KE Unit 5 Notes

KNOWLEDGE ENGINEERING
UNIT - 5 NOTES
LEARNING AND RULE LEARNING

8 Learning for Knowledge-based Agents
The previous chapters introduced the main knowledge elements from the knowledge base
of an agent, which are all based on the notion of concept. This chapter presents the basic
operations involved in learning, including comparing the generality of concepts, general-
izing concepts, and specializing concepts. We start with a brief overview of several
machine learning strategies that are particularly useful for knowledge-based agents.
8.1 INTRODUCTION TO MACHINE LEARNING
8.1.1 What Is Learning?

The following are definitions of learning given by some of the most prominent researchers
in this field:
“Learning denotes changes in the system that are adaptive in the sense that they
enable the system to do the same task or tasks drawn from the same population more
efficiently and more effectively the next time” (Simon, 1983, p. 28).
“‘Learning’ is making useful changes in the workings of our minds” (Minsky, 1986,
p. 120).
“Learning is constructing or modifying representations of what is being experienced”
(Michalski, 1986, p. 10).
“A computer program is said to learn from experience E with respect to some class
of tasks T and performance measure P, if its performance at tasks in T, as measured by
P, improves with experience E.” (Mitchell, 1997, p. 2).
Given the preceding definitions, we may characterize learning as denoting the way in
which people and computers:
Acquire, discover, and organize knowledge by building, modifying, and organizing

internal representations of some external reality.
Acquire skills by gradually improving their motor or cognitive abilities through repeated
practice, sometimes involving little or no conscious thought.
There are two complementary dimensions of learning: competence and efficiency. A system
is improving its competence if it learns to solve a broader class of problems and to make
fewer mistakes in problem solving. The system is improving its efficiency if it learns to solve
the problems from its area of competence faster or by using fewer resources.
222
13:53:43,
.009
8.1. Introduction to Machine Learning 223
Machine learning is the domain of artificial intelligence that is concerned with building
adaptive computer systems that are able to improve their performance (competence and/or
efficiency) through learning from input data, from a user, or from their own problem-solving
experience.
Research in machine learning has led to the development of many basic learning
strategies, each characterized by the employment of a certain type of:
Inference (e.g., deduction, induction, abduction, analogy)

Computational or representational mechanism (e.g., rules, trees, neural networks)
Learning goal (e.g., to learn a concept, discover a formula, acquire new facts, acquire
new knowledge about an entity, refine an entity)
The following are some of the most representative learning strategies:
Rote learning
Version space learning
Decision trees induction
Clustering
Rule induction (e.g., Learning rule sets, Inductive logic programming)
Instance-based strategies (e.g., K-nearest neighbors, Locally weighted regression, Col-
laborative filtering, Case-based reasoning and learning, Learning by analogy)
Bayesian learning (e.g., Naïve Bayes learning, Bayesian network learning)
Neural networks and Deep learning
Model ensembles (e.g., Bagging, Boosting, ECOC, Staking)
Support vector machines
Explanation-based learning
Abductive learning
Reinforcement learning
Genetic algorithms and evolutionary computation
Apprenticeship learning
Multistrategy learning
In the next sections, we will briefly introduce four learning strategies that are particularly
useful for agent teaching and learning.
8.1.2 Inductive Learning from Examples

The goal of inductive learning from examples is to learn a general description of a concept,
for instance, the concept of “cup,” by analyzing positive examples of cups (i.e., objects that
are cups) and negative examples of cups (i.e., objects that are not cups).
The learning agent will attempt to find out what is common to the cups and what
distinguishes them from non-cups. For instance, in the illustration from Figure 8.1, the
agent may learn that a cup should have a handle because all the positive examples of cups
have handles, and the negative examples of cups do not have handles. However, the color
does not seem to be important for a cup because a particular color is encountered for both
cups and non-cups.
Learning a good concept description through this learning strategy requires a very large
number of positive and negative examples. On the other hand, this is the only information
13:53:43,
.009
224 Chapter 8. Learning for Knowledge-based Agents
Positive examples of cups: P1 P2 ...
Negative examples of cups: N1 N2 …
Description of the cup concept: has-handle(x), ...
Figure 8.1. Illustration of inductive concept learning from examples.
An example of a cup
cup(o1): color(o1, white), made-of(o1, plastic), light-mat(plastic),
has-handle(o1), has-flat-bottom(o1), up-concave(o1), ...
Proof that o1 is a cup Generalized proof
cup(o1) cup(x)
lif table(o1) stable(o1) open-vessel(o1) lif table(x) stable(x) open-vessel(x)

... ... ... ...
light(o1) graspable(o1) light(x) graspable(x)
light-mat(plastic) light-mat(y)
made-of(o1,plastic) has-handle(o1) made-of(x, y) has-handle(x)
Learned rule
∀x ∀y made-of(x, y), light-mat(y), has-handle(x) ... cup(x)
Figure 8.2. Illustration of explanation-based learning.
the agent needs. For instance, the agent does not require any prior knowledge to perform
this type of learning.
The result of this learning strategy is the increase of the problem-solving competence of
the agent. Indeed, the agent will learn to perform tasks it was not able to perform before,
such as recognizing the cups from a set of objects.
8.1.3 Explanation-based Learning

The goal of explanation-based learning is to improve the agent’s efficiency in problem
solving. The agent is able to perform some task, for example, to recognize cups, but in
an inefficient manner. Through explanation-based learning, the agent learns to perform its
tasks faster, as illustrated in the following.
Let us assume that the agent receives the description of the cup o1 from the top part of
Figure 8.2. Using its prior knowledge, it will recognize that this object is a cup by
performing a complex reasoning process demonstrating that o1 is indeed a cup. This
process is illustrated by the proof tree from the left-hand side of Figure 8.2.
The agent’s reasoning process proceeds as follows:
o1 is made of plastic, which is a light material. Therefore, o1 is light.

o1 has a handle, and it is therefore graspable.
13:53:43,
.009
8.1. Introduction to Machine Learning 225
Being light and graspable, o1 is liftable.

...
Being liftable, stable, and an open vessel, o1 is a cup.
Notice that the agent used the fact that o1 has a handle in order to prove that o1 is a cup.
This means that having a handle is an important feature. On the other hand, the agent did
not use the color of o1 to prove that o1 is a cup. This means that color is not important.
Notice how the agent reaches the same conclusions as in inductive learning from
examples, but through a different line of reasoning, and based on a different type of
information (i.e., prior knowledge instead of multiple examples).
The next step in the learning process is to generalize the proof tree from the left-hand
side of Figure 8.2 into the general tree from the right-hand side. This is done by using the
agent’s prior knowledge of how to generalize the individual inferences from the specific tree.
While the tree from the left-hand side proves that the specific object o1 is a cup, the tree
from the right-hand side proves that any object x that satisfies the leaves of the general tree
is a cup. Thus the agent has learned the general cup recognition rule from the bottom of
Figure 8.2.
To recognize that another object, o2, is a cup, the agent needs only to check that it
satisfies the rule, that is, to check for the presence of these features discovered as
important (i.e., light-mat, has-handle, etc.). The agent no longer needs to build a complex
proof tree. Therefore, cup recognition is done much faster.
Finally, notice that the agent needs only one example from which to learn. However, it
needs a lot of prior knowledge to prove that this example is a cup. Providing such prior
knowledge to the agent is a very complex task.
8.1.4 Learning by Analogy

Learning by analogy is the process of learning new knowledge about some entity by
transferring it from a known similar entity. The central intuition supporting the learning
by analogy paradigm is that if two entities are similar in some respects, then they could be
similar in other respects as well.
An important result of the learning by analogy research (Winston, 1980; Carbonell, 1983,
1986; Gentner, 1983; Davies and Russell, 1990; Forbus et al., 1994; Veloso, 1994) is that the
analogy involves mapping some underlying causal network of relations between analogous
situations. A causal network of relations generally means a set of relations related by higher-
order causal-like relations, such as “physical-cause(ri, rj),” “logically-implies(ri, rj),” “enables
(ri, rj),” “explains(ri, rj),” “justifies(ri, rj),” “determines(ri, rj),” and so on. The idea is that such
similar “causes” are expected to have similar effects. This idea is illustrated in Figure 8.3.
Because A’ is similar to A, which causes B, A’ is expected to cause something (B’) that is
similar to B. Thus analogy involves mapping some underlying “causal network of rela-
tions” between a (better-known) source entity A and a (less-known) target entity B, with
the goal of transferring knowledge from A to B.
For instance, we can teach students about the structure of the hydrogen atom by using the
analogy with the solar system (see Figure 8.4): The hydrogen atom is like our Solar System.
We are telling the students that the hydrogen atom has a similar structure with that of
the Solar System, where the electrons revolve around the nucleus as the planets revolve
around the sun.
13:53:43,
.009
Source similar Target

A A’
causes causes?
similar
B B’
A causes B and A is similar to A’.

Therefore, it is possible that A’ causes something (B’), which is similar to B.
Figure 8.3. Analogical reasoning.
yellow
has as color
Sun nucleus
has as mass has as temp has as mass
mass Sun at tracts temp Sun mass nucleus at tracts
greater than greater than greater than

causes
revolves causes?
around revolves
mass Earth temp Earth mass electron around
has as mass has as temp has as mass
Earth electron
The Solar System The hydrogen atom

Figure 8.4. Illustration of analogical learning.
The students may then infer that other features of the Solar System are also features of
the hydrogen atom. For instance, in the Solar System, the greater mass of the sun and its
attraction of the planets cause the planets to revolve around it. Therefore, the students may
hypothesize that this causal relationship is also true in the case of the hydrogen atom: The
greater mass of the nucleus and its attraction of the electrons cause the electrons to revolve
around the nucleus. This is indeed true and represents a very interesting discovery.
The main problem with analogical reasoning is that not all the features of the Solar
System are true for the hydrogen atom. For instance, the sun is yellow, but the nucleus is
not. Therefore, the information derived by analogy has to be verified.
8.1.5 Multistrategy Learning

As illustrated in Table 8.1, the individual learning strategies have complementary strengths
and weaknesses. For instance, inductive learning from example requires a lot of examples
while explanation-based learning requires only one example. On the other hand, inductive
learning from examples does not require any prior knowledge while explanation-based
learning requires complete prior knowledge.
Multistrategy learning attempts to integrate synergistically such complementary learn-
ing strategies in order to take advantage of their relative strengths to compensate for
13:53:43,
.009
8.2. Concepts 227
Table 8.1 Illustration of the Complementariness of the Learning Strategies
Inductive learning Explanation-based Multistrategy

from examples learning learning
Examples needed many one several

Prior knowledge very little complete incomplete
Type of inference induction deduction induction, deduction, . . .
Effect on agent’s improves competence improves efficiency improves competence
behavior and/or efficiently
positive example concept descript ion

negative exception + ?O1 instance of PhD student
negat ive example - is interested in ?O2
- ?O2 instance of area of expertise
positive exception +
Figure 8.5. Examples and exceptions of a concept.
their relative weaknesses. Multistrategy learning is concerned with developing learning

agents that synergistically integrate two or more learning strategies in order to solve
learning problems that are beyond the capabilities of the individual learning
strategies that are integrated (Tecuci and Michalski, 1991; Tecuci 1993; Michalski and
Tecuci, 1994).
An example of multistrategy learning system is Disciple which integrates learning from
examples, learning from explanations, and learning by analogy and experimentation, as
will be discussed in Chapters 9 and 10.
8.2 CONCEPTS
8.2.1 Concepts, Examples, and Exceptions

As discussed in Section 5.2, a concept represents a set of instances.
With respect to the description of a concept, an instance can be a positive example of
the concept, a negative example, a positive exception, or a negative exception, as illus-
trated in Figure 8.5.
A positive example is an instance of the concept that is covered by the description of the
concept.
A negative example is an instance that does not belong to the concept and is not covered
by the description of the concept.
A positive exception is a positive example that is not covered by the description of the
concept, and any generalization of the description that would cover it would also cover some
negative examples.
13:53:43,
.009
A negative exception is a negative example that is covered by the description of the

concept, and any specialization of the description that would uncover it would also uncover
some positive examples.
The generalization and the specialization of concepts will be discussed in the following
sections.
8.2.2 Examples and Exceptions of a Partially Learned Concept

Figure 8.6 shows the representation of a partially learned concept consisting of the
Plausible Version Space of the Main Condition (with the plausible upper bound in light
green and the plausible lower bound in dark green) and the Plausible Version Space of an
Except-When Condition (with the plausible upper bound in light red and the plausible
lower bound in dark red). In general, a partially learned concept has a plausible version
space of a Main condition, and none, one, or several plausible version spaces of Except-
When conditions.
A partially learned concept may have known positive and negative examples. For the
other instances of the representation space, one may estimate their nature based on their
actual position with respect to the plausible bounds of the concept, as illustrated in
Figure 8.6 and defined in the following.
An instance covered by the plausible lower bound of the Main condition of the concept
and not covered by the plausible version space of any Except-When condition is most
likely a positive example of the concept.
An instance covered by the plausible upper bound of the Main condition of the
concept, but not covered by the plausible lower bound of the Main condition and not
covered by the plausible version space of any Except-When condition, is likely to be
a positive example of the concept.
An instance covered by the plausible lower bound of one of the Except-When
plausible version space conditions of the concept is most likely a negative example of
the concept.
Finally, an instance covered by the plausible upper bound of an Except-When plaus-
ible version space condition of the concept is likely to be a negative example of the
concept.
Plausible Version Space Plausible Version Space of

of the Main Condition the Except-When Condition
Positive
exception +
Negative
exception –
+
Positive
example
Negative –
example
Most likely a More likely a More likely a Most likely a
positive example positive example negative example negative example
Figure 8.6. Examples and exceptions of a partially learned concept.
13:53:43,
.009
8.3. Generalization and Specialization Rules 229
In the next sections, we will describe in detail the basic learning operations dealing with
concepts: the generalization of concepts, the specialization concepts, and the comparison of
the generality of the concepts.
8.3 GENERALIZATION AND SPECIALIZATION RULES
A concept was defined as representing a set of instances. In order to show that a concept
P is more general than a concept Q, this definition would require the computation and
comparison of the (possibly infinite) sets of the instances of P and Q. In this section, we
will introduce generalization and specialization rules that will allow one to prove that a
concept P is more general than another concept Q by manipulating the descriptions of
P and Q, without computing the sets of instances that they represent.
A generalization rule is a rule that transforms (the description of) a concept into (the
description of) a more general concept. The generalization rules are usually inductive
transformations. The inductive transformations are not truth preserving but falsity pre-
serving. That is, if P is true and is inductively generalized to Q, then the truth of Q is not
guaranteed. However, if P is false, then Q is also false.
A specialization rule is a rule that transforms a concept into a less general concept. The
reverse of any generalization rule is a specialization rule. Specialization rules are deduct-
ive, truth-preserving transformations.
A reformulation rule transforms a concept into another, logically equivalent concept.
Reformulation rules are also deductive, truth-preserving transformations.
If one can transform concept P into concept Q by applying a sequence of generalization
rules, then Q is more general than P.
Consider the phrase, “Students who have majored in computer science at George
Mason University between 2007 and 2008.” The following are some of the phrases that
are obvious generalizations of this phrase:
“Students who have majored in computer science between 2007 and 2008”
“Students who have majored in computer science between 2000 and 2012”
“Students who have majored in computer science at George Mason University”
“Students who have majored in computer science”
Some of the phrases that are specializations of the preceding phrase follow:
“Graduate students who have majored in computer science at George Mason Univer-
sity between 2007 and 2008”
“Students who have majored in computer science at George Mason University in 2007”
“Undergraduate students who have majored in both computer science and mathemat-
ics at George Mason University in 2008”
It will be easy to demonstrate the generalization relationships between the preceding

phrases by using some of the following generalization rules, which will be described in the
following sections:
Turning constants into variables

Turning occurrences of a variable into different variables
Climbing the generalization hierarchies
13:53:43,
.009
Dropping conditions
Extending intervals
Extending ordered sets of intervals
Extending discrete sets
Using feature definitions
Using inference rules
8.3.1 Turning Constants into Variables

The turning constants into variables generalization rule consists in generalizing an expres-
sion by replacing a constant with a variable. For example, expression [8.1] represents the
following concept: “The set of professors with 55 publications.”
E1 = ?O1 instance of professor [8.1]

number of publications 55
By replacing 55 with the variable ?N1, which can take any value, we generalize this
concept to the one shown in [8.2]: “The set of professors with any number of publications.”
In particular, ?N1 could be 55. Therefore the second concept includes the first one.

number of publications ?N1
Conversely, by replacing ?N1 with 55, we specialize the concept [8.2] to the concept
[8.1]. The important thing to notice here is that by a simple syntactic operation (turning a
number into a variable), we can generalize a concept. This is one way in which an agent
generalizes concepts.
8.3.2 Turning Occurrences of a Variable into Different Variables

According to this rule, the expression [8.3] may be generalized to the expression [8.4]
by turning the two occurrences of the variable ?O3 in E1 into two variables, ?O31 and ?
O32:
E1 = ?O1 instance of paper [8.3]

is authored by ?O3
?O2 instance of paper
is authored by ?O3
?O3 instance of professor
E2 = ?O1 instance of paper [8.4]
is authored by ?O31
?O2 instance of paper
is authored by ?O32
E1 may be interpreted as representing the concept: “the papers ?O1 and ?O2 authored
by the professor ?O3.” E2 may be interpreted as representing the concept: “the papers ?O1
13:53:43,
.009
and ?O2 authored by the professors ?O31 and ?O32, respectively.” In particular, ?O31 and
?O32 may represent the same professor. Therefore, the second set includes the first one,
and the second expression is more general than the first one.
8.3.3 Climbing the Generalization Hierarchies

One can generalize an expression by replacing a concept from its description with a more general
concept, according to some generalization hierarchy. For instance, the expression [8.5] may be
generalized to [8.6] by replacing the concept assistant professor with the more general concept
professor (see the generalization hierarchy in Figure 5.4, p. 158). The reverse operation of
replacing a concept with a less general one leads to the specialization of an expression.
E1 = ?O1 instance of assistant professor [8.5]

has as employer ?O2
?O2 instance of state university
has as employer ?O2
8.3.4 Dropping Conditions

The agent can generalize a concept by dropping a condition, that is, by dropping a constraint
that its instances must satisfy. For example, the expression [8.7] may be generalized to [8.8]
by removing the constraint on the professor to be employed by a state university.

has as employer ?O2
8.3.5 Extending Intervals

A number may be generalized to an interval containing it. For example, the expression [8.9]
(the set of professors with 55 publications) may be generalized to [8.10] (the set of
professors with 50 to 60 publications), which in turn may be generalized to [8.11] (the
set of professors with 25 to 75 publications).


?N1 is in [50, 60]
?N1 is in [25, 75]
13:53:43,
.009
8.3.6 Extending Ordered Sets of Intervals

An ordered set of intervals is an ordered set of symbols where each symbol corresponds to
an interval. For example, “infant” corresponds to the interval “(0.0, 1.0),” and “toddler”
corresponds to the interval “[1.0, 4.5).” Obviously infant < toddler. Such an ordered set of
intervals may be regarded as an ordered generalization hierarchy, where the nodes are
ordered from left to right, as illustrated in Figure 8.7.
Using such an ordered set of intervals, one may generalize the expression [8.12]
(persons from youth to teen) to the expression [8.13] (persons from youth to mature)
by replacing the symbolic interval “[youth – teen]” with the larger interval “[youth –
mature].”
E1 = ?O1 instance of person [8.12]

has as age [youth – teen]
has as age [youth – mature]
8.3.7 Extending Symbolic Probabilities

A symbolic probability may be generalized by replacing it with a symbolic probability
interval containing it. For example, “very likely” may be generalized to the interval “[likely –
very likely]” (which includes the values “likely” and “very likely”). This interval can be
further generalized by replacing it with a larger interval, such as “[likely – certain].”
8.3.8 Extending Discrete Sets

An expression may be generalized by replacing a discrete set with a larger discrete set that
includes the first set. For example, the expression [8.14] (the set of flags with white color, or
red color, or both) may be generalized to the expression [8.15] (the set of flags with white
color, or red color, or blue color, or any combination of these colors).
E1 = ?O1 instance of flag [8.14]

includes colors from {white, red}
E2 = ?O1 instance of flag [8.15]
includes colors from {white, red, blue}
human age
infant toddler youth teen mature elder
(0.0, 1.0) [1.0, 4.5) [4.5, 12.5) [12.5, 19.5) [19.5, 65.5) [65.5, 150.0]
Figure 8.7. Ordered set of intervals as an ordered generalization hierarchy.
13:53:43,
.009
8.3.9 Using Feature Definitions

This rule generalizes an expression containing a feature, such as, “A feature B,” by replacing
A and B with the domain and the range of feature, respectively.
This rule is illustrated by the generalization of the expression [8.16] (professors who are
experts in computer science) to the expression [8.17] (persons who are experts in some
area of expertise).

is expert in ?O2
?O2 instance of Computer Science

is expert in ?O2
?O2 instance of area of expertise
In the preceding example, professor was generalized to person (the domain of is expert
in), and Computer Science was generalized to area of expertise (the range of is expert in).
In general, a generalization rule indicates how one can generalize a concept through a
simple syntactic transformation without suggesting the actual generalization to perform,
but only a set of possible generalizations. The using feature definition rule actually suggests
a generalization to perform, which will be useful during learning. In fact, it indicates the
most general generalization that can be performed. Indeed, ?O1 in [8.16] has the feature is
expert in. Therefore, it has to be in the domain of this feature, which is person. Similarly,
?O2 is the value of the feature is expert in. Therefore, it has to be in the range of this
feature, which is area of expertise.
8.3.10 Using Inference Rules

Given an inference rule of the form “A ➔ B,” one may generalize an expression by replacing A with
B. For example, using the theorem [8.18] one may generalize the expression [8.19] (students
and their PhD advisors) to the expression [8.20] (students and the professors they know).
8X, 8Y ((X has as PhD advisor Y) ➔ (X knows Y)) [8.18]

E1 = ?O1 instance of student [8.19]
has as PhD advisor ?O2
E2 = ?O1 instance of student [8.20]
knows ?O2
Indeed, by applying the preceding inference rule, one may transform E1 into the
equivalent expression E’1 shown in [8.21]. Then, by dropping the relation has as PhD
advisor, one generalizes E’1 into E2.
E’1 = ?O1 instance of student [8.21]

has as PhD advisor ?O2
knows ?O2
13:53:43,
.009
8.4 TYPES OF GENERALIZATIONS AND SPECIALIZATIONS
Up to this point we have only defined when a concept is more general than another
concept. Learning agents, however, would need to generalize sets of examples and
concepts. In the following we define some of these generalizations.
8.4.1 Definition of Generalization

As defined in Section 5.3, a concept P is said to be more general than another concept
Q if and only if the set of instances represented by P includes the set of instances
represented by Q.
If a concept Q can be transformed into another concept P by applying a sequence of
generalization rules, then P is more general than Q. Indeed, the application of each
successive generalization rule generalizes the concept Q, as shown in the previous
sections. As an illustration, consider the concept C1 in [8.22] and the concept C2 in [8.23].
C1 = ?O1 instance of assistant professor [8.22]

is employed by George Mason University
C2 = ?O1 instance of professor [8.23]
?N1 is in [10, 35]
To show that [8.23] is more general than [8.22] it is enough to show that [8.22] can be
transformed into [8.23] by applying a sequence of generalization rules. The sequence is the
following one:
Generalize assistant professor to professor (climbing the generalization hierarchy).

Generalize 10 to [10, 35] (extending a number to an interval).
Drop “?O1 is employed by George Mason University” (dropping condition).
8.4.2 Minimal Generalization

The concept P is a minimal generalization of a different concept Q if and only if P is a
generalization of Q, and P is not more general than any other generalization of Q.
Consider the generalization hierarchy from Figure 8.8 and the concepts from Figure 8.9.
G3 is a minimal generalization of G2 because there is no concept in Figure 8.8 which is
more general than student and less general than person.
However, G3 is not a minimal generalization of G1. Indeed, the concept G4 in [8.24] is
more general than G1 and less general than G3.
G4 = ?O1 instance of employee [8.24]

is interested in ?O2
If the minimal generalization of a concept is unique, it is called the least general

generalization.
13:53:43,
.009
8.4. Types of Generalizations and Specializations 235
person
employee student
university employee graduate student
graduate graduate
research teaching
assistant assistant
Figure 8.8. Fragment of a generalization hierarchy.
G3 ?O1 instance of person

?O2 instance of area of exper tise
G1 ?O1 instance of university employee G2 ?O1 instance of student

is interested in ?O2 is interested in ?O2
?O2 instance of area of expertise ?O2 instance of area of expertise
C1 ?O1 instance of graduate research assistant

Figure 8.9. Illustration of a minimal generalization.
G1 ?O1 instance of university employee

S1 ?O1 instance of graduate research assistant S2 ?O1 instance of graduate teaching assistant
C1 ?O1 instance of graduate research assistant C2 ?O1 instance of graduate teaching assistant
requires programming requires fieldwork
Figure 8.10. Illustration of a minimal specialization.
8.4.3 Minimal Specialization

The concept Q is a minimal specialization of a different concept P if and only if Q is a
specialization of P, and no other specialization of P is more general than Q. For example,
consider the generalization hierarchy from Figure 8.8 and the concepts from Figure 8.10.
Clearly C1 and C2 are not minimal specializations of G1. However, both S1 and S2 are
minimal specializations of G1.
13:53:43,
.009
8.4.4 Generalization of Two Concepts

The concept Cg is a generalization of the concepts C1 and C2 if and only if Cg is more general
than C1 and Cg is more general than C2. One may show that if both the concept C1 and the
concept C2 can be transformed into the concept Cg by applying generalization rules, then
Cg is a generalization of C1 and C2.
As an illustration, consider the concept C1 described by the expression [8.25], and the
concept C2 described by the expression [8.26]. One may easily show that the concept
C described by the expression [8.27] is a generalization of C1 and C2. Indeed, by general-
izing assistant professor to professor (through climbing the generalization hierarchy), by
generalizing 10 to [10, 35] (through replacing a number with an interval containing it), and
by dropping the condition “?O1 is employed by George Mason University,” one generalizes
C1 to C. Similarly, by generalizing associate professor to professor, and 35 to [10, 35], one
generalizes C2 into C. Therefore, C is a generalization of C1 and C2.
C1 = ?O1 instance of assistant professor [8.25]

C2 = ?O1 instance of associate professor [8.26]
C = ?O1 instance of professor [8.27]
?N1 is in [10, 35]
In general, to determine a generalization of two concepts, each represented as a clause,

such as [8.25] and [8.26], one first matches the features of the two concepts. Then one
applies the dropping condition rule to remove the unmatched features (and possibly even
matched features). Finally, one applies other generalization rules to determine the gener-
alizations of the matched feature values. Notice that usually there are several generaliza-
tions of two concepts.
In a similar way, one can determine a generalization G of two more complex expres-
sions E1 and E2, each consisting of a conjunction of clauses. G will consist of the
conjunction of the generalizations of some of the corresponding clauses in the two
expressions E1 and E2.
8.4.5 Minimal Generalization of Two Concepts

The concept G is a minimal generalization of the concepts C1 and C2 if and only if G is a
generalization of C1 and C2, and G is not more general than any other generalization of
C1 and C2.
To determine a minimal generalization of two clauses, one has to keep all the common
features of the clauses and determine a minimal generalization of each of the matched
feature values. In a similar way, one determines the minimal generalization of two
conjunctions of clauses by matching the clauses and determining the minimal generaliza-
tions of the matched clauses. These procedures are correct if we assume that there are no
other common features due to theorems. Otherwise, all the common features will have to
first be made explicit by applying the theorems.
13:53:43,
.009
8.4. Types of Generalizations and Specializations 237
Notice, however, that there may be more than one minimal generalization of two
expressions. For instance, according to the generalization hierarchy from the middle of
Figure 8.8, there are two minimal generalizations of graduate research assistant and
graduate teaching assistant. They are university employee and graduate student. Conse-
quently, there are two minimal generalizations of S1 and S2 in Figure 8.11: mG1 and
mG2. The generalization mG1 was obtained by generalizing graduate research assistant and
graduate teaching assistant to university employee. mG2 was obtained in a similar fashion,
except that graduate research assistant and graduate teaching assistant were generalized to
graduate student. Neither mG1 nor mG2 is more general than the other. However, G3 is
more general than each of them.
Disciple agents employ minimal generalizations, also called maximally specific gener-
alizations (Plotkin, 1970; Kodratoff and Ganascia, 1986). They also employ maximal
generalizations, also called maximally general generalizations (Tecuci and Kodratoff,
1990; Tecuci, 1992; Tecuci 1998).
8.4.6 Specialization of Two Concepts

The concept Cs is a specialization of the concepts C1 and C2 if and only if Cs is less general
than C1 and Cs is less general than C2. For example, research professor is a specialization of
researcher and professor.
One can easily show that if both the concept C1 and the concept C2 can be transformed
into the concept Cs by applying specialization rules (or Cs can be transformed into C1, and
it can also be transformed into C2 by applying generalization rules), then Cs is a special-
ization of C1 and C2.
8.4.7 Minimal Specialization of Two Concepts

The concept C is a minimal specialization of two concepts C1 and C2 if and only if C is
a specialization of C1 and C2 and no other specialization of C1 and C2 is more general
than C.
G3 ?O1 instance of person

mG1 ?O1 instance of university employee mG2 ?O1 instance of graduate student
S1 ?O1 instance of graduate research assistant S2 ?O1 instance of graduate teaching assistant
Figure 8.11. Several generalizations of the concepts S1 and S2.
13:53:43,
.009
The minimal specialization of two clauses consists of the minimal specialization of the
matched feature-value pairs, and of all the unmatched feature-value pairs. This procedure
assumes that no new clause feature can be made explicit by applying theorems. Otherwise,
one has first to make all the features explicit.
The minimal specialization of two conjunctions of clauses C1 and C2 consists of the
conjunction of the minimal specializations of each of the matched clauses of C1 and C2,
and of all the unmatched clauses from C1 and C2.
Figure 8.12 shows several specializations of the concepts G1 and G2. mS1 and mS2 are
two minimal specializations of G1 and G2 because graduate research assistant and graduate
teaching assistant are two minimal specializations of university employee and graduate
student.
Notice that in all the preceding definitions and illustrations, we have assumed that the
clauses to be generalized correspond to the same variables. If this assumption is not
satisfied, then one would need first to match the variables and then compute the general-
izations. In general, this process is computationally expensive because one will need to try
different matchings.
8.5 INDUCTIVE CONCEPT LEARNING FROM EXAMPLES
Inductive concept learning from examples has already been introduced in Section 8.1.2. In
this section, we will discuss various aspects of this learning strategy that are relevant to
agent teaching and learning. The problem of inductive concept learning from examples
can be more precisely defined as indicated in Table 8.2.
The bias of the learning agent is any basis for choosing one generalization over another,
other than strict consistency with the observed training examples (Mitchell, 1997). In the
following, we will consider two agents that employ two different preference biases: a
cautious learner that always prefers minimal generalizations, and an aggressive learner
that always prefers maximal generalizations.
Let us consider the positive examples [8.28] and [8.29], and the negative example [8.30]
of a concept to be learned by these two agents in the context of the generalization
hierarchies from Figure 8.13.
G1 ?O1 instance of university employee G2 ?O1 instance of graduate student

mS1 ?O1 instance of graduate research assistant mS2 ?O1 instance of graduate teaching assistant
C1 ?O1 instance of graduate research assistant C2 ?O1 instance of graduate teaching assistant
requires programming requires fieldwork
Figure 8.12. Several specializations of the concepts G1 and G2.
13:53:43,
.009
8.5. Inductive Concept Learning from Examples 239
Table 8.2 The Problem of Inductive Concept Learning from Examples
Given
A language of instances.
A language of generalizations.
A set of positive examples (E1, . . ., En) of a concept.
A set of negative (or counter) examples (C1, . . ., Cm) of the same concept.
A learning bias.
Other background knowledge.
Determine
A concept description that is a generalization of the positive
examples and that does not cover any of the negative examples.
Purpose of concept learning
Predict if an instance is a positive example of the learned concept.
person
university
staf f member professor
state private
university university
computer full associate assistant
technician professor professor professor
George Janet Mark George Mason University Stanford

Dean Rice White University of Virginia University
Figure 8.13. Two generalization hierarchies.
Positive examples:
Mark White instance of assistant professor [8.28]
Janet Rice instance of assistant professor [8.29]

is employed by University of Virginia
Negative example:
George Dean instance of computer technician [8.30]
is employed by Stanford University
What concept might be learned by the cautious learner from the positive examples
[8.28] and [8.29], and the negative example [8.30]? The cautious learner would learn a
minimal generalization of the positive examples, which does not cover the negative
example. Such a minimal generalization might be the expression [8.31], “an assistant
professor employed by a state university,” obtained by minimally generalizing George
Mason University and University of Virginia to state university.
13:53:43,
.009
?O1 instance of assistant professor [8.31]

is employed by ?O2
The concept learned by the cautious learner is represented in Figure 8.14 as the
minimal ellipse that covers the positive examples without covering the negative example.
Assuming a complete ontology, the learned concept is included into the actual concept.
How will the cautious learner classify each of the instances represented in Figure 8.14
as black dots? It will classify the dot covered by the learned concept as positive example,
and the two dots that are not covered by the learned concept as negative examples.
How confident are you in the classification, when the learner predicts that an instance
is a positive example? When a cautious learner classifies an instance as a positive example
of a concept, this classification is correct because an instance covered by the learned
concept is also covered by the actual concept.
But how confident are you in the classification, when the learner predicts that an
instance is a negative example? The learner may make mistakes when classifying an
instance as a negative example, such as the black dot that is covered by the actual concept
but not by the learned concept. This type of error is called “error of omission” because
some positive examples are omitted – that is, they are classified as negative examples.
Let us now consider the concept that might be learned by the aggressive learner from
the positive examples [8.28] and [8.29], and the negative example [8.30]. The aggressive
learner will learn a maximal generalization of the positive examples that does not cover the
negative example. Such a maximal generalization might be the expression [8.32], “a
professor employed by a university.” This is obtained by generalizing assistant professor
to professor (the most general generalization that does not cover computer technician from
the negative example) and by maximally generalizing George Mason University and Univer-
sity of Virginia to university. Although university covers Stanford University, this is fine
because the obtained concept [8.32] still does not cover the negative example [8.30].
?O1 instance of professor [8.32]

is employed by ?O2
?O2 instance of university
An alternative maximally general generalization of the positive examples [8.28] and

[8.29], which does not cover the negative example [8.30], is [8.33].
?O1 instance of person [8.33]

is employed by ?O2
Concept learned by
the cautious learner
-
Actual concept
+
+
Figure 8.14. Learning and classifications by a cautious learner.
13:53:43,
.009
8.5. Inductive Concept Learning from Examples 241
The concept learned by the aggressive learner is represented in Figure 8.15 as the
maximal ellipse that covers the positive examples without covering the negative example.
Assuming a complete ontology, the learned concept includes the actual concept.
How will the aggressive learner classify each of the instances represented in Figure 8.15
as black dots? It will classify the dot that is outside the learned concept as a negative
example, and the other dots as positive examples.
How confident are you in the classification when the learner predicts that an instance is
negative example? When the learner predicts that an instance is a negative example, this
classification is correct because that instance is not covered by the actual concept, which is
itself covered by the learned concept.
But, how confident are you in the classification when the learner predicts that an
instance is a positive example? The learner may make mistakes when predicting that an
instance is a positive example, as is the case with the dot covered by the learned concept,
but not by the actual concept. This type of error is called “error of commission” because
some negative examples are committed – that is, they are classified as positive examples.
Notice the interesting fact that the aggressive learner is correct when it classifies
instances as negative examples (they are indeed outside the actual concept because they
are outside the concept learned by the aggressive learner) while the cautious learner is
correct when it classifies instances as positive examples (they are inside the actual concept
because they are inside the concept learned by the cautious learner). How could one
synergistically integrate these two learning strategies to take advantage of their comple-
mentariness? An obvious solution is to use both strategies, learning both a minimal and a
maximal generalization from the examples, as illustrated in Figure 8.16.
What class will be predicted by a dual-strategy learner for the instances represented as
black dots in Figure 8.16? The dot covered by the concept learned by the cautious learner
Concept learned by the aggressive learner
-
Actual concept
+
+
Figure 8.15. Learning and classifications by an aggressive learner.
Concept learned by the aggressive learner

-
Actual concept
+
+
Concept learned by the cautious learner
Figure 8.16. Learning and classifications by a dual-strategy learner.
13:53:43,
.009
will be classified, with high confidence, as a positive example. The dot that is not covered
by the concept learned by the aggressive learner will be classified, again with high
confidence, as a negative example. The dual-strategy learner will indicate that it cannot
classify the other two dots.
8.6 LEARNING WITH AN INCOMPLETE REPRESENTATION

LANGUAGE
Let us consider the ontology from Figure 8.17. What is the maximal generalization of the
positive examples John Doe and Jane Austin that does not cover the given negative example
Bob Sharp, in the case where graduate research assistant is included into the ontology? The
maximal generalization is faculty member.
But what is the maximal generalization in the case where graduate research assistant is
missing from the ontology? In this case, the maximal generalization is employee, which is,
in fact, an overgeneralization.
What is the minimal specialization of person that does not cover Bob Sharp in the case
where graduate research assistant is included into the ontology? It is faculty member.
But what is the minimal specialization in the case where graduate research assistant is
missing from the ontology? In this case, the minimal specialization is employee, which
is an underspecialization.
person
subconcept of
employee student
subconcept of
subconcept of
university employee
graduate
subconcept of student
undergraduate
staff member faculty member student
subconcept of
subconcept of
subconcept of
instructor professor PhD advisor graduate
research teaching
assistant assistant
subconcept of BS student
assistant associate full PhD student MS student

professor professor professor
instance of instance of instance of instance of instance of
John Smith John Doe Jane Ausn Bob Sharp Joan Dean
_
+ +
Figure 8.17. Plausible generalizations and specializations due to ontology incompleteness.
13:53:43,
.009
8.7. Formal Definition of Generalization 243
Notice that the incompleteness of the ontology causes the learner both to overgeneral-
ize and underspecialize. In view of the preceding observations, what can be said about the
relationships between the concepts learned using minimal and maximal generalizations
and the actual concept when the ontology and the representation language are incom-
plete? The minimal and maximal generalizations are only approximations of the actual
concept, as shown in Figure 8.18.
Why is the concept learned with an aggressive strategy more general than the one
learned with a cautious strategy? Because they are based on the same ontology and
generalization rules.
8.7 FORMAL DEFINITION OF GENERALIZATION
8.7.1 Formal Representation Language for Concepts

A knowledge representation language defines the syntax and the semantics for expressing
knowledge in a form that an agent can use. We define a formal representation language for
concepts as follows:
Let be a set of variables. For convenience in identifying variables, their names start
with “?,” as in, for instance, ?O1. Variables are used to denote unspecified instances of
concepts.
Let be a set of constants. Examples of constants are the numbers (such as “5”), strings
(such as “programming”), symbolic probability values (such as “very likely”), and
instances (such as “John Doe”). We define a term to be either a variable or a constant.
Let be a set of features. The set includes the domain independent features
“instance of,” “subconcept of,” and “direct subconcept of,” as well as other domain-
specific features, such as “is interested in.”
Let be an object ontology consisting of a set of concepts and instances defined using the
clause representation [7.2] presented in Section 7.2, where the feature values (vi1 . . . vim)
are constants, concepts, instances, or intervals (numeric or symbolic). That is, there are no
variables in the definition of a concept or an instance from , such as the following one:
instance-k instance of concept-i

feature-i1 vi1
...
feature-in vim
With complete ontology With incomplete ontology
Learned by aggressive learner Learned by aggressive learner

Actual concept
Actual concept
Learned by Learned by
cauous learner cauous learner
Figure 8.18. The influence of the ontology on concept learning.
13:53:43,
.009
The concepts and the instances from are related by the generalization relations
“instance of” and “subconcept of.” includes the concept object, which represents all
the instances from the application domain and is therefore more general than any
other object concept.
Let be the set of the theorems and the properties of the features, variables, and
constants.
Two properties of any feature are its domain and its range. Other features may have
special properties. For instance, the relation subconcept of is transitive (see Section
5.7). Also, a concept or an instance inherits the features of the concepts that are more
general than it (see Section 5.8).
Let be a set of connectors. includes the logical connectors AND (∧), OR (∨), and
NOT (Except–When), the connectors “{” and “}” for defining alternative values of a
feature, the connectors “[” and “]” as well as “(” and “)” for defining a numeric or a
symbolic interval, the delimiter “,” (a comma), and the symbols “Plausible Upper
Bound” and “Plausible Lower Bound.”
We call the tuple a representation language for concepts.

The basic representation unit (BRU) for a concept in has the form of a tuple (?O1,
?O2,. . ., ?On), where each ?Oi is a clause:
?O1 instance of concept-i

feature-11 ?Oj
...
feature-1i ?Ol
...
?On instance of concept-n
feature-n1 ?Op
...
feature-nk ?Ot
In the preceding expression, each of concept-i . . . concept-n is either an object concept from
the object ontology (such as PhD student), a numeric interval (such as [50, 60]), a set of
numbers (such as {1, 3, 5}), a set of strings (such as {white, red, blue}), a symbolic probability
interval (such as [likely - very likely]), or an ordered set of intervals (such as [youth - mature]).
?Oj, . . . , ?Ol,. . ., ?Op, . . . , ?Ot are distinct variables from the sequence (?O1, ?O2, . . . , ?On).
When concept-n is a set or interval such as “[50, 60],” we use “is in” instead of “instance of.”
A more complex concept is defined as a conjunctive expression “BRU ∧ not BRU1 ∧ . . . ∧
not BRUp,” where “BRU” and each “BRUk (k = 1, . . . , p)” is a conjunction of clauses. This is
illustrated by the following example, which represents the set of instances of the tuple
(?O1, ?O2, ?O3), where ?O1 is a professor employed by a university ?O2 in a long-term
position ?O3, such that it is not true that ?O1 plans to retire from ?O2 or to move to some
other organization:

has as employer ?O2
has as position ?O3
?O3 instance of long-term faculty position
13:53:43,
.009
8.7. Formal Definition of Generalization 245
Not
plans to retire from ?O2
Not
plans to move to ?O4
?O4 instance of organization
We generally use “Except-When” instead of “Not.”

In the following sections, we will provide a formal definition of generalization in the
representation language , based on substitutions.
A substitution is a function σ = (x1 t1, . . . , xn tn), where each xi (i = 1,. . .,n) is a
variable and each ti (i = 1,. . .,n) is a term.
If li is an expression in the representation language , then σli is the expression
obtained by substituting each variable xi from li with the term ti.
8.7.2 Term Generalization

In the representation language , a term is a constant (e.g., number, string, symbolic
interval, or instance) or a variable. An unrestricted variable ?X is more general than any
constant and is as general as any other unrestricted variable (such as ?Y).
8.7.3 Clause Generalization

Let us consider two concepts described by the following two clauses, C1 and C2, respect-
ively, where v1, v2, v11, . . ., v2n are variables, b1 and b2 are concepts, and f11, . . ., f2n are
features from the ontology.
C1 = v1 instance of b1
f11 v11
...
f1m v1m
C2 = v2 instance of b2
f21 v21
...
f2n v2n
We say that the clause C1 is more general than the clause C2 if there exists a substitution
σ such that:
σv1 = v2
b1 = b2
8i2{1,. . .,m}, 9 j2{1,. . .,n} such that f1i = f2j and σv1i = v2j.
For example, the concept
C1 = ?X instance of student
enrolled at George Mason University
13:53:43,
.009
is more general than the concept
C2 = ?Y instance of student
enrolled at George Mason University
has as sex female
Indeed, let σ = (?X ?Y). As one can see, σC1 is a part of C2, that is, each feature of
σC1 is also a feature of C2. The first concept represents the set of all students enrolled at
George Mason University, while the second one represents the set of all female students
enrolled at George Mason University. Obviously the first set includes the second one, and
therefore the first concept is more general than the second one.
Let us notice, however, that this definition of generalization does not take into account
the theorems and properties of the representation language . In general, one needs to
use these theorems and properties to transform the clauses C1 and C2 into equivalent
clauses C’1 and C’2, respectively, by making explicit all the properties of these clauses.
Then one shows that C’1 is more general than C’2. Therefore, the definition of the more
general than relation in is the following one:
A clause C1 is more general than another clause C2 if and only if there exist C’1, C’2,
and a substitution σ, such that:
C0 1 ¼L C1
C0 2 ¼L C2
σv1 ¼L v2
b1 is more general than b2 in
8 i2f1; . . . ; mg, 9 j2f1; . . . ; ng such that f 0 1i ¼L f 0 2j and σv0 1i ¼L v0 2j :
In the following sections, we will always assume that the equality is in and we will no
longer indicate this.
8.7.4 BRU Generalization

As discussed in Section 8.7.1, a basic representation unit (BRU) is a conjunction of clauses.
An example of BRU is the following one where, for notational convenience, we have
dropped the AND connector between the clauses:
?O1 instance of course

has as reading ?O2
?O2 instance of publication
has as author ?O3
Therefore, anytime there is a sequence of clauses, they are to be considered as being

connected by AND (∧).
Let us consider two concepts, A and B, defined by the following expressions
A ¼ A1 ∧A2 ∧ . . . ∧An
B ¼ B1 ∧B2 ∧ . . . ∧Bm
where each Ai (i = 1, ...,n) and each Bj (j = 1, ...,m) is a clause.

13:53:43,
.009
8.8. Review Questions 247
A is more general than B if and only if there exist A’, B’, and σ such that:
A0 ¼ A, A0 ¼ A0 1 ∧A0 2 ∧ . . . ∧A0 p
B0 ¼ B, B0 ¼ B0 1 ∧B0 2 ∧ . . . ∧B0 q
8 i2f1; . . . ; pg, 9 j2f1; . . . ; qg such that σA0 i ¼ B0 j :
Otherwise stated, one transforms the concepts A and B, using the theorems and the
properties of the representation language, so as to make each clause from A’ more general
than a corresponding clause from B’. Notice that some clauses from B’ may be “left over,”
that is, they are not matched by any clause of A’.
8.7.5 Generalization of Concepts with Negations

By concept with negations, we mean an expression of the following form:
BRU∧not BRU1 ∧ . . . ∧not BRUp
where each BRU is a conjunction of clauses.

Let us consider two concepts with negations, A and B, defined by the following
expressions:
A ¼ BRUa ∧not BRUa1 ∧ . . . ∧not BRUap
B ¼ BRUb ∧not BRUb1 ∧ . . . ∧not BRUbq
A is more general than B if and only if there exist A0 , B0 , and σ, such that:
A0 ¼ A, A0 ¼ BRU0 a ∧not BRU0 a1 ∧ . . . ∧not BRU0 ap
B0 ¼ B, B0 ¼ BRU0 b ∧not BRU0 b1 ∧ . . . ∧not BRU0 bq
σBRU0 a is more general than BRU0 b
8 i2f1; . . . ; pg, 9 j2f1; . . . ; qg such that BRU0 bj is more general than σBRU0 ai :
8.7.6 Substitutions and the Generalization Rules

One can use the definition of generalization based on substitution to prove that the
generalization rules transform concepts into more general concepts.
As an illustration, let us consider the turning constants into variables generalization
rule (see Section 8.3.1) that transformed the expression E1 from [8.1] on p. 230, into the
expression E2 from [8.2] on p. 230. E2 is indeed a generalization of E1 because E1 = σE2,
where σ = (?N1 55).
8.8 REVIEW QUESTIONS
8.1. What is a positive example of a concept? What is a negative example of a concept?
8.2. What is a generalization rule? What is a specialization rule? What is a

reformulation rule?
13:53:43,
.009
8.3. Name all the generalization rules you know.

8.4. Define and illustrate the following generalization rule: climbing generalization
hierarchies.
8.5. Briefly describe and illustrate the “turning constants into variables” generalization rule.
8.6. Define and illustrate the “dropping conditions” generalization rule.
8.7. Define the minimal generalization G of two concepts A and B. Then define their
least general generalization. Do these generalizations always exist?
8.8. What is a negative exception? What is a positive exception?
8.9. Indicate three different generalizations of the phrase “History books in the Fenwick
Library,” and demonstrate that each of them is more general than the given phrase.
8.10. Give an example of a natural language phrase C that has a concept interpretation.
(a) Formulate a phrase G that is a generalization of C, and use the generalization
rules to demonstrate that G is a generalization of C.
(b) Formulate a phrase S that is a specialization of C, and use the generalization
rules to demonstrate that S is a specialization of C.
(c) Formulate a phrase D that is neither a generalization of C nor a specialization of C.
8.11. P3 in Figure 8.19 is a natural language sentence that represents a concept. Fill in each
of the other boxes with sentences such that the arrows from the figure represent all the
“subconcept of” relationships between these concepts. This means, for instance, that
P3 is less general than P5, P3 is more general than P1, and P3 is more general than P2.
However, P3 is not more general than P4, and P4 is not more general than P3. Justify
each shown “subconcept of” relationship by labeling it with the name of the general-
ization rule(s) that transform the less general concept into the more general concept.
You should use at least three different generalization rules while solving this problem.
8.12. Draw a picture representing a plausible version space, as well as a positive
example, a negative example, a positive exception, and a negative exception. Then
briefly define each of these notions.
8.13. Consider the cells consisting of two bodies, each body having two attributes,
number of nuclei (1 or 2), and color (yellow or green). The relative position of
P5
P3 The U.S. athletes who have competed in the P4

2012 Olympic games in individual competitions.
P1 P2
Figure 8.19. Determining concepts that satisfy given “subconcept of” relationships.
13:53:43,
.009
the bodies is not relevant because they can move inside the cell. For example,
((1 green) (2 yellow)) is the same with ((2 yellow) (1 green)) and represents a cell
where one body has one nucleus and is green, while the other body has two nuclei
and is yellow. You should also assume that any generalization of a cell is also
described as a pair of pairs ((s t) (u v)).
(a) Indicate all the possible generalizations of the cell from Figure 8.20 and the
generalization relations between them.
(b) Determine the number of the distinct sets of instances and the number of the
concept descriptions from this problem.
(c) Consider the cell descriptions from Figure 8.21 and determine the following
minimal generalizations: g(E1, E2), g(E2, E3), g(E3, E1), g(E1, E2, E3).
8.14. Consider the ontology fragment from the loudspeaker manufacturing domain,
shown in Figure 5.23 (p. 171), and the following expressions:
E1: ?X instance of membrane E2: ?X instance of mechanical chassis

made of ?M made of ?M
?M instance of paper ?M instance of metal
?Z instance of contact adhesive ?Z instance of mowicoll
glues ?M glues ?M
state fluid
(a) Find the minimal generalizations of E1 and E2.

(b) Find two generalizations of E1 and E2 that are not minimal generalizations.
(c) Consider one of the generalizations found at (b) and demonstrate that it is a
generalization of E1 and E2 but it is not a minimal generalization.
(d) What would be the least general generalization of E1 and E2? Does it exist?
(e) Indicate a specialization of E1.
8.15. Consider the ontology fragment from the loudspeaker manufacturing domain,
shown in Figure 5.24 (p. 172). Notice that each most specific concept, such as dust
or air press, has an instance, such as dust1 or air press1.
Consider also the following two expressions:
+ ((1 green) (2 yellow))
Figure 8.20. An example of a cell.
E1: ((1 green) (1 green)) E2: ((1 yellow) (2 green)) E3: ((1 green) (2 green))
Figure 8.21. The descriptions of three cells.
13:53:43,
.009
E1: ?X instance of soft cleaner

removes ?Z
?Y instance of membrane
?Z instance of waste material
E2: ?X instance of air sucker

removes ?Z
?Y instance of membrane
?Z instance of dust
Use the generalization rules to show that E1 is more general than E2.
8.16. Determine a generalization of the following two expressions in the context of the
ontology fragment from Figure 5.24 (p. 172):
E1: ?X instance of entrefer

may have ?Y
?Y instance of dust
?Z instance of air sucker
removes ?Y
E2: ?X instance of membrane
may have ?Y
?Y instance of surplus adhesive
?Z instance of alcohol
removes ?Y
8.17. Consider the background knowledge represented by the generalization hier-
archies shown in Figure 8.22.
Consider also the following concept:
E: ?O instance of object
color yellow
shape circle
radius 5
Indicate five different generalization rules. For each such rule, determine an
expression Eg that is more general than E according to that rule.
any color any shape
polygon round
warm color cold color
triangle rectangle circle ellipse
red orange yellow blue green black

square
Figure 8.22. Generalization hierarchies for color and shape.
13:53:43,
.009
adhesive toxic substance inflammable object
scotch tape super glue mowicoll contact adhesive
loudspeaker component material
membrane chassis assembly bolt caoutchouc paper metal
Figure 8.23. Ontology fragments.
8.18. Consider the following two concepts:
C1: ?X instance of screw C2: ?X instance of nut

head ?M cost 6
cost 5
Indicate two different generalizations of them.
8.19. Consider the following two concepts G1 and G2, and the ontology fragments in
Figure 8.23. Indicate four specializations of G1 and G2 (including a minimal
specialization).
G1:?X instance of loudspeaker component G2: ?X instance of loudspeaker component

made of ?M made of ?M
?M instance of material ?M instance of material
?Z instance of adhesive ?Z instance of inflammable object
glues ?M glues ?M
8.20. Illustrate the clause generalization defined in Section 8.7.3 with an example from
the PhD Advisor Assessment domain.
8.21. Illustrate the BRU generalization defined in Section 8.7.4 with an example from the
PhD Advisor Assessment domain.
8.22. Illustrate the generalization of concepts with negations defined in Section 8.7.5 by
using an example from the PhD Advisor Assessment domain.
8.23. Use the definition of generalization based on substitution to prove that each of the
generalization rules discussed in Section 8.3 transforms a concept into a more
general concept.
13:53:43,
.009
9 Rule Learning
In this and the next chapter on rule refinement, we will refer to both problems and
hypotheses, interchangeably, to emphasize the fact that the learning methods presented
are equally applicable in the context of hypotheses analysis and problem solving.
9.1 MODELING, LEARNING, AND PROBLEM SOLVING
Figure 9.1 summarizes the interactions between the subject matter expert and the learning
agent that involve modeling, learning, and problem solving.
The expert formulates the problem to be solved (or the hypothesis to be analyzed), and
the agent uses its knowledge to generate a (problem-solving or argumentation) tree to be
verified by the expert.
Several cases are possible. If the problem is not completely solved, the expert will extend
the tree with additional reductions and provide solutions for the leaf problems/hypotheses.
P1
Problem Question
Answer
1 1 1
P1 S1 Pn
Mixed-Initiative
Question
Problem Solving Question
Answer
Answer
2 2 2 2
P1 S1 Pm Sm
Question
Reasoning Tree Question
Answer
Answer Ontology + Rules
3 3 3
Extend Tree Accept step P1 S1 Sp
3
Sp
1
Pn
Modeling Reject step Rule Refinement Refined

Question
Answer with Positive Rules
2
Pn1
2
Pnk
Example
Rule Refinement
Rule with Negative Refined Rules
Learning Example
Explanation-based Refined Ontology

Ontology Extension
Learned Rules
Figure 9.1. Mixed-initiative modeling, learning, and problem solving.
252
13:59:50,
.010
9.2. Rule Learning and Refinement 253
From each new reduction provided by the expert, the agent will learn a new rule, as will be
presented in the following sections.
If the expert rejects any of the reasoning steps generated by the agent, then an explan-
ation of why that reduction is wrong needs to be determined, and the rule that generated it
will be refined to no longer generate the wrong reasoning step.
If the expert accepts a reasoning step as correct, then the rule that generated it may be
generalized. The following section illustrates these interactions.
9.2 AN ILLUSTRATION OF RULE LEARNING AND REFINEMENT
As will be discussed in the following, the subject matter expert helps the agent to learn
by providing examples and explanations, and the agent helps the expert to teach it by
presenting attempted solutions.
First, as illustrated in Figure 9.2, the expert formulates the problem to solve or the
hypothesis to analyze which, in this illustration, is the following hypothesis:
John Doe would be a good PhD advisor for Bob Sharp.
In this case, we will assume that the agent does not know how to assess this hypothesis.
Therefore, the expert has to teach the agent how to assess it. The expert will start by
developing a reduction tree, as discussed in Chapter 4 and illustrated in the middle of
Figure 9.2. The initial hypothesis is first reduced to three simpler hypotheses, guided
by a question/answer pair. Then each of the subhypotheses is further reduced, either
to a solution/assessment or to an elementary hypothesis to be assessed based on
evidence. For example, the bottom part of Figure 9.2 shows the reduction of the first
subhypothesis to an assessment.
After the reasoning tree has been developed, the subject matter expert interacts with
the agent, helping it “understand” why each reduction step is correct, as will be discussed
in Section 9.5. As a result, from each reduction step the agent learns a plausible version
1. Modeling 2. Learning
Expert explains Agent learns
how to solve a general
specific problem reducon rules
Rule1
Rule2
Figure 9.2. Modeling and rule learning.
13:59:50,
.010
254 Chapter 9. Rule Learning
space rule, as a justified generalization of it. This is illustrated in the right-hand side of
Figure 9.2 and discussed in Section 9.7. These rules are not shown to the expert, but they
may be viewed with the Rule Browser.
The agent can now use the learned rules to assess by itself similar hypotheses formu-
lated by the expert, as illustrated in Figure 9.3, where the expert formulated the following
hypothesis:
Dan Smith would be a good PhD advisor for Bob Sharp.
The reduction tree shown in Figure 9.3 was generated by the agent. Notice how the agent
concluded that Bob Sharp is interested in an area of expertise of Dan Smith, which is
Information Security, by applying the rule learned from John Doe and Bob Sharp, who share
a common interest in Artificial Intelligence.
The expert has to inspect each reduction generated by the agent and indicate whether it
is correct or not. Because the reductions from Figure 9.3 are correct, the agent generalizes
the lower bound conditions of the applied rules, if the reductions were generated based on
the upper bound conditions of these rules.
The bottom part of Figure 9.4 shows a reduction generated by the agent that is rejected
by the expert. While Dan Smith has indeed a tenured position, which is a long-term faculty
position, he plans to retire. It is therefore wrong to conclude that it is almost certain that he
will stay on the faculty of George Mason University for the duration of the dissertation of
Bob Sharp.
Such failure explanations are either proposed by the agent and accepted by the expert,
or are provided by the expert, as discussed in Section 9.5.2.
Based on this failure explanation, the agent specializes the rule that generated this
reduction by adding an Except-When plausible version space condition, as illustrated in
the right-hand side of Figure 9.4. From now on, the agent will check not only that the
faculty member has a long-term position (the main condition of the rule), but also that he
or she does not plan to retire (the Except-When condition). The refined rule is not shown
to the expert, but it may be viewed with the Rule Browser.
3. Solving 5. Refinement
Agent applies Agent refines

learned rules rules with
to solve a new new positive
problem examples
4. Critiquing
Expert accepts
reasoning
Figure 9.3. Problem solving, critiquing, and rule refinement.
13:59:50,
.010
9.2. Rule Learning and Refinement 255
Agent refines
rule with
negative
example
2. Critiquing
Incorrect
because
Dan Smith
plans to retire
- +
Figure 9.4. Problem solving, critiquing, and refinement.
Agent refines rule with

negative example
2. Critiquing
- +
Incorrect
because
Jane Austin
+ -
plans to move -
Figure 9.5. Problem solving, critiquing, and refinement.
Figure 9.5 shows another reasoning tree generated by the agent for an expert-
formulated hypothesis. Again the expert rejects one of the reasoning steps: Although
Jane Austin has a tenured position and does not plan to retire, she plans to move from
George Mason University and will not stay on the faculty for the duration of the dissertation
of Bob Sharp.
Based on this failure explanation, the agent specializes the rule that generated the
reduction by adding an additional Except-When plausible version space condition, as
shown in the right-hand side of Figure 9.5. From now on, the agent will check not only that
the faculty member has a long-term position, but also that he or she does not plan to retire
or move from the university.
13:59:50,
.010
The refined rule is shown in Figure 9.6. Notice that this is a quite complex rule that was
learned based only on one positive example, two negative examples, and their explan-
ations. The rule may be further refined based on additional examples.
The following sections describe in more detail the rule-learning and refinement pro-
cesses. Before that, however, let us notice a significant difference between the develop-
ment of a knowledge-based learning agent and the development of a (nonlearning)
knowledge-based agent. As discussed in Sections 1.6.3.1 and 3.1, after the knowledge base
of the (nonlearning) agent is developed by the knowledge engineer, the agent is tested
with various problems. The expert has to analyze the solutions generated by the agent, and
the knowledge engineer has to modify the rules manually to eliminate any identified
problems, testing the modified rules again.
+
-
-
Figure 9.6. Partially learned rule.
13:59:50,
.010
9.3. The Rule-Learning Problem 257
In the case of a learning agent, both rule learning and rule refinement take place as part
of agent teaching. Testing of the agent is included into this process. This process will also
continue as part of knowledge base maintenance. If we would like to extend the agent
to solve new problems, we simply need to teach it more. Thus, in the case of a learning
agent, such as Disciple-EBR, there is no longer a distinction between knowledge base
development and knowledge base maintenance. This is very important because it is well
known that knowledge base maintenance (and system maintenance, in general) is much
more challenging and time consuming than knowledge base (system) development.
Thus knowledge base development and maintenance are less complex and much faster
in the case of a learning agent.
9.3 THE RULE-LEARNING PROBLEM
The rule-learning problem is defined in Table 9.1 and is illustrated in Figures 9.7 and 9.8.
The agent receives an example of a problem or hypothesis reduction and learns a plausible
version space rule that is an analogy-based generalization of the example. There is no
restriction with respect to what the example actually represents. However, it has to be
described as a problem or hypothesis that is reduced to one or several subproblems,
elementary hypotheses, or solutions. Therefore, this example may also be referred to as a
problem-solving episode. For instance, the example shown in the top part of Figure 9.8
reduces a specific hypothesis to its assessment or solution, guided by a question and
its answer.
The expert who is training the agent will interact with it to help it understand why the
example is a correct reduction. The understanding is done in the context of the agent’s
ontology, a fragment of which is shown in Figure 9.7.
The result of the rule-learning process is a general plausible version space rule that will
allow the agent to solve problems by analogy with the example from which the rule was
learned. The plausible version space rule learned from the example at the top of Figure 9.8
is shown at the bottom part of the figure. It is an IF-THEN structure that specifies the
conditions under which the problem from the IF part has the solution from the THEN part.
The rule is only partially learned because, instead of a single applicability condition, it has
two conditions:
Table 9.1 The Rule-Learning Problem
GIVEN
A knowledge base that includes an ontology and a set of (previously learned) rules
An example of a problem reduction expressed with the concepts and instances from the agent’s
knowledge base
An expert who will interact with the agent to help it understand why the example is correct
DETERMINE
A plausible version space rule, where the upper bound is a maximal generalization of the
example, and the lower bound is a minimal generalization that does not contain any specific
instance
An extended ontology, if any extension is needed for the understanding of the example
13:59:50,
.010
object
actor position area of expertise
person organization university position subconcept of
employee educational Computer

student faculty position
organization Science
university employee subconcept of long-term instance of

graduate faculty position
student university
faculty member
Artificial Information
subconcept of instance of tenured position
Intelligence Security
professor
George Mason Indiana
PhD advisor
University University has as position
subconcept of domain person
PhD student range position
associate plans to retire from
is expert in
professor instance of domain person
domain person
range area of expertise range organization
instance of
is interested in plans to move to
is expert in is interested in domain person domain person
John Doe Bob Sharp range area of expertise range organization
Figure 9.7. Fragment of the agent’s ontology.
A plausible upper bound condition that is a maximal generalization of the instances and
constants from the example (e.g., Bob Sharp, certain), in the context of the agent’s ontology
A plausible lower bound condition that is a minimal generalization that does not
contain any specific instance
The relationships among the variables ?O1, ?O2, and ?O3 are the same for both conditions
and are therefore shown only once in Figure 9.8, under the conditions.
Completely learning the rule means learning an exact condition, where the plausible
upper bound is identical with the plausible lower bound.
During rule learning, the agent might also extend the ontology with new features or
concepts, if they are needed for the understanding the example.
9.4 OVERVIEW OF THE RULE-LEARNING METHOD
An overview of the rule-learning method is presented in Figure 9.9 and in Table 9.2. As in
explanation-based learning (DeJong and Mooney, 1986; Mitchell et al., 1986), it consists of
two phases: an explanation phase and a generalization phase. However, in the explanation
phase the agent does not automatically build a deductive proof tree but an explanation
structure through mixed-initiative understanding. Also, the generalization is not a deduct-
ive one, but an analogy-based one.
In the following, we will describe this learning method in more detail and illustrate it.
First we will present the mixed-initiative process of explanation generation and example
13:59:50,
.010
9.4. Overview of the Rule-Learning Method 259
Rule learning
Figure 9.8. Rule learned from an example.
Mixed-Initiative Analogy-based
Understanding Generalization
Plausible version
space rule
Example of a
reduction step
Explanation
NLP, Analogy
Knowledge Base
Figure 9.9. The rule-learning method.
13:59:50,
.010
Table 9.2 Basic Steps of the Rule-Learning Method
(1) Mixed-Initiative Understanding (Explanation Generation)

Through a mixed-initiative interaction with the subject matter expert, determine the set of
relationships EX from the agent’s ontology that collectively explain why the example is correct.
EX represents the explanation of the example E. In general, these relationships express the
meaning of the question/answer pair from the example E, as well as other conditions that need
to be satisfied by the instances and constants from the example E. During this process, new
objects and features may be elicited from the expert and added to the ontology, if the ontology
is incomplete.
(2) Example Reformulation
Generate a variable for each instance and each constant (i.e., number, string, or symbolic
probability) that appears in the example E and its explanation EX. Then use these variables to
create an instance I of the concept C representing the applicability condition of the rule R to be
learned. C is the concept to be learned as part of rule learning and refinement. Finally, reformulate
the example as a very specific IF-THEN rule with I as its applicability condition. The IF problem, the
question, the answer, and the THEN subproblems (or solutions) are obtained by replacing each
instance or constant from the example E with the corresponding variable.
(3) Analogy-based Generalizations
Generate the plausible upper bound condition of the rule R as the maximal generalization of I in
the context of the agent’s ontology. Generate the plausible lower bound condition of the rule R as
the minimal generalization of I, which does not contain any specific instance.
(4) Rule Analysis
If there is any variable from the THEN part of a rule that is not linked to some variable from the
other parts of the rule, or if the rule has too many instances in the knowledge base, then interact
with the expert to extend the explanation of the example and update the rule if new explanation
pieces are found.
(5) Determining the Generalized Example
Generate a minimal generalization of the example E and its explanation EX that does not contain
any specific instance, and associate it with the learned rule.
understanding, which is part of the first phase. Then we will present and justify the
generalization method, which is based on analogical reasoning.
9.5 MIXED-INITIATIVE EXAMPLE UNDERSTANDING
9.5.1 What Is an Explanation of an Example?

The mixed-initiative example understanding process is illustrated in Figure 9.10. The
expert has defined the example during the modeling process. Now the expert and the
agent have to collaborate to learn a general rule from this specific example.
The first step consists in finding the set of relationships from the agent’s ontology
that collectively explain why the example is correct. In general, these relation-
ships express the meaning of the question/answer pair from the example, as well
as other conditions that need to be satisfied by the instances and constants from
the example.
13:59:50,
.010
9.5. Mixed-Initiative Example Understanding 261
Consider the following question/answer pair:
Is Bob Sharp interested in an area of expertise of John Doe?

Yes, Artificial Intelligence.
The following is an approximate representation of the meaning of the question/answer

pair using the terms from the agent’s ontology:
Bob Sharp is interested in Artificial Intelligence

John Doe is expert in Artificial Intelligence
Adding “Probability of solution is always certain” to the preceding relationships results in

the explanation of the example, which is shown on the left-hand side of Figure 9.10. This is
a formal explanation of the example and consists of a set of relationships (or explanation
pieces) involving the instances and the constants from the example.
One can distinguish between different types of explanation pieces, as shown in
Table 9.3. Formally, the explanation of the example consists of several explanation pieces.
Each explanation piece corresponds to a path in the ontology between an object in the
example and another object or constant, as illustrated in Figure 9.10. In principle, the path
could have any length. In practice, however, one has to limit the maximum length of the
path for a certain type of explanation, in order to reduce the combinatorial explosion in
the generation of plausible explanations by the agent.
Table 9.3 Types of Explanation Pieces
Association – a relation between two objects from the example

Correlation – a common feature of two objects from the example
Property – a property of an object from the example
Relation – a relation between an object from the example and one from the knowledge base
Generalization – a generalization of an object from the example
Specific value – a specific value for an entity from the example (constant or generic instance)
Bob Sharp John Doe
is interested in is expert in
Artif icial Intelligence

Probability of solution
is always certain
Figure 9.10. Example understanding.
13:59:50,
.010
Because the agent’s ontology is incomplete, sometimes the explanation includes only
an approximate representation of the meaning of the question/answer (natural language)
sentences.
The next section presents the explanation generation method.
9.5.2 Explanation Generation

Fully automatic explanation generation by the agent, while highly desirable, is problematic
because it would require human-like natural language understanding capabilities from an
agent that has only an incomplete ontology.
Manual explanation definition by an expert who is not a knowledge engineer is also
problematic. For one thing, he or she would need to use the formal language of the agent.
But this would not be enough. The expert would also need to know the names of the
potentially many thousands of concepts and features from the agent’s ontology (such as “is
interested in”).
While defining the formal explanation of a reduction step is beyond the individual
capabilities of both the expert and the agent, it is not beyond their joint complementary
capabilities. Therefore, finding such explanation pieces is modeled as a mixed-initiative
process (Tecuci et al., 2007a), based on the following observations:
It is easier for an expert to understand sentences in the formal language of the agent
than it is to produce such formal sentences
It is easier for the agent to generate formal sentences than it is to understand sentences
in the natural language of the expert
In essence, the agent will use basic natural language processing, various heuristics,
analogical reasoning, and help from the expert in order to identify and propose a set of
plausible explanation pieces, ordered by their plausibility of being correct explanations.
Then the expert will select the correct ones from the generated list.
The left-hand side of Figure 9.11 shows an example to be understood, and the upper-
right part of Figure 9.11 shows all the instances and constants from the example. The agent
will look for plausible explanation pieces of the types from Table 9.3, involving those
instances and constants. The most plausible explanation pieces identified, in plausibility
Figure 9.11. Explanation generation.
13:59:50,
.010
9.5. Mixed-Initiative Example Understanding 263
order, are shown in the bottom-right of Figure 9.11. Notice that the two most plausible
explanation pieces from Figure 9.11 are the correct explanation pieces shown in Figure 9.10.
The expert will have to select each of them and click on the Accept button. As a result, the
agent will move them in the Explanations pane from the left side of Figure 9.11.
Notice in the upper-right of Figure 9.11 that all the objects and constants from the example
are selected. Consequently, the agent generates the most plausible explanations pieces
related to all these objects and displays those with the highest plausibility. The expert may
click on the See More button, asking the agent to display the next set of plausible explanations.
The expert may also deselect some of the objects and constants, asking the agent to
generate only plausible explanations involving the selected elements. For example,
Figure 9.12 illustrates a situation where only the constant “certain” is selected. As a result,
the agent generated only the explanation “The value is specifically certain,” which means
that this value should be kept as such (i.e., not generalized) in the learned rule.
The expert may also provide a new explanation, even using new instances, concepts, or
features. In such a case, the expert should first define the new elements in the ontology.
After that, the expert may guide the agent to generate the desired explanations.
If the example contains any generic instance, such as “Artificial Intelligence,” the agent
will automatically select the explanation piece “Artificial Intelligence is Artificial Intelligence”
(see Explanation pane on the left side of Figure 9.11), meaning that this instance will
appear as such in the learned rule. If the expert wants Artificial Intelligence to be general-
ized, he or she should simply remove that explanation by clicking on it and on the Remove
button at its right.
The expert may also define explanations involving functions and comparisons, as will
be discussed in Sections 9.12.3 and 9.12.4.
Notice, however, that the explanation of the example may still be incomplete for at least
three reasons:
The ontology of the agent may be incomplete, and therefore the agent may not be able
to propose all the explanation pieces of the example simply because they are not
present in the ontology
1. Select an object
2. Click on “Search”
3. The agent generates

plausible explanations
Figure 9.12. Guiding explanation generation.
13:59:50,
.010
The agent shows the plausible explanation pieces incrementally, as guided by the
expert, and if one of the actual explanation pieces is not among the first ones shown, it
may not be seen and selected by the expert
It is often the case that the human expert forgets to provide explanations that corres-
pond to common-sense knowledge that also is not represented in the question/
answer pair
The incompleteness of the explanation is not, however, a significant problem because the
explanation may be further extended during the rule refinement process, as discussed
in Chapter 10.
To conclude, Table 9.4 summarizes the mixed-initiative explanation generation method.
Once the expert is satisfied with the identified explanation pieces, the agent will generate the
rule, as discussed in the following sections.
9.6 EXAMPLE REFORMULATION
As indicated in Table 9.2 (p. 260), once the explanation of the example is found, the agent
generates a very specific IF-THEN rule with an applicability condition that covers only that
example. The top part of Figure 9.13 shows an example, and the bottom part shows the
generated specific rule that covers only that example. Notice that each instance (e.g., Bob
Sharp) and each constant (e.g., certain) is replaced with a variable (i.e., ?O1, ?Sl1).
However, the applicability condition restricts the possible values of these variables to
those from the example (e.g., “?O1 is Bob Sharp”). The applicability condition also includes
the properties and the relationships from the explanation. Therefore, the rule from the
bottom of Figure 9.13 will cover only the example from the top of Figure 9.13. This rule will
be further generalized to the rule from Figure 9.8, which has a plausible upper bound
condition and a plausible lower bound condition, as discussed in the next section. In
particular, the plausible upper bound condition will be obtained as the maximal general-
ization of the specific condition in the context of the agent’s ontology. Similarly, the
plausible lower bound condition will be obtained as the minimal generalization of the
specific condition that does not contain any specific instance.
Table 9.4 Mixed-Initiative Explanation Generation
Let E be an example.
Repeat
The expert focuses the agent’s attention by selecting some of the instances and constants from
the example.
The agent proposes what it determines to be the most plausible explanation pieces related to
the selected entities, ordered by their plausibility.
The expert chooses the relevant explanation pieces.
The expert may ask for the generation of additional explanation pieces related to the selected
instances and constants, may select different ones, or may directly specify explanation pieces.
until the expert is satisfied with the explanation of the example.
13:59:50,
.010
9.7. Analogy-based Generalization 265
Example
Specific rule
IF ?O1 is interested in an area of expertise of ?O2.
Q: Is ?O1 interested in an area of expertise of ?O2 ?

A: Yes,?O3.
Condition
?O1 is Bob Sharp
?O2 is John Doe
is expert in ?O3
?O3 is Artificial Intelligence
?SI1 is-exactly certain
THEN It is ?SI1 that ?O1 is interested in an area of expertise of ?O2.
Figure 9.13. Specific rule covering only the example.
9.7 ANALOGY-BASED GENERALIZATION
9.7.1 Analogical Problem Solving Based on Explanation Similarity

The goal of the learning agent is to learn a general rule that will enable it to solve new
problems by analogy with the example provided by the expert. As discussed in Section
8.1.4, analogy involves mapping some underlying “causal network of relations” between a
source and a target, with the goal of transferring knowledge from the source to the target.
This applies to analogical problem solving as illustrated in Figure 9.14.
The explanation in the upper-left part of Figure 9.14 explains why the input example
in the lower-left part is correct. The expression from the upper-right part of Figure 9.14 is
similar with this explanation because both have the same structure. Therefore, one may
infer, by analogy, that this expression explains an example that is similar to the initial
example. This example is shown in the bottom-right part of Figure 9.14.
The more general question is: When should we consider that two expressions or explan-
ations are similar? In general, one considers that two expressions, explanations, or situ-
ations are similar if they match within a certain predefined threshold (Winston, 1980). As
shown in Kodratoff and Ganascia (1986), generalization may be reduced to structural
matching. Therefore, we may consider that two expressions (explanations) are similar if
13:59:50,
.010
explanation similar explanation

Bob Sharp John Doe Peter Jones Dan Smith
is interested in is expert in similar is interested in is expert in
Artificial Intelligence Information Security
Probability of solution is always certain Probability of solution is always certain
explains explains?
initial example similar example
I need to I need to
Bob Sharp is interested in an area of Peter Jones is interested in an area of
expertise of John Doe. expertise of Dan Smith.
similar
Therefore I conclude that Therefore I conclude that
It is certain that Bob Sharp is interested It is certain that Peter Jones is interested
in an area of expertise of John Doe. in an area of expertise of Dan Smith.
Figure 9.14. Analogical problem solving based on explanation similarity.
they are both less general then a given expression that represents the analogy criterion.
Consequently, the preceding question may be rephrased as:
Given the explanation EX of an example E, which generalization of EX should be
considered an analogy criterion, enabling the agent to generate reductions that are analo-
gous to E?
There are two interesting answers of this question, one given by a cautious learner, and
the other given by an aggressive learner, as discussed in the next sections.
9.7.2 Upper Bound Condition as a Maximally General

Analogy Criterion
An aggressive learner always prefers maximal generalizations, as discussed in Section 8.5.
Consequently, it will determine the analogy criterion as a maximal generalization of the
explanation, as illustrated in Figure 9.15 and explained in the following.
The explanation of the initial example corresponds to the applicability condition of the
specific rule representing the example (see Figure 9.13). Thus the aggressive learner will
learn a rule by maximally generalizing the specific applicability condition, as illustrated in
the upper right side of Figure 9.16.
The specific instance Bob Sharp is generalized to the most general concept from the
ontology that covers it and satisfies all the constraints imposed by the domains and the
ranges of the features related to it in the specific condition (explanation). As shown in the
left-hand side of Figure 9.16, the most general generalization of Bob Sharp in the ontology
is object. On the other hand, in the specific condition (explanation), Bob Sharp has the
feature “is interested in,” the domain of which is person. Thus the maximal generalization
of Bob Sharp is: object \ domain(is interested in) = object \ person = person.
In general, if an instance in the explanation has the features f11, . . . , f1m, and appears as
value of the features f21, . . . , f2n, then its maximal general generalization is:
object \ domain(f11) \ . . . \ domain(f1m) \ range(f21) \ . . . \ range(f2n)
13:59:50,
.010
Analogy person person

criterion
area of
instance of expertise instance of
?O1 ?O2
instance of
is interested in is expert in
?O3
Probability of solution is always certain
less general than less general than

Figure 9.15. Upper bound condition as a maximally general analogy criterion.
object
Most general generalization
?O1 is person
actor area of UB is interested in ?O3
expertise ?O2 is person
person UB
is expert in ?O3
subconcept of
?O3 is area of expertise
employee student ?SI1 is in [certain – certain]
Computer
Science Maximal
university employee generalization
instance of Specif ic condit ion
faculty member graduate ?O1 is Bob Sharp
student Artificial is interested in ?O3
subconcept of
Intelligence LB ?O2 is John Doe
professor PhD advisor ?O3 is expert in ?O3
LB ?O3 is Artificial Intelligence
subconcept of subconcept of
is expert in ?SI1 is exactly certain
associate LB PhD student LB domain person
professor range area of expert ise
instance of instance of is interested in
John Doe Bob Sharp domain person
range area of expertise
?O2 ?O1
Figure 9.16. Maximal generalization of the specific applicability condition.
Now consider John Doe. Its most general generalization is object \ domain(is expert in) =
object \ person = person.
Consider now Artificial Intelligence. It appears as a value of the features “is interested in”
and “is expert in.” Therefore, its maximal generalization is: object \ range(is interested in) \
range(is expert in) = object \ area of expertise \ area of expertise = area of expertise.
13:59:50,
.010
On the other hand, the maximal generalization of “certain” is the interval with a single
value “[certain – certain]” because ?Sl1 is restricted to this value by the feature “is exactly.”
Let us consider again the example from the top part of Figure 9.13, but now let us
assume that “The value is specifically certain” was not identified as an explanation piece.
That is, the explanation of the example consists only of the following pieces:
Bob Sharp is interested in Artificial Intelligence

John Doe is expert in Artificial Intelligence
In this case, the generated specific condition is the one from the bottom of Figure 9.17, and
its maximal generalization is the one from the top of Figure 9.17. Notice that the maximal
generalization of “certain” is the entire interval [no support – certain] because there is no
restriction on the possible values of ?Sl1.
9.7.3 Lower Bound Condition as a Minimally General

Analogy Criterion
A cautious learner prefers minimal generalizations, as discussed in Section 8.5. Conse-
quently, it will determine the analogy criterion as a minimal generalization of the explan-
ation, as illustrated in Figure 9.18.
The explanation of the initial example corresponds to the applicability condition of the
specific rule representing the example (see Figure 9.13). Thus the cautious learner will
learn a rule by minimally generalizing the specific applicability condition, as illustrated in
Figure 9.19.
The specific instance Bob Sharp is minimally generalized to the most specific concept
from the ontology that covers it, that is, to PhD student. This generalization is allowed
because the domain of the feature “is interested in” is person, which includes PhD student.
Most general generalization

?O1 is person
?O2 is person
is expert in ?O3
?O3 is area of expertise
?SI1 is in [no support – certain]
Maximal
generalization
Specific condition
?O1 is Bob Sharp
?O2 is John Doe
is expert in ?O3
?SI1 is certain
Figure 9.17. Maximal generalization of a symbolic probability value when no explanation is identified.
13:59:50,
.010
Analogy associate
criterion PhD student PhD advisor
professor
instance of instance of
Artificial Intelligence
?O1 ?O2
is interested in is is expert in
?O3
Probability of solution is always certain
less general than less general than

Figure 9.18. Lower bound condition as a minimally general analogy criterion.
Most specific generalization

?O1 is PhD student
?O2 is (PhD advisor, associate professor)
is expert in ?O3
?SI1 is in [certain – certain]
Minimal
generalization
Specific condition
?O1 is Bob Sharp
?O2 is John Doe
is expert in ?O3
?SI1 is exactly certain
Figure 9.19. Minimal generalization of the specific applicability condition.
Similarly, the specific instance John Doe is minimally generalized to PhD advisor or
associate professor, because these are the minimal generalizations of John Doe and neither
is more specific than the other. Additionally, both these concepts are subconcepts of
person, the domain of is expert in.
Because Artificial Intelligence is a generic instance, it can appear in the learned rule (as
opposed to the specific instances Bob Sharp and John Doe). Therefore, its minimal general-
ization is Artificial Intelligence itself. Similarly, the constants (such as certain) can also appear
in the learned rule, and they are kept as such in the minimal generalization.
13:59:50,
.010
Notice that if you want an instance to appear in the condition of a learned rule, it needs
to be defined as a generic instance. Specific instances are always generalized to concepts
and will never appear in the condition.
9.8 RULE GENERATION AND ANALYSIS
The partially learned rule is shown in the bottom part of Figure 9.8 (p. 259). Notice that the
features are listed only once under the bounds because they are the same for both bounds.
The generated rule is analyzed to determine whether there are any variables in the
THEN part that are not linked to some variable from the IF part. If such an unlinked
variable exists, then it can be instantiated to any value, leading to solutions that make no
sense. Therefore, the agent will interact with the expert to find an additional explanation
that will create the missing link and update the rule accordingly.
The generated rule is also analyzed to determine whether it has too many instances in
the knowledge base, which is also an indication that its explanation is incomplete and
needs to be extended.
9.9 GENERALIZED EXAMPLES
The rule learned from an example and its explanation depends on the ontology of the
agent at the time the rule was generated. If the ontology changes, the rule may need to be
updated, as will be discussed in Chapter 10. For example, the minimal generalization of a
specific instance will change if a new concept is inserted between that instance and the
concept above it. To enable the agent to update its rules automatically when relevant
changes occur in the ontology, minimal generalizations of the examples and their explan-
ations are associated with the learned rules.
Why is the agent maintaining minimal generalizations of examples instead of the
examples themselves? Because the examples exist only in Scenario KBs, where the specific
instances are defined, while the rules are maintained in the Domain KB. If a scenario is no
longer available, the corresponding examples are no longer defined. However, generalized
examples (which do not contain specific instances) will always be defined in the Domain
KB. Thus the generalized examples represent a way to maintain a history of how a rule was
learned, independent of the scenarios. They are also a compact way of preserving this
history because one generalized example may correspond to many actual examples.
Figure 9.20 shows the minimal generalization of the example and its explanation from
which the rule in Figure 9.8 (p. 259) was learned.
One should notice that the minimal generalization of the example shown at the top part
of Figure 9.20 is not the same with the plausible lower bound condition of the learned rule
from Figure 9.8. Consider the specific instance John Doe from the example.
In the ontology, John Doe is both a direct instance of PhD advisor and of associate
professor (see Figure 9.16). In the lower bound condition of the rule, John Doe is general-
ized to PhD advisor or associate professor, indicated as (PhD advisor, associate professor),
because each of these two concepts is a minimal generalization of John Doe in the
ontology. Thus the agent maintains the two concepts as part of the lower bound of the
rule’s version space: one corresponding to PhD advisor, and the other corresponding to
13:59:50,
.010
9.10. Hypothesis Learning 271
Generalized Example
?O1 is PhD student
?O2 is PhD advisor
is associate professor
is expert in ?O3
?SI1 is in [certain - certain]
Covered positive examples: 1
Covered negative examples: 0
Minimal example
generalization
Example and its explanation
?O1 is Bob Sharp
?O2 is John Doe
is expert in ?O3
?SI1 is exactly certain
Figure 9.20. Determining the generalized example.
associate professor. During further learning, the agent will choose one of these generaliza-
tions or a more general generalization that covers both of them.
In the minimal generalization of the example, John Doe is generalized to PhD advisor
and associate professor because this is the best representation of the minimal generaliza-
tion of the example that can be used to regenerate the rule, when changes are made to the
ontology. This minimal generalization is expressed as follows:
?O2 is PhD advisor
Initially, the generalized example shown at the top of Figure 9.20 covers only one
specific example. However, when a new (positive or negative) example is used to refine
the rule, the agent checks whether it is already covered by an existing generalized example
and records this information. Because a generalized example may cover any number of
specific positive and negative examples, its description also includes the number of
specific examples covered, as shown in the top part of Figure 9.20.
Cases of using generalized examples to regenerate previously learned rules are pre-
sented in Section 10.2.
9.10 HYPOTHESIS LEARNING
In addition to learning a general reduction rule from a specific reduction example, Disciple-
EBR also learns general hypotheses (or problems). The left-hand side of Figure 9.21 shows
13:59:50,
.010
Hypothesis Generalizaon
1. Modeling 2. Learning
Agent learns general

rule and hypotheses
Reducon Rule
Expert explains how to

solve a specific problem
Reducon Example
.010
13:59:50,
Hypothesis Generalizaon Hypothesis Generalizaon Hypothesis Generalizaon
Figure 9.21. A reduction rule and four hypotheses learned from a specific hypothesis reduction.
9.10. Hypothesis Learning 273
an example of a reduction where a hypothesis is reduced to three subhypotheses.

From this example, Disciple-EBR learns five general knowledge pieces that share the
same variables:
A general hypothesis reduction rule that is a generalization of the reduction example

where each instance was replaced with a variable
A generalization of the top hypothesis with a plausible version space condition
Generalizations of the three subhypotheses, each with its own plausible version space
condition
The hypothesis learning method, shown in Table 9.5, is very similar with the rule-learning
method.
Figure 9.22 illustrates the automatic learning of a general hypothesis from the specific
hypothesis, “John Doe would be a good PhD advisor for Bob Sharp,” when no explanation is
provided.
The specific instances, John Doe and Bob Sharp, are replaced with the variables ?O1 and
?O2, respectively, as in the reduction rule.
The lower bounds of these variables are obtained as the minimal generalizations of
John Doe and Bob Sharp, according to the agent’s ontology from the left-hand side of
Figure 9.22, because both of them are specific instances. Notice that there are two minimal
generalizations of John Doe: PhD advisor and associate professor. The minimal generaliza-
tion of Bob Sharp is PhD student.
The upper bounds are obtained as the maximum generalizations of John Doe and
Bob Sharp, according to the agent’s ontology from the left-hand side of Figure 9.22. They
are both object.
During the explanation generation process, the user may wish to restrict the general-
ization of the hypothesis by providing the following explanations:
Table 9.5 Basic Steps of the Hypothesis Learning Method
(1) Mixed-Initiative Understanding (Explanation Generation)

Through a mixed-initiative interaction with the subject matter expert, determine the set of
constraints EX that need to be satisfied by the instances and constants from the example
hypothesis EH in order for the hypothesis statement to make sense. These include the types of the
instances from the hypothesis.
(2) Hypothesis Reformulation
Use the constraints EX and the variables generated during the learning of the corresponding
reduction rule to create an instance I of the concept C representing the applicability condition of
the hypothesis H to be learned. Then reformulate the specific hypothesis as a hypothesis pattern
with I as its applicability condition.
(3) Analogy-based Generalizations
Generate the plausible upper bound condition of the general hypothesis H as the maximal
generalization of I in the context of the agent’s ontology. Generate the plausible lower bound
condition of the hypothesis H as the minimal generalization of I that does not contain any specific
instance.
(4) Determining the Generalized Example
Generate a minimal generalization of the example hypothesis EH and its explanation EX that does
not contain any specific instance and associate it with the learned hypothesis.
13:59:50,
.010
object UB
Most specific generalization Most general generalization
?O1 is (PhD advisor, associate professor) ?O1 is object
actor
?O2 is PhD student ?O2 is object
person Min Max

generalization generalization
employee student Specific condition
?O1 is John Doe
university employee ?O2 is Bob Sharp
faculty member graduate Hypothesis Example

student
subconcept of
professor PhD advisor LB Hypothesis Generalization

subconcept of
subconcept of
associate
LB PhD student LB
professor
John Doe Bob Sharp

?O1 ?O2
Figure 9.22. Generation of the applicability conditions of a hypothesis.
John Doe is faculty member

Bob Sharp is student
In this case, the lower bound condition of the learned hypothesis remains the same, but
the upper bound condition becomes:
?O1 is faculty member

?O2 is student
In general, hypothesis learning is done automatically, as a byproduct of rule learning, but

the explanations found for learning the reduction rule will not be used for learning the
hypothesis. Indeed, these explanations indicate only why this particular reduction is
correct. There may be other reductions of the same hypothesis, which will have different
explanations. If desired, the user may provide specific explanations for the hypothesis that
limit the possible instantiations of the learned hypothesis to those that make sense, as
illustrated previously.
Notice that for each specific hypothesis, the agent may learn a single general hypothesis
but several reduction rules. Then, when trying to reduce a hypothesis to simpler hypoth-
eses, the agent checks that both the applicability condition of the corresponding general
hypothesis and that of the considered rule are satisfied. It is therefore important that the
applicability condition of the general hypothesis is general enough to enable the use of all
the learned reduction rules associated with that general hypothesis.
A general hypothesis may be learned in several circumstances. As illustrated in
Figure 9.21, a hypothesis may be learned as a byproduct of rule learning. It may also be
learned during the modeling and formalization of a reasoning tree by using the Learn
13:59:50,
.010
9.11. Hands On: Rule and Hypotheses Learning 275
Hypothesis Pattern or the Learn Tree Patterns commands that were introduced in
Section 4.10. Finally, it may be learned by specifically invoking hypothesis (problem)
learning when working with the Mixed-Initiative Reasoner. But it is only this last situation
that also allows the definitions of explanations, as will be discussed later in this section. In
all the other situations, a hypothesis is automatically learned, with no explanations, as was
illustrated in Figure 9.22.
The overall user–agent interactions during the hypothesis explanation process are
illustrated in Figure 9.23 and described in Operation 9.1. Here it is assumed that the
reasoning tree was already formalized and thus a hypothesis pattern was already learned.
If it was not learned, the pattern will be automatically learned before the explanations are
identified.
Operation 9.1. Define explanations for a hypothesis

Select the Scenario workspace.
Click on the Reasoning menu and select Mixed-Initiative Reasoner.
Select the hypothesis to analyze, and click on the Select button on the right side of
the window.
In the Reasoning Hierarchy viewer of the Mixed-Initiative Reasoner, select a hypoth-
esis by clicking on it.
In the right panel, click on the Modify Explanations button.
Explanation generation is automatically invoked, the left pane showing the hypothesis
in the Reasoning Step viewer, and the right pane showing the selected entities and a
corresponding ordered list of plausible explanations (see the upper-right part of
Figure 9.23).
Inspect the explanation pieces proposed by the agent in the right panel, click on a
correct explanation piece (if any), and then click on the Accept button.
Repeat explanation selection and acceptance to select all the needed explanation pieces.
If necessary, click on the See More button, asking the agent to generate additional
possible explanation pieces related to the selected entities.
If necessary, select a subset of instances and constants and click on the Search button,
asking the agent to generate explanations related to the selected entities.
All the accepted explanations are added to the hypothesis (see the bottom-left part of
Figure 9.23). You may delete any of them by clicking on it and then clicking on the
Remove button.
Click on the Finish button to end the learning of the current hypothesis.
9.11 HANDS ON: RULE AND HYPOTHESES LEARNING
This case study will guide you to use Disciple-EBR to learn rules and hypotheses from
examples. More specifically, you will learn how to:
Invoke the rule learning module

Interact with the agent to find the explanation of the example
Learn the rule (and the associated hypotheses)
The overall user–agent interactions during the rule- (and hypotheses-) learning process
are illustrated in Figure 9.24 and described in Operation 9.2. It is assumed that the
13:59:50,
.010
1. In “Reasoning Hierarchy” (a) Select Hypothesis 4. In “Reasoning Step” (b) Accept Explanation Pieces
6. Click on “Accept”
5. Select a relevant
3. Click on explanation piece
“Modify Explanations”
2. Click on a
hypothesis to select it
(d) Finish Hypothesis Learning (c) Accept Explanation Pieces

.010
13:59:50,
7. Repeat steps 5 and 6 to select

8. The accepted explanation pieces all the needed explanation pieces.
were added to the example 9. Click on “Finish” to end the
hypothesis learning process
Figure 9.23. Overview of the user–agent interactions during the hypothesis explanation process.
1. In “Reasoning Hierarchy” 4. In “Reasoning Step” (b) Accept Explanation Pieces
(a) Select Reasoning Step
6. Select a relevant
explanation piece 7. Click on “Accept”
3. Click on
2. Click on question/answer “Learn Condition” 5. If necessary, select and remove any
to select the reasoning step automatically selected explanation piece
(d) Finish Rules Learning (c) Accept Explanation Pieces

.010
8. Repeat steps 6 and 7 to select

13:59:50,
all the needed explanation pieces.
10. Click on “Finish” to end

the rule learning process
9. The accepted explanation pieces
were added to the example
Figure 9.24. Overview of the user–agent interactions during rule (and hypotheses) learning.
reasoning tree is formalized. If not, one can easily formalize it in the Evidence workspace
before invoking rule learning by simply right-clicking on the top node and selecting
Learn Tree.
Operation 9.2. Learn rule

Click on the Reasoning menu and select Mixed-Initiative Reasoner.
Select the hypothesis to analyze, and click on the Select button on the right side of
the window.
In the Reasoning Hierarchy viewer, select a reasoning step by clicking on the corres-
ponding Q/A node (see the upper-left part of Figure 9.24).
In the right panel, click on the Learn Condition button.
Explanation generation is automatically invoked, the left pane showing the reduction
in the Reasoning Step viewer, and the right pane showing the selected entities and a
corresponding ordered list of plausible explanations (see the upper-right part of
Figure 9.24).
If the example contains generic instances, check whether you want to remove any of
the corresponding explanations that were automatically accepted by the agent. To
remove such an explanation piece, click on it in the left panel and then click on the
Remove button.
Inspect the explanation pieces proposed by the agent in the right panel, click on a
correct explanation piece (if any), and then click on the Accept button.
Repeat explanation selection and acceptance to select all the needed explanation
pieces.
If necessary, click on the See More button, asking the agent to generate additional
possible explanation pieces related to the selected entities.
If necessary, select a subset of instances and constants and click on the Search button,
asking the agent to generate explanations related to the selected entities.
All the accepted explanations are added to the reduction (see the bottom-left part of
Figure 9.24). You may delete any of them by clicking on them and then clicking on the
Remove button.
Click on the Finish button to end the learning of the current rule.
Continue with selecting another reasoning step and learning the corresponding
rule.
You are now ready to perform a rule-learning case study. There are two of them, a shorter
one and a longer one. In the shorter case study, you will guide the agent to learn the rule
from Figure 9.8, as discussed in the previous sections. In the longer case study, you will
guide the agent to learn several rules, including the rule from Figure 9.8.
Start Disciple-EBR, select one of the case study knowledge bases (either “11-Rule-Learning-
short/Scen” or “11-Rule-Learning/Scen”), and proceed as indicated in the instructions at the
bottom of the opened window.
A learned rule can be displayed as indicated in the following operation.
Operation 9.3. Display a learned rule with the Rule Viewer

In the Scenario workspace, in the Reasoning Hierarchy viewer select a reasoning step
for which a rule has already been learned by clicking on the corresponding Q/A node.
At the bottom of the right panel, click on the Reduction Rule button to see the rule that
generated the reduction step.
13:59:50,
.010
9.12. Explanation Generation Operations 279
At the top of the Rule Viewer, notice the name of the rule (e.g., DDR.00018). You may
also display or delete this rule with the Rule Browser, as described in Operations 10.4
and 10.5.
Click on the X button of the Rule Viewer to close it.
Operation 9.4. Display a learned hypothesis with the Problem Viewer

In the Scenario workspace, in the Reasoning Hierarchy viewer click on a hypothesis to
select one for which a general hypothesis has already been learned.
In the right panel, click on the View Problem button to see the learned hypothesis.
Click on the X button of the Problem Viewer to close it.
9.12 EXPLANATION GENERATION OPERATIONS
An important part of rule learning is finding an explanation of the example that is as

complete as possible. The following sections discuss various operations that support
this process.
9.12.1 Guiding Explanation Generation

You can guide the agent in explanation generation by selecting the instances and con-
stants from the example for which explanation pieces will be proposed. As a result, the
agent generates a list of potential explanations, ordered by their plausibility, from which
you can select the correct ones. The corresponding user–agent interaction is described in
Operation 9.5 and illustrated in Figure 9.25.
Operation 9.5. Guide the generation of explanations

In the Scenario workspace, during rule learning, in the Reasoning Hierarchy viewer,
select a reasoning step by clicking on the corresponding Q/A node.
1. Select an object
2. Click on “Search”
3. The agent generates

plausible explanations
Figure 9.25. Guiding explanation generation.
13:59:50,
.010
Select one or several entities from the “Elements to search for” pane, such as certain in
Figure 9.25. You may also need to deselect some entities by clicking on them.
Click on the Search button, asking the agent to generate explanation pieces related to
the selected entities.
Select an explanation piece and click on the Accept button.
Click on the See More button to see more of the generated explanation pieces.
Repeat the preceding steps until all the desired explanations are generated and selected.
9.12.2 Fixing Values

You may block the generalizations of numbers (e.g., “5”), symbolic probability values (e.g.,
“certain”), or generic instances (e.g., “Artificial Intelligence”) by selecting explanations of the
following forms, as described in Operation 9.6:
The value is specifically 5

The value is specifically certain
Artificial Intelligence is Artificial Intelligence
Operation 9.6. Generate explanations with fixed values

Select one or several generic instances or constants (e.g., certain in Figure 9.25) in the
“Elements to search for” pane. No specific instances should be selected.
If a selected element is a constant, then the agent will generate a possible explanation
piece of the form “The value is specifically <selected constant>.”
If a selected element is a generic instance, then the agent will generate a possible
explanation piece of the form “<selected instance> is <selected instance>,” as well as
other explanations containing the selected instance.
Accept the explanation “The value is specifically <selected constant>” (or “<selected
instance> is <selected instance>”) to fix that value in the learned rule.
Notice that these types of explanation pieces are generated when only constants or generic
instances are selected in the “Elements to search for” pane. For example, only “certain”
was selected in the “Elements to search for” pane in the upper-right part of Figure 9.25,
and therefore the potential explanation piece “The value is specifically certain” was
generated.
Notice also that explanations such as “Artificial Intelligence is Artificial Intelligence” can be
generated only for generic instances. They cannot be generated for specific instances
because these instances are always generalized in the learned rules.
9.12.3 Explanations with Functions

We will illustrate these explanations based on the reasoning tree in Figure 9.26. Notice that
all the numbers are correctly recognized as numbers and appear in green.
The Q/A pair from the first reduction explains how the price of the Apple iPad 16GB is
computed. The corresponding explanation pieces express the meaning of the Q/A pair by
13:59:50,
.010
Figure 9.26. Reasoning tree involving numbers.
The function is defined with

the Expression Editor, using
numbers from the example
Figure 9.27. Explanations with functions.
using relationships from the agent’s ontology. They are the following ones, and are shown
also in the left-hand side of Figure 9.27:
Apple iPad 16GB has as price 495.0

MicroCenter Fairfax is located in Virginia
Virginia has as state sale tax 0.05
13:59:50,
.010
Additionally, you have to teach the agent how the price is actually computed. You invoke
the Expression Editor by clicking on the Edit Expression button, which displays a pane to
define the expression (see the bottom right of Figure 9.27). Then you fill in the left side of
the equality with the price, and the right side with the expression that leads to this price, by
using the numbers from the example:
519:75 ¼ 495:0 0:05 þ 495:0
Each of these numbers is generalized to a variable, and a general relationship between

these variables is learned (i.e., “?N1 = ?N2 * ?N3 + ?N2”). The learned rule, containing this
expression in its applicability condition, is shown in Figure 9.28.
In general, to define an explanation containing a function, you have to follow the steps
from Operation 9.7.
Figure 9.28. Learned rule with a learned function in the applicability condition.
13:59:50,
.010
Operation 9.7. Define explanations containing functions

In the “Possible explanations” pane, click on the Edit Expression button.
The agent displays the Expression Editor.
In the left part of the Expression Editor, start typing the result of the expression (e.g., the
actual price of “519.75” shown in Figure 9.27) and select it from the completion pop-up.
In the right part of the Expression Editor, type the expression used to compute the
result (i.e., “495.0 * 0.05 + 495.0”), making sure you select the values from the completion
pop-up.
Click on the OK button under the Expression Editor to define the expression.
9.12.4 Explanations with Comparisons

We will illustrate these explanations based on the reasoning tree from Figure 9.26.
Notice that the Q/A pair corresponding to the bottom reduction explains why Mike has
enough money to buy the Apple iPad 16GB. The following is the corresponding explanation
piece:
Mike has as available funds 650.35
Additionally, you have to indicate that 650.35 is greater than 519.75. You click on the
Create New… button, which opens a window allowing you to define a new explanation as
an object-feature-value triplet (see the bottom of Figure 9.29). In the left editor, you
start typing the amount of money Mike has (i.e., 650.35) and select it from the completion
The comparison is defined with

the Create Explanation Editor,
using numbers from the example
Figure 9.29. Defining explanations with comparisons.
13:59:50,
.010
pop-up. In the center editor, you type >=. Then, in the right editor, you start typing the
actual cost (519.75) and select it from the completion pop-up. Finally, you click on the OK
button in the Create explanation window to select this explanation:
650:35 >¼ 519:75
Each of these numbers is generalized to a variable, and a general relationship between

these variables is learned (i.e., “?N2 >= ?N1”). The learned rule, containing this relation-
ship in its applicability condition, is shown in Figure 9.30.
In general, to define an explanation containing a comparison you have to follow the
steps from Operation 9.8.
Operation 9.8. Define explanations containing comparisons

In the “Possible explanations” pane, click on the Create New… button.
The agent opens a window for defining an explanation as an object-feature-value
triplet (see the bottom of Figure 9.29).
In the left editor, type the number corresponding to the left side of the comparison
(e.g., “650.35” in Figure 9.29) and select it from the completion pop-up.
Figure 9.30. Learned rule with a condition containing a comparison.
13:59:50,
.010
9.13. Guidelines for Rule and Hypothesis Learning 285
In the center editor, type the comparison operator (<, <=, =, !=, >=, or >).
In the right editor, type the number corresponding to the right side of the comparison.
Click on the OK button in the Create explanation window to accept the explanation.
9.12.5 Hands On: Explanations with Functions and Comparisons

The objective of this case study is to learn how to define explanations involving functions
(as discussed in Section 9.12.3) and comparisons (as discussed in Section 9.12.4).
Start Disciple-EBR, select the case study knowledge base “12-Explanations/Scen,” and
proceed as indicated in the instructions at the bottom of the opened window.
9.13 GUIDELINES FOR RULE AND HYPOTHESIS LEARNING
Guideline 9.1. Properly identify all the entities in the example before
starting rule learning
Before starting rule learning, make sure that all the elements are properly recognized as
instances, numbers, symbolic intervals, or strings. This is important because only the
entities with one of these types will be replaced with variables as part of rule learning, as
shown in the top part of Figure 9.31. Recognizing concepts is also recommended, but it is
optional, since concepts are not generalized. However, recognizing them helps the agent
in explanation generation.
Notice the case from the middle part of Figure 9.31. Because “the United States”
appears as text (in black) and not as instance (in blue), it will not be replaced with a
variable in the learned rule. A similar case is shown at the bottom of Figure 9.31. Because
600.0 is not recognized as number (in green), it will appear as such in the learned rule,
instead of being generalized to a variable.
concept instance Learning concept variable
The price of Apple iPad is under The price of ?O1 is under

600.0 dollars in the United States. ?N1 dollars in ?O2.
number instance variable variable

600.0 dollars in the United States. ?N1 dollars in the United States.
text (not recognized as an instance) same text (not generalized)

600.0 dollars in the United States. 600.0 dollars in the United States.
text (not recognized as a number) same text (not generalized)
Figure 9.31. Variables generation during rule learning.
13:59:50,
.010
Guideline 9.2. Avoid learning from examples that are too specific
It is important to teach the agent with good examples from which it can learn general
rules. A poor example is illustrated in the upper-left part of Figure 9.32. In this case, the
amount of money that Mike has is the same as the price of the Apple iPad 16GB. As a result,
both occurrences of 519.75 are generalized to the same variable ?N1, and the agent will
learn a rule that will apply only to cases where the amount of money of the buyer is exactly
the same as the price of the product (see the upper-right part of Figure 9.32).
You need instead to teach the agent with an example where the numbers are different,
such as the one from the bottom-left part of Figure 9.32. In this case, the agent will
generalize the two numbers to two different variables. Notice that the learned rule will
also apply to cases where ?N1 = ?N2.
Guideline 9.3. Use modeling-based ontology extension before

starting rule learning
Let us consider the reduction step from the left-hand side of Figure 9.33.
If you have used modeling-based ontology specification, then the ontology contains the
basic explanation piece of this reduction step:
John Doe has as number of publications 53
Therefore, before starting rule learning, review the modeling and check that you have
defined the features suggested by the Q/A pair in the ontology.
What are the other explanation pieces for this reduction step? You will need to define
two additional explanation pieces involving comparisons, as well as explanation pieces
fixing the values 41, 53, and very likely, as shown in the left-hand side of Figure 9.34. The
learned rule is shown in the right-hand side of Figure 9.34.
Guideline 9.4. Carefully define the domains and the ranges

of the features
Thoughtfully defined features significantly simplify and speed up the rule-learning process
because their domains and ranges are used in determining the minimal and maximal
generalizations of the examples, as discussed in Sections 9.7.2 and 9.7.3.
Does Mike have enough money Learning Does ?O1 have enough money
to buy Apple iPad 16GB? to buy ?O2?
Yes, because Mike has 519.75 Yes, because ?O1 has ?N1
dollars and Apple iPad 16GB dollars and ?O2
costs 519.75 dollars. costs ?N1 dollars.
Does Bob have enough money Does ?O1 have enough money
to buy Apple iPad 16GB? to buy ?O2?
Yes, because Bob has 620.25 Yes, because ?O1 has ?N1
dollars and Apple iPad 16GB dollars and ?O2
costs 519.75 dollars. costs ?N2 dollars.
Figure 9.32. Variable generation for identical entities.
13:59:50,
.010
9.13. Guidelines for Rule and Hypothesis Learning 287
Figure 9.33. Specification of the ontology knowledge for learning.
Select or define the necessary explanations
Figure 9.34. Another example of rule learning.
When you define a new feature, make sure you define both the domain and the range.
Do not leave the ones generated automatically. The automatically generated range (Any
Element) is too general, and features with that range are not even used in explanations.
Therefore, make sure that you select a more specific domain and range, such as a concept,
a number interval, or a symbolic interval.
If you have already defined facts involving that feature, you need to remove them first
before you can change the domain and the range of the feature.
13:59:50,
.010
Guideline 9.5. Provide hints to guide explanation generation

You may guide explanation generation by selecting one or several entities from the
example, asking the agent to propose only plausible explanation pieces involving those
entities, as was discussed in Section 9.12.1. This is a very effective way of reducing the
combinatorial explosion of explanation generation.
Guideline 9.6. Avoid learning rules without explanations

When you invoke Learn Condition and then Finish without selecting any explanation, the
agent automatically generates a plausible version space rule with a very general upper
bound condition. The rule is then automatically applied to reduce problems. There is a
danger that the rule will generate too many reductions, which may even block the
system. To alleviate this situation, the current implementation of Disciple-EBR uses the
following heuristic:
Generalization heuristic: When no explanation is provided for a reduction step, the
upper bound of the learned rule is obtained by generalizing each instance not to “object,” but
to a direct child of it.
Even with this heuristic, the plausible upper bound condition is still very general. To
alleviate this situation further, consider selecting explanations of the form “<specific
instance> is <concept>,” such as “John Doe is PhD advisor.” You can always provide
such explanations that limit the possible values in similar reductions to those that
make sense.
An additional explanation type to consider is “<generic instance> is <generic
instance>,” such as “Artificial Intelligence is Artificial Intelligence.” However, make sure that
such an explanation makes sense. Do you want “Artificial Intelligence” in all the similar
reductions? If yes, then this explanation is necessary. Otherwise, use “Artificial Intelligence
is <some concept>.”
Yet another explanation type to consider is, “The value is specifically <some value>,”
such as, “The value is specifically 5.0,” or, “The value is specifically certain.” However,
make sure that these explanations make sense. Do you want 5.0 in all the similar reduc-
tions? Do you want “certain” in all the similar reductions?
Guideline 9.7. Recognize concepts in the reasoning tree

When editing a node in the reasoning tree, Disciple-EBR proposes completions of the
current word based on the names of the instances, concepts, and symbolic probabilities
from the ontology. Selecting one of the proposed completions introduces it into the edited
statement with a color characteristic to its type (e.g., dark blue for concepts).
Even if the concepts were not recognized at the time of the development of the
reasoning tree, they can later be recognized automatically by right-clicking on the top
hypothesis and selecting Recognize Concepts in Tree. However, only the concepts that
appear exactly as in the ontology will be recognized, and only in the nodes that have not
yet been formalized.
Having the concepts recognized in the reasoning tree helps the agent in identifying the
relevant explanation pieces during the explanation generation process.
13:59:50,
.010
Operation 9.9. Recognize concepts in the reasoning tree

In the Evidence workspace, click on the Reasoner menu at the top of the window.
In the right panel, right-click on the top hypothesis and select Recognize Concepts
in Tree.
9.14 PROJECT ASSIGNMENT 6
Learn rules from the reasoning trees developed in the previous project assignments.
9.1. Define the rule-learning problem.
9.2. What are the basic steps of rule learning?
9.3. Consider the following expression, where both Jane Austin and Bob Sharp are
specific instances:
?O1 is Jane Austin

?O2 is Bob Sharp
has as advisor ?O1
Find its minimal generalization that does not contain any instance, in the context
of the ontological knowledge from Figure 9.35. Find also its maximal generalization.
9.4. Consider the following explanation of a reduction:
Dan Smith plans to retire from George Mason University.
(a) Reformulate this explanation as a concept with variables.

(b) Determine the minimal generalization of the concept, in the context of the
ontology from Figure 9.36, where all the instances are specific instances.
(c) Determine the maximal generalization of the concept.
9.5. Consider the following explanation of a reduction:
Jane Austin plans to move to Indiana University.
(a) Reformulate this explanation as a concept with variables.

(b) Determine the minimal generalization of the concept, in the context of the
ontology from Figure 9.36, where all the instances are specific instances.
(c) Determine the maximal generalization of the concept.
9.6. Consider the ontological knowledge from Figure 9.37, where Dana Jones, Rutgers
University, and Indiana University are specific instances.
(a) What are the minimal generalization and the maximal generalization of the
following expression?
?O1 is Dana Jones

?O2 is Indiana University
13:59:50,
.010
object
subconcept of
has as advisor
actor research area domain student
range faculty member
subconcept of
organization person
subconcept of
employee student
subconcept of
university employee subconcept of
subconcept of
graduate undergraduate
faculty member staff member student student
subconcept of
instructor professor PhD advisor graduate graduate subconcept of
research teaching
subconcept of assistant assistant BS student
assistant associate full

professor professor professor PhD student MS student
instance of instance of instance of instance of instance of
John Smith John Doe Jane Austin Bob Sharp Joan Dean
Figure 9.35. Ontology fragment.
object plans to rere from

domain person
range organizaon
actor
plans to move to
person organizaon domain person
range organizaon
employee
educaonal organizaon
university employee subconcept of
university
faculty member
instance of
subconcept of
professor PhD advisor University University
subconcept of
full professor plans to

rere from plans to move to
ce of
instance instance of
Dan Smith Jane Ausn
13:59:50,
.010
is faculty member at object

domain faculty member
range university
actor
place
plans to move to
domain person person organizaon
range organizaon
employee educaonal subconcept of
organizaon
university employee subconcept of fun place
university
faculty member instance of
subconcept of instance of
Rutgers Indiana
professor PhD advisor University University
subconcept of plans to move to

full professor is faculty
member at
instance of
Dana Jones
Figure 9.38. A reduction example and its explanation.
(b) What are the minimal generalization and the maximal generalization of the
following expression?
?O1 is Dana Jones

9.7. Consider the example problem reduction and its explanation from Figure 9.38.
Which is the specific rule condition covering only this example? What rule will be
learned from this example and its explanation, assuming the ontology fragment
13:59:50,
.010
from Figure 9.39? What general problem will be learned from the specific IF
problem of this reduction?
Notice that some of the instances are specific (e.g., Aum Shinrikyo and Masami
Tsuchiya), while others are generic (e.g., chemistry).
9.8. Consider the problem reduction example and its explanation from Figure 9.40.
Which is the specific rule covering only this example? What rule will be learned
from this example and its explanation, assuming the ontology fragment from
Figure 9.41? What general problems will be learned from this example? Assume
that all the instances are specific instances.
has as member
domain organization
range person
has master degree in

domain person
range expertise area
Figure 9.40. A problem reduction example and its explanation.
13:59:50,
.010
object
is tesmony by is tesmony about
subconcept of domain evidence domain evidence
range source range evidence
actor source evidence
tesmonial evidence
person
direct tesmonial
author terrorist
item of evidence evidence
instance of instance of subconcept of
subconcept of
Hamid Mir Osama bin Laden tesmonial tesmonial evidence
evidence obtained non-elementary elementary piece based on direct
interview at second hand piece of evidence of evidence observaon
EVD-Dawn-Mir-01-01 EVD-Dawn-Mir-01-01c
9.9. Compare the rule-learning process with the traditional knowledge acquisition
approach, where a knowledge engineer defines such a rule by interacting with
a subject matter expert. Identify as many similarities and differences as possible,
and justify the relative strengths and weaknesses of the two approaches, but be as
concise as possible.
13:59:50,
.010
10 Rule Refinement
10.1 INCREMENTAL RULE REFINEMENT
10.1.1 The Rule Refinement Problem

The triggering event for rule refinement is the discovery of a new positive or negative
example of the rule. There are several possible origins for such an example:
It is generated by the agent during its regular problem solving

It is generated by the agent through active experimentation for rule testing
It is provided by the expert
It is obtained from an external source (e.g., a repository of examples)
Regardless of the origin of the example, the goal of the agent is to refine the rule to be
consistent with the example. A possible effect of rule refinement is the extension of the
ontology.
The rule refinement problem is defined in Table 10.1 and an overview of the rule
refinement method is presented in the next section.
Table 10.1 The Rule Refinement Problem
GIVEN
A plausible version space reduction rule
A positive or a negative example of the rule (i.e., a correct or an incorrect reduction)
A knowledge base that includes an ontology and a set of (previously learned) reduction rules
An expert who will interact with the agent, helping it understand why the example is positive
(correct) or negative (incorrect)
DETERMINE
An improved rule that covers the example if it is positive, and does not cover the example if it is
negative
An extended ontology, if this is needed for rule refinement
294
13:54:13,
.011
10.1. Incremental Rule Refinement 295
10.1.2 Overview of the Rule Refinement Method

Figure 10.1 shows an overview of rule refinement during problem solving. The agent
applies existing partially learned rules to solve the current problem or assess the current
hypothesis. As discussed in Section 9.7, the reductions generated by a rule R are analogous
with the reduction from which the rule was initially learned. Thus this process can be
understood as learning by analogy and experimentation, as indicated in the top part
of Figure 10.1.
The subject matter expert has to verify the generated reduction and accept it if it is
correct, or reject it otherwise. In both cases, the rule will be improved to become
consistent with this new example. We therefore refer to this learning phase as learning
from examples, as indicated in the bottom part of Figure 10.1.
If the generated reduction is correct, then it is a new positive example of the rule, which
may need to be generalized to cover it, by minimally generalizing its applicability condi-
tion. There are various ways to generalize the applicability condition, depending on the
position of the positive example with respect to it. For instance, if the example is covered
by the plausible upper bound of the main condition (the light green ellipse), but it is not
covered by the plausible lower bound (the dark green ellipse), then the plausible lower
bound is generalized as little as possible to cover this new positive example, while still
remaining less general than (or included into) the plausible upper bound.
If the generated example is incorrect, then the expert has to interact with the agent to
help it understand why the reduction is wrong. This interaction is similar to that taking
place during the explanation generation phase of rule learning (see Section 9.5). The
identified failure explanation will be used to refine the rule. We therefore refer to this
phase as learning from explanations, as indicated in the left part of Figure 10.1.
There are various strategies to refine the rule based on a negative example and its
failure explanation (if identified), as illustrated in the top part of Figure 10.1. For instance,
if the negative example is covered only by the upper bound of the main condition, then
this bound is minimally specialized to no longer cover the negative example, while still
remaining more general than the plausible lower bound. Or, a new Except-When plausible
Learning by Analogy
Knowledge Base and Experimentation Rule
Rule’s condition IF we have to solve
– <Problem>
– +
+ Main
PVS Condition
Except-When
Failure PVS Condition
explanation
Examples of problem reductions
THEN solve
generated by the agent <Subproblem 1>
…
<Subproblem m>
Incorrect Correct
example example
Learning from
Explanations Learning from Examples
Figure 10.1. Multistrategy rule refinement.
13:54:13,
.011
296 Chapter 10. Rule Refinement
version space condition may be learned, starting from that negative example and its failure
explanation. This plausible version space Except-When condition is represented by the red
ellipses at the top of Figure 10.1.
The refined rule is shown in the right-hand side of Figure 10.1. The applicability
condition of a partially learned rule consists of a main applicability condition and zero,
one, or more Except-When conditions. The way the rule is refined based on a new
example depends on the type of the example (i.e., positive or negative), on its position
with respect to the current conditions of the rule, and on the type of the explanation of
the example (if identified). The refinement strategies will be discussed in more detail in
the next sections by considering a rule with a main condition and an Except-When
condition, as shown in Figure 10.2. We will consider all the possible nine positions of the
example with respect to the bounds of these conditions. Notice that the presented
methods will similarly apply when there is no Except-When condition or more than
one Except-When condition.
We will first illustrate rule refinement with a positive example and then we will present
the general method.
10.1.3 Rule Refinement with Positive Examples

10.1.3.1 Illustration of Rule Refinement with a Positive Example
The upper-left part of Figure 10.3 shows an example generated by the agent based on the
partially learned rule shown in the right-hand side of the figure. This example is generated
because it satisfies the plausible upper bound condition of the rule. Indeed, the condition
corresponding to this example, shown in the bottom-left part of Figure 10.3, is covered the
plausible upper bound condition: Bob Sharp is a person, Dan Smith is a person, Information
Security is an area of expertise, and certain is in the interval [certain – certain]. Additionally,
the entities from the example are in the relationships from the rule’s condition, that is,
“Bob Sharp is interested in Information Security,” and “Dan Smith is expert in Information
Security.”
Notice also that the example is not covered by the plausible lower bound condition of
the rule because Dan Smith (corresponding to ?O2) is a full professor, and Information
MU: Main Condition

. XU: Except-When Condition
. 4 Plausible Upper Bound

Plausible Upper Bound
.
5
XL: Except-When Condition
ML: Main Condition
6 Plausible Lower Bound
.
Plausible Lower Bound
. 1
..
7
. 8
Universe of
.
2 9
Instances
3
Figure 10.2. Partially learned condition and various positions of a new example.
13:54:13,
.011
Positive example that satisfies the upper bound but not the lower bound
Rule to be refined through the generalization of the lower bound

.011
Condition corresponding to the example

13:54:13,
?O1 is Bob Sharp

?O2 is Dan Smith

is expert in ?O3
?O3 is Informaon Security
Figure 10.3. Positive example generated by the agent.

Security (corresponding to ?O3) is different from Artificial Intelligence. Therefore, this

example corresponds to Case 2 in Figure 10.2.
Because the example is accepted as correct by the expert, the agent will minimally
generalize the plausible lower bound condition of the rule to cover it, while still keeping
this condition less general than the plausible upper bound condition, as shown in
Figure 10.4.
Notice that, with respect to the variable ?O3, Artificial Intelligence and Information
Security are generalized to Computer Science, which is the least general concept covering
both of them, and is less general than area of expertise, the corresponding concept from
the plausible upper bound. With respect to the variable ?O2, PhD advisor does not need to
be generalized because it already covers Dan Smith. However, associate professor is
minimally generalized to professor to cover Dan Smith, which is a full professor. This is
possible because professor is less general than person, the concept from the plausible
upper bound. The refined rule is shown in Figure 10.5.
10.1.3.2 The Method of Rule Refinement with a Positive Example

The general rule refinement method illustrated by the preceding example (i.e., Case 2 in
Figure 10.2) is presented in Table 10.2. It is similar with the part of the Candidate Elimination
Algorithm (Mitchell, 1997) corresponding to the treatment of a positive example.
The method described in Table 10.2 corresponds to the more complex case where the
entity EX from the example is an instance, such as Bob Sharp or Information Security, and
the “climbing the generalization hierarchies” rule is used (see Section 8.3.3). Similar
methods apply when the entity EX corresponds to a number, string, or symbolic probabil-
ity, except that other generalization rules are used, as discussed in Section 8.3.
One of the most difficult problems in computing a good generalization of some
expressions is to establish the objects to be matched (Kodratoff and Ganascia, 1986). This
area of expertise person

Generalized plausible lower bound condition
?O1 is PhD student
is interested in ?O3 professor
Computer Science PhD advisor
?O2 is (PhD advisor, professor) subconcept of
is expert in ?O3
instance of full professor
?O3 is Computer Science
associate professor
Artificial Intelligence ?SI1 is-exactly certain instance of
Information Security Dan Smith
minimal generalization
Condition corresponding to the example Rule’s plausible lower bound condition

?O1 is Bob Sharp ?O1 is PhD student
?O2 is Dan Smith ?O2 is (PhD advisor, associate professor)
is expert in ?O3 is expert in ?O3
?O3 is Information Security ?O3 is Artificial Intelligence
?SI1 is-exactly certain ?SI1 is-exactly certain
Figure 10.4. Minimal generalization of the rule’s plausible lower bound condition.
13:54:13,
.011
Table 10.2 The Example-based Incremental Inductive Generalization Method
Let R be a plausible version space rule, U its main plausible upper bound condition, L its main plausible
lower bound condition, and P a positive example of R covered by U and not covered by L.
Repeat for each variable ?X from U

Let the concept corresponding to ?X from U be UX = {u1 . . . um}.
Let the concept corresponding to ?X from L be LX = {l1 . . . ln}.
Each concept ui from UX is a maximal generalization of all the known positive examples of ?X
that does not cover any of the known negative examples of ?X and is more general than (or
as general as) at least one concept lk from LX. Also, ui is not covered by any other element uk
of UX.
Each concept li is a minimal generalization of the known positive examples of ?X that does
not cover any of the known negative examples of ?X and is less general than (or as general
as) at least one concept uk from UX. Also li does not cover any other element lk of LX.
Let EX be the new positive example of ?X from the rule’s example P. EX is covered by at least
one element ui of UX.
Remove from UX any element that does not cover EX.
Repeat for each li from LX that does not cover EX
Remove li from LX.
Add to LX all minimal generalizations of li and EX that are less general or at most as general
as an element of UX.
Remove from LX all the elements that are more general than or as general as other
elements from LX.
end
end
Determine Pg, the minimal generalization of the example P (see Section 9.9).
Return the generalized rule R with the updated conditions U and L, and Pg in the list of
generalized examples of R.
Refined rule
Condition corresponding to the example

?O1 is Bob Sharp
?O2 is Dan Smith

is expert in ?O3
?O3 is Information Security
Rule’s condition
Figure 10.5. Rule generalization based on a positive example.
13:54:13,
.011
problem, however, is trivial in Disciple-EBR because both the plausible lower bound
condition and the condition corresponding to the example have exactly the same struc-
ture, and the corresponding variables have the same names, as shown in Figure 10.5. This
is a direct consequence of the fact that the example is generated from the plausible upper
bound condition of the rule.
10.1.3.3 Summary of Rule Refinement with a Positive Example

The general method for rule refinement with a positive example, covering all the nine
cases from Figure 10.2, is presented in Table 10.3. Notice that a positive example covered
by an Except-When condition leads to the specialization of that condition to no longer
cover it. This is similar with a negative example covered by the main applicability condi-
tion, which is discussed in the next section.
Next we will illustrate rule refinement with negative examples, and then we will present
the general refinement method.
10.1.4 Rule Refinement with Negative Examples

10.1.4.1 Illustration of Rule Refinement with Except-When Conditions
The left side of Figure 10.6 shows an example generated by the agent based on the partially
learned rule shown in the right-hand side of Figure 10.6. This example is generated
because it satisfies the plausible upper bound condition of the rule. However, the example
is rejected as incorrect by the expert, with the following explanation:
Dan Smith plans to retire from George Mason University
Based on this failure explanation, the agent generates an Except-When plausible version
space condition by applying the method described in Sections 9.6 and 9.7. First it reformu-
lates the explanation as a specific condition by using the corresponding variables from the
rule, or by generating new variables (see also the bottom-left part of Figure 10.7):
?O1 is Dan Smith

?O2 is George Mason University
Then the agent generates a plausible version space by determining maximal and minimal
generalizations of the preceding condition. Finally, the agent adds it to the rule as an
Except-When plausible version space condition, as shown in the bottom-right part of
Figure 10.7. The Except-When condition should not be satisfied to apply the rule. Thus, in
order to conclude that a professor will stay on the faculty for the duration of the disserta-
tion of a student, the professor should have a long-term position (the main condition) and
it should not be the case that the professor plans to retire from the university (the Except-
When condition).
Figure 10.8 shows the further refinement of the rule with an additional negative
example. This example satisfies the rule in Figure 10.7. Indeed, Jane Austin has a long-term
position and she does not plan to retire from George Mason University. Nevertheless, the
expert rejects the reasoning represented by this example because Jane Austin plans to
move to Indiana University. Therefore, she will not stay on the faculty of George Mason
University for the duration of the dissertation of Bob Sharp.
13:54:13,
.011
Table 10.3 Rule Refinement with a Positive Example

MU XU
1. If the positive example E is covered by ML and is not
covered by XU (case 1), then the rule does not need to ML
XL
be refined because the example is correctly classified

1
as positive by the current rule.
MU XU XU
MU
2. If E is covered by MU, is not covered by ML, and is not
covered by XU (case 2), then minimally generalize ML ML
XL
ML
XL
to cover E and remain less general than MU. Remove

from MU the elements that do not cover E.
2 +
MU XU MU XU
3. If E is not covered by MU (cases 3 and 5), or if E is 5 +
covered by XL (cases 5, 6, and 7), then keep E as a ML 6 XL
ML + XL
positive exception of the rule. Alternatively, learn a 7 +

new rule starting from this example.
3 +
MU + XU
4. If E is not covered by MU but is covered by XU (case
XL
4), then E is kept as a positive exception (or a new rule ML
is learned). Additionally, interact with the expert to MU 4 XU

find an explanation of the form, “The problem
XL Ii is Ci
reduction step is correct because Ii is Ci,” where Ii is an ML
+
instance from the example E and Ci is a concept from MU XU
the ontology. If such an explanation is found, then XU XL

ML
is minimally specialized to no longer cover E.
5. If E is covered by ML and XU, but it is not covered by

MU XU
XL (case 8), then interact with the expert to find an
explanation of the form, “The problem reduction step ML
XL
is correct because Ii is Ci,” where Ii is an instance from XU

MU
the example E and Ci is a concept from the ontology. If +
XL Ii is Ci
such an explanation is found, then XU is minimally ML
XU
specialized to no longer cover E. Otherwise, E is kept MU
as a positive exception. 8
ML
XL
6. If E is covered by MU and XU, it is not covered by ML,

and it is not covered by XL (case 9), then minimally XU
MU
generalize ML to cover E and remain less general than
MU. Also remove from MU the elements that do not ML
XL
cover E. Then interact with the expert to find an XU

MU
explanation of the form, “The problem reduction step
XL Ii is Ci .+
is correct because Ii is Ci,” where Ii is an instance from ML
XU
the example E and Ci is a concept from the ontology. If MU
such an explanation is found, then XU is minimally 9 ML

XL
specialized to no longer cover E. Otherwise, E is kept as

a positive exception.
+
As in the case of the previous negative example, the failure explanation
Jane Austin plans to move to Indiana University
is rewritten as a specific condition
?O1 is Jane Austin

13:54:13,
.011
Negative example generated by the rule Rule that generated the negative example
.011
13:54:13,
Failure explanation
Dan Smith plans to retire from George Mason University
Figure 10.6. Negative example generated by the agent.

Negative example Specialized rule
.011
+
-
13:54:13,
Failure explanation
Dan Smith plans to rere from George Mason University
Rewrite as
Specific Except-When condition
?O1 is Dan Smith
?O2 is George Mason University
Figure 10.7. Rule specialization with a negative example.

Specialized rule
Negative example
.011
13:54:13,
+
-
Failure explanation -
Jane Austin plans to move to Indiana University
Rewrite as
Specific Except-When condition
?O1 is Jane Austin
Max generalization
plans to move to ?O5 Min generalization
Figure 10.8. Rule refinement with an additional negative example.

Notice that the agent has introduced a new variable ?O5 because Indiana University does
not correspond to any entity from the previous form of the rule (as opposed to Jane Austin
who corresponds to ?O1).
Then the agent generates a plausible version space by determining maximal and
minimal generalizations of the preceding condition. Finally, the agent adds it to the rule
as an additional Except-When plausible version space condition, as shown at the bottom-
right part of Figure 10.8.
10.1.4.2 The Method of Rule Refinement with Except-When Conditions

The basic steps for learning an Except-When condition based on the explanation of a
negative example are summarized in Table 10.4.
10.1.4.3 Illustration of Rule Refinement through Condition

Specialization
The preceding two sections presented the refinement of a rule through the addition of
Except-When plausible version space conditions. This section will illustrate rule refinement
through the specialization of the upper bound of the main applicability condition of the rule.
Figure 10.9 illustrates the mixed-initiative development of a reasoning tree. The expert
formulated the hypothesis:
Jill Knox would be a good PhD advisor for Peter Jones.
Then the agent developed a partial reasoning tree, but it was unable to assess one of the
subhypotheses:
Jill Knox will stay on the faculty of George Mason University for the duration of the
dissertation of Peter Jones.
Table 10.4 Basic Steps for Learning an Except-When Condition
Let R be a plausible version space rule, N an instance of R rejected by the expert as an incorrect
reasoning step (a negative example of R), and EX an explanation of why N is incorrect (a failure
explanation).
(1) Reformulation of the Failure Explanation
Generate a new variable for each instance and each constant (i.e., number, string, or symbolic
probability) that appears in the failure explanation EX but does not appears in the negative
example N. Use the new variables and the rule’s variables to reformulate the failure explanation EX
as an instance I of the concept EC representing an Except-When condition of the rule R.
(2) Analogy-based Generalizations of the Failure Explanation
Generate the plausible upper bound XU of the concept EC as the maximal generalization of I in the
context of the agent’s ontology.
Generate the plausible lower bound LU of the concept EC as the minimal generalization of I that
does not contain any specific instance.
(3) Rule Refinement with an Except-When Plausible Version Space Condition
Add an Except-When plausible version space condition (XU, LU) to the existing conditions of the
rule R. This condition should not be satisfied for the rule to be applicable in a given situation.
13:54:13,
.011
Therefore, the expert defined the reduction of this hypothesis, which includes its assess-
ment, as shown at the bottom of Figure 10.9.
Based on this example, the agent learned a general rule, as illustrated in the right-hand
side of Figure 10.9 and as discussed in Chapter 9. The rule is shown in Figure 10.10.
1. Solving
Agent applies
learned rules to
solve new
problems
3. Learning
2. Modeling
Agent learns
a new rule
Expert extends the

reasoning tree
Figure 10.9. Problem solving, modeling, and learning.
Figure 10.10. Partially learned rule.
13:54:13,
.011
This and the other learned rules enabled the agent to develop the reasoning tree from
Figure 10.11 for assessing the following hypothesis:
Bill Bones would be a good PhD advisor for June Allison.
However, the expert rejected the bottom reasoning step as incorrect. Indeed, the correct
answer to the question, “Is Bill Bones likely stay on the faculty of George Mason University
for the duration of the PhD dissertation of June Allison?” is, “No,” not, “Yes,” because there
is no support for Bill Bones getting tenure.
The user–agent interaction during example understanding is illustrated in Figure 10.12.
The agent identified an entity in the example (the symbolic probability “no support”) that
would enable it to specialize the upper bound of the main condition of the rule to no
longer cover the negative example. Therefore, it proposed the failure explanation shown at
the right-hand side of Figure 10.12:
The example is incorrect because of the value no support.
The expert accepted this explanation by clicking on OK, and the rule was automatically
specialized as indicated in Figure 10.13. More precisely, the upper bound of the main
condition for the variable ?Sl1 was minimally specialized from the interval [no support –
certain] to the interval [likely – certain], in order to no longer cover the value no support,
while continuing to cover the interval representing the lower bound, which is [almost
certain – almost certain].
10.1.4.4 The Method of Rule Refinement through Condition

Specialization
The previous example illustrated the specialization of an upper bound of a condition with
a negative example. The corresponding general method is presented in Table 10.5.
3. Solving
Agent applies
learned rules
5. Refinement
to solve new
problems
Agent refines
rule with
negative example
4. Critiquing
Incorrect because + -
of the “no support”
probability
Figure 10.11. Solving, critiquing, and rule refinement.
13:54:13,
.011
Figure 10.12. Explaining why the problem reduction is incorrect.
Specialized rule
Failure explanation
Min specialization
[no support – certain]
Figure 10.13. Specialization of the upper bound of a plausible version space condition.
10.1.4.5 Summary of Rule Refinement with a Negative Example

We have illustrated only some of the nine cases of negative examples from Figure 10.2
(p. 296). The general method for rule refinement with a negative example, covering all the
cases, is presented in Table 10.6.
13:54:13,
.011
10.2. Learning with an Evolving Ontology 309
Table 10.5 The Example-based Incremental Inductive Specialization Method
Let R be a plausible version space rule, U the plausible upper bound of the main condition, L the
plausible lower bound of the main condition, N a negative example covered by U and not covered
by L, and C an entity from N that is blamed for the failure.
1. Let ?X be the variable from the rule’s conditions that corresponds to the blamed
entity C.
Let UX and LX be the classes of ?X in the two bounds.
If each concept from LX covers C
then Continue with step 2.
else Continue with step 3.
2. The rule cannot be specialized to uncover the current negative example.
The negative example N is associated with the rule as a negative exception.
Return the rule R.
3. There are concepts in LX that do not cover C. The rule can be specialized to uncover
N by specializing UX, which is known to be more general than C.
3.1. Remove from LX any element that covers C.
3.2. Repeat for each element ui of UX that covers C
Remove ui from UX.
Add to UX all minimal specializations of ui that do not cover C and are more general than
or at least as general as a concept from LX.
Remove from UX all the concepts that are less general than or as general as other
concepts from UX.
end
4. Return the specialized rule R.
10.2 LEARNING WITH AN EVOLVING ONTOLOGY
10.2.1 The Rule Regeneration Problem

The applicability conditions of the learned rules depend of the agent’s ontology, which
is used as a generalization hierarchy. When the ontology changes in significant ways,
the rules need to be updated accordingly. Let us consider that the ontology fragment
shown in Figure 9.7 (p. 258) has evolved into the one from Figure 10.14. The changed
elements are underlined. In particular, PhD advisor was moved from under professor to
above professor. A new concept, expert, was introduced between PhD advisor and
person. Additionally, the domains of the features is expert in, has as position, and plans
to retire from have been changed from person to expert, employee, and employee,
respectively. These are all changes that affect the minimal and the maximal generaliza-
tions of the examples from which the rules have been learned. Therefore, the rules
need to be updated.
In Disciple-EBR, the adaptation of the rules to a changed ontology is done by simply
automatically regenerating the rules based on the minimal generalizations of the examples
13:54:13,
.011
Table 10.6 Rule Refinement with a Negative Example

MU XU
1. If the negative example N is covered by ML and it is XL

not covered by XU (case 1), then interact with the ML
XU
subject matter expert to find an explanation of why MU _
N is an incorrect problem reduction step. If an XL

ML EX
explanation EX is found, then generate a new
1 MU XU
Except-When plausible version space condition and
add it to the rule. Otherwise, keep N as a negative ML
XL
exception. _
2. If N is covered by MU but it is not covered by ML MU XU

and by XU (case 2), then interact with the expert to
XL
find an explanation of why N is an incorrect ML
problem reduction step. If an explanation EX is MU XU
found and it has the form “because of v,” then XL “because of v” _

ML
minimally specialize MU to no longer cover N while
XU
still remaining more general than ML. Otherwise, if MU
another type of explanation EX is found, then learn a 2 EX XL

ML
new Except-When plausible version space condition
based on it, and add it to the rule.
_
3. If N is not covered by MU (cases 3, 5), or it is

covered by XL (cases 5, 6, 7), then the rule does not MU 4 XU MU _ XU
5 _
need to be refined because the example is correctly XL _ XL
ML 6 ML
classified as negative by the current rule. If N is not
7 _
covered by MU, is not covered by XL, and is covered
by XU (case 4), then minimally generalize XL to
_
cover N and remain less general than XU. 3
MU XU MU XU
4. If N is covered by ML and by XU, but it is not covered
by XL (case 8), or N is covered by MU and by XU, but ML
XL
ML
XL
it is not covered by ML and XL (case 9), then

minimally generalize XL to cover N while still 8 _
9 _
remaining less general than XU.
and their explanations in the context of the updated ontology. This is, in fact, the reason
why the generalized examples are maintained with each rule, as discussed in Section 9.9.
The rule regeneration problem is presented in Table 10.7. Notice that not all the
changes of an ontology lead to changes in the previously learned rules. For example,
adding a new concept that has no instance, or adding a new instance, will not affect the
previously learned rules. Also, renaming a concept or a feature in the ontology automatic-
ally renames it in the learned rules, and no additional adaptation is necessary.
10.2.2 On-Demand Rule Regeneration

Rules adaptation in response to changes in the ontology is an expensive operation if it
means regenerating all the rules after each significant change of the ontology. Therefore,
the question is: When should the rules be regenerated? The solution adopted in Disciple-
EBR is on-demand regeneration. That is, a rule is regenerated only when the agent needs
to apply it in problem solving, and only if the current version of the ontology is different
from the one when the rule was last refined or regenerated. Thus, the agent associates with
13:54:13,
.011
the ontology a version, and each time a rule is learned or refined, it associates the version
of the ontology with the rule. The version of the ontology is incremented each time a
significant change – that is, a change that may affect the conditions of the previously
learned rules – is made. Then, before using a rule in problem solving, the agent checks
the rule’s ontology version with the current version of the ontology. If the versions are
the same, the rule is up to date and can be used. Otherwise, the agent regenerates the
rule based on the current ontology and also updates the rule’s ontology version to the
current version of the ontology. The on-demand rule regeneration method is presented
in Table 10.8.
Table 10.7 The Rule Regeneration Problem
GIVEN
A plausible version space reduction rule R corresponding to a version v of the ontology
Minimal generalizations of the examples and explanations from which the rule R was learned, in
the context of the version v of the ontology
An updated ontology with a new version v’
DETERMINE
An updated rule that corresponds to the same generalized examples, but in the context of the
new version v’ of the ontology
Updated minimal generalizations of the specific examples from the current scenario, if any
object
actor position area of expertise
person organization university position subconcept of
educational
expert employee student organization faculty position Computer Science
university employee subconcept of instance of

long-term
faculty position
faculty member university
tenured position Artificial Information
graduate instance of Intelligence Security
PhD advisor student
subconcept of University University has as position
professor domain employee
PhD student range position
subconcept of
is expert in plans to retire from
associate full domain expert domain employee
professor professor range area of expertise range organization
instance of instance of is is interested in plans to move to
is expert in interested domain person domain person
John Doe Bob Sharp in range area of expertise range organization
Figure 10.14. Evolution of the ontology fragment from Figure 9.7.
13:54:13,
.011
We will first illustrate the regeneration of the rules presented in the previous sections
and then provide the general regeneration method.
10.2.3 Illustration of the Rule Regeneration Method

Let us first consider the rule from the right-hand side of Figure 10.5 (p. 299). This rule was
learned from two positive examples and their explanations, based on the ontology from
Figure 9.7 (p. 258), as discussed in Section 10.1.3.1. This rule can be automatically relearned
from its generalized examples, in the context of the updated ontology, as illustrated in
Figure 10.15. The bottom part of Figure 10.15 shows the generalized examples of the rule,
which were determined in the context of the ontology from Figure 9.7 (p. 258), as discussed
Table 10.8 On-Demand Rule Regeneration
Let R be a plausible version space rule, and O the current ontology with version v.
If R’s ontology version is v, the same as the version of the current ontology O
then Return R (no regeneration is needed).
else Regenerate rule R (see Table 10.9).
Set R’s ontology version to v.
Return R
Plausible Lower Plausible Upper

Bound Condition (LB) Bound Condition (UB)
?O1 is PhD student ?O1 is person
?O2 is professor ?O2 is expert
?O3 is Computer Science ?O3 is area of expertise
?SI1 is in [certain - certain] ?SI1 is in [certain - certain]
Min Max
Example generalization Example generalization

?O1 is PhD student ?O1 is PhD student
?O2 is PhD advisor ?O2 is PhD advisor
is associate professor is full professor
?O3 is Artificial Intelligence ?O3 is Information Security
?SI1 is exactly certain ?SI1 is exactly certain
Covered positive examples: 1 Covered positive examples: 1
Covered negative examples: 0 Covered negative examples: 0
Figure 10.15. The updated conditions of the rule from Figure 10.5 in the context of the ontology
from Figure 10.14.
13:54:13,
.011
in Section 9.9. The top part of Figure 10.15 shows the updated bounds of the rule, in the
context of the updated ontology from Figure 10.14. The new plausible lower bound condi-
tion is the minimal generalization of the generalized examples, in the context of the updated
ontology. Similarly, the new plausible upper bound condition is the maximal generalization
of the generalized examples in the context of the updated ontology.
Notice that in the updated plausible lower bound condition (shown in the upper-left
part of Figure 10.15), ?O2 is now a professor, instead of a professor or PhD advisor. Indeed,
note the following expression from the first generalized example shown in the lower-left
part of Figure 10.15:
?O2 is PhD advisor [10.1]

is expert in ?O3
Based on the updated ontology, where associate professor is a subconcept of PhD advisor,
this expression is now equivalent to the following:
?O2 is associate professor [10.2]

is expert in ?O3
Similarly, note the following expression from the second generalized example shown in
the lower-right part of Figure 10.15:
?O2 is PhD advisor [10.3]

is full professor
is expert in ?O3
Because full professor is now a subconcept of PhD advisor, this expression is now equivalent
to the following:
?O2 is full professor [10.4]

is expert in ?O3
Then the minimal generalization of the expressions [10.2] and [10.4] is the following
expression because professor is the minimal generalization of associate professor and full
professor:
?O2 is professor [10.5]

is expert in ?O3
Also, in the updated plausible upper bound condition (shown in the upper-right part of
Figure 10.15), ?O2 is an expert instead of a person, because this is the maximal generaliza-
tion of associate professor and full professor, which is included into the domain of the is
expert in feature of ?O2, which is expert.
Let us now consider the rule from the right-hand part of Figure 10.8 (p. 304). The
minimal generalizations of the examples from which this rule was learned are shown
under the updated conditions in Figure 10.16. They were determined based on the
ontology from Figure 9.7. You remember that this rule was learned from one positive
example and two negative examples. However, each of the negative examples was used as
a positive example of an Except-When plausible version space condition. That is why each
of the generalized examples in Figure 10.16 has a positive example.
13:54:13,
.011
Except-When Condition 1
?O1 is full professor ?O1 is employee
plans to retire from ?O2 plans to retire from ?O2
?O2 is university ?O2 is organization
Min Max
Main Condition generalization generalization
Bound Condition (LB) Bound Condition (UB) Example generalization
?O1 is full professor
?O1 is associate professor ?O1 is employee
is PhD advisor
has as position ?O4 has as position ?O4
?O2 is university ?O2 is actor
?O2 is university
?O3 is PhD student ?O3 is actor
?O4 is tenured position ?O4 is long-term faculty position Covered positive examples: 1
?SI1 is in [almost certain - almost certain] ?SI1 is in [almost certain - almost certain] Covered negative examples: 0
Min Max Except-When Condition 2

generalization generalization Plausible Lower
.011
Plausible Upper
Example Generalization Bound Condition (LB) Bound Condition (UB)
?O1 is associate professor ?O1 is full professor ?O1 is person
plans to move to ?O5 plans to move to ?O5
13:54:13,
is PhD advisor
has as position ?O4 ?O5 is university ?O5 is organization
?O2 is university
?O3 is PhD student Min Max
?O4 is tenured position generalization generalization
?SI1 is exactly almost certain Example Generalization
Covered positive examples: 1 ?O1 is full professor
Covered negative examples: 0 is PhD advisor
?O5 is university

Figure 10.16. The updated conditions of the rule from Figure 10.8, in the context of the ontology from Figure 10.14.
The lower and upper bounds of the rule in Figure 10.8 (p. 304) were updated by
computing the minimal and maximal generalizations of these generalized examples in
the context of the updated ontology from Figure 10.14. Let us first consider the updated
version space of the main condition. Notice that in the lower bound, ?O1 is now associate
professor, instead of PhD advisor or associate professor (see Figure 10.8). Also, in the upper
bound, ?O1 is employee instead of person. The version spaces of the Except-When condi-
tions have also been updated. In the first Except-When condition, ?O1 is now full professor
in the lower bound and employee in the upper bound, instead of PhD advisor or full
professor and person, respectively. Similarly, in the second Except-When condition, ?O1
is now full professor in the lower bound, instead of PhD advisor or full professor.
Finally, let us consider the rule from Figure 10.13 (p. 308), which was learned based on
the ontology from Figure 9.7, as discussed in Section 10.1.4. The minimal generalizations
of the positive and negative examples from which this rule was learned are shown at the
bottom of Figure 10.17. They were determined based on the ontology from Figure 9.7.
Notice that these generalized examples include all the explanations from which the rule
was learned. In particular, the explanation that fixed the value of “tenure-track position” is
represented as “?O4 is exactly tenure-track position” and that which excluded the value “no
support” is represented as “?SI1 is-not no support in main condition.”

?O1 is assistant professor ?O1 is professor
probability of tenure ?SI1 probability of tenure ?SI1
?O2 is university ?O2 is actor
?O3 is PhD student ?O3 is actor
?O4 is tenure-track position ?O4 is tenure-track position
?O5 is tenured position ?O5 is tenured position
?SI1 is in [almost certain - almost certain] ?SI1 is in [likely - almost certain]
Min Max
Example generalization Example generalization

?O1 is assistant professor ?O1 is assistant professor
probability of tenure ?SI1 probability of tenure ?SI1
?O2 is university ?O2 is university
?O3 is PhD student ?O3 is PhD student
?O4 is exactly tenure-track position ?O4 is exactly tenure-track position
?O5 is exactly tenured position ?O5 is exactly tenured position
?SI1 is almost certain ?SI1 is no support
is-not no support in main condition
Covered negative examples: 0 Covered positive examples: 0
Figure 10.17. The updated conditions of the rule from Figure 10.13 in the context of the ontology
from Figure 10.14.
13:54:13,
.011
The new lower and upper bounds of the rule in the context of the updated ontology
from Figure 10.14 are shown at the top of Figure 10.17. Notice that in this case, the
regenerated rule is actually the same with the previous rule. The changes made to the
ontology did not affect this rule. However, the agent did recompute it because it cannot
know, a priori, whether the rule will be changed or not. The only change made to the rule
is to register that it was determined based on the new version of the ontology.
10.2.4 The Rule Regeneration Method

The previous section has illustrated the automatic rule regeneration method. Table 10.9
and the ones following it describe the actual method in a more formal way.
10.3 HYPOTHESIS REFINEMENT
Hypothesis refinement is performed using methods that are very similar to the preceding
methods for rule refinement, as briefly summarized in this section.
Remember that general hypotheses are automatically learned as a byproduct of reduc-
tion rule learning, if they have not been previously learned. When a reduction rule is
refined with a positive example, each included hypothesis is also automatically refined
with its corresponding positive example. Indeed, when you say that a reduction is correct,
you are also implicitly saying that each of the included hypotheses is correct.
However, when a reduction rule is refined with a negative example, the hypotheses
are not affected. Indeed, a negative reduction example means that the corresponding
reduction is not correct, not that the any of the involved hypotheses is incorrect. For this
reason, an explanation of why a specific reduction is incorrect does not automatically
apply to the hypotheses from that reduction. If you want to say that a specific hypothesis
is incorrect, you have to select it and click on the Incorrect problem button. Then the
Table 10.9 Rule Regeneration Method
Let O be the current ontology with version v, and R a plausible version space rule with a different
version.
1. Recompute the formal parameters P for rule R (see Table 10.10).
2. Refresh examples for rule R (see Table 10.11).
3. If R is no longer valid (i.e., the rule has no longer any generalized positive example)
then Return null
4. Recompute plausible version space for rule R (see Table 10.12).
5. Repeat for each specific example EX of R
If the upper bound of the main plausible version space condition (PVS) of R does not cover
EX and EX is a positive example
then make EX a positive exception.
If the upper bound of PVS of R does cover EX and EX is a negative example
then make EX a negative exception.
end
6. Return R
13:54:13,
.011
10.4. Characterization of Rule Refinement 317
Table 10.10 Method for Recomputing the Formal Parameters of a Rule
Let R be a plausible version space rule having the list of parameters P.

Recompute the parameters P of R as follows:
1. P = IF task parameters [ question parameters [ answer parameters
2. Repeat for each THEN statement of R
P = P [ THEN statement parameters
end
3. Repeat for each explanation of R
P = P [ explanation parameters
end
4. Return P
Table 10.11 Method for Refreshing the Examples of a Rule
Let R be a plausible version space rule.

1. Repeat for each specific example EX of the rule R
Keep only the values corresponding to the current list of parameters in the rule R.
end
2. Repeat for each generalized example GE of the rule R
Keep only the values corresponding to the current list of parameters in the rule R.
end
3. Repeat for each specific example EX of the rule R
If the stored generalized example is no longer a minimal generalization then
Unregister the example from the generalized example.
If the generalized example has no examples any more
then delete the generalized example.
Register the example with its new minimal generalization.
end
4. If the rule R has no generalized or specific positive examples any more
then Remove the rule R.
selected hypothesis will be refined basically using the same methods as those for rule
refinement.
Just as a refined rule, a refined hypothesis may include, in addition to the plausible
version space of the main condition, one or several plausible version spaces of Except-When
conditions, and generalized positive and negative examples. When the ontology is changed,
the hypotheses can be automatically regenerated based on the associated generalized
examples. They are actually regenerated when the corresponding rules are regenerated.
10.4 CHARACTERIZATION OF RULE LEARNING

AND REFINEMENT
The presented rule learning and refinement methods have the following characteristics:
They use multistrategy learning that synergistically integrates mixed-initiative learn-

ing from examples, from explanations, and by analogy, to take advantage of the
13:54:13,
.011
Table 10.12 Method for Recomputing the Plausible Version Space of a Rule
Let R be a plausible version space rule having the main condition M and the list of the Except-
When conditions LX.
1. Let MP be the list of the parameters from the main condition M.
Let IP be the list of parameters from the natural language part of the rule R, referred to as
informal parameters.
MP = IP
2. Repeat for each explanation of a positive example EP in the rule R
MP = MP [ the new parameters from EP
end
3. Let LGS be the list of the generalized examples that have at least one specific positive
example.
Let LEP be the list of the explanations EP of the positive examples in the rule R.
Create the multivariable condition MC based on MP, LGS, LEP, IP (see Table 10.13).
4. Let LX be the list of Except-When conditions.
LX = [ ]
5. Repeat for each group EEp of Except-When explanations in R.
Let EP be the list of parameters used in EEp.
Let EX be the list of the generalized negative examples associated with EEp that have
at least one specific negative example.
Create the multivariable condition XC based on MP = EP, LGS = EX, LEP = EEp, and
IP = ∅ (see Table 10.13).
LX = LX [ XC
end
6. Return R
complementary strengths of these individual learning strategies and compensate for

their relative weaknesses.
The methods efficiently capture the expert’s tacit knowledge, significantly reducing the
complexity of developing cognitive assistants.
They use the explanation of the first positive example and analogical reasoning to
generate a much smaller initial version space than Mitchell’s classical version space
method (Mitchell, 1997).
The methods efficiently search the version space, guided by explanations obtained
through mixed-initiative interaction with the user (both the upper bounds and
the lower bounds being both generalized and specialized to converge toward
one another).
They learn from only a few examples, in the context of an incomplete and evolving
ontology.
These methods enable agents to learn even in the presence of exceptions.
They keep minimally generalized examples and explanations to regenerate the rules
automatically when the ontology changes.
The methods have been successfully applied to develop cognitive assistants for many
complex real-world domains (e.g., intelligence analysis, military strategy and planning,
education, collaborative emergency response planning), because they enable agents to
learn within a complex representation language.
13:54:13,
.011
10.5. Hands On: Rule Refinement 319
Table 10.13 Method for Creating the Multivariable Condition Structure
Let R be a plausible version space rule, MP be the list of the parameters from the main condition,
LGS be the list of generalized examples that have at least one specific positive example, LEP be the
list of the explanations EP of the positive examples in the rule R, and IP be the list of informal
parameters of R.
1. Let A be the list of the generalized explanations fragments (such as “?Oi is interested in
?Oj”) from the generalized examples of R.
Compute A based on MP and LEP.
2. Let D be the domains of the variables from MP, each domain consisting of a lower bound
and an upper bound.
D=[]
3. Repeat for each parameter ?Oi in MP
Determine the list GE = {ge1, . . ., geg} of the concepts from LGS corresponding to ?Oi
(e.g., “assistant professor” from “?Oi is assistant professor”).
Determine the list PC = {PC1, . . ., PCp} of the concepts to which ?Oi must belong,
corresponding to positively constraining explanations
(e.g., “?Oi is exactly tenured position”).
Determine the list NC = {NC1, . . ., NCn} of the concepts to which ?Oi must not belong,
corresponding to negatively constraining explanations (e.g., “Sl1 is-not no support”).
Create the domains Do from the examples GE and the constraints PC and NC
(see Table 10.14).
D = D [ Do
end
4. Create the multivariable condition structure from MP, D, and A.
Return MVC
10.5 HANDS ON: RULE REFINEMENT
There are two rule refinement case studies, a shorter one and a longer one. In the shorter
case study, you will guide the agent to refine the rule from Figure 9.8 (p. 259), as discussed
in the previous sections. In the longer case study, you will guide the agent to refine several
rules, including the rule from Figure 9.8. You may perform the short case study, or the long
one, or both of them.
Start Disciple-EBR, select the case study knowledge base (either “13-Rule-Refinement-
short/Scen” or “13-Rule-Refinement/Scen”), and proceed as indicated in the instructions
at the bottom of the opened window.
The following are the basic operations for rule refinement, as well as additional operations
that are useful for knowledge base refinement, such as changing a generated reasoning step
into a modeling step, visualizing a rule with the Rule Editor, and deleting a rule.
Operation 10.1. Refine rule with positive example

In the Scenario workspace, during rule refinement, in the Reasoning Hierarchy viewer,
select a reasoning step for which a rule has already been learned by clicking on the
corresponding Q/A node.
In the right panel, click on the Correct Reduction button.
13:54:13,
.011
Table 10.14 Method for Creating Domains from Examples and Constraints
Let ?Oi be a parameter for which we compute a plausible version space.

GE = {ge1, . . ., geg} be the set of generalized examples corresponding to variable ?Oi.
PC = {PC1, . . ., PCp} be the set of concepts to which ?Oi must belong, representing positive
constraints.
NC = {NC1, . . ., NCn} be the set of concepts to which ?Oi must not belong, representing
negative constraints.
1. Compute UB, the maximal generalization of the generalized examples in GE.
2. Compute LB, the minimal generalization of the generalized examples in GE.
3. Compute UB, the minimal specialization of UB based on all the positive constraints PC.
4. Compute LB by selecting from LB the elements covered by UB.
5. Repeat for each NCi in NC
Compute LB, the minimal specialization of LB that is disjoint with
the negative constraint NCi.
Compute UB, the minimal specialization of UB covering LB that is disjoint with the
negative constraint NCi.
end
6. Create the list (UB, LB) as the domain Do for ?Oi.
7. Return Do
Operation 10.2. Refine rule with negative example

In the Scenario workspace, during rule refinement, in the Reasoning Hierarchy viewer,
In the right panel, click on the Incorrect Reduction button.
Follow the agent’s instructions to select or provide a failure explanation.
Operation 10.3. Replace a generated reasoning step with

a modeling step
In the Scenario workspace, in the Reasoning Hierarchy viewer, right-click on the Q/A
node of the reasoning step that was generated through the application of a learned rule
and select Replace with modeling.
As a result, the reasoning step is replaced with a modeled one that can be changed.
Additionally, the initial reasoning step becomes a negative exception of the rule.
Operation 10.4. View a rule with the Rule Browser

Click on the Rules menu and select Rule Browser.
The left pane displays the names of the rules.
Click on one of these rules (e.g., DDR.00018) and the right pane displays its descrip-
tion in a given format (e.g., “Formal Description”) when you select it in the
corresponding tab.
Generally, you would first find the name of a rule of interest from a reasoning tree (e.g.,
DDR.00018), as described in Operation 9.3, and then you can view or delete it with the
Rule Browser.
Click on the X button of the Rule Browser, to close it.
13:54:13,
.011
10.6. Guidelines for Rule Refinement 321
Operation 10.5. Delete a rule with the Rule Browser

Generally, you would first find the name of the rule to be deleted (e.g., DDR.00018) by
inspecting the reasoning tree, as described in Operation 9.3.
In the Scenario workspace, click on the Rules menu and select Rule Browser.
The left pane displays the names of the rules.
Click on the name of the rule to be deleted (e.g., DDR.00018), and the right pane
displays its description in a given format (e.g., “Formal Description”) when you select it
in the corresponding tab.
At the bottom of the right pane, click on the Edit Rule button. As a result, new options
are displayed.
At the bottom of the right pane, click on the Delete Rule button.
Click on the X button of the Rule Browser to close it.
10.6 GUIDELINES FOR RULE REFINEMENT
Guideline 10.1. Assess similar hypotheses to refine the rules

Formulate hypotheses (or problems) that are similar to those from which the rules were
learned. Refine the rules with positive examples by selecting (clicking on) the corresponding
Q/A pair, and then clicking on Correct Reduction (see the right-hand side of Figure 10.18).
You can also click on Modify Explanations to interact with the agent and update the
explanations (e.g., by adding new ones and/or deleting some of the existing ones).
Refine the rules with negative examples by selecting (clicking on) the corresponding
Q/A pair and then clicking on Incorrect Reduction (see the right-hand side of Figure 10.18).
Then interact with the agent to identify an explanation of why the reduction is incorrect.
Guideline 10.2. Extend the ontology to define failure explanations

If necessary, extend the ontology to provide explanations for negative examples. However,
do not limit yourself to a minimal extension of the ontology (that consists of only the
failure explanation) if you anticipate that related concepts and features may be needed in
the future for related hypotheses or problems.
Figure 10.18. Rule refinement with a selected example.
13:54:13,
.011
10.7 PROJECT ASSIGNMENT 7
Refine the learned reduction rules by assessing hypotheses that are similar to the ones
considered in the previous assignments.
10.1. Define the rule refinement problem.
10.2. What are the basic steps of rule refinement?
10.3. Consider the version space from Figure 10.19. In light of the refinement strategies
studied in this chapter, how will the plausible version space be changed as a result
of a new negative example labeled 1? Draw the new version space(s).
10.4. Consider the version space from Figure 10.20. In light of the refinement strategies
studied in this chapter, what are three alternative ways in which this version space
may be changed as a result of the negative example 2?
Upper Bound of the

Upper Bound of Except-When Condition
the Main Condition
Lower Bound of the

Lower Bound of
Except-When Condition
the Main Condition
. 1
Universe of
Instances
Figure 10.19. Version space and a negative example covered by the lower bound of the main
condition.
Upper Bound of the

Upper Bound of Except-When Condition
the Main Condition
Lower Bound of the

Lower Bound of
Except-When Condition
the Main Condition
. 2
Universe of
Instances
Figure 10.20. Version space and a negative example covered by the upper bound of the main
condition.
13:54:13,
.011
Positive Example 1
We need to
Determine a strategic center of gravity for a
member of Allied Forces 1943.
Explanation
Which is a member of Allied Forces 1943?
Allied Forces 1943 has as member US 1943
US 1943
Therefore we need to
Determine a strategic center of gravity for US 1943.
Figure 10.21. A reduction example and its explanation.
object
… force force goal
opposing force multimember force single-member force group
multigroup multistate single-state single-group illegal legal

force force force force group group
… … … …
multistate multistate
alliance coalition
dominant partner equal partners

multistate alliance multistate alliance
US 1943
UK 1943
Allied Forces 1943
has as member
European Axis 1943 Germany 1943
has as member
domain multimember force
Finland 1943 range force
Figure 10.22. Ontology fragment from the center of gravity analysis domain. Dotted links
indicate instance of relationships while continuous unnamed links indicate subconcept of
relationships.
10.5. (a) Consider the example and its explanation from Figure 10.21. What rule will
be learned from them, assuming the ontology from Figure 10.22, where all
the instances are considered specific instances?
(b) Consider the additional positive example from Figure 10.23. Indicate the
refined rule.
(c) Consider the negative example, its failure explanation, and the additional
ontological knowledge from Figure 10.24. Indicate the refined rule.
13:54:13,
.011
Positive Example 2
We need to
Determine a strategic center of gravity for a member of
European Axis 1943.
Which is a member of European Axis 1943?
Germany 1943
Therefore we need to
Determine a strategic center of gravity for Germany 1943.
Figure 10.23. An additional positive example.
Negative Example 2 Ontology Fragment

We need to contribution
Determine a center of gravity for a member of
European Axis 1943. subconcept of
Which is a member of European Axis 1943? military contribution
Finland 1943 subconcept of
Therefore we need to minor military intermediary military major military

Determine a center of gravity for Finland 1943. contribution contribution contribuon
Failure Explanation instance of

Finland 1943 has as military contribution has as military contribution
military contribution of Finland 1943 military contribution domain force
is minor military contribution of Finland 1943 range military contribution
Figure 10.24. Negative example, failure explanation, and additional ontology fragment.
war material and transports of US 1943
is a major generator of
industrial capacity of US 1943
has as industrial factor

US 1943
explains
IF the problem to solve is

Identify each strategic COG candidate with
respect to the industrial civilization of US 1943.
THEN
industrial capacity of US 1943 is a strategic
COG candidate for US 1943.
Figure 10.25. An example and its explanation.
10.6. Consider the example and its explanation shown in Figure 10.25. Find the plaus-
ible version space rule that will be learned based on the ontology fragments from
Figures 10.26, 10.27, and 10.28, where all the instances are defined as generic
instances.
13:54:13,
.011
is a major generator of economic factor

domain economic factor
range product
has as industrial factor
domain force
industrial commerce information
range industrial factor
factor authority network or system
raw material
industrial transportation
industrial farm implement center factor
authority industry
strategic raw
material industrial
farm implement capacity
industry of Italy 1943 transportation transportation
center network or system
oil chromium copper and
bauxite of Germany 1943
is critical to the production of
war material of Germany 1943 industrial capacity of US 1943

is obtained
from has as strategic
raw material industrial capacity of Germany 1943 is a major generator of
has as industrial factor has as industrial factor
Balkans
Germany 1943 US 1943 war material and transports of US 1943
Figure 10.26. Ontology of economic factors. Dotted links indicate instance of relationships while
continuous unnamed links indicate subconcept of relationships.
object
… …
force
group
multigroup force opposing force multistate force single-state force single-group force
US 1943 Germany 1943

component state component state
Allied Forces 1943 European axis 1943
component state component state
Britain 1943 Italy 1943
Figure 10.27. An ontology of forces. Dotted links indicate instance of relationships while continuous
unnamed links indicate subconcept of relationships.
10.7. Minimally generalize the rule from the left side of Figure 10.29 in order to cover
the positive example from the right side of Figure 10.29, considering the back-
ground knowledge from Figures 10.26, 10.27, and 10.28.
13:54:13,
.011
object
… …
resource or infrastructure element
resource
strategically essential resource

or infrastructure element raw material
product
strategic
raw material
strategically essential strategically essential non-strategically essential

infrastructure element goods or material goods or services
war material and fuel war material and transports farm implements
main airport main seaport

war material and fuel war material and farm implements
sole airport sole seaport of Germany 1943 transports of US 1943 of Italy 1943
Figure 10.28. An ontology of resources. Dotted links indicate instance of relationships while con-
tinuous unnamed links indicate subconcept of relationships.
Rule
IF Positive example that satisfies the upper bound

Identify each strategic COG candidate with respect
to the industrial civilization of ?O1. IF the task to accomplish is
Identify each strategic COG candidate with respect
Plausible Upper Bound Condition
to the industrial civilization of Germany 1943.
?O1 is force
has as industrial factor ?O2 THEN accomplish the task
?O2 is industrial factor industrial capacity of Germany 1943 is a strategic
is a major generator of ?O3 COG candidate for Germany 1943.
?O3 is product
Plausible Lower Bound Condition

?O1 is US 1943 Explanation
has as industrial factor ?O2 Germany 1943
?O2 is industrial capacity of US 1943 has as industrial factor
is a major generator of ?O3 industrial capacity of Germany 1943
?O3 is war material and transports of US 1943
industrial capacity of Germany 1943
THEN is a major generator of
?O2 is a strategic COG candidate for ?O1. war material and fuel of Germany 1943
Figure 10.29. Rule and positive example.
10.8. Minimally specialize the rule from the left side of Figure 10.30, in order to cover
the positive example from the right side of Figure 10.30, considering the back-
ground knowledge from Figures 10.26, 10.27, and 10.28.
13:54:13,
.011
Table 10.15 Rule Refinement with Learning Agent versus Rule Refinement by
Knowledge Engineer
Rule Refinement with Rule Refinement by

Learning Agent Knowledge Engineer
Description (highlighting
differences and similarities)
Strengths
Weaknesses
Rule
IF
Identify each strategic COG candidate with respect Negative example that satisfies the upper bound
to the industrial civilization of ?O1.
IF the task to accomplish is
Plausible Upper Bound Condition Identify each strategic COG candidate with
?O1 is force respect to the industrial civilization of Italy 1943.
has as industrial factor ?O2
?O2 is industrial factor THEN accomplish the task
is a major generator of ?O3 farm implement industry of Italy 1943 is a
?O3 is product strategic COG candidate for Italy 1943.
Plausible Lower Bound Condition

?O1 is single state force Explanation
has as industrial factor ?O2
Italy 1943
?O2 is industrial capacity has as industrial factor
is a major generator of ?O3 farm implement industry of Italy 1943
?O3 is strategically essential goods or material
farm implement industry of Italy 1943
THEN is a major generator of
?O2 is a strategic COG candidate for ?O1. farm implements of Italy 1943
Figure 10.30. Rule and negative example.
10.9. Consider the problem reduction step and its explanation from Figure 10.25, as
well as the ontology of economic factors from Figure 10.26. Show the correspond-
ing analogy criterion generated by a cautious learner, and an analogous reduction
made by that cautious learner.
10.10. Indicate five qualitatively different possible uses of an ontology in an instructable

agent such as Disciple-EBR.
10.11. Compare the learning-based rule refinement process discussed in this chapter
with the traditional knowledge acquisition approach discussed in Section 3.1.4 by
filling in Table 10.15. Identify similarities and differences and justify the relative
strengths and weaknesses of the two approaches.
13:54:13,
.011
Main Plausible Version Except-When Plausible

Space Condition Version Space Condition
I2
I3
I1
I4
Figure 10.31. A partially learned concept and several instances.
10.12. Consider the partially learned concept and the four instances from Figure 10.31.
Order the instances by the plausibility of being positive examples of this concept
and justify the ordering.
13:54:13,
.011

KE Unit 5 Notes

Uploaded by

Copyright:

Available Formats

KE Unit 5 Notes

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

KE Unit 5 Notes

Uploaded by

Copyright:

Available Formats

KNOWLEDGE ENGINEERING

LEARNING AND RULE LEARNING

8.1 INTRODUCTION TO MACHINE LEARNING

8.1.1 What Is Learning?

Acquire, discover, and organize knowledge by building, modifying, and organizing

Inference (e.g., deduction, induction, abduction, analogy)

The following are some of the most representative learning strategies:

8.1.2 Inductive Learning from Examples

Positive examples of cups: P1 P2 ...

Negative examples of cups: N1 N2 …

Description of the cup concept: has-handle(x), ...

Figure 8.1. Illustration of inductive concept learning from examples.

lif table(o1) stable(o1) open-vessel(o1) lif table(x) stable(x) open-vessel(x)

Figure 8.2. Illustration of explanation-based learning.

8.1.3 Explanation-based Learning

o1 is made of plastic, which is a light material. Therefore, o1 is light.

Being light and graspable, o1 is liftable.

8.1.4 Learning by Analogy

Source similar Target

A causes B and A is similar to A’.

mass Sun at tracts temp Sun mass nucleus at tracts

greater than greater than greater than

The Solar System The hydrogen atom

8.1.5 Multistrategy Learning

Table 8.1 Illustration of the Complementariness of the Learning Strategies

Inductive learning Explanation-based Multistrategy

Examples needed many one several

positive example concept descript ion

their relative weaknesses. Multistrategy learning is concerned with developing learning

8.2.1 Concepts, Examples, and Exceptions

A negative exception is a negative example that is covered by the description of the

8.2.2 Examples and Exceptions of a Partially Learned Concept

Plausible Version Space Plausible Version Space of

Figure 8.6. Examples and exceptions of a partially learned concept.

8.3 GENERALIZATION AND SPECIALIZATION RULES

It will be easy to demonstrate the generalization relationships between the preceding

Turning constants into variables

8.3.1 Turning Constants into Variables

E1 = ?O1 instance of professor [8.1]

E2 = ?O1 instance of professor [8.2]

8.3.2 Turning Occurrences of a Variable into Different Variables

E1 = ?O1 instance of paper [8.3]

8.3.3 Climbing the Generalization Hierarchies

E1 = ?O1 instance of assistant professor [8.5]

8.3.4 Dropping Conditions

E1 = ?O1 instance of assistant professor [8.7]

E2 = ?O1 instance of assistant professor [8.8]

8.3.5 Extending Intervals

E1 = ?O1 instance of professor [8.9]

number of publications ?N1

8.3.6 Extending Ordered Sets of Intervals

E1 = ?O1 instance of person [8.12]

8.3.7 Extending Symbolic Probabilities

8.3.8 Extending Discrete Sets

E1 = ?O1 instance of ﬂag [8.14]

infant toddler youth teen mature elder

8.3.9 Using Feature Deﬁnitions

E1 = ?O1 instance of professor [8.16]