing hybrid systems of the era as being either "piecewise" (using different methods for different rules) or "ad hoc" (lacking a sound theoretical ground). His goal was to establish a "likelihood calculus" that respects the qualitative differences in how facts influence conclusions.

The Role System: Categorizing Causality

At the heart of Liu’s framework is the "Role System," which identifies six primitive causal roles a condition can play relative to a conclusion. This system allows for a more nuanced representation of knowledge than simple "If-Then" rules:

  • Associational: Represents a standard empirical link between facts.
  • Supportive/Adverse: Secondary conditions that strengthen or weaken a conclusion but do not drive the rule independently.
  • Sufficient: A condition that, on its own, warrants a conclusion.
  • Necessary: A condition that must be present (or at least not disproved) for a conclusion to hold.
  • Contrary: A condition that explicitly excludes a conclusion.
  • Exception (UNLESS): A mechanism for "defeasible" reasoning, allowing the system to handle exceptions (e.g., "Birds fly, unless they are penguins") within a numerical framework.

Extended Dempster-Shafer Theory

To handle the "strength" of these inferences, Liu builds upon an extended version of Dempster-Shafer (D-S) theory. Unlike Bayesian probability, which requires a probability for every single outcome, D-S theory allows for a "degree of ignorance." Liu utilizes a frame of discernment (a set of possible answers) and basic probability assignments (bpa).

By using belief intervals (e.g., $[v_1, v_2]$), the system can distinguish between a total lack of evidence (represented as $[0, 0]$) and evidence that supports a negation (represented as $[0, 1]$). This distinction is vital for expert systems that must operate with incomplete information.

Mathematical Foundation and Belief Propagation

A key technical contribution of the paper is Theorem 1 (Propagation of Beliefs). Rather than simply defining how beliefs should propagate through a system, Liu provides a mathematical derivation. He shows that the belief in a conclusion can be calculated as the sum of the belief in the rule multiplied by the belief in the condition. This effectively extends Bayesian conditioning into the D-S framework, allowing for a more robust and theoretically grounded propagation of uncertainty.

Validation through Illustrative Scenarios

As a theoretical AI paper from 1986, Liu validates his framework using several "toy" problems and reformulated examples from contemporary literature rather than modern large-scale datasets:

  • Animal Camouflage: Liu demonstrates how the Role System produces more nuanced conclusions than generic rules. For instance, the system can calculate exactly how much the belief in "defense-by-camouflage" drops when an animal's color is discovered to be different from its environment.
  • The Submarine Example: A naval expert scenario involving the location of a hidden vessel based on varying types of sensor data.
  • Legal Alibi: A scenario where a witness serves as a "strong necessary" condition for a suspect’s claim, illustrating how the system handles necessity.

Significance and Limitations

Liu’s work demonstrates that his framework is a generalization of previous models; under specific conditions (such as zero ignorance), his formulas reduce to classical Bayesian posterior probabilities or Yen’s partition-based probabilities. This unification suggests that causal roles and plausible reasoning are not mutually exclusive but are two independent axes—nature and strength—of the same inference.

However, the paper acknowledges significant hurdles. The inclusion of "exception roles" introduces non-monotonic reasoning, which makes inference significantly more complicated. Liu notes that "dependency-directed backtracking"—the process of updating beliefs when an exception is found—presents a complex efficiency issue. Furthermore, the paper focuses on propagating belief through single rules, leaving the "parallel combination" of beliefs from multiple independent rules for future research.

Summary updated

BibTeX
@inproceedings{liu1986causal,
  author = {Gerald Liu},
  title = {{Causal and Plausible Reasoning in Expert Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure