Explanation-Based Learning¶
Overview¶
Explanation-based learning (EBL) does not learn new concepts. Instead, it learns new connections among existing concepts by building causal explanations. EBL transfers knowledge from familiar situations to novel ones, making it a form of speed-up learning — the agent becomes more capable without acquiring fundamentally new knowledge.
EBL is closely tied to creativity: when usual tools or methods are unavailable, an agent improvises by finding novel uses for known objects through explanation.
Concept Space and Prior Knowledge¶
EBL operates over a concept space — a network of concepts and their interrelationships. Each concept is represented with:
Structural features — observable properties (e.g., “has a flat bottom,” “is concave,” “has a handle”)
Functional features — behavioral properties (e.g., “is stable,” “is liftable,” “carries liquid”)
Causal connections — links explaining why a functional feature holds (e.g., “the brick is stable because its bottom is flat”)
Example prior knowledge:
Brick: stable because bottom is flat; heavy
Briefcase: liftable because it has a handle and is light; useful because it contains papers
Bowl: carries liquid because it is concave
Glass: enables drinking because it carries liquid and is liftable; pretty
Not all structural features participate in causal explanations — only those linked by causal connections matter for EBL.
Abstraction¶
From a known concept, the agent abstracts a transferable causal pattern by:
Retaining only features involved in causal relationships (dropping causally irrelevant features)
Replacing the specific concept with a generic placeholder (e.g., “bowl” → “object”)
Example: “The bowl carries liquid because it is concave” abstracts to “An object carries liquid because it is concave.”
Transfer and Explanation Construction¶
Given a novel object and a target concept to prove (e.g., “Is this object a cup?”), the agent:
Identifies the conditions the target concept requires (e.g., a cup must be stable and enable drinking)
Searches prior knowledge for abstractions whose causal patterns match observable features of the new object
Chains these abstractions into a causal proof connecting the object’s features to the target concept’s requirements
For example, to prove an object is a cup:
Stability: the object has a flat bottom → stable (transferred from brick)
Enables drinking: the object is concave → carries liquid; has a handle and is light → liftable; carries liquid + liftable → enables drinking (transferred from bowl, briefcase, glass)
The proof the agent builds depends on what background knowledge is available — different knowledge yields different valid proofs.
Proving an Object Is a Mug¶
A mug requires three properties: stable, enables drinking, and protects against heat. The first two can be proved as with a cup. For heat protection, the agent needs a concept like a pot (limits heat transfer because it has thick sides and is made of clay) or an oven mitt / wooden spoon — whichever is available in memory.
This illustrates that the minimal knowledge needed is goal-driven: the agent asks “what do I need to prove X?” and searches for exactly that knowledge.
Everyday Improvisation¶
EBL explains everyday creative problem-solving:
Using a coffee mug as a paperweight (heavy + flat bottom → stable → acts as weight)
Using a chair to prop open a door
Using an eraser as a doorstop
In each case, the agent builds a causal explanation for why a familiar object can serve an unfamiliar function.
Cognitive Connection¶
EBL is central to cognitive science’s goal of human-level intelligence; it models how humans generate and use explanations
Humans can only explain consciously accessible processes — we cannot easily explain implicit skills (e.g., motor coordination)
Generating explanations can deepen understanding by exposing causal structure, but the explanation process may differ from the original reasoning process
For AI systems to earn trust, they must explain both their answers and the reasoning behind them — explanation is fundamental to trust