Implementing a Model of Human Plausible Reasoning

Michelle Baker, Mark H. Burstein, Allan M. Collins

TL;DR

The Plausible Reasoning Simulation System (PRSS) is a cognitive model designed to replicate how humans make educated guesses when faced with incomplete or inconsistent information. Moving beyond the rigid logic of traditional expert systems, PRSS utilizes a library of "statement transforms"—such as similarity and generalization—to construct multiple, often conflicting, arguments for and against a proposition before weighing them to reach a final conclusion.

The Challenge of Real-World Uncertainty

In the late 1980s, AI expert systems often struggled with the ambiguity of the real world. Most relied on statistical frameworks like Bayesian networks, which required precise probability data that was frequently unavailable. Furthermore, these systems typically operated under a "closed-world assumption," where any fact not explicitly in the database was deemed false.

Humans, by contrast, excel at reasoning through what they don't know. For example, if asked whether coffee grows in a specific Colombian region, a person might not know the fact directly but can infer an answer based on the region's climate, its similarity to known coffee-producing areas, or the absence of contradictory information. The authors of this paper sought to bridge this gap by implementing a formal theory of "plausible reasoning" originally proposed by Collins and Michalski.

Methodology: The PRSS Architecture

The PRSS simulation tool treats reasoning as an active search for evidence rather than a linear logical proof. Its architecture is built on several key pillars:

Dual-Path Verification

Unlike standard systems that attempt to prove a statement true, PRSS simultaneously seeks "proofs" for both the truth and the falsity of a proposition. This allows the system to handle conflicting evidence gracefully.

Knowledge Representation

Data is stored in a multiply-indexed semantic network. Each entry consists of a descriptor (e.g., Climate), an argument (e.g., England), and a referent (e.g., Temperate).

Statement Transforms

The core of the reasoning engine lies in eight types of transforms based on four primary operators: Generalization (GEN), Specialization (SPEC), Similarity (SIM), and Dissimilarity (DIS). These operators can be applied to either the argument (the subject) or the referent (the value) to derive new plausible facts.

Examples of statement transforms applied to the referent (the value).
Examples of statement transforms applied to the argument (the subject).

Certainty Parameters

To quantify uncertainty, the model tracks four distinct parameters for every piece of knowledge:

  1. Certainty: The degree of belief in the fact's truth.
  2. Frequency: The occurrence rate of a referent within a set (e.g., the percentage of Africa that is tropical).
  3. Dominance: The relative "size" or importance of a subset (e.g., how significant Surrey is within England).
  4. Likelihood: The conditional probability of a dependency (e.g., the likelihood of heavy rainfall given a specific crop).

Datasets and Evaluation

The research utilized two primary sources:

  • Human Protocols: A corpus of human responses to everyday questions, collected by Collins in 1978, served as the basis for the underlying theory.
  • Domain-Specific Knowledge Bases: Small, structured databases containing geographical and botanical facts (e.g., crop requirements and regional climates) were used to test the system's inferential capabilities.

Key Insights

  • Reasoning as Evidence Weighing: Human certainty is not a monolithic probability score but the result of weighing multiple independent, and sometimes contradictory, lines of reasoning.
  • Functional Similarity: Similarity is used as a functional tool rather than a simple fuzzy match. For instance, knowing that Holland and England share a latitude allows the system to plausibly infer they share a climate.
  • Functional Dependencies: The model emphasizes reasoning about the relationships between properties, such as the dependency between latitude and temperature.
  • Negative Inference: The system can conclude a statement is likely false if a required condition is missing or if a similar "landmark" case behaves differently.

Results and System Traces

The paper demonstrates PRSS through qualitative "traces" of its reasoning:

  • The Coffee Example: When queried about coffee growth in the Llanos, PRSS identifies a negative implication (lack of rainfall) and a positive similarity transform (proximity to Brazil). The system correctly identifies the evidence as "evenly mixed," mimicking human hesitation.
  • The England Climate Example: The system generates three independent proofs for England's temperate climate—via inheritance from Europe, latitude implications, and similarity to Holland—resulting in a "High" certainty conclusion.

Limitations and Future Directions

The authors identify several areas for improvement:

  • Computational Complexity: As the knowledge base grows, the number of possible inferences could expand exponentially, necessitating better "filtering" of weak arguments.
  • Similarity Models: The current "urn model" for similarity is relatively simple; more sophisticated techniques are needed to handle multi-contextual features.
  • Relation Complexity: The current implementation is restricted to binary relations, with plans to extend to n-ary relations and more complex structured objects.

Summary updated

BibTeX
@inproceedings{baker1987implementing,
  author = {Michelle Baker and Mark H. Burstein and Allan M. Collins},
  title = {{Implementing a Model of Human Plausible Reasoning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1987},
}
Figure