Information Technology Reference
In-Depth Information
to receive the explanation of concrete instancewhile table GENERAL substitute
the generalized explanation. Let
Ŏ
be the substitution of SPECIFIC,
ȳ
be the
substitution of GENERAL, then the generalization is formally presented as:
for each equality between expression
e
1
and
e
2
in the explanation structure
If
e
1
is the antecedent of a domain rule and
e
2
is the consequent of a
domain rule
Then
let φ = the most general unifier of
2
Ŏ
let
Ŏ
=
Ŏ
φ // update SPECIFIC substitution
let
Ō
= the most general unifier of
e
1
Ŏ
and
e
2
ȳ
let
ȳ
=
ȳŌ
// update GENERAL substitution
Else let φ = the most general unifier of
e
1
ȳ
and
e
e
2
Ŏ
let
Ŏ
=
Ŏ
φ // update SPECIFIC substitution
Based on the above illustration, Mooney and Bennett proposed an algorithm
of Explanation Generalization using Global Substitutions EGGS, which is similar
to abstract STRIPS.
Algorithm 9.1
Explanation Generalization using Global Substitutions (EGGS)
(DeJong and Mooney, 1986)
e
1
Ŏ
and
1.
let
ȳ
be the null substitution {}.
2.
in the explanation structure do
3. let
p
i
' be the result of applying
ȳ
to
p
i.
4. let
p
j
' be the result of applying
ȳ
to
p
j
.
5. let
ȶ
be the MGU of
p
i
' and
p
j
'.
6. let
ȳ
be
ȳȶ
.
7. for each pattern
p
k
in the explanation structure do
8. replace
p
k
with
p
k
ȳ
.
for each equality between
p
i
and
p
j
9.5 Explanation-Based Specialization
Minton and Carbonell from the university of Carnegie-Mellon proposed a
learning method called Explanation-Based Specialization (abbreviated as EBS)
(Minton, 1987) in 1987. A learning system PRODIGY is developed by means of
EBS, which overcomes shortcomings in EBG. EBS learns from multi-object
concept learning, the process of explanation is the detailed illustration of every
goal concept. After the process of explanation, the description of goal concept is
converted to a corresponding control rule, which can be used for selecting proper
node, sub-goal, operators and constraints. The concrete method is as follows:
