Generalizing Plans Through Causal Link Preservation

15 concepts8 papers2 proposals

When an intelligent system observes a successful solution to a problem, it faces a fundamental challenge: how much of that solution should it remember? If it memorizes the exact sequence of steps, the knowledge is too brittle to reuse. If it generalizes too aggressively, it might create a "rule" that fails in practice.

The research direction of causal-link-preserving abstraction addresses this by shifting the focus from what happened to why it worked. By identifying the underlying logical structure—the causal links—that connects actions to goals, systems can transform a single specific instance into a robust, reusable template.

The Logic of Explanation

Traditional machine learning often relies on seeing many examples to find statistical patterns. In contrast, Explanation-Based Learning (EBL) allows a system to learn from a single example by constructing an "explanation" of why the example succeeded.

DeJong argued in Acquiring Schemata Through Understanding and Generalizing Plans that generalization should be driven by the domain's rules. Instead of just copying a plan, the system replaces specific entities with the most abstract categories that still satisfy the original causal constraints. For instance, if a plan requires "breaking a glass object," the system generalizes "glass" not to "any object," but to the most abstract category that is still "breakable."

This process involves a systematic pruning of details. O'Rorke describes this in Generalization for Explanation-Based Schema Acquisition as categorizing constraints into essential, optional, or extraneous types. By removing extraneous constraints—those specific to the training instance but irrelevant to the goal—the system distills a specific solution into a reusable schema. Mooney and DeJong further demonstrated in Learning Schemata for Natural Language Processing that this logical approach allows systems to learn complex conceptual structures, like narrative scripts, from a single story.

Preserving the Producer-User Relationship

The core of a plan is not the sequence of steps, but the causal links between them. A causal link exists when one step (the producer) creates a condition that a subsequent step (the user) requires.

Anderson and Farley propose in Plan Abstraction Based on Operator Generalization that effective abstraction is achieved by finding the highest level of action generalization that still preserves these producer-user relationships. If we generalize an action too much, it might no longer produce the specific condition the next step needs.

This becomes particularly complex when steps can occur in different orders. Kambhampati and Kedar highlight in Explanation-Based Generalization of Partially Ordered Plans that generalizing based on a single fixed sequence can be dangerous. A generalized plan must account for "destructive interactions"—where a step in one valid ordering might accidentally undo a condition needed by another. By using formal truth criteria to protect causal links, the authors ensure that the resulting template remains valid across all possible execution sequences.

From Constants to Symbolic Constraints

To move from a specific design to a general rule, a system must also abstract numerical values into relational conditions. In the context of VLSI design, Mitchell et al. developed LEAP, a system that transforms specific signal values into general symbolic constraints. As described in LEAP: A Learning Apprentice for VLSI Design, this allows the system to learn rules that define their own operational limits, ensuring they are only applied in contexts where the underlying logic holds.

Modern research has extended this to the field of Constraint Programming (CP). Often, a model that works for a small problem (like scheduling five nurses) fails when the problem size changes. Kumar et al. address this in Learning Constraint Programming Models from Data Using Generate-And-Aggregate by learning symbolic bounds. Instead of learning that a value must be "less than 5," the system learns a mathematical expression that adapts to the problem's parameters.

Building on this, Berden et al. introduced GENCON in Generalizing Constraint Models in Constraint Acquisition. The authors use machine learning to classify and reconstruct higher-level "constraint specifications." These specifications capture the parameterized nature of requirements, allowing the system to automatically generate new, valid constraints for problem instances it has never seen before.

Why It Works: The Power of "Why"

Causal-link-preserving abstraction works because it treats generalization as a search for logical necessity. By identifying the minimal set of conditions required to guarantee success, these systems avoid the two main pitfalls of automated learning:

  1. Over-fitting: Memorizing incidental details (like the specific name of a variable or a specific numerical threshold).
  2. Invalidity: Creating broad rules that fail because they ignore the subtle dependencies between steps.

Whether applied to natural language, robotic planning, or industrial design, the principle remains the same: to learn a general truth from a specific experience, one must protect the causal links that made that experience a success.

Key Foundational Work

Acquiring Schemata Through Understanding and Generalizing Plans (1983) — contributed 1 concept elements to this direction.
Generalization for Explanation-Based Schema Acquisition (1984) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_link_preserving_abstraction_8f3515ed,
  title = {causal-link-preserving abstraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-link-preserving-abstraction/rdr_8f3515ed3ae84d7da2be1e24570e7e68},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_link_preserving_abstraction_8f3515ed_v1,
  title = {causal-link-preserving abstraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-link-preserving-abstraction/rdr_8f3515ed3ae84d7da2be1e24570e7e68/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake