Can EBL learn from a single example?

Yes, in suitable settings with enough prior domain knowledge to explain and generalize the example.

Does the learned rule always add new logical knowledge?

No. In a deductive setting it may compile an existing consequence into a more useful form.

Is every example feature retained?

No. Features irrelevant to the explanation can be omitted, while necessary relationships and conditions must remain.

AI Glossary / Graphs and symbolic AI / Intermediate

Explanation-based learning

Definition

Explanation-based learning, or EBL, uses existing domain knowledge to explain an example and derive a reusable rule from that explanation. Instead of generalizing only from similarities across many observations, it examines why a particular example satisfies a concept or supports a successful solution. The learned rule retains conditions that make the explanation work.

What it means

Suppose a system already knows rules for determining whether a task can be completed. It solves one case by combining those rules. EBL can turn that successful reasoning into a more direct condition for recognizing similar cases later.

The explanation is typically a structured derivation grounded in the domain theory, not simply a persuasive paragraph. Generalization replaces irrelevant specifics while preserving the relationships and assumptions needed by that derivation.

In a deductive version, the resulting rule may already follow logically from existing knowledge. Learning then changes how efficiently that knowledge can be used rather than adding a new empirical law. Broader EBL research also considers refinement and imperfect theories, so the precise learning guarantees depend on the formulation.

How it works

Specify the target concept, a training example, and relevant domain rules. The system constructs an explanation showing how the example meets the target using those rules.

It identifies the conditions supporting the explanation and generalizes the example's particulars. A named object can become a variable, but dependencies between objects must remain where the proof needs them. Details not used in the explanation need not appear in the learned condition.

The result should be usable by the system's performance component. Classical accounts call this an operationality requirement: the learned description should use conditions the system can evaluate appropriately.

A generalized explanation usually gives sufficient conditions for the target through that reasoning path. It need not capture every possible way of meeting the target. Storing and applying the rule can save future reasoning, but excessive learned rules can also create matching and maintenance costs.

Three practical examples

1. A symbolic planner compiles conditions from a successful action sequence into a reusable planning rule for related situations.

2. A theorem-proving system generalizes a proof pattern so that later instances can be recognized through a shorter derivation.

3. A rule-based configuration tool extracts a direct sufficient condition for compatibility from a longer explanation involving component properties and interface rules.

Think of it this way

Think of solving one worked mathematics problem and then identifying the conditions that made the solution method applicable. You keep the reasoning pattern while replacing incidental numbers or names.

The analogy has limits. EBL needs an explicit representation of what supports the explanation. A human's plausible story about why something succeeded is not necessarily a valid derivation, and a shortcut is reliable only within the assumptions it preserves.

Worked example

Consider a toy rule system for transporting an item with a carrier. Its rules say:

An item fits a carrier if the item's width is no greater than the carrier's opening. The carrier can hold the item if the item's mass is no greater than the carrier's capacity. A move is feasible if the item fits, the carrier can hold it, and the destination is reachable.

For item p, the width is 4 units and mass is 3 units. Carrier c has opening 5 and capacity 6. Destination d is reachable. The item is also colored blue, but color appears in none of the rules.

The explanation establishes 4 ≤ 5, 3 ≤ 6, and reachability, then concludes the move is feasible. Generalizing that explanation gives a direct sufficient condition: item width ≤ carrier opening, item mass ≤ carrier capacity, and destination reachable.

The name p and the color blue are unnecessary. The comparison between this item's dimensions and this carrier's limits is essential and must remain connected correctly. A red item with width 4.5 and mass 5 satisfies the same condition for carrier c and destination d. An item with width 5.5 does not, regardless of color.

This simplified model omits physical transport requirements. The learned rule inherits that scope.

Illustrative case study

Fictional case: Spruce Assembly repeatedly proves that certain component combinations satisfy a configuration requirement. The team adds explanation-based learning to store direct conditions from those proofs.

After an interface rule changes, some stored shortcuts still reflect the earlier theory. The team introduces dependency tracking so affected rules can be reconsidered. The case illustrates that reusing explanations creates a maintenance responsibility: a valid shortcut under yesterday's theory may be invalid under today's revised assumptions.

Common misconceptions and mistakes

Do not confuse EBL with producing a natural-language explanation for a model prediction. Explainable AI and explanation-based learning can interact, but they are different concepts.

Do not generalize every feature of a successful example or discard every detail indiscriminately. The explanation determines which conditions matter.

Avoid presenting sufficient conditions as a complete definition of the target. A different proof may establish the target for cases the learned rule does not cover.

Do not assume that learning from one example requires little prior knowledge. Classical EBL can use one example precisely because a substantial domain theory supplies the justification.

Business use

EBL can be useful where explicit domain rules support expensive repeated reasoning and the same explanatory patterns recur. Its value often lies in making existing knowledge easier to apply.

Measure total performance, including the cost of matching learned rules and maintaining them. Record which theory versions and assumptions support each rule. When the environment is noisy or the domain theory incomplete, distinguish a deductively justified shortcut from a heuristic generalization that needs empirical evaluation.

Check your understanding

In the transport example, why can the learned rule omit the item's color but not its width comparison?

Answer: Color is unused in the derivation. The width comparison establishes fit, which is required by the proof of feasibility.

Related terms

Operationality
Knowledge compilation
Generalization
Explanation-based generalization
domain theory
proof tree
sufficient condition
symbolic learning

Learn next

Study proof trees and sufficient conditions. Explain one small rule-based example, mark every premise actually used, and generalize it with variables. Then find a valid target example that requires a different proof to see why one learned rule may not cover the whole concept.

Sources

Primary source: Gerald DeJong and Raymond Mooney, Explanation-Based Learning: An Alternative View: https://www.cs.utexas.edu/~ai-lab/pubs/ebl-mlj-86.pdf

The paper distinguishes classical generalization from broader EBL. The worked example and fictional case are original.

Previous: Explanation Next: Explanatory model