h t t p s : / / d o i . o r g / 1 0 . 1 0 2 3 / A : 1 0 2 2 6 9 1 1 2 0 8 0 7

Ingest: Explanation-Based Generalization: A Unifying View

Type: kb/sources/types/ingest-report.md

Classification

This is a scientific paper whose main contribution is a formal conceptual unification: it specifies a general explanation-based generalization method, reconstructs several earlier systems through that method, and derives an open research agenda rather than reporting a controlled empirical evaluation. Authors: Tom M. Mitchell, Richard M. Keller, and Smadar T. Kedar-Cabelli, machine-learning researchers at Rutgers who developed parts of the lineage being unified and therefore write as technically informed proponents of the approach.

Summary

The paper defines explanation-based generalization (EBG) from four explicit inputs: a goal concept, one positive training example, a domain theory, and an agent-and-task-relative operationality criterion. EBG proves why the example satisfies the goal concept and regresses the goal through that proof to derive an operational sufficient condition; the theory selects relevant features and licenses each inference, while the example focuses search on one useful re-expression. The result is normally a proof-relative specialization of the goal concept, not an inductive discovery of a complete definition. The paper identifies incomplete, intractable, and inconsistent theories as central open problems and sketches combinations with empirical learning, but it does not supply a method that uses empirical failure to localize and repair theory defects, model the learner's own software organization, or demonstrate that a revised theory changes a later learning episode.

Quotes

No source quotes have been retained yet.

Connections Found

The paper is the historical mechanism anchor for the lineage account for reflective theory refinement: it establishes how an explicit theory of an external task can govern interpretation, search, and generalization from experience. It is also a boundary on that synthesis, because its domain theory is supplied rather than empirically repaired and is not a causally connected representation of the learner's own software organization. EITHER is the direct technical successor for the missing fallible-theory step: it uses proof structure and labeled failures to localize and repair defects in an external Horn theory. EBG's operationality criterion also provides a narrow historical antecedent for the agent-and-task-relative core of actionable methodology, without establishing that definition's fuller access and authority conditions.

Extractable Value

  1. An explicit theory constrains both inference and generalization -- The domain theory must prove the example's membership in the goal concept, and regression through that proof derives the conditions under which the same explanation remains sufficient. Theory is therefore the mechanism that selects relevant features and licenses the generalized rule, not merely extra context supplied to an empirical learner. [quick-win]
  2. A training example focuses the search over operational re-expressions -- The goal concept could in principle be operationalized without examples, but an example selects transformations relevant to the learner's encountered environment and avoids enumerating every possible usable reformulation. This establishes a concrete way that theory and experience can jointly guide search. [quick-win]
  3. The learned rule is justified only within a supplied proof path -- EBG normally returns a sufficient specialization corresponding to the explanation constructed for one example; it need not recover all ways of satisfying the goal concept. Empirical success of that rule would test its usefulness and coverage, not by itself validate the completeness or truth of the underlying theory. [quick-win]
  4. Fallible theory is an explicit boundary, not a solved mechanism -- The paper distinguishes incomplete theories that can support only plausible explanations, complete but intractable theories that require approximation, and inconsistent theories that can yield conflicting explanations. It calls for methods that improve such theories but gives no procedure for assigning a failed prediction to a particular rule or repairing the rule. [quick-win]
  5. The proposed empirical combinations do not revise the domain theory -- The surveyed combinations use empirical learning to propose generalizations, merge generalized examples, or evaluate intermediate operational concept definitions; theory then explains, filters, or helps transform those candidates. Even METALEX's diagnostic execution feeds back into operationalization of the goal concept, not into correction of the domain theory that supports the explanation. [deep-dive]
  6. The theory concerns an external problem domain rather than the learner itself -- Its rules describe stacking, artifact function, calculus search, games, or story plans, while the goal concept and operationality criterion are supplied to the learner. The paper does not represent or revise the learner's inference procedure, evaluator, software architecture, or purposes as objects in the same theory. [quick-win]
  7. Changed later learning after theory revision remains unestablished -- The authors infer that learning new domain rules could improve learning performance, but they neither implement theory revision nor run a later episode in which a repaired theory changes explanation construction, search, or generalization. That longitudinal causal claim requires a separate experiment. [experiment]

Limitations (our opinion)

The paper's guarantees are conditional on the supplied theory, goal concept, proof procedure, and operationality criterion. A deductively justified generalization can still be wrong about the world when the theory is wrong, narrow when the selected explanation covers only one route to the goal, or unstable when defaults permit inconsistent explanations. The worked stacking, cup, calculus, game, and story examples clarify the formal account but do not provide controlled evidence that EBG improves predictive accuracy, search cost, or learning efficiency against empirical alternatives.

The discussion of imperfect theories is a research taxonomy and set of proposals. It does not define an empirical contradiction signal, a credit-assignment method for locating the responsible theory fragment, an admitted repair space, or a validation rule for revised theories. Likewise, the combinations with similarity-based learning do not show that evidence can overturn the domain theory itself, and the METALEX discussion concerns feedback on an operational concept definition rather than repair of explanatory domain knowledge.

Finally, nothing in the paper establishes reflective theory learning. The domain theory represents an external task, while the learner's own proof machinery, task formulation, operationality test, and software organization remain fixed outside the update surface. The source therefore supports the external-domain antecedent of theory-mediated learning, not a system that revises a theory of itself or carries such a revision into later self-modification.

Update the lineage account for reflective theory refinement so its EBG paragraph cites this ingest with (snapshot required) and states that EBG establishes theory-guided inference and operationalization for a supplied external domain theory while leaving empirical defect localization, theory repair, reflective self-targeting, and changed later learning unestablished.