Explanation-Based Learning - Advanced Artificial Intelligence - page 336

9.3.2 Interchange with explanation and generalization

In 1987, Hirsh from the University of Stanford proposed a new Explanation
based

Generalization (Hirsh, 1987), which is making explanation and

generalization by turns.
1. Logical description of problem

The learning system provides various ways for inference which include forward,
backward and inductive reasoning. The logical representation makes the semantic
of EBG clearer and provides more convenient language environment. For
instance, suppose the learner needs to learn goal concept Safe-to-Stack (V1, V2),
the domain theory can be represented as:
Fact knowledge
On (obj1, obj2)
Isa (obj2, Endtable)
Color (obj1, red)
Color (obj2, blue)
Volume (obj1, 1)
Density (obj1, 0.1)
Domain rule:
Not (Fragile(y)) ŗ Safe-to-Stack(x,y)
Lighter(x, y) ŗ Safe-to-Stack(x,y)
Volume (p1, v1) ∧ Density (p1, d1) ∧ X (v1, d1, w1) ŗ Weight (p1, w1)
Isa (p1, Endtable) ŗ Weight (p1, 5)
Weight(p1,w1) ∧ Weight(p2,w2) ∧ (w1,w2) ŗ Lighter (p1,p2)
2. Generate explanation structure
In order to proof that the example above satisfies goal concept, the system begins
to process backward inference that is to decompose the goal in terms of existing
facts in knowledge base. Every time a rule is used, the rule is applied to the goal
concept which is variable. In this way, the explanation structure of the example is
yielded while at the same time the variable generalization explanation is
produced as well.
3. Generate the control rule
Conjunction of all the leaf nodes in the explanation structure is considered as
antecedent and the goal concept in the vertex is taken as consequent, the
intermediate composed of the explanation structure is omitted. Therefore the
generalized production rule is created. The solving process is fast and the
efficiency is high when this control rule is applied to resolve similar problems.