Explanation-Based Learning (EBL)

Explanation-Based Learning generalises from just one training example — instead of needing many examples like induction, EBL uses a domain theory (existing background knowledge) to explain why the example is an instance of the target concept, and then generalises that explanation.

EBL vs Induction

The Four Inputs to EBL

The EBL Process

Worked Example: "Safe to Stack"

Training example: a specific cup is safe to stack on a specific table.

A domain theory contains general facts such as: an object is stackable on another if the supporting object's top is flat and the supported object's weight is less than the support's strength.

EBL constructs the specific proof for this one example (this particular cup, this particular table), then generalises the proof by replacing the specific objects with variables wherever the proof did not depend on their exact identity — yielding a reusable rule: any lightweight object with a flat base is safe to stack on any object whose top is flat and strong enough. Only the properties that the proof actually used (flat top, sufficient strength, low weight) survive into the generalised rule — irrelevant details of the one example (colour, exact size) are dropped.

Why "Explanation-Based" Rather Than "Example-Based"

The real work is done by the domain theory's explanation, not by statistically pooling many examples — the single example merely triggers the system to construct and then generalise a proof it could, in principle, have constructed without any example at all, given enough time. EBL is therefore sometimes described as "learning by deduction" rather than learning by generalisation from data.

Strengths and Limitations