Philosophy of Mind', pp. 81±108. Atascadero, CA: Nemirow L (1990) Physicalism and the cognitive role of
Ridgeview. acquaintance. In: Lycan W (ed.) Mind and Cognition,
Lycan W (1990) What is the `subjectivity' of the mental? pp. 490±499. Oxford: Blackwell.
In: Tomberlin J (ed) Philosophical Perspectives, vol. IV
`Action Theory and the Philosophy of Mind',
pp. 109±130. Atascadero, CA: Ridgeview.
CONTENTS
Introduction Semantic networks
Representing common-sense knowledge Schemas, frames, and scripts
Predicate calculus and other logical representation Pictorial representations
schemes Connectionist representations: local and distributed
Procedural representations Managing change
Production systems
decide that to get from our house to the airport, we example, the Cyc Project has devoted many staff-
should walk to our car and drive to the years to the organization of a computer-usable
airport rather than, for example, drive to our car representation of all the knowledge that is not con-
and then walk to the airport. tained in encyclopedias (thus the name `Cyc,' from
In the 1960s and 1970s, much knowledge repre- `encyclopedia') but is assumed to be already
sentation research was concerned with represent- known by people who read them, and Lycos is
ing and using the kind of information we get from using such an ontology to organize searches of the
reading and talking to other people; that is, the World Wide Web. (See Spatial Representation and
information that is often expressed in natural lan- Reasoning)
guages, and that underlies our ability to under- All these threads continue into the 2000s.
stand and use natural languages. For example, we
probably understand each of the sentences in the
first column of Table 1 as shown in the second
PREDICATE CALCULUS AND OTHER
column, by adding our `background knowledge'
LOGICAL REPRESENTATION
to what the sentences explicitly say. Moreover, SCHEMES
our understanding of English includes our being In the late 1800s and early 1900s, various formal
able to make the following inferences. (See Natural systems were developed by people who hoped to
Language Processing; Meaning; Semantic turn human reasoning into a kind of calculation.
Memory: Computational Models) From our perspective, we can now see that what
these people were engaged in was research in
Every student studies hard: Therefore every smart
student studies: knowledge representation. The formal systems
they developed were systems of logic, a topic
On Tuesday evening, Jack either went to the
which has been studied since the days of Plato
movies, played bridge, or studied: On Tuesday
and Aristotle. We may consider logic to be the
evening, Jack played bridge: Therefore, Jack
study of correct reasoning. The systems of logic
neither went to the movies nor studied on
developed in the late 1800s and early 1900s consist
Tuesday evening:
1
of three basic components:
In the 1970s and 1980s, researchers became in-
. syntax: the specification of a set of atomic symbols, and
creasingly concerned with knowledge about spe-
the grammatical rules for combining them into well-
cific domains in which human experts operate, formed expressions;
such as medical diagnosis and the identification of . semantics: the specification of the meaning of the
chemical compounds from mass spectrometry atomic symbols, and the rules for determining the
data, and also with the other extreme ± knowledge meanings of well-formed expressions from the mean-
about the everyday world that everyone knows, ings of their parts;
such as the fact that when you tip over a glass of . proof theory: the specification of a set of rules, called
water, the water will spill on the floor. (See Expert `rules of inference', which, given an initial collection,
Systems; Expertise) called a `proof', of well-formed expressions, called
In the 1980s and 1990s, these concerns focused on `axioms', specify what other well-formed expressions
the details of specific subdomains of everyday can be added to the proof. (See Inference using Formal
Logics)
knowledge, such as theories of time and space,
and also on the general structure of our knowledge There are two kinds of `meaning' determined by
of everyday terms, leading to the construction the semantics of a system of logic. In one, we might
of large and general purpose `ontologies'. For say that the meaning of G is the claim made by the
sentence, `The moon is made of green cheese.' For `Students study hard' and Q represents `Students get
this notion of meaning, the meaning of :G, as good grades'. First-order logics (predicate logics)
shown in Table 2, would be the same as `It is not continue the analysis down to objects, classes,
the case that the moon is made of green cheese' or properties, and relations, with the aid of the quan-
`The moon is not made of green cheese.' The other tifiers shown in Table 3. So in some first-order logic,
sense of `meaning' is a truth value. Different `Students who study hard get good grades' might be
systems of logic have different truth values, and represented as 8x (Student(x) ^ study (x, hard) ) get
even different numbers of truth values. There are (x, grades, good)). Some first-order logics allow func-
two-valued logics, three-valued logics, four-valued tions. In one of them, this sentence might be repre-
logics, and even logics with an infinite number of sented as 8x (Student(x) ^ study (x, hard) ) get (x,
truth values. Two-valued logics usually call their good (grades))). Second-order logics allow functions,
truth values `True' and `False'. Some logicians classes, properties, and relations, themselves, to be
would say that in such a two-valued logic any the arguments of other functions, classes, proper-
sentence either means True or False. Less strictly, ties, and relations. In one of them, this sentence
one might say that in such a logic, the semantics might be represented as 8x (Student (x) ^ hard
assigns a truth value of True or False to every (study) (x) ) get(x, good (grades))). (See Representa-
sentence. In this notion of meaning, if some sen- tions Using Formal Logics)
tence P happened to be (or be assigned the truth A logical sentence that always evaluates to True
value of) True, then :P would be (or be assigned regardless of the meanings of its atomic parts is
the truth value of) False, and if P were False, then called valid. In most standard logics, the sentence
:P would be True. So, if G meant (by the first sense P ^ :P ) Q is valid, meaning that a contradiction
of meaning) `The moon is made of green cheese,' implies anything whatsoever, but in some logics,
then G would be (or mean, in the second sense of called `paraconsistent' logics, that sentence is not
meaning) False, so :G would be (or mean, or have valid. In most standard logics the sentence P _ :P
the truth value of) True. is valid, meaning that any sentence is either True
Although the proof theory considers only the or False, but in some logics, called `intuitionistic'
syntax of the expressions, not their semantics, it is logics, that sentence is not valid.
usually the case that if the semantics assigns a truth Someone who uses a propositional logic to for-
value of True to the axioms, all expressions that the malize some domain of interest chooses the prop-
rules of inference add to the proof will also be True. osition symbols to be used to represent the
Logics that have this property are called sound. sentences of the domain, and their semantics ±
Soundness seems to capture the notion of correct what sentence each symbol will represent. Some-
reasoning, which is what the study of logic is all one who uses a predicate logic to formalize some
about. domain of interest chooses the syntax and seman-
Many different logics have been described and tics of the individual constants that represent
investigated. Propositional (or `sentential') logics objects in the domain, the function symbols that
do not analyze information below the level of the represent functions in the domain, and the predi-
proposition (or sentence), but use `propositional cate symbols that represent classes, properties, and
connectives,' such as are shown in Table 2 to relations in the domain. The logic itself determines
build more complex sentences from simpler sen- the propositional connectives and the quantifiers,
tences. For example, the sentence `Students who and how they are to be used, along with function
study hard get good grades' could not be represented and predicate application, to determine the non-
in more detail than P ) Q, where P represents atomic expressions, and their meaning. The rules
of inference also operate only on nonatomic expres-
sions, and pay attention only to the logical con-
stants. This is the sense in which people consider
Table 2. A set of propositional connectives and their
meaning these logics to be `formal' logics that pay attention
only to the form, and not to the content, of the Clinton, but how many of us can truly say, `I know
logical expressions. When selecting a logic to use, Bill Clinton'? Knowing how is knowledge of how
one is choosing the formal apparatus supplied by to do things, for example how to swim or ride a
that logic. bicycle.
Any knowledge representation and reasoning There has not been much work in knowledge
system consists of two parts ± a knowledge repre- representation on knowing who, and everyone
sentation language and a reasoning component. would agree that a procedural representation is
If the knowledge representation language is well- appropriate for knowing how, though more on this
defined, it will have a well-defined syntax to de- later. The declarative/procedural controversy was
scribe the atomic symbols and how well-defined about how to represent knowing that. The declarati-
symbol structures may be constructed from them, vists were in favor of representing propositions
and a well-defined semantics to describe what the that are known (believed) by some agent as a
atomic symbols and the symbol structures are sup- symbol structure with declarative semantics, for
posed to mean. The reasoning component is a pro- example a well-formed expression of propositional
gram, often called an `inference engine', that, given or predicate logic, stored in the agent's knowledge
a `knowledge base' of symbol structures, adds add- base. The proceduralists were in favor of represent-
itional symbol structures to that knowledge base ing such propositions as small programs. For
according to the rules implemented in the program. example, when the early (simulated) robot
Clearly these components ± syntax, semantics, in- SHRDLU was told, `I own blocks which are not
ference engine, knowledge base ± correspond to the red, but I don't own anything which supports a
components of logics ± syntax, semantics, proof pyramid', it represented that information as two
theory, proof. So we may view any knowledge small procedures in the PLANNER programming
representation and reasoning system as a logic, language. When, later, SHRDLU was asked `Do I
and ask what kind of logic it is, what formal appar- own anything in the box?', it ran those two proced-
atus it supplies, and whether or not it is sound. The ures to determine the answer. The problem with
user of a knowledge representation and reasoning maintaining this distinction between declarative
system, like the user of a logic, must choose a and procedural representations of knowing that is
system, and then design the representations that the well-known equivalence of data and program.
are not at the level of knowledge representation A procedure can be written in a declarative pro-
constructs that the system deals with. In the know- gramming language such as Lisp or Prolog, and
ledge representation world, this person is called a can thereby be viewed as a symbol structure with
`knowledge engineer'. declarative semantics. On the other hand, when a
declarative representation is used by the inference
engine to draw some inference, the declarative
PROCEDURAL REPRESENTATIONS
representation may be viewed as a program in the
In the mid-1970s, knowledge representation re- programming language interpreted by the infer-
searchers were embroiled in what was called the ence engine. In this sense, whether a representation
`declarative/procedural controversy'. Although is declarative or procedural depends on how one
this controversy has largely been resolved (in a views it. (See SHRDLU)
sort of compromise), it is worthwhile under- We might resurrect the declarative/procedural
standing these two approaches to knowledge distinction by considering our own knowledge of
representation. how to do things. Many of us know how to ride a
Firstly, we must recognize that there are several bicycle. However, few of us can describe how to
kinds of knowing, among which are knowing that, ride a bicycle, for example in order to instruct
knowing who, and knowing how. Knowing that is the someone else. We might consider this knowledge
kind of knowledge we have of propositions. For procedural only. In this view, all knowledge may be
example, we may know that Seattle is north of viewed as procedural knowledge, but only know-
San Francisco. Knowing who is acquaintance with a ledge that can be expressed in a declarative
person, animal, object, etc. We may say that we language by the knower may be viewed as declara-
know a person even though we might not know tive knowledge. As another example, the restaur-
some important facts about that person, for ant script (see below) is a representation of what
example their birthdate. On the other hand, we typically happens in a restaurant. There have been
may know many facts about a person without programs that, supplied with the restaurant script,
being able to say we know that person. For could fill in details about what happened in res-
example, many of us know many facts about Bill taurant stories. For example, given the story, `John
Knowledge Representation 675
went to a restaurant and ordered a steak', such a size of working memory and the allowed actions
program could infer that John was seated and were restricted. However, they have been a popu-
given a menu between entering and ordering. lar architecture for expert systems, for which pur-
However, most of these programs could not pose those restrictions were lifted.
answer questions about the restaurant script itself, It should be noted that the symbol structures in
such as `What typically happens in a restaurant working memory and the patterns and actions in
after the patron is seated?' It is, therefore, reason- rule memory must be formulated in some know-
able to say that for these programs the restaurant ledge representation language. What the produc-
script is not represented declaratively, but only tion system architecture provides is a particular
procedurally. (See Story Understanding) style of reasoning and acting using that language.
has-part
Animal head
isa isa
can
Sing Canary Penguin Charlie
instance instance
Tweety Opus
of the features common to different description but penguins are birds that swim instead, is to
logics, a parent can be defined as a person with at separate flying birds from swimming birds:
least one child who is also a person, as follows.
cprim BIRD
and ANIMAL
cdef PARENT
and PERSON
c-some
c-atleast 2 Part WING
Child PERSON
3
c-atmost 2 Part WING
cdef FLYING-BIRD
and BIRD
Here, PARENT is the concept being defined,
c-some Moves-by FLYING
PERSON is a concept which, presumably, has al-
cprim PENGUIN
and BIRD
ready been defined, and Child is a role. This is also
c-some Moves-by SWIMMING
an example of a concept defined with necessary
cprim CANARY
and FLYING-BIRD
and sufficient conditions. That is, if Ken is said to
c-some Can SING
5
be a PARENT, it is necessary that Ken be a PERSON
with at least one Child who is a PERSON. So the (See Concepts, Philosophical Issues about; Con-
description logic system can infer that Ken is ceptual Representations in Psychology; Natural
a person, has at least one child, and that child is a Kinds and Artifacts)
person. On the other hand this same definition says All these KL constructs define concepts, and are
that if Judi is a PERSON with at least one Child considered part of the description logic termino-
who is a PERSON, that is sufficient information to logical component. To actually make assertions
conclude that Judi is a Parent. Natural kinds, such about individuals, most description logics also
as birds, fish, and animals, cannot be given neces- have an assertional component. Assertions in the
sary and sufficient conditions, so primitive concepts assertional component are usually written in a
can be defined with only necessary conditions. syntax that looks like normal first-order predicate
The KL definitions of ANIMAL and FISH from logic in which defined concepts can be used as
Figure 1 are: unary predicates and defined relations can be
used as binary relations. For example, we might
cprim ANIMAL
and top have:
c-some Part HEAD
CANARY
Tweety WING
Tweety-left-wing
c-atmost 1 Part HEAD
PENGUIN
Opus
cprim FISH
and ANIMAL
SWIMMING
Opus-swimming-style
6
c-some Part FIN
c-some Moves-by SWIMMING
4 Part
Tweety, Tweety-left-wing
Moves-by
Opus, Opus-swimming-style
7
This says that every FISH has one head, by inherit-
ance from ANIMAL, and, in addition, has one or Besides the confused semantics of their relations,
more FINs. Since description logic roles accumu- another deficiency of inheritance networks is that
late in this way, the only way to say that birds fly, since information can only be represented about
Knowledge Representation 677
nodes, one cannot represent information about re- the `bedroom frame' would contain pointers to the
lations, such as that the `isa' relation is transitive. representation of objects you would expect to be in
Nor can one represent information about beliefs, a bedroom, etc. As frame representation systems
such as that the encyclopedia is the source of the were implemented, they became more similar to
belief that canaries are birds. Description logics semantic networks, but with the labelled arcs, now
do represent information about relations, but they called `slots', pushed into the nodes, now called
do not represent information about beliefs. This `frames', and the nodes pointed to by the arcs, now
deficiency is solved by propositional semantic called `slot fillers'. For example, Figure 3 shows the
networks, in which nodes are used to represent information of Figure 1 as a frame system.
beliefs (propositions) as well as the individuals, One feature frame systems tend to have that
categories, and properties represented by nodes in semantic networks do not is procedural attachment.
inheritence networks. Figure 2 illustrates a propos- Instead of a slot being filled by a pointer to a frame
itional semantic net in which `M1!' represents the or a set of such pointers, the slot could be filled by
proposition that canaries are birds, `M2!' represents an if-needed or an if-added procedure. If a slot con-
the proposition that `isa' is a transitive relation, and taining an if-needed procedure is accessed, the pro-
`M3!' represents the proposition that the source of cedure is executed and is expected to compute and
`M1!' is the encyclopedia. return the slot filler. If a slot containing an if-added
procedure is filled, the procedure is executed and is
expected to fill other slots that depend on the new
SCHEMAS, FRAMES, AND SCRIPTS
information being added to this slot. If-needed and
Some researchers felt that semantic networks used if-added procedures are procedural versions of
a representation that was too fine-grained and too inference by backward chaining and forward
passive. Instead, they argued for representational chaining, respectively.
structures that contain more information about the Scripts, like frame systems, were designed to be
entities being represented, and also incorporate structured representations, but of activities rather
active processes. They adapted the notion of sche- than objects. For example, the restaurant script con-
mas (or `schemata') from psychological literature. tains a representation of all the activities that typic-
The two most widely used schema representation ally occur when one visits a restaurant. If you read a
systems are frames and scripts. (See Schemas in story about someone going to a restaurant and
Psychology) ordering a steak, you fill in the information
Frames were originally proposed as a represen- about being seated and being given a menu from
tation of structured visual information about com- your restaurant script. (See Natural Language Pro-
plex objects. For example, if you open a door to an cessing: Models of Roger Schank and his
office, you expect to see certain things, such as a Students)
desk, chairs, etc. You would be surprised to see a
tennis court, a beach, and a swimming pool in the
office. On the other hand, if you opened a door to a
PICTORIAL REPRESENTATIONS
bedroom, you would expect to see a bed, a chest of Some people think mostly linguistically; others
drawers, etc. The proposal was that the `office think mostly in pictures. Everyone can probably
frame' would contain pointers to the representa- do both, even though they usually do one or the
tions of objects you would expect to be in an office, other. Try this: Think of an elephant. Which way is
M3 !
Prop Source
M2 !
M1! Encyclopedia
Animal
has-part: head
Bird Fish
isa: Animal isa: Animal
has-part: wing has-part: fin
moves-by: flying moves-by: swimming
Canary Penguin
Charlie
isa: Bird isa: Bird
instance: Fish
can: sing moves-by: swimming
Tweety Opus
instance: Canary instance: Penguin
it facing? If, when you thought of an elephant, you (desk, chair), Left(sofa, desk), Left (chair, stool), where
pictured one in your mind, you should have a Left (x, y) means that x is to the left of y. To decide
definite answer to that question. the left-to-right arrangement of the stool and the
Just as people can represent entities in their sofa requires a certain amount of search and infer-
minds either linguistically or pictorially, we can ence. However if, instead, the Left relation were
use linguistic and pictorial representations in represented by order in a list, the four relations
other media, including computers. The distinction would be captured by the list (sofa, desk, chair,
is also often termed digital vs. analog, as in digital stool), and the left-to-right arrangement of the
clocks vs. analog clocks. (See Mental Imagery, stool and the sofa could be decided by a linear
Philosophical Issues about) search. Some researchers have created systems
The best way to distinguish analog from digital that can reason about diagrams or visual scenes
representations is to compare the domain of by representing them in two-dimensional data
the representation (syntax) to the domain of what structures where it is easy to rotate or otherwise
is represented (semantics). An analog representa- manipulate the figures. (See Analogical Reasoning,
tion has a syntactic operation that is a direct ana- Psychology of)
logue of a semantic representation. Consider
clocks. What is represented is time. On an analog
CONNECTIONIST REPRESENTATIONS:
clock, the representation is the rotation of the clock
LOCAL AND DISTRIBUTED
hands around the circle of the clock face. The dif-
ference between the representation of 10.15 a.m. Connectionist representations are designed to
and that of 10.30 a.m. is a 90 degree rotation of the model the brain by using a large collection of inter-
minute hand, which is one quarter of the complete communicating units, each of which is a model of
360 degree rotation. The complete 360 degree rota- a neuron. These units are organized in layers: an
tion represents one hour, and one quarter of a input layer, an output layer, and one or more
rotation represents one quarter of an hour. On a `hidden' layers. Each unit maintains an activation
digital clock, however, the times 10.15 a.m. and level and connections to other units, which may be
10.30 a.m. are represented with different numerals. on the same layer (in some versions) or on layers
Nothing about the difference between the two sets closer to the output layer. Each connection is also
of numerals indicates what the difference in the given some weight, which might be negative or
represented times is, unless one moves to the sep- positive. The network as a whole makes some deci-
arate semantic domain of numbers, and subtracts sion or characterizes its input. Input is achieved by
15 from 30. adjusting the activation level of the units in the
Analogue representations can be constructed in input layer. When the activation of a unit exceeds
computers by using a data structure whose oper- some threshold (which may be different for differ-
ations are analogues of the relations being repre- ent units), an activation is passed to all outgoing
sented. For example, consider a predicate logic connections, where it is adjusted by the connection
representation of items arranged in a row: Left weights, and passed to the connected units, etc. The
Knowledge Representation 679
final decision or characterization is read off the Removing default conclusions that have been
units in the output layer. Networks are trained defeated by more specific information is just one
by adjusting the weights on the connections by possible reason that information might have to be
one of several possible feedback mechanisms. (See removed from a knowledge base. If the knowledge
Connectionism; A00068; A00163) base is a model of the world, or a model of some
Local connectionist representations are distin- agent's beliefs about the world, it may be that the
guished by the requirement that each decision or world changes because of the action of the agent or
characterization is represented by a single unit, and some other agent. If the knowledge base is a model
each input unit also represents some concept of the of some agent, or a collection of beliefs input by
input. For example, in a lexical decision task, each some agent or agents, it may be that the agent or
input unit might represent a particular letter in a agents have changed their beliefs. If the knowledge
particular position, and each output unit might base is a collection of facts and `laws' of some
represent a particular word. developing theory, it might be found that some of
In a distributed connectionist representation, the facts and laws are contradictory, and the theory
each represented decision or characterization is must be revised. Removing information from a
represented, not by a single unit, but by a pattern knowledge base seldom involves merely removing
of unit activations. Distributed representations a single proposition (fact, rule, law). If additional
have been found to generalize what they have propositions have been derived from the one to be
learned better than local representations do. removed, they might need to be found and re-
Connectionist representations are considered moved also. If the proposition to be removed was
subsymbolic rather than symbolic representations. derived from other propositions in the knowledge
As such, they are not as capable of representing and base, or could be rederived from them, they must
reasoning about beliefs as the other representation be found, and at least one of them must be removed
techniques discussed in this article. (See Symbolic or else the removed proposition could be reintro-
versus Subsymbolic; Bayesian Belief Networks; duced. The first systems that knowledge represen-
Language, Connectionist and Symbolic Repre- tation researchers implemented to handle these
sentations of) complications of removing information from
knowledge bases were called `truth maintenance
systems'. More formal studies, carried out by com-
MANAGING CHANGE puter scientists, logicians, and philosophers go
Consider again some of the information in Figures under the name `belief revision'.
1 and 3, namely that birds fly, but penguins do not. Using belief revision or truth maintenance to
If you learn that Opus is a bird, you are justified in deal with a changing world is appropriate if the
concluding that Opus can fly. However, if you then knowledge base always represents the current time,
learn that Opus is a penguin, you must reject your and should be changed as time moves. However,
conclusion that Opus can fly. This is an example of this eliminates the possibility of representing what
an interesting phenomenon where a new piece of was the case at past times. To do that, time must be
information causes the rejection of a previous con- represented explicitly, and propositions that hold
clusion. It is sometimes said that, in this case, the only for some specific time must indicate so expli-
new piece of information defeats the old conclusion. citly. To do this, specialized logics including tem-
This phenomenon often occurs in the presence of poral logics and modal logics have been used.
general information to which there are exceptions. Another logic for this purpose, popular among
The general information is sometimes referred to as knowledge representation researchers, is situation
default knowledge, and conclusions drawn from the calculus, in which predicates that can change are
general information are sometimes said to be defeas- given an extra argument that ranges over situ-
ible. From the point of view of classical propos- ations. For example, if a particular book is on a
itional and predicate logic, this situation is most particular table in a particular situation, this
unusual, since these logics are monotonic, meaning might be represented as On(book 1, table3, S5). In
that if a conclusion can be drawn from some set of situation calculus, an action is considered to be a
beliefs, it can also be drawn from any superset of function from the situation before the action is per-
those beliefs. (Just ignore the extra beliefs.) At- formed to the situation afterward. (Some versions
tempts to formalize defeasible reasoning have of situation calculus use slight variations of this.)
been made by knowledge representation research- For example, the action pickup(book1, S5) might rep-
ers, and this remains an active area of research. (See resent the situation that exists after picking up
Non-monotonic Logic) book1 in situation S5. We would then have :On
680 Knowledge Representation
(book1, table3, pickup(book1, S5)). Stating the effects GaÈrdenfors P (ed.) (1992) Belief Revision. Cambridge, UK:
of actions is fairly straightforward. However, Cambridge University Press.
stating what is not changed by an action is more IwanÂska èM and Shapiro SC (eds) (2000) Natural
involved. For example, if it is the case that In(table3, Language Processing and Knowledge Representation:
Language for Knowledge and Knowledge for Language.
room2, S5), is it the case that In(table3, room2, pickup
Menlo Park, CA: AAAI Press/MIT Press.
(book1, S5))? The problem of specifying what is not Kramer B and Mylopoulos J (1992) Knowledge
changed by an action has been called the `frame representation. In: Shapiro SC (ed.) Encyclopedia of
problem', not to be confused with the frames used Artificial Intelligence, 2nd edn, pp. 743±759. New York,
as schema representations. (See Frame Problem, NY: John Wiley.
The) Lehmann F (ed.) (1992) Semantic Networks in Artificial
Intelligence. Oxford, UK: Pergamon Press.
Further Reading Levesque HJ and Lakemeyer G (2000) The Logic of
Knowledge Bases. Cambridge, MA: MIT Press.
Addanki S (1992) Connectionism. In: Shapiro SC (ed.) Lifschitz V (ed.) (1990) Formalizing Common Sense: Papers
Encyclopedia of Artificial Intelligence, 2nd edn, by John McCarthy. Norwood, NJ: Ablex.
pp. 268±274. New York, NY: John Wiley. Reichgelt H (1991) Knowledge Representation: An AI
Bobrow DG and Collins A (eds) (1975) Representation and Perspective. Norwood, NJ: Ablex.
Understanding: Studies in Cognitive Science. New York, Rumelhart DE and McClelland JL (eds) (1986) Parallel
NY: Academic Press. Distributed Processing 2 vols. Cambridge, MA: MIT
Brachman RJ and Levesque HJ (eds) (1985) Readings in Press.
Knowledge Representation. San Mateo, CA: Morgan Sowa JF (ed.) (1991) Principles of Semantic Networks:
Kaufmann. Explorations in the Representation of Knowledge. Los Altos,
Cercone N and McCalla G (eds) (1987) The Knowledge CA: Morgan Kaufmann.
Frontier: Essays in the Representation of Knowledge. New Sowa JF (2000) Knowledge Representation: Logical,
York, NY: Springer-Verlag. Philosophical, and Computational Foundations. Pacific
Davis E (1990) Representations of Commonsense Knowledge. Grove, CA: Brooks/Cole.
San Mateo, CA: Morgan Kaufmann.
Hobbs JR and Moore RC (eds) (1985) Formal Theories of the
Commonsense World. Norwood, NJ: Ablex.
CONTENTS
The basics of representation Using representations
Types of representations
`Knowledge representation' is an umbrella term for Theaetetus, the Greek philosopher Plato described
the methods by which information is stored in the memory as a wax tablet, in which information is
mind for later use. stored as impressions in the wax. In this proposal,
the information remained in memory for as long
as the impression remained in the wax. Later in
THE BASICS OF REPRESENTATION the same work, Plato suggested that memory is like
From the beginning of the study of psychology, phil- an aviary with birds flying around it. Retrieving
osophers and psychologists have been interested in information was like grabbing a bird from the
the way information is stored in the mind. In his aviary.