A Hybrid Framework for Representing Uncertain Knowledge

Alessandro SaffiottiAAAI 1990

This paper, originally presented at AAAI-90, represents a foundational effort in symbolic AI to reconcile two historically separate subfields: Knowledge Representation (KR) and Uncertain Reasoning (UR). By proposing a formal bridge between structured logic and numerical calculus, the author addresses a long-standing divide in how machines process information.

The Problem: A Divided AI Landscape

In the late 1980s, AI research was split. The KR community focused on structured, categorical relationships (e.g., "All birds are animals"), while the UR community focused on numerical degrees of belief (e.g., "There is an 80% chance it will rain").

The author argues that these two fields suffered from a lack of "cross-fertilization." KR systems struggled with "shades of gray," often failing to handle exceptions gracefully. Conversely, UR systems frequently ignored the rich, structured nature of the knowledge they were assigning probabilities to, treating complex statements as simple atomic events. This paper seeks a unified framework where uncertainty is treated as metaknowledge—knowledge about the validity of our categorical knowledge.

The Hybrid Framework: A Functional Approach

The author adopts a "functional approach" to KR, inspired by Hector Levesque’s work. Instead of focusing on internal data structures, the framework defines how a "Belief Set" (BS) behaves through three core operations:

  1. Empty: Initializes an empty belief set.
  2. Tell: Adds new knowledge (which may have an associated degree of certainty) to the set.
  3. Ask: Queries the belief set to determine the degree of certainty for a specific formula.
The functional mappings for the 'Ask' and 'Tell' operations within the belief set framework.

As shown in the diagram above, these operations map linguistic formulas and belief judgments into updated states of knowledge. The breakthrough of this hybrid approach is the decomposition of these operations into two distinct modules.

The KR and UR Modules

The framework decouples the linguistic processing from the uncertainty calculus using a two-step architecture:

  • The KR Step: A Knowledge Representation module maps a linguistic formula (e.g., "Enzo is rich") to its intension. An intension is an abstract representation of meaning—either a set of "possible worlds" where the statement is true (semantic view) or a set of "possible arguments" that prove the statement (syntactic view).
  • The UR Step: An Uncertain Reasoning module evaluates the "belief judgment" of that intension. This is where the numerical calculus (Probability, Dempster-Shafer, etc.) occurs, independent of the original language's syntax.
The modular architecture showing how knowledge is first processed by a KR module before being evaluated by a UR module.

By using the intension as a "glue," developers can swap out the KR language (like First-Order Logic) or the UR calculus (like Bayesian networks) without redesigning the entire system. This modularity allows for high flexibility in system design.

Key Insights and Ideas

  • Uncertainty as Metaknowledge: The core philosophy is that knowledge itself is categorical; it is our confidence in that knowledge that is a matter of degree.
  • The Intension Bridge: By treating meaning as an abstract object (the intension), the framework decouples the structure of knowledge from the logic of uncertainty.
  • Epistemic vs. Ontological: The framework focuses on epistemic uncertainty (our lack of certain knowledge about a certain reality). Interestingly, the author argues that even "vagueness" (like the word "rich") can be modeled as epistemic uncertainty regarding the adequacy of a description.
  • Dual Perspectives: The paper proves the framework works whether one prefers a Semantic perspective (Kripke possible worlds) or a Syntactic perspective (formal proofs and deductions).

Instantiations and Results

The paper provides three concrete "instantiations" to prove the framework's flexibility:

  1. FOL + {yes, no}: A "collapsed" case showing the framework can replicate standard First-Order Logic.
  2. FOL + Dempster-Shafer (DS): A complex case where structured logic is paired with DS theory to handle evidence and degrees of belief.
  3. ZERO + DS: A custom, minimal KR system that distinguishes between "definitions" (isa-links) and "facts," demonstrating that the framework can handle inheritance alongside uncertainty.

Limitations and Future Work

Despite its theoretical elegance, the author acknowledges significant computational complexity. Generating all "possible arguments" for an intension is often computationally intractable in large systems. The paper suggests using Assumption-Based Truth Maintenance Systems (ATMS) to manage fragments of arguments as a potential path toward practical implementation. Ultimately, this work serves as a mathematical foundation for the "hybrid" systems that would follow in the decades to come.

Summary updated

BibTeX
@inproceedings{saffiotti1990hybrid,
  author = {Alessandro Saffiotti},
  title = {{A Hybrid Framework for Representing Uncertain Knowledge}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1990},
}
Figure