04, G.Tecuci, Learning Agents Center CS 785 Fall 2004 Learning Agents Center and Computer Science Department George Mason University Gheorghe Tecuci [email protected]http://lac.gmu.edu /
CS 785 Fall 2004. Knowledge Acquisition and Problem Solving. Agent Teaching and Multistrategy Learning. Gheorghe Tecuci [email protected] http://lac.gmu.edu/. Learning Agents Center and Computer Science Department George Mason University. Overview. What is Machine Learning. - PowerPoint PPT Presentation
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
PowerPoint PresentationCS 785 Fall 2004
Overview
Demo: Agent Teaching and Rule Learning
The Rule Learning Problem in Disciple
Strategies for Explanation Generation
What is Learning?
acquire and organize knowledge (by building, modifying and
organizing internal representations of some external
reality),
discover new knowledge and theories (by creating hypotheses that
explain some data or phenomena), or
acquire skills (by gradually improving their motor or cognitive
skills through repeated practice, sometimes involving little or no
conscious thought).
Learning results in changes in the agent (or mind) that improve its
competence and/or efficiency.
Learning is a very general term denoting the way in which people
and computers:
2004, G.Tecuci, Learning Agents Center
The Disciple agent is concerned with the first type of learning:
acquiring and organizing knowledge from a subject matter expert (by
building, modifying and organizing internal representations of some
external reality). The external reality is a strategic scenario and
how a subject matter expert reasons to identify and test strategic
COG candidates for that scenario.
2004, G.Tecuci, Learning Agents Center
The architecture of an intelligent agent
Problem Solving
in the knowledge base.
Implements a general problem solving method that uses the knowledge
from the knowledge base to interpret the input and provide an
appropriate output.
Data structures that represent the objects from the application
domain,
general laws governing them, actions that can be performed with
them, etc.
Ontology
Rules/Cases/…
Overview
Demo: Agent Teaching and Rule Learning
The Rule Learning Problem in Disciple
Strategies for Explanation Generation
Generalization and specialization rules
A generalization rule is a rule that transforms an expression into
a more general expression.
A specialization rule is a rule that transforms an expression into
a less general expression.
The reverse of any generalization rule is a specialization
rule.
Fundamental to learning are the processes of generalization and
specialization.
We will present several basic rules for generalizing or
specializing expressions representing concepts. These rule are used
to generalize concepts or to specialize concepts.
2004, G.Tecuci, Learning Agents Center
Turning constants into variables
Generalizes an expression by replacing a constant with a
variable.
?O1 is multi_group_force
The set of multi_group_forces with any number of subgroups.
Allied_forces_operation_Husky
Axis_forces_Sicily
Japan_1944_Armed_Forces
2004, G.Tecuci, Learning Agents Center
The top expression represents the following concept: the set of
multi group forces with 5 subgroups. This set contains, for
instance, Axis_forces_Sicily from the Sicily_1943 scenario. By
replacing 5 with a variable ?N1 that can take any value, we
generalize this concept to the following one: the set of multi
group forces with any number of subgroups. In particular ?N1 could
be 5. Therefore the second concept includes the first one.
Conversely, by replacing ?N1 with 5, we specialize the bottom
concept to the top one.
The important thing to notice here is that by a simple syntactic
operation (transforming a number into a variable) we can generalize
a concept. This is how an agent generalizes concepts.
2004, G.Tecuci, Learning Agents Center
Climbing/descending the generalization hierarchy
Generalizes an expression by replacing a concept with a more
general one.
?O1 is single_state_force
The set of single state forces governed by representative
democracies
democratic_government representative_democracy
democratic_government
representative_democracy
parliamentary_democracy
2004, G.Tecuci, Learning Agents Center
One can also generalize an expression by replacing a concept from
its description with a more general concept, according to some
generalization hierarchy. The reverse operation, of replacing a
concept with a less general one, leads to the specialization of an
expression.
The agent can also generalize a concept by dropping a condition.
That is, by dropping a constraint that its instances must
satisfy.
This rule is illustrated in the next slide.
2004, G.Tecuci, Learning Agents Center
Dropping/adding condition
Generalizes an expression by removing a constraint from its
description.
?O1 is multi_member_force
The set of multi-member forces that have international
legitimacy.
The set of multi-member forces (that may or may not have
international legitimacy).
2004, G.Tecuci, Learning Agents Center
Generalizing/specializing numbers
Generalizes an expression by replacing a number with an interval,
or by replacing an interval with a larger interval.
?O1 is multi_group_force
The set of multi_group_forces with exactly 5 subgroups.
The set of multi_group_forces with at least 3 subgroups and at most
7 subgroups.
The set of multi_group_forces with at most 10 subgroups.
2004, G.Tecuci, Learning Agents Center
A concept may also be generalized by replacing a number with an
interval containing it, or by replacing an interval with a larger
interval. The reverse operations specialize the concept.
Yet another generalization rule, which is illustrated in the next
slide, is to add alternatives.
According to the expression from the top of this slide, ?O1 is any
alliance. Therefore this expression represents the following
concept: the set of all alliances.
This concept can be generalized by adding another alternative for
?O1, namely the alternative of being a coalition. Now ?O1 could be
either an alliance or coalition. Consequently, the expression from
the bottom of this slide represents the following more general
concept: the set of all alliances and coalitions.
2004, G.Tecuci, Learning Agents Center
Adding/removing alternatives
The set including both the alliances and the coalitions.
Generalizes an expression by replacing a concept C1 with the union
(C1 U C2), which is a more general concept.
2004, G.Tecuci, Learning Agents Center
Overview
Demo: Agent Teaching and Rule Learning
The Rule Learning Problem in Disciple
Strategies for Explanation Generation
Representative learning strategies
2004, G.Tecuci, Learning Agents Center
Empirical inductive learning from examples
The learning problem
• a set of positive examples (E1, ..., En) of a concept
• a set of negative examples (C1, ... , Cm) of the same
concept
• a learning bias
• other background knowledge
• a concept description which is a generalization of the
positive
examples that does not cover any of the negative examples
Purpose of concept learning:
Predict if an instance is an example of the learned concept.
2004, G.Tecuci, Learning Agents Center
Learning from examples
Compares the positive and the negative examples of a concept, in
terms of their similarities and differences, and learns the concept
as a generalized description of the similarities of the positive
examples. This allows the agent to recognize other entities as
being instances of the learned concept.
Requires many examples
Illustration:
Negative examples of cups: N1 …
Description of the cup concept: has-handle(x), ...
2004, G.Tecuci, Learning Agents Center
The goal of this learning strategy 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 this
illustration, 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 the same color is
encountered for both cups and non-cups.
To learn a good concept description through this learning strategy
requires a very large set of positive and negative examples. On the
other hand, this is the only information that the agent needs. That
is, the agent does not require 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 do
things it was not able to do before. In this illustration it will
learn to recognize cups, something that it was not able to do
before.
2004, G.Tecuci, Learning Agents Center
Explanation-based learning (EBL)
The EBL problem
has-handle(o1), has-flat-bottom(o1), up-concave(o1),...
(e.g. features present in the examples)
• BK: cup(x) Ü liftable(x), stable(x), open-vessel(x).
liftable(x) Ü light(x), graspable(x).
up-concave(x).
Explanation-based learning
cup(o1): color(o1, white), made-of(o1, plastic),
light-mat(plastic), has-handle(o1), has-flat-bottom(o1),
up-concave(o1),...
Learns to recognize more efficiently the examples of a concept by
proving that a specific instance is an example of it, and thus
identifying the characteristic features of the concept.
The proof identifies the characteristic features:
Proof generalization generalizes them:
• color(o1,white) is not needed to prove cup(o1)
• made-of(o1, plastic) is needed to prove cup(o1)
• made-of(o1, plastic) is generalized
cup(o1)
stable(o1)
liftable(o1)
graspable(o1)
light(o1)
light-mat(plastic)
made-of(o1,plastic)
2004, G.Tecuci, Learning Agents Center
The goal of this learning strategy is to improve the efficiency in
problem solving. The agent is able to perform some task but in an
inefficient way. We would like to teach the agent to perform the
task faster. Consider, for instance, an agent that is able to
recognize cups. The agent receives a description of a cup that
includes many features. The agent will recognize that this object
is a cup by performing a complex reasoning process, based on its
prior knowledge. This process is illustrated by the proof tree from
the left hand side of this slide. The object o1 is made of plastic
which is a light material. Therefore o1 is light. o1 has a handle
and therefore it is graspable. Being light and graspable, it is
liftable. And so on … being liftable, stable and an open vessel, it
is a cup. However, the agent can learn from this process to
recognize a cup faster. 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 learning from examples, but through a
different line of reasoning, and based on a different type of
information.
The next step in the learning process is to generalize the tree
from the left hand side into the tree from the right hand side.
While the tree from the left hand side proves that the specific
object o1 is a cup, the tree from the right hand side shows that
any object x that is made of some light material y, has a handle
and some other features is a cup. Therefore, to recognize that an
object o2 is a cup, the agent only needs to look for the presence
of these features discovered as important. It 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 to learn
from. 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.
2004, G.Tecuci, Learning Agents Center
• Needs only one example
General features of Explanation-based learning
2004, G.Tecuci, Learning Agents Center
Learning by analogy
• ACCESS: find a known entity that is analogous with the input
entity.
• MATCHING: match the two entities and hypothesize knowledge.
• EVALUATION: test the hypotheses.
• LEARNING: store or generalize the new knowledge.
Learns new knowledge about an input entity by transferring it from
a known similar entity.
The learning problem
The learning method
Learning by analogy: illustration
Illustration: The hydrogen atom is like our solar system.
Learning by analogy is the process of learning new knowledge about
some entity by transferring it from a known entity. For instance, I
can teach students about the structure of the hydrogen atom by
using the analogy with the solar system. I am telling the students
that the hydrogen atom has a similar structure with the solar
system, where the electrons revolve around the nucleus as the
planets revolve around the sun.
In this illustration, the ACCESS problem is already solved. I have
already told the students what the source is (the solar
system).
2004, G.Tecuci, Learning Agents Center
Learning by analogy is the process of learning new knowledge about
some entity by transferring it from a known entity.
For instance, I can teach students about the structure of the
hydrogen atom by using the analogy with the solar system. I am
telling the students that the hydrogen atom has a similar structure
with the solar system, where the electrons revolve around the
nucleus as the planets revolve around the sun.
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, we may
conclude that this 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 sun. This is indeed true
and represents a very interesting discovery.
The main problem with analogical reasoning is that not all the
facts related to the solar system are true for the hydrogen atom.
For instance, the sun is yellow, but the nucleus is not. Therefore,
facts derived by analogy have to be verified.
A general heuristic is that similar causes have similar effects.
That is, if A is similar to A’ and A causes B. Then we would expect
A’ to cause B’ which should be similar to B.
2004, G.Tecuci, Learning Agents Center
Learning by analogy: illustration
Illustration: The hydrogen atom is like our solar system.
The Sun has a greater mass than the Earth and attracts it, causing
the Earth to revolve around the Sun.
The nucleus also has a greater mass then the electron and attracts
it.
Therefore it is plausible that the electron also revolves around
the nucleus.
General idea of analogical transfer: similar causes have similar
effects.
The main problem with analogical reasoning is that not all the
facts related to the solar system are true for the hydrogen atom.
Therefore, facts derived by analogy have to be verified.
A general heuristic is that similar causes have similar effects.
That is, if A is similar to A’ and A causes B. Then we would expect
A’ to cause B’ which should be similar to B.
2004, G.Tecuci, Learning Agents Center
Multistrategy learning
Multistrategy learning is concerned with developing learning agents
that synergistically integrate two or more learning strategies in
order to solve learning tasks that are beyond the capabilities of
the individual learning strategies that are integrated.
2004, G.Tecuci, Learning Agents Center
Examples
Learning
The individual learning strategies have complementary strengths and
weaknesses. For instance learning from example requires a lot of
example while explanation-based learning requires only one example.
On the other hand, learning from examples does not require any
prior knowledge while explanation-based learning requires a lot of
prior knowledge.
Multistrategy learning attempts to synergistically integrate such
complementary learning strategies, in order to take advantage of
their relative strengths to compensate for their relative
weaknesses.
The Disciple agent uses a multistrategy learning strategy, as will
be presented in the following.
2004, G.Tecuci, Learning Agents Center
Overview
Demo: Agent Teaching and Rule Learning
The Rule Learning Problem in Disciple
Strategies for Explanation Generation
GIVEN:
• an example of a problem solving episode;
• a knowledge base that includes an object ontology and a set of
problem solving rules;
• an expert that understands why the given example is correct and
may answer agent’s questions.
DETERMINE:
generalization of the specific problem solving episode.
2004, G.Tecuci, Learning Agents Center
Input example
Which is a member of Allied_Forces_1943?
I need to
Therefore I need to
Identify and test a strategic COG candidate corresponding to a
member of the Allied_Forces_1943
This is an example of a problem solving step from which the agent
will learn a general problem solving rule.
2004, G.Tecuci, Learning Agents Center
IF
Identify and test a strategic COG candidate corresponding to a
member of a force
The force is ?O1
Identify and test a strategic COG candidate for a force
The force is ?O2
Plausible Upper Bound Condition
IF
Identify and test a strategic COG candidate corresponding to a
member of the ?O1
Question
Answer
?O2
THEN
INFORMAL STRUCTURE OF THE RULE
FORMAL STRUCTURE OF THE RULE
2004, G.Tecuci, Learning Agents Center
This is the rule that is learned from the input example. It has
both a formal structure (used for formal reasoning), and an
informal structure (used to communicate more naturally with the
user).
Let us consider the formal structure of the rule. This is an
IF-THEN structure that specifies the condition under which the task
from the IF part can be reduced to the task from the THEN part.
This rule, however, is only partially learned. Indeed, instead of a
single applicability condition, it has two conditions:
1) a plausible upper bound condition which is more general than the
exact (but not yet known) condition, and
2) a plausible lower bound condition which is less general than the
exact condition.
Completely learning the rule means learning an exact condition.
However, for now we will show how the agent learns this rule from
the input example shown on a previous slide. The basic steps of the
learning method are those from the next side.
2004, G.Tecuci, Learning Agents Center
Overview
Demo: Agent Teaching and Rule Learning
The Rule Learning Problem in Disciple
Strategies for Explanation Generation
Example of a
Basic steps of the rule learning method
3. Generalize the example and the explanation into a plausible
version space rule.
1. Formalize and learn the tasks
2. Find a formal explanation of why the example is correct.
This explanation is an approximation of the question and the
answer, in the object ontology.
2004, G.Tecuci, Learning Agents Center
1. Formalize the tasks
I need to
Therefore I need to
Identify and test a strategic COG candidate corresponding to a
member of the Allied_Forces_1943
Identify and test a strategic COG candidate for a force
The force is US_1943
Therefore I need to
Identify and test a strategic COG candidate corresponding to a
member of a force
The force is Allied_Forces_1943
Sample formalization rule:
obtain the task name by replacing each specific instance with a
more general concept
for each replaced instance define a task feature of the form “The
concept is instance”
2004, G.Tecuci, Learning Agents Center
Because the tasks from the modeling are in unrestricted English
Disciple cannot reason with them. We need to formalize these
tasks.
For each task we need to define an abstract phrase that indicates
what this task is about (the task name), and a list of specific
phrases that give all the details about the task (the task
features). The task name should not contain any instance (such as
Allied_Forces_1943). All these instances should appear in the task
features. In general, the task name may be obtained from the
English expression in the left hand side by simply replacing each
specific object with a more abstract concept. Then we will add a
corresponding task feature that specifies the value for this
abstract concept.
2004, G.Tecuci, Learning Agents Center
Identify the strategic COG candidates
for the Sicily_1943 scenario
for Anglo_allies_1943
in the Sicily_1943 scenario?
Identify the strategic COG candidates for
the Anglo_allies_1943 which is
for a scenario
for an opposing force
Identify the strategic COG candidates for
an opposing force which is
a multi-member force
Any other formalization is acceptable if:
the task name does not contain any instance or constant;
each instance from the informal task appears in a feature of the
formalized task.
These are some other examples of task formalizations.
2004, G.Tecuci, Learning Agents Center
Task learning
Identify and test a strategic COG candidate for ?O1
INFORMAL STRUCTURE OF THE TASK
Identify and test a strategic COG candidate for US_1943
Identify and test a strategic COG candidate for a force
The force is US_1943
Identify and test a strategic COG candidate for a force
The force is ?O1
2004, G.Tecuci, Learning Agents Center
The top part of this slide shows the English expression and the
formalized expression of a specific task.
From the English expression of the specific task the agent learns
the informal structure of the general task by replacing the
specific instance US_1943, with the variable ?O1.
From the formalized expression of the specific task, the agent
learns the formal structure of the general task. The formal
structure also specifies the conditions that ?O1 should satisfy.
However, the agent cannot formulate the exact condition, but only
two bounds for the exact condition that will have to be learned.
The plausible lower bound condition is more restrictive, allowing
?O1 to only be a single-state force. This condition is obtained by
replacing US_1943 with its most specific generalization in the
object ontology.
The plausible upper bound condition is less restrictive. ?O1 could
be any force. This condition is obtained by replacing US_1943 with
the most general sub-concept of <object> which is more
general than US_1943. The plausible upper bound condition allows
the agent to generate more tasks, because now ?O1 can be replaced
with any instance of force. However, there is no guarantee that the
generated task is a correct expression.
The agent will continue to improve the learned task, generalizing
the plausible lower bound condition and specializing the plausible
upper bound condition until they become identical and each object
that satisfies the obtained condition leads to a correct task
expression.
2004, G.Tecuci, Learning Agents Center
2. Find an explanation of why the example is correct
US_1943
has_as_member
Allied_Forces_1943
The explanation is an approximation of the question and the answer,
in the object ontology.
US_1943
Which is a member of Allied_Forces_1943?
I need to
Therefore I need to
Identify and test a strategic COG candidate corresponding to a
member of the Allied_Forces_1943
2004, G.Tecuci, Learning Agents Center
The expert has defined the example during the modeling
process.
During the task formalization process, the expert and the agent
have collaborated to formalize the tasks. Now the expert and the
agent have to collaborate to also formalize the question and the
answer. This formalization is the explanation from the bottom of
this slide. It consists of a relation between two elements from the
agent's ontology:
“Allied_Forces_1943 has_as_member US_1943”
It states, in Disciple’s language, that US_1943 is a member of
Allied_Forces_1943.
An expert can understand such formal expressions because they
actually correspond to his own explanations. However, he cannot be
expected to be able to define them because he is not a knowledge
engineer. For one thing, he would need to use the formal language
of the agent. But this would not be enough. He would also need to
know the names of the potentially many thousands of concepts and
features from the agent’s ontology (such as “has_as_member”).
While defining the formal explanation of this task reduction step
is beyond the individual capabilities of the expert and the agent,
it is not beyond their joint capabilities. Finding such explanation
pieces is a mixed-initiative process involving the expert and the
agent. In essence, the agent will use analogical reasoning and help
from the expert to identify and propose a set of plausible
explanation pieces from which the expert will have to select the
correct ones.
Once the expert is satisfied with the identified explanation
pieces, the agent will generate a general rule.
2004, G.Tecuci, Learning Agents Center
Identify and test a strategic COG candidate corresponding to a
member of a force
The force is Allied_Forces_1943
Identify and test a strategic COG candidate for a force
The force is US_1943
Condition
IF
Identify and test a strategic COG candidate corresponding to a
member of a force
The force is ?O1
Identify and test a strategic COG candidate for a force
The force is ?O2
Plausible Upper Bound Condition
2004, G.Tecuci, Learning Agents Center
Notice that the explanation is first re-written as a task
condition, and then two generalizations of this condition are
created: a most conservative one (the plausible lower bound
condition) and a most aggressive one (the plausible upper bound
condition).
The plausible lower bound is the minimal generalization of the
condition from the left hand side of the slide.
Similarly, the most general generalization is the plausible upper
bound.
2004, G.Tecuci, Learning Agents Center
Analogical reasoning
similar example
Identify and test a strategic COG candidate for a force
The force is Germany_1943
Therefore I need to
Identify and test a strategic COG candidate corresponding to a
member of a force
The force is European_Axis_1943
Identify and test a strategic COG candidate for a force
The force is US_1943
Therefore I need to
Identify and test a strategic COG candidate corresponding to a
member of a force
The force is Allied_Forces_1943
2004, G.Tecuci, Learning Agents Center
The agent uses analogical reasoning to generalize the example and
its explanation into a plausible version space rule. This slide
provides a justification for the generalization procedure used by
the agent.
Let us consider that the expert has provided to the agent the task
reduction example from the bottom left of this slide.
This reduction is correct because “Allied_Forces_1943 has_as_member
US_1943”.
Now let us consider the European_Axis_1943 which has as member
Germany_1943. Using the same logic as above, one can create the
task reduction example from the bottom right of the slide.
This is a type of analogical reasoning that the agent performs. The
explanation from the left hand side of this slide explains the task
reduction from the left hand side. This explanation is similar with
the explanation from the right hand side of this slide (they have
the same structure, being both less general than the analogy
criterion from the top of this slide). Therefore one could expect
that this explanation from the right hand side of the slide would
explain an example that would be similar with the initial example.
This example is the one from the right hand side of the
slide.
To summarize: The expert provided the example from the left hand
side of this slide and helped the agent to find its explanation.
Using analogical reasoning the agent can perform by itself the
reasoning from the bottom right hand side of the slide.
2004, G.Tecuci, Learning Agents Center
Generalization by analogy
DOMAIN(has_as_member) RANGE(The force is) =
RANGE(has_as_member) RANGE(The force is) = force force =
force
Knowledge-base constraints on the generalization:
initial example
Identify and test a strategic COG candidate for a force
The force is US_1943
Therefore I need to
Identify and test a strategic COG candidate corresponding to a
member of a force
The force is Allied_Forces_1943
Identify and test a strategic COG candidate for a force
The force is ?O2
Therefore I need to
Identify and test a strategic COG candidate corresponding to a
member of a force
The force is ?O1
2004, G.Tecuci, Learning Agents Center
Notice that in the previous illustration we could have used any
other forces ?O1 and ?O2 instead of European_Axis_1943 and
Germany_1943. As long as ?O1 has as member ?O2, the agent would
hypothesize that in order to identify and test a strategic COG for
?O1 one could identify and test a strategic COG for ?O2.
The agent uses various constraints from the knowledge base to
restrict the values that the variables ?O1 and ?O2 could take. For
instance, ?O1 should have the feature “has_as_member” and the
domain of this feature (i.e. the set of objects that may have this
feature) is multi_member_force. Therefore ?O1 should be a
multi_member_force.
Also, ?O1 is the value of the task feature “The force is” the range
of which is “force”. Therefore ?O1 should also be a force. From
these two restrictions, we conclude that ?O1 should be a
multi_member_force.
Using this kind of reasoning, the agent generalizes the example
from the left hand side of this slide to the expression from the
right hand side of this slide.
2004, G.Tecuci, Learning Agents Center
Universe of
Characterization of the learned rule
2004, G.Tecuci, Learning Agents Center
As discussed previously, the plausible upper bound condition of the
learned rule is an analogy criterion that allows the agent to solve
problems by analogy with the example from which the rule was
learned. Because analogy is only a plausible reasoning process,
some of the examples covered by the rule may be wrong. The
plausible upper bound of the rule is therefore only an
approximation of a hypothetical exact condition that will cover
only positive examples of the rule. That is why it is called
plausible upper bound.
The plausible lower bound condition of the rule covers the input
example that is known to be correct. However, the bound is a
minimal generalization performed in the context of an incomplete
ontology (generalization hierarchy). Therefore it is also a
plausible bound.
The previous slide shows the most likely relation between the
plausible lower bound, the plausible upper bound and the
hypothetical exact condition of the rule. Notice that there are
instances of the plausible upper bound that are not instances of
the hypothetical exact condition of the rule. This means that the
learned rule could also generate wrong solutions to some problems,
as already mentioned. Also, there are instances of the hypothetical
exact condition that are not instances of the plausible upper
bound. This means that the plausible upper bound does not cover all
the cases in which the solution provided by the rule would be
correct.
Similarly, there may be cases that are covered by the plausible
lower bound, without being covered by the hypothetical exact
condition. All these situations are a consequence of the fact that
the explanation of the initial example might be incomplete, and
that the representation language for learning (which is based on
the object ontology) might also be incomplete. These results are
consistent with what one would expect from an agent performing
analogical reasoning.
2004, G.Tecuci, Learning Agents Center
Overview
Demo: Agent Teaching and Rule Learning
The Rule Learning Problem in Disciple
Strategies for Explanation Generation
Input example
THEN
Identify the strategic COG candidates with respect to the
industrial civilization of US_1943
Who or what is a strategically
critical industrial civilization
element in US_1943?
Industrial_capacity_of_US_1943
This slide shows an example of a problem solving step from which
the agent will learn a rule. The domain used for illustration is
Center of Gravity analysis.
2004, G.Tecuci, Learning Agents Center
IF
Identify the strategic COG candidates with respect to the
industrial civilization of a force
The force is ?O1
THEN
A strategic COG relevant factor is strategic COG candidate for a
force
The force is ?O1
Plausible Upper Bound Condition
to the industrial civilization of ?O1
Question
civilization element in ?O1 ?
INFORMAL STRUCTURE OF THE RULE
FORMAL STRUCTURE OF THE RULE
This is the rule that is learned from the previous example. It has
both a formal structure (used for formal reasoning), and an
informal structure (used to communicate more naturally with the
user). Let us consider the formal structure of the rule. This is an
IF-THEN structure that specifies the condition under which the task
from the IF part can be reduced to the task from the THEN part.
This rule, however, is only partially learned. Indeed, instead of a
single applicability condition, it has two conditions: 1) a
plausible upper bound condition which is more general than the
exact (but not yet known) condition, and 2) a plausible lower bound
condition which is less general than the exact condition.
Completely learning the rule means learning this exact condition.
However, for now we will show how the agent learns the rule in this
slide from the example shown on the previous slide.
2004, G.Tecuci, Learning Agents Center
Explanation of the example
Identify the strategic COG candidates with respect to the
industrial civilization of a force
The force is US_1943
A strategic COG relevant factor is strategic COG candidate for a
force
The force is US_1943
explanation:
critical industrial civilization
element in US_1943?
Identify the strategic COG candidates with respect to the
industrial civilization of US_1943
2004, G.Tecuci, Learning Agents Center
The expert has defined the example during the modeling process. The
left hand side of this slide shows how the expert reasons using
task reduction. All the expressions are in natural language. The
agent, however, does not reason with natural language. It needs
formal logical expressions, like the ones shown in the right hand
side of this slide.
During the task formalization process, the expert and the agent
have collaborated to formalize the tasks. Now the expert and the
agent have to collaborate to also formalize the question and the
answer. This formalization is the explanation from the right hand
side of this slide. It consists of various relations between
certain elements from the agent's ontology:
US_1943 has_as_industrial_factor
Industrial_capacity_of_US_1943
Industrial_capacity_of_US_1943 is_a_major_generator_of
War_materiel_and_transports_of_US_1943
They state, in Disciple’s language, that US_1943 has as industrial
factor its industrial capacity, which is a major generator of war
materiel and transports.
An expert can understand these formal expressions because they
actually correspond to his own explanations. However, he cannot be
expected to be able to define them because he is not a knowledge
engineer. For one thing, he would need to use the formal language
of the agent. But this would not be enough. He would also need to
know the names of the potentially many thousands of concepts and
features from the agent’s ontology (such as
“is_a_major_generator_of”).
While defining the formal explanation of this task reduction step
is beyond the individual capabilities of the expert and the agent,
it is not beyond their joint capabilities. Finding these
explanation pieces is a mixed-initiative process involving the
expert and the agent. In essence, the agent will use analogical
reasoning and help from the expert to identify and propose a set of
plausible explanation pieces from which the expert will have to
select the correct ones.
Once the expert is satisfied with the identified explanation
pieces, the agent will generate a general rule.
2004, G.Tecuci, Learning Agents Center
What is the form of the explanation?
The explanation is a sequence of object relationships that
correspond to fragments of the object ontology.
explanation:
2004, G.Tecuci, Learning Agents Center
General heuristics for explanation generation
Look for the relationships between the objects from the question
and the answer
Look for the relationships between an object from the IF task and
an object from the question or the answer
2004, G.Tecuci, Learning Agents Center
War_materiel_and_transports_of_US_1943 …
IF
THEN
Identify the strategic COG candidates with respect to the
industrial civilization of US_1943
Who or what is a strategically
critical industrial civilization
element in US_1943?
The expert selects an object from the example.
The agent generates a list of plausible explanations containing
that object.
The expert selects the correct explanation(s).
Industrial_capacity_of_US_1943 IS …
The agent may need help to propose additional necessary
explanations.
For instance, the previously found explanation
“US_1943 -- has_as_industrial_factor -->
Industrial_capacity_of_US_1943”
does not express the fact that Industrial_capacity_of_US_1943 is a
strategically critical element of US_1943. To help the agent to
generate an formal explanation that expresses this fact, the expert
may provide two types of hint.
One hint is to select one object from the example. In this case the
object would be Industrial_capacity_of_US_1943. The agent will then
generate plausible explanations containing this object, as
illustrated in this slide. The expert will then need to select the
correct explanation, as shown in this slide.
Another type of hint is to select two objects from the example. In
this case the agent will show the relationships between these two
objects. For instance, if the agent would have not generated the
explanation
“US_1943 -- has_as_industrial_factor -->
Industrial_capacity_of_US_1943”,
the expert could have simply selected these objects. Then the agent
would have generated the above explanation.
2004, G.Tecuci, Learning Agents Center
IF
THEN
a multi-member force?
for Anglo_allies_1943
a multi-member force
The expert provides a hint: Anglo_allies_1943
The agent generates a list of plausible explanations containing
that object.
The expert selects one expression that ends in “…” and clicks on
EXPAND:
Anglo_allies_1943 IS …
Anglo_allies_1943 --- component_state US_1943 …
The agent generates the list of possible expansions of the
expression and the experts select the explanation:
Anglo_allies_1943 IS Strategic_COG_relevant_factor
Anglo_allies_1943 IS Force
Anglo_allies_1943 IS Military_factor
Anglo_allies_1943 IS Multi_member_force
Anglo_allies_1943 IS Multi_state_force
Anglo_allies_1943 IS Multi_state_alliance
Anglo_allies_1943 IS Equal_partners_multi_state_alliance
2004, G.Tecuci, Learning Agents Center
In this example, according to the question and the answer from the
upper left of the slide, the explanation should express the fact
that Anglo_allies_1943 is a multi member force. The expert first
gives a hint to the agent by selecting “Anglo_allies_1943”. Then
the agent generates several explanations that start with
Anglo_allies_1943.
In order to reduce the number of plausible explanations that the
expert has to analyze, the agent may generate only the beginning of
a plausible explanation, followed by “…” The expert may select such
an expression and my ask the agent to expand it, as illustrated in
this slide.
Indeed, the expert selected “Anglo_allies_1943 IS …”, because these
are the explanations that express the fact that Anglo_allies_1943
is of a certain type.
Then the expert requests the agent to expand this explanation and
the agent lists all the types of Anglo_allies_1943. After that the
expert selects the correct explanation.
The next slide also illustrates the hint refinement process, this
time for a different type of explanation.
It is important to stress that the hint may also be an object from
a previously selected explanation, not necessarily from the initial
example.
2004, G.Tecuci, Learning Agents Center
Hint refinement (another example)
has_as_political_leader
PM_Churchill
has_as_ruling_political_party
Conservative_party
Hint:
Generate:
Britain_1943 -- has_as_governing_body
governing_body_of_Britain_1943 --
IF
Identify the strategic COG candidates with respect to the
industrial civilization of a force
The force is ?O1
THEN
A strategic COG relevant factor is strategic COG candidate for a
force
The force is ?O1
Plausible Upper Bound Condition
to the industrial civilization of ?O1
Question
civilization element in ?O1 ?
INFORMAL STRUCTURE OF THE RULE
FORMAL STRUCTURE OF THE RULE
Using this form of generalization based on analogy, the agent
generalizes the example provided by the expert and its explanation
to this plausible version space rule. Notice that the IF task and
the THEN task are those from the example, where the instances have
been replaced with variables. The lower bound condition of this
rule requires the variable to take only the values from the example
(for instance, ?O1 should be US_1943). We know that for these
values of the variables the task reduction is correct (it is the
task reduction indicated by the expert).
However, the plausible upper bound condition of this rule allows
the variables to take any value that satisfies the analogy
criterion (for instance, ?O1 could also take the value
Germany_1943). The task reduction generated in this way may
sometimes be correct and sometimes may be wrong, and has to be
confirmed or rejected by the expert. In any case the agent will
learn from the expert as will be explained in the following.
Basically the lower bound will be generalized to cover the examples
that are correct. Also the upper bound will be specialized to
eliminate the examples that are not correct. In this way the two
bounds will converge toward one another until they will become
identical.
2004, G.Tecuci, Learning Agents Center
Analogical reasoning heuristic
1. Look for a rule Rk that reduce the current task T1.
2. Extract the explanations Eg from the rule Rk.
3. Look for explanations of the current task reduction that are
similar with Eg.
Example to be explained:
Explanation Eg
PUB condition
PLB condition
THEN accomplish T1a,…T1d
Previously learned rule Rk:
2004, G.Tecuci, Learning Agents Center
Justification of the heuristic
This heuristic is based on the observation that the explanations of
the alternative reductions of a task tend to have similar
structures. The same factors are considered, but the relationships
between them are different.
T1
T1a
T1b
T1e
Ab
Aa
Ae
Question:
Answers:
Explanations:
Eb
Ea
Ee
Q
Another analogical reasoning heuristic
1. Look for a rule Rk that reduce a similar task to similar
subtasks.
2. Extract the explanations Eg from the rule Rk.
3. Look for explanations of the current task reduction that are
similar with Eg.
Another heuristic is to generate explanations of a task reduction
by analogy with the explanations corresponding to the reduction of
a similar task.
2004, G.Tecuci, Learning Agents Center
This heuristic is based on the observation that similar problem
solving episodes tend to have similar explanations:
Justification of the heuristic
2004, G.Tecuci, Learning Agents Center
Yet another analogical reasoning heuristic
The plausible explanations found by the agent can be ordered by
their plausibility (based on the heuristics used).
1. Look for a rule Rk that reduces a task that is similar to the
current task even if the subtasks are not similar.
2. Extract the explanations Eg from the rule Rk.
3. Look for explanations of the current task reduction that are
similar with Eg.
The goal of using these heuristics is to have the agent propose
explanations ordered by their plausibility with the expert choosing
the right ones, rather than requiring the expert to define
them.
2004, G.Tecuci, Learning Agents Center
No explanation necessary
candidates should I
consider for Britain_1943?
I consider strategic
Anglo_allies_1943
with respect to the governing element
of Britain_1943, a member of
Anglo_allies_1943
Sometimes no formal explanation is necessary. In this example, for
instance, each time I want to identify the strategic COG candidate
for a state, such as Britain, I would like to also consider the
candidates with respect the governing element of this state.
We need to invoke Rule Learning, but then quit it without selecting
any explanation. The agent will generalize this example to a
rule.
2004, G.Tecuci, Learning Agents Center
Overview
Demo: Agent Teaching and Rule Learning
The Rule Learning Problem in Disciple
Strategies for Explanation Generation
Disciple-RKF/COG:
Agent Teaching and Rule Learning
2004, G.Tecuci, Learning Agents Center
We will now demonstrate how Disciple learns general tasks and rules
from the expert’s reasoning.
2004, G.Tecuci, Learning Agents Center
First the expert and Disciple have to formalize the English
statements of the tasks.
This is done in the Formalization mode.
2004, G.Tecuci, Learning Agents Center
In the formalization mode the tool shows:
The modeling
in English
The formalized
2004, G.Tecuci, Learning Agents Center
When the expert clicks on “Formalize” Disciple will propose a
formalization of the task
2004, G.Tecuci, Learning Agents Center
The expert may accept it or
he may edit it
2004, G.Tecuci, Learning Agents Center
After the tasks from the current task reduction step are formalized
the expert may explain this example to Disciple, which will learn a
rule from it
2004, G.Tecuci, Learning Agents Center
The Rule Learning tool helps the expert to explain the example to
Disciple
The tool shows the English form of the example
Disciple uses analogical reasoning and other heuristics to propose
plausible explanations pieces that justify the task reduction
step.
The expert selects those explanation pieces that correspond to the
meaning of the question-answer pair from the task reduction example
and clicks on “Accept”
2004, G.Tecuci, Learning Agents Center
The expert may direct Disciple to generate explanation pieces
related to certain objects from the example
2004, G.Tecuci, Learning Agents Center
The expert may direct Disciple to generate explanation pieces
related to certain objects from the example
2004, G.Tecuci, Learning Agents Center
2004, G.Tecuci, Learning Agents Center
When the expert is satisfied with the identified explanation pieces
he can click on “End learning”. Disciple will then create a general
rule corresponding to this example and its explanation.
2004, G.Tecuci, Learning Agents Center
During rule refinement the two conditions will converge toward one
another, ultimately leading to a rule with a single
condition.
Notice that it has a plausible upper bound condition and a
plausible lower bound condition.
This is the general task reduction rule learned by Disciple.
2004, G.Tecuci, Learning Agents Center
This is the general task reduction rule learned by Disciple.
In addition to the formal structure of the rule, which is used in
problem solving and learning, Disciple maintains also an informal
structure of the rule.
The informal structure is used in the communication with the
user.
2004, G.Tecuci, Learning Agents Center
After learning a rule from the current task reduction step Disciple
returns to the formalization mode.
2004, G.Tecuci, Learning Agents Center
Following the same procedure, Disciple will learn another rule from
the this task reduction step:
The expert selects the next task reduction step by clicking on the
corresponding question-answer pair
2004, G.Tecuci, Learning Agents Center
2004, G.Tecuci, Learning Agents Center
When the expert clicks on “End learning” Disciple analyzes the
accepted explanations to determine whether additional explanations
are needed.
If Disciple determines that additional explanations are needed it
will asks the expert to provide them
2004, G.Tecuci, Learning Agents Center
2004, G.Tecuci, Learning Agents Center
After learning a new rule from the current task reduction step
Disciple returns to the formalization mode.
2004, G.Tecuci, Learning Agents Center
Following the same procedure, Disciple learned another rule from
the last task reduction step:
2004, G.Tecuci, Learning Agents Center
Recommended reading
Tecuci G., Building Intelligent Agents: A Theory, Methodology, Tool
and Case Studies, Academic Press, 1998.
Tecuci G., Boicu M., Marcu D., Stanescu B., Boicu C. and Comello
J., Training and Using Disciple Agents: A Case Study in the
Military Center of Gravity Analysis Domain, in AI Magazine, AAAI
Press, Menlo Park, California, 2002.
http://lalab.gmu.edu/publications/default.htm
Tecuci G., Boicu M., Bowman M., and Marcu M., with a commentary by
Burke M.: An Innovative Application from the DARPA Knowledge Bases
Programs: Rapid Development of a High Performance Knowledge Base
for Course of Action Critiquing, in AI Magazine, 22, 2, 2001, pp.
43-61. AAAI Press, Menlo Park, California, 2001.
http://lalab.gmu.edu/publications/default.htm Describes the course
of action domain.
sun
planet
yellow
mass
mass
temperature
greater
color
revolves
à