Granularity

Jerry R. HobbsIJCAI 1985

In the mid-1980s, artificial intelligence research faced a daunting "scaling up" problem. As knowledge bases expanded to encompass the complexities of the real world, reasoning within those systems became computationally intractable. Jerry Hobbs’ seminal work, "The Art of Ignoring Detail: A Theory of Granularity," addresses this by proposing a formal framework for how agents—both human and artificial—simplify complex environments into manageable "local theories."

The Problem of Computational Intractability

Humans possess a remarkable ability to ignore irrelevant details. When planning a cross-country trip, we treat a road as a one-dimensional line. When crossing that same road, we perceive it as a two-dimensional surface. When repairing a pothole, it becomes a three-dimensional volume. This shifting of "grain size" allows us to reason efficiently without being overwhelmed by the total complexity of the universe.

The challenge for AI is to represent these shifts formally. How can a machine determine what is "relevant" to a specific task and ignore the rest, while maintaining the ability to reconnect with a more detailed reality when necessary?

A Formal Logic of Granularity

Hobbs utilizes first-order logic to define a Global Theory ($T_0$), representing the world in its full complexity. To derive a simplified Local Theory ($T_1$), he introduces several key formalisms:

  1. The Indistinguishability Relation ($\sim$): This is defined by a subset of "relevant" predicates ($R$). Two objects are considered indistinguishable if no predicate in the relevant set can differentiate between them.
  2. Simplification via Mapping ($K$): This mapping collapses the global domain into equivalence classes based on the indistinguishability relation, creating a coarse-grained model.
  3. Partial Predicates and Idealization: Real-world boundaries are often "fuzzy" (e.g., the perception of temperature). However, fuzzy boundaries can break mathematical transitivity. Hobbs proposes Idealization—the deliberate imposition of sharp boundaries (e.g., grouping temperatures into "the 50s" vs. "the 60s")—to maintain a discrete and tractable propositional system.
  4. Articulation Axioms: These serve as the logical "glue" connecting different granularities. They allow an agent to translate a problem from a coarse-grained view to a fine-grained one and back again, ensuring that the local theory remains grounded in the global reality.

Key Insights: Intelligence as Abstraction

Hobbs’ framework suggests that intelligence is not merely the accumulation of facts, but the ability to map complexities into simple, tractable theories. Several core ideas emerge from this perspective:

  • Utility Over Faithfulness: In many cases, the "utility" of a local theory is more important than its absolute "faithfulness" to the global theory. We intentionally sacrifice precision to gain the ability to compute an answer within a reasonable timeframe.
  • Relevance as a Filter: Indistinguishability is entirely dependent on the goal. For a car dealership, two identical models are the same "type"; for a driver involved in a collision, they are distinct physical "tokens."
  • The Value of "Toy" Domains: Hobbs provides a defense for "microworlds" in AI research. He suggests that these simplified domains are not just academic simplifications but represent the actual "local theories" that a sophisticated AI must switch between during complex reasoning.

Limitations and Future Directions

While the paper provides a robust logical foundation, it leaves several questions open. The Relevance Problem—determining exactly which predicates are relevant for a given goal—remains a significant challenge that the paper does not fully solve. Furthermore, the process of "Idealization" forces sharp distinctions that may lead to errors at the boundaries of categories.

Finally, the Knowledge Engineering Burden is substantial. Building a knowledge base that includes articulation axioms across many grain sizes requires significant manual effort. While Hobbs argues that this approach leads to tractability, the paper does not provide an automated algorithm for generating these local theories from a global one.

Summary updated

BibTeX
@inproceedings{hobbs1985granularity,
  author = {Jerry R. Hobbs},
  title = {{Granularity}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1985},
}

Loading related content...

Figure