Knowledge Representation Techniques

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

KNOWLEDGE REPRESENTATION TECHNIQUES

By Captain Wilfried Honekamp

March 2004
CONTENTS i

CONTENTS

1 INTRODUCTION............................................................................................................ 1

2 KNOWLEDGE REPRESENTATION TECHNIQUES ............................................... 2

2.1 SEMANTIC NETS .......................................................................................................... 2


2.2 FRAMES AND SCRIPTS .................................................................................................. 2
2.3 LOGIC .......................................................................................................................... 3
2.4 FACTS AND RULES ....................................................................................................... 3
2.5 STATISTIC .................................................................................................................... 4
2.6 NEURAL NETWORKS ................................................................................................... 5
2.7 MIXED APPROACHES ................................................................................................... 6

3 CONCLUSIONS............................................................................................................... 7

4 BIBLIOGRAPHY ............................................................................................................ 8
LIST OF FIGURES ii

LIST OF FIGURES

Figure 1: A Semantic Net........................................................................................................... 2


Figure 2: Two frames describing a human being ....................................................................... 2
Figure 3: Shank’s famous restaurant script ................................................................................ 3
Figure 4: Adjustable Rule Set .................................................................................................... 4
Figure 5: A Bayes Belief Network ............................................................................................. 5
Figure 6: Simplified illustration of a neural network ................................................................. 5
Figure 7: Parameters of a perceptron ......................................................................................... 6
INTRODUCTION 1

1 INTRODUCTION
Knowledge Representation is the method used to encode knowledge in Intelligent Systems.
An Intelligent System is an engineered system which has a purpose and uses techniques of
artificial intelligence to fulfil its task. This essay is about advantages and drawbacks of
knowledge representation techniques. Therefore in the next chapter a variety of possible tech-
niques is described and assessed. In the last chapter then the results are summarised and some
conclusions are drawn.
KNOWLEDGE REPRESENTATION TECHNIQUES 2

2 KNOWLEDGE REPRESENTATION TECHNIQUES


2.1 Semantic Nets
In Semantic Nets the knowledge is represented by a Directed Acyclic Graph (DAG). In-
stances are represented as nodes and relationships, influences or dependencies are depicted as
Arrows. Figure 1 depicts a simple Semantic Net.

Figure 1: A Semantic Net1

All features are inheritable, thus an emu is not only a bird but also has wings and can fly. This
eases the implementation in modern developing environments, above all if these are object
oriented. The modular structure and the coherence make Semantic Nets simple to be modified
and reused. As it is a graphical method it is easy to understand. The method is especially suit-
able for binary relationships. One drawback is the lack of standards for the links between in-
stances. Complex relationships and structures can only be depicted with huge graphs. Uncer-
tainty and incomplete information can hardly be represented. The expenditure of work to ex-
press relationships is relatively high and the visual appeal can be misleading in case of the
representation of a node. They can be concepts, classes and nodes.

2.2 Frames and scripts


In frames instances are represented as classes with slots of attributes. These attributes can be
linked to other frames to depict influences, dependencies or connections. Frames use inheri-
tance like Semantic Nets. An example of connected frames is given in Figure 2.

Figure 2: Two frames describing a human being

1
Taken from: Marshall, Dave: Artificial Intelligence 2 COURSEWARE, Lecture notes + integrated exercises,
solutions and marking, 2004
KNOWLEDGE REPRESENTATION TECHNIQUES 3

Scripts describe a particular process in simple steps. Figure 3 depicts the probably most fa-
mous script of someone having a meal at a restaurant.

THE RESTAURANT SCRIPT


1) Actor goes to a restaurant.
2) Actor is seated.
3) Actor orders a meal from waiter.
4) Waiter brings the meal to actor.
5) Actor eats the meal.
6) Actor gives money to the restaurant.
7) Actor leaves the restaurant.
Figure 3: Shank’s famous restaurant script2

Frames and scripts are appropriate for the representation of stereotypical knowledge. They
can depict hierarchies by connections and uncertainty by unfilled slots. Slots in frames and
parts of scripts can be inserted and deleted which provides high flexibility. Default values can
be used to represent common sense. The execution of frames and scripts is relatively efficient.
On the other hand the precise structure is hard to select to achieve an optimal performance.
Furthermore complex concepts, influences and dependencies are difficult to represent.

2.3 Logic
All varieties of logic base on the binary or propositional logic knowing only the states true
and false. These can be combined by operators like or, and, not or Implication. An extension
of the propositional logic is the predicate calculus which includes variables, functions and
quantifiers. A further extension is the Fuzzy Logic which uses ranges of values, so called
Fuzzy Sets, to calculate a determined value with blurry information. Logic often seems a
natural way to express certain notions, it is precise flexible and modular. That is why pro-
gramming languages and Expert System Shell are often based on logic. The major drawback
is the separation of representing and processing.

2.4 Facts and rules


Simplified this can be described as an If Then function where the If part consists of a bunch of
parameters and AND or OR operators. Expert systems can be extended by learning compo-
nents, so rules can be added or deleted. Expert systems, for example, contain an inference
machine which uses adjustable rule sets (R) to analyse facts (X) and steering (P) parameters to
draw conclusions (Y). This is illustrated in Figure 2.

2
Taken from: HUMA 201: HUMA 201 Metaphors in English and Chinese, 2004
KNOWLEDGE REPRESENTATION TECHNIQUES 4

Figure 4: Adjustable Rule Set3

Rules can include certainty factors, priorities and fuzzy states. They are easy to extend and
small knowledge bases are easy to maintain. Furthermore they can provide an explanation
component. The similarity to human reasoning makes this technique the most understandable
and suitable for Expert Systems. Additionally there are several conflict resolutions strategies
for the inference engine available. One problem with rule sets is the speed. Matching a condi-
tion, selecting a rule and firing needs much more time then other methods and is so inherently
inefficient. A second drawback occurs when knowledge bases grow significantly large
enough. Then they are hard to maintain and debug not only because the order of rules which
is often crucial. Finally the control knowledge is frequently mixed up with the domain knowl-
edge. So the context in which a rule applies is to be considered.

2.5 Statistic
Empirical or calculated data can be an example of the use of statistical information is a Bayes-
ian Belief Network. Foundation of a Bayesian Belief Network is the probability of its events
and the conditional probability of an event chain. The probability of an event B occurring
given that an event A has already occurred is expressed mathematically as P(B|A) = P(B∩A) /
P(A). Bayes found a way of answering the question using the flip conditional probability of
event A given that event B has occurred. His theorem provides the probability of the truth of a
hypothesis, H, given some evidence, E, and is expressed mathematically as P(H|E) = P(H) *
P(E|H) / P(E), where P(H|E) is the reverse probability that H is true given E, P(H) is the prior
probability that H is true, P(E|H) is the probability of observing E when H is true and P(E) is
the probability of E occurring.4 Figure 3 describes this in connection with a medical diagnose.
For each event the probability of being true or false is given.

3
Taken from: Tolk, Andreas: Human Behaviour Representation – Recent Developments, 2002
4
Further described in: Brosnan, Amanda: Use of Bayesian Belief Networks for Enemy Course of Action As-
sessment at the Tactical Level, Masters Thesis, Shrivenham, 2003
KNOWLEDGE REPRESENTATION TECHNIQUES 5

Figure 5: A Bayes Belief Network5

Using the conditional probabilities of the event chains, a diagnose can be made, which disease
is the most likely. These systems can handle uncertainty and provide information even with-
out any input. Therefore this technique is suitable especially for Decision Support or Expert
Systems. The disadvantage is the ease of misinterpretation this technique provides. If a system
suggests one solution with a probability of sixty percent and another with forty percent one
can be mislead to think the solution with the higher probability has to be the correct one.

2.6 Neural Networks


Another approach to represent knowledge is the use of neural networks. These mathematical
models of connections of neurons, human brain cells are depicted in Figure 4.

Figure 6: Simplified illustration of a neural network6

5
Taken from: Looney, Carl: CS 773c - Intelligent Systems Applications: Information Fusion, 2001
6
Taken from: Z Solutions: Light Description of Neural Networks
KNOWLEDGE REPRESENTATION TECHNIQUES 6

By customising input weights (w), threshold (Ω), activation function (F) or output function
(F’) of particular perceptrons (mathematical neurons, see Figure 5), the network can be trai-
ned or when implemented, learn unauthorised.

Figure 7: Parameters of a perceptron7

Neural networks are especially suitable for pattern matching. Known constellations of pa-
rameters can be recognised and determined output can be given. In contrast to rule based sys-
tems neural networks can be robust and error tolerant, so similar situations can be identified.
Some of the most famous neural networks are back propagation nets, Hopfield Network and
the Boltzmann-Machine.8 Neural networks are extremely suitable for pattern matching and
searching because of their capability of being trained. The disadvantage is the complexity.
The creation of neural nets assumes a deep knowledge of their function.

2.7 Mixed Approaches


Certainly all previous mentioned ways can be combined. This combination represents a pow-
erful but complicated solution. The problem is the data exchange between the different meth-
ods. Here potent interfaces must be created.

7
Taken from (and modified): Z Solutions: Light Description of Neural Networks
8
Further described in: Honekamp, Wilfried: Intelligente Systeme. Skriptum zur Vorlesung, Fachbereich Wirt-
schaft, Hochschule Bremen, 2002
CONCLUSIONS 7

3 CONCLUSIONS
A variety of different techniques exist to represent knowledge of which each has its advan-
tages and disadvantages. Mainly application and domain are the crucial factors which one has
to be chosen. Semantic Nets are the most visual representation and suitable for binary and
simple relationships in a small domain. Frames and scripts provide a high degree of flexibility
and are to be used in fast changing but not complex domains. Facts and rules are similar to
human reasoning and thus appropriate for Decision Support and Expert Systems dealing with
certain knowledge. Uncertainty in this context can be represented by statistical data and func-
tions like Bayesian Belief Networks. Relationships are best to be represented with logical
connections and when they are blurry with Fuzzy Logic. Complex patterns are the best do-
main for neural networks. Finally there can be mixed approaches of two or more techniques to
suit special domains and application.
BIBLIOGRAPHY 8

4 BIBLIOGRAPHY

Barr, Avron and Fei- The Handbook of Artificial Intelligence, Addison-Wesley,


genbaum, Edward Reading 1986
Brosnan, Amanda Use of Bayesian Belief Networks for Enemy Course of Action
Assessment at the Tactical Level, Masters Thesis, Shrivenham,
2003
Honekamp, Wilfried Intelligente Systeme. Skriptum zur Vorlesung, Fachbereich
Wirtschaft, Hochschule Bremen, 2002
http://betriebssysteme.keyspace.info/kiskript.zip, 27.02.2004
HUMA 201 HUMA 201 Metaphors in English and Chinese, 2004
http://teaching.ust.hk/~huma201/ai.html, 01.03.2004
Looney, Carl CS 773c - Intelligent Systems Applications: Information Fu-
sion, 2001
http://ultima.cs.unr.edu/cs773c/u8/unit8773c.htm, 29.02.2004
Marshall, Dave Artificial Intelligence 2 COURSEWARE, Lecture notes +
integrated exercises, solutions and marking, 2004
www.cs.cf.ac.uk/Dave/AI2/node61.html, 01.03.2004
McNaught, Ken Bayesian Belief Networks, Presentation, Shrivenham, 2004
Sastry, Venkat Knowledge Representation, Presentation, Shrivenham, 2004
Tolk, Andreas Human Behaviour Representation – Recent Developments,
2002
http://www.vmasc.odu.edu/Publications/TolkHBR%20Technol
ogies%202002.pdf, 28.02.2004
Z Solutions
Light Description of Neural Networks
http://www.zsolutions.com/light.htm, 29.02.2004

You might also like