http://orion.journals.ac.

za/
47

ORiON, Vol. 4, No. 1, pp. 47-64 ISSN 0259-191X

PROBLEM SOLVING
USING
ARTIFICIAL INTELLIGENCE TECHNIQUES

by
A.R. Greef and R. Reinecke
Centre for Robotics
Department of Industrial Engineering
University of Stellenbosch

ABSTRACT

Real-world problems often do not lend themselves to an algo-

rithmic solution. Humans, however, cope with these problems
despite their fallible problem solving techniques. Instead of
trying to construct algorithms to solve problems AI
researchers have concentrated on using the more successful
methods used by humans. This paper reviews the area of prob-
lem solving in the field of Artificial Intelligence. This
includes problem representation for comJ?utation, "weak" meth-
ods of searching for a problems solution, knowledge represen-
tations that facilitate more efficient search strategies and
planning - an advanced problem solving technique.

0.0 INTRODUCTION
What is Artificial Intelligence (AI) and why should it
be of use to Operations Researchers?
Barr and Feigenbaum [1] answer the first part of the
question by defining AI to be "the part of Computer Science
concerned with designing intelligent computer systems, that
is, systems that exhibit characteristics we associate with
intelligent human behaviour". Margaret Boden [2) explains
that "computers are its (AI) tools, because its theories are
expressed as computer programmes that enable machines to do
things that would require intelligence if done by people".
The second part of the question is ans•vered when consid-
ering that much of Operations Research is concerned with
 http://orion.journals.ac.za/

48

planning activities or actions required to solve problems.

The field of AI covers a number of areas of which problem
solving is one. Research in this area has been involved with
finding solutions to problems that do not lend themselves to
an algorithmic approach, a characteristic that many problems
display. Furthermore, in many instances the knowledge or
information about a problem domain is difficult or unnatural
to represent in data structures such as arrays or sets of
numbers. Often it is more "natural" to represent a problem
as a set of English language sentences or statements describ-
ing the problem domain. In these instances AI programming
techniques are useful.
There are two difficulties associated with the algorith-
mic (step-by-step) approach to solving a class of problems
(Graham [3]):
a) there are some classes of problems for which there is no
algorithm that will solve every problem in a particular
class and,
b) even if there is an algorithm that will solve every prob-
lem in a particular class, the algorithm may be so ineffi-
cient as to be unsuitable for practical problems (pro-
grammes that run in exponential time are considered to be
impractical).
Humans, however, often cope with a class of problems
even though no all-encompassing algorithm exists and if one
does exist, it runs in exponential time. This is done in
spite of the fact that the problem solving techniques
employed have been found to be fallible, approximate and
based on quick-and-dirty methods. Instead of trying to con-
struct algorithms to solve problems AI researchers have con-
centrated on the more successful methods used by humans. As
humans can understand and express problem solving knowledge
more easily using symbols and natural language, symbolic pro-
gramming languages have been developed for AI programming
which tend to be based on symbol processing rather than num-
ber crunching.
This paper reviews the area of problem solving in the
field of AI. Part 1 illustrates methods for representing a
problem for coding into a computer. Part 2 gives an overview
 http://orion.journals.ac.za/

49

of the "weak" methods that can be employed for finding the

solution to a problem. Part 3 describes the production sys-
tem architecture that most AI programmes adhere to. Part 4
presents methods of representing knowledge to enrich the
problem representations outlined in Part 1. Finally, Part 5
contains an overview of planning, an advanced problem solving
technique.

1.0 PROBLEM REPRESENTATION

There are two well-used methods for representing a prob-
lem (see Winston (4] for details). The first method treats
the problem as a network of inherently ordered states. The
second sees the problem as a series of sub-problems which in
turn have sub-sub problems and so on until a problem with an
immediate solution is reached.
Formally these two problem representation methods are
known as:
a) the state-space representation, and
b) the problem-reduction representation.
The choice of which representation to use for a particu-
lar problem is commented on by Winston (4]:
"In a practical sense, however, some representa-
tions emphasize things that are more important to
solving a class of problems. one scheme, there-
fore, is more powerful than another because it
offers more convenience to the user even though
theoretically, both can do the same work."

1.1 THE STATE-SPACE PROBLEM REPRESENTATION

A state-space problem representation consists of states
which are frozen conditions of a problem at each stage of its
solution. This representation should be used for problems
whose solutions are characterized by a succession of states.
An example is the "shortest path" problem (see Figure 1).
This problem can be represented by a network with nodes
representing cities and links representing distances between
the cities.
The problem is that of finding the shortest possible
route that a traveller can take in travelling from city "S"
 http://orion.journals.ac.za/

50

to city "G". Each node in Figure 1 depicts a state of the

salesman in relation to his start and goal position. The
problem's solutio~ can now be given as a suitable succession
of states that need to be attained in going from node "S" to
node"G".

FIGURE 1 THE SHORTEST PATH PROBLEM

When scanning networks for a solution it is possible to

become trapped in a cyclic loop (for example S-D-A-S-D-A in
Fig. 1). This can be detected and stopped by extra pro-
grammed mechanisms, however, it may be easier to eliminate
cyclic paths from the network, the resulting structure is
then called a state-space graph or tree (see Fig. 2).

FIGURE 2 A STATE-SPACE GRAPH FOR THE SHORTEST PATH PROBLEM

 http://orion.journals.ac.za/

51

Although this increases the size of the problem repre-

sentational data structure it reduces the amount of difficult
code needed to check for cyclic loops. Networks are made
into trees by tracing all the possible paths to the point
that they re-enter previously visited nodes or they reach a
node with no exit path. By convention the tree has nodes
connected by branches. Branches directly connect parent nodes
with children nodes. The top node, the one that has no par-
ent, is called the root node.

1.2 THE PROBLEM-REDUCTION PROBLEM REPRESENTATION

Using the problem-reduction (or goal-reduction) method,
the problem is structured as a set of sub-problems (or sub-
goals). The problem's solution is seen as the achievement of
sub-goals in going from the initial goal to the final goal.
Ah example problem represented using problem-reduction is the
a-puzzle. A tray contains eight tiles numbered 1 to 8 and a
space where the ninth tile has been left out as shown in Fig-
ure 3.
The tiles are initially positioned in a random state in
the tray (Fig. 3(a)). Solving the puzzle requires the order-
ing of the tiles as shown in Figure 3(b). The tiles can be
moved by sliding them either up, down, left or right as the
space permits.

(a) (b)
FIGURE 3 : THE 8-PUZZLE PROBLEM

Applying the problem-reduction representation breaks the

problem of ordering the initial a-puzzle into three sub-
problems that are easier to solve, namely:
1) The problem of getting the first row in order.
2) The problem of getting the second row in order and
keeping the first row in order, given that the first
 http://orion.journals.ac.za/
52

row was in order.

3) The problem of getting the third row in order and
keeping the second and first rows in order, given that
the second and first rows were in order.
Each of the above sub-problems can be further decomposed
into the sub-sub-problem of moving each tile into its respec-
tive place in the row being ordered. Problem-reduction con-
tinues in this manner until an immediate problem (eg. move a
tile either left, right, up or down) is achieved.
The above problem decomposition is one of many ways to
solve the 8-puzzle. Ordering columns instead of rows is
another alternative. These multiple methods of reducing a
problem into smaller problems can be depicted as a general-
ized tree called an AND/OR graph (Fig. 4).

2 4 7 1 2 3
1 3 --4 5 6
5 6 8 7 8

~ BY COLUMNS

2 4 7
1 3
5 6 8
1 2 3
f--? ? ?
? ? ?
~~ 1 2 3 1 2 3
4 5 6 f-- 4 5 6
? ? ? 7 8

/ ~ NO.DE 1 2 3 1 2 3 ~
AND NODE ? ? 7-4 5 6
? ? ? ? ? ?
l~
r
FIGURE 4 : AN AND/OR GRAPH FOR THE 8-PUZZLE

The AND/OR graph consists of nodes representing the

decomposed sub-problems and arcs linking the sub-problems in
the order that they must be considered. The start node cor-
responds to the initial problem. If a parent node is solved
by solving any one of its children nodes the parent node is
called an "OR" node (the start node is an "OR" node) . If a
parent node is only solved when all of its children nodes are
 http://orion.journals.ac.za/

53

solved then the parent node is called an "AND" node. To dis-

tinguish "AND" nodes from "OR" nodes, the arcs leading to
"AND" node children are joined by a line as shown in Figure
4.

2. "WEAK" PROBLEM SOLVING TECHNIQUES

Having represented a problem using one of the methods in
Part 1, its solution .can then be formulated as ·a search for a
sequence of nodes in a graph that connects the root node with
a terminal node. This sequence of nodes is the solution to
the problem.
If a tree search is conducted by starting with the ini-
tial problem configuration(s) at the root of the tree and
continues by tracing the solution path to the final problem
configuration at a terminal node, then the system conducts
forward reasoning (forward chaining). If the search begins
at the final problem configuration and ends at the initial
problem configuration, the system conducts backwards reason-
ing (backward chaining).
Two exhaustive search methods are (see Rich [5] for oth-
ers) :
(a) depth-first search and
(b) breadth-first search.
A depth-first search considers each node in a down-up
and left-to-right fashion. Breadth-first search traverses
across the nodes, also from left-to-right, considering all
the nodes on one level before considering any on a lower
level. The two search techniques are shown in Figure 5.
These brute-force techniques can be used effectively
with small search spaces, however with most practical prob-
lems a combinational explosion results with the nodes of the
search tree growing exponentially. Another problem associ-
ated with brute-force search is that they are not guaranteed
to find the problem solution of least cost although they do
find some solution. In these cases searches can be made more
efficient by employing some "heuristic" that will guide the
search in the right direction.
The word "heuristic" comes from the Greek word which
means "serving to discover". A heuristic is any hint or rule
 http://orion.journals.ac.za/
54

of thumb which helps to guide the search for a problem. A

search guided by heuristics is called a heuristic search.
(This should not be confused with the term "heuristics" as
used in knowledge representation) .

DEPTH-FIRST SEARCH BREADTH-FIRST SEARCH

FIGURE 5 : "WEAK" SEARCH TECHNIQUES

A heuristic which can be used to find a solution for the

"shortest path" problem in Part 1 is to always choose cities
that have the least distance , as the crow flies, to the goal
city, as the next step in the search path. A number of gen-
eralized heuristic search methods are presented in AI books
such as those of Winston [4] and Rich [5].

3.0 PRODUCTION SYSTEMS

Most AI systems have a clear separation between the
standard computational components of data, operations and
control. Various generalizations of this computational for-
malism are known as production systems. The major elements
of an AI production system are (Nilson [6]) :
a) a global data base;
b) a set of production rules; and
c) a control system.
The data base contains a description of a problem repre-
sentation and whatever other information that is appropriate.
It may be as simple as a small matrix of numbers or as com-
 http://orion.journals.ac.za/
55

plex as a large relational file structure. Some parts of the

data base may be permanent, while others may pertain only to
the solution of the current problem being solved. The data
base is changed during the problem solving process to reflect
the problems current status.
A production rule is of the form: IF a certain condition
is satisfied THEN do the following. The rule consists of a
left hand side (a pattern) that determines the applicability
of the rule, and a right hand side that describes the action
to be performed if the rule is applied. Production rules are
matched against the problem state in the global data base. A
forward chaining system matches the left hand side of a rule
against the data base and updates the data base as dictated
by the right hand side of the rule. A backward chaining sys-
tem attempts to match the right hand side of the rules to the
d·ata base and updates the data base as dictated by the left
hand side of the rule. Production rules are then the opera-
tors that define the conditions necessary to move from one
state or sub-problem (see Part 1) to another during the trav-
erse of a graph representing a problem.
Under direction of the control system, the computer
searches through a list of productions, trying to match each
one against the data base. The control strategy specifies
the order in which the rules are compared to the data base
and a way of resolving the conflicts (conflict resolution)
that arise when several rules match at once. A typical con-
trol strategy is to use a depth-first search to specify the
order in which the nodes of a problem graph are to be consid-
ered. A conflict resolution strategy would be to consider
the first rule that is found. The control system continues
to apply production rules until a goal state exists in the
data base or until it is discovered that the problem has no
solution. Furthermore, it keeps track of the problems solu-
tion path.

4.0 KNOWLEDGE REPRESENTATION

In order to solve complex problems using artificial
intelligence, a large amount of knowledge and some mechanisms
for manipulating that knowledge to create solutions, is
 http://orion.journals.ac.za/

56

needed. So far only a very general method of manipulating

knowledge - search, has been presented. Due to their gener-
ality, these search techniques are limited and only form the
skeleton of practical AI systems. More specific knowledge
representation models allow for more powerful inference mech-
anisms to guide the search for a solution. Emphasis is
placed on enriching the knowledge representations using
semantics and heuristics that constrain the search for a
solution. Heuristics are now embedded in the knowledge repre-
sentation rather than in the control system as was previously
done.
A knowledge base is the collective term for a production
system's global data base and set of production rules.
Objects, facts and relations describing a problem domain are
encoded into the data base. This is termed declarative knowl-
edge. The operators that act on the data base contain knowl-
edge on what to do if certain conditions arise in the data
base. This is termed procedural knowledge.
There are two schools of thought concerning how knowl-
edge should be represented (Winograd [7]). One school sup-
ports the theory that knowledge is stored in the human brain
in a declarative manner and that there are general procedures
that manipulate the knowledge. The other school supports the
theory that knowledge is contained as procedures (production
rules) only and there is no declarative knowledge. Often the
best representation schemes use both declarative and proce-
dural knowledge.
The following sections give a brief description of the
knowledge representation methods.

4.1 SEMANTIC NETWORKS

Semantic networks were first proposed by Quillian [8] as
a psychological model depicting how declarative knowledge is
encoded in human memory. A semantic network is a collection
of objects called nodes. The nodes are connected together by
arcs or links which represent actions, events or relations
between the objects. A node-and-link network is not neces-
sarily a semantic network, however. A true semantic network
must have semantics. Figure 6 shows a portion of a semantic
 http://orion.journals.ac.za/

57

network depicting the relationships between different classes

of machines. In this case lathe-20 is of the class lathe
which is a-kind-of (A-K-0) metal remover.
As soon as an object has been identified as lathe-20 for
example, it can inherit all the properties of its super-
classes lathe, metal-remover and CNC-machine. Inheritance
information can be found by searching the network and follow-
ing the A-K-0 links. For example, to find out what lathe-20
is "controlled-by" the A-K-0 links are traced to the CNC-
machine object node that has a relation showing that all
these machines are controlled by a computer. A simple
breadth-first search will accomplish this.

FIGURE 6 A SEMANTIC NETWORK REPRESENTATION

Default values can be expressed as in the case of the

number of axes that the milling machine and lathe have. Here
the lathe sub-class has a default value of two axes: however,
the metal-remover superclass has a default value of three
axes. This is the value inherited by the mill sub-class.
Semantic networks represent declarative knowledge con-
cerning a problem, the solution to which is found using gen-
eralized search procedures. The links in the network make
class-based inheritance opportunities explicit, thus facili-
tating the work of class-oriented inference procedures.

4.2 FRAME REPRESENTATIONS

Minsky [9] first proposed the frame representation of
knowledge. A frame is a collection of semantic network nodes
and links that together describe a stereotyped object, act or
 http://orion.journals.ac.za/

58

event. Stereotyping allows an element of predictability and

expectation to be built into the representation. A frame-has
slots that may contain default values, pointers to other
frames, sets of rules, or procedures by which values are
found (Harmon and King (10]).
Figure 7 shows part of a frame representation for
instructing a robot to move a block from position A to posi-
tion B. The frames are depicted as nodes in a graph.
The frame "MOVE-BLOCK" has two types of slot: the loca-
tion slot, representing an expected condition and containing
a value which gives the position of the block; and the if-
needed slot which calculates a value for the block's mass if
it is required. The latter slot filler is called procedural
attachment. This inclusion of procedures in frames joins
together the declarative and procedural representations of
knowledge.

FRAME MOVE-BLOCK FRAME CLEAR-BLOCK

TYPE EVENT TYPE EVENT
IF -PRESENT 1 BLOCK
LOCATION 1 POS-A
DESTINATION 1 POS-B
IF -NEEDED : BLOCK MASS VOL x DEN

t
FRAME PICK-BLOCK
--o
FRAME PUT-BLOCK
TYPE MOVE-BLOCK TYPE MOVE-BLOCK
GRIPPER 1 BIG GRIP HAND 1 CLOSED
HAND 1 OPEN ROBOT 1 IDLE
ROBOT • IDLE DO-ACTION 1 EXECUTE PUT PROG.
DO-ACTION 1 EXECUTE PICK PROG.

FIGURE 7 A FRAME REPRESENTATION

Solving a problem that is represented using frames

requires a search (depth-first, for example) to traverse the
problem representational graph. In this case however, a node
is only included into a problem solution if all the frame
slots are matched in the global data base.

4. 3 RULES
A rule-based system encodes only procedural knowledge of
the form:
 http://orion.journals.ac.za/
59

IF (condition] premise
AND (condition]
THEN (action 1] conclusion
and (action 2]
Rules are used with either attribute-value (A-V) or
object-attribute-value (0-A-V) representations. A-V and
0-A-V relationships are a specialized case of the semantic
network approach. Exotic links are removed leaving two
simple relationships. The attribute <- value link is an
"IS-A" link and the object -> attribute is a "HAS-A" link.

FIGURE 8 REPRESENTATION USING RULES

Rule-based systems have enjoyed the most success in the

construction of Expert Systems. MYCIN (in O'Shea (11]) uses
the 0-A-V rule representation and most PC based Expert System
shells use the A-V form of rule representation. An example
of an A-V rule is:
ATTRIBUTE PREDICATE VALUE
IF milling machine is requesting service
& correct gripper is attached to robot
& part location is input buffer
THEN message sent to robot is "EXE MOVE PART"
The above rule states exactly the correct action to be
taken if the premise is true. In some domains it is advanta-
geous to assign a certainty factor to the rule which repre-
sents the certainty of the inference contained in the rule
 http://orion.journals.ac.za/
60

(eg. MYCIN and PROSPECTOR Expert Systems). By using these

numbers it is possible to combine several sources of incon-
clusive information to form an almost certain conclusion.
Rule-based knowledge representations can be diagrammati-
cally represented with AND/OR graphs as shown in Figure 8.
Finding the solution a problem is then performed by searching
the graph in a forward chaining or backward chaining manner.

4.4 REPRESENTATION USING LOGIC

Real-world facts, relations and actions can be repre-
sented by statements written as well-formed-formulas (wff's)
in the First Order Predicate Calculus. Logic representations
provide a useful method of reasoning with knowledge namely,
theorem proving. There are two basic statements (wff's) in
logic programming:
A. - an assertion or piece of declarative knowledge
A:- Bl &... & Bn- which is a piece of procedural knowl-
edge stating that the sub-goals Bl to Bn must be
satisfied in order to satisfy the sub-goal A.
A fact can be represented in logic as a binary (two-
argument) relation. For example:
OBJECT(BOOK, RED)
states the fact that the object which is a book is the colour
red. Action representations have n-argument relations. For
example:
GIVES(BOOK, JOHN, MARY)
states that John carries out the action of giving a book to
Mary. A procedure can be written as
HAS(BOOK, MARY):-
OBJECT(BOOK, RED) AND GIVES(BOOK, JOHN, MARY)
which states that Mary has the book, which is an object,
after John carries out the action of giving the book to her.
There are a number of inferencing mechanisms which can
be used to infer new conditions from a knowledge base of log-
ical wff's. Modes Pones is one which uses the deduction that
if A is true and A implies B, then B is true. Resolution is
another inferencing mechanism which tries to prove that the
negation of a query statement is inconsistent with the facts
contained in the knowledge base. If it is inconsistent then
 -- l
http://orion.journals.ac.za/
61

the statement is true; otherwise, it is false.

Wff's representing a problem can be mapped onto an
AND/OR tree. The problems solution is then given as the
sequence of logical operators applied to the data during the
theorem proving process.

4.5 STRONGER PROBLEM SOLVING STRATEGIES

The control strategy determines the manner in which the
problem representational graph is to be searched for a solu-
tion. When dealing with complex problems a more efficient
search for a solution is facilitated by placing knowledge,
specific to the problem, into the production system rules or
operators. Heuristics are now coded into the procedural
knowledge rather than in the control part of the system.
Intelligent procedures working in conjunction with "weak"
search techniques prune the search tree rapidly and lead to
more efficient problem solving systems.
There are a number of ways in which a knowledge repre-
sentation can hold information about which procedures (opera-
tors or rules) to consider during the traverse of a solution
space :
1) Action-centred control systems have procedures that
know which other sub-procedures can be considered by the con-
trol system. MYCIN, for example, has rules that contain the
rule numbers of further rules that must be considered next if
the current rules premise exists in the data base.
2) Object-centred control systems use class descrip-
tions to determine which set of rules apply at any given
time. For example, if the problem is to determine which mode
of transport to use in travelling 50 km, then it is more
feasible to go by car, taxi or train and not by bicycle or
plane. A class of procedures is thus considered under cer-
tain conditions. GPS (General Problem Solver) uses this con-
trol.
3) Request-centred control systems have procedures
which know their own purpose and volunteer for consideration
when they can be used. This is used in blackboard systems as
in the Hearsay II Expert System.
 http://orion.journals.ac.za/
62

5.0 ADVANCED PROBLEM SOLVING - PLANNING

In order to solve most nontrivial problems, it is neces-
sary to combine some of the "weak" search methods with one or
more knowledge representation techniques. It is also neces-
sary to divide the large problem into smaller problems that
are solved separately. The separate solutions are then com-
bined to form the full problem's solution. This is unlike
the 8-puzzle presented in Part 1 where the problem is simple
enough to allow the complete state description at each node
of a graph. The methods used to decompose the original prob-
lem into sub-parts, and the ways of recording and handling
interactions between the sub-parts as they are detected, are
often collected under the heading of planning.
A linear planning system determines a complete solution
for a sub-part before considering another as used in the
STRIPS (Rich [5]) programme. For some problems, however, it
is necessary to perform some work on one goal, then on
another and then some more on the first. Systems which solve
these problems are called non-linear planners. If a planner
needs to solve particularly difficult problems it may be
advantageous to consider the problem at different levels of
abstraction. For example, when planning a route to get from
one city to a particular place in another city, it would be
best to consult a large scale national highway map initially.
The lower level detail of a street map for the city could be
consulted at a later stage. Systems that consider problems
in this manner are called hierarchical planners.
There are more detailed planning systems ~hich are spe-
cific to the particular types of problem that they solve, a
discussion of which is not necessary here. All planning sys-
tems, however, derive their power from a great deal of domain
specific knowledge. Due to this fact the most sophisticated
planners are found in Expert Systems where a constrained
problem is usually considered.

6. 0 CONCLUSION
Artificial Intelligence problem solving techniques pro-
vide a range of tools which can be combined in any number of
combinations to construct programmes for solving problems
 http://orion.journals.ac.za/
63

that do not lend themselves to algorithmic solutions. Most

real world problems fit into this category.
The most difficult part of applying these techniques is
in deciding which to use in solving a particular problem.
Initially, a problem representation - state-space or problem-
reduction - must be chosen to make explicit those. factors
that can be utilized when searching a solution space. For
small problems it is sufficient to use a simple knowledge
representation (not necessarily symbolic based) with a heur-
istically guided control strategy which will efficiently
search the problem graph. The larger and more complex the
problem, the more sophisticated the knowledge representation
scheme needed.
Most of the AI problem solving strategies have evolved
by being designed to solve specific problems which have
arisen in particular domains. As a result, it would be
highly unlikely that one type of knowledge representation and
control strategy would suit any other large, real-world prob-
lem. It is more appropriate then to look for similarities
between well established problem solvers and the problem to
be solved, and then to tailor the techniques accordingly.
Many strategy combinations could fit a problem to be solved;
however, as is the goal of AI researchers, it is necessary to
pursue that combination which will yield the most efficient
solutions in terms of computational time and resources.

REFERENCES
[1] A. BARR, E.A. FEIGENBAUM, "The Handbook of Artificial
Intelligence", Volume 1, William Kaufmann, 1981.

(2] M. BODEN, "Artificial Intelligence and Natural Man", The

Harvester Press Limited, USA, 1977.

(3] N. GRAHAM, "Artificial Intelligence. Making Machines

Think.", Tab Books Inc., U.S.A., 1979.

[4] P.H. WINSTON, "Artificial Intelligence", (2nd ed), Addi-

son-Wesley Publishing Company, Inc., U.S.A., 1984.
 http://orion.journals.ac.za/
64

[5] E. RICH, "Artificial Intelligence", McGraw-Hill, Japan,

1983.

[6] N.J. NILSON, "Principles of Artificial Intelligence",

Springer-Verlag, Berlin, 1982.

[7] T. WINOGRAD, "Frame Representations and the Declara-

tive/Procedural Controversy", in BORROW, G D (Ed),
"Representation and Understanding", Academic Press, New
York, 1975, pp. 185-210.

[8] M.R. QUILLIAN, "Semantic Memory", in MINSKY, M (Ed),

"Semantic Information Processing", MIT Press, England,
1986.

(9) M. MINSKY, "A Framework for Representing Knowledge" in

WINSTON, PH (Ed), "The Psychology of Computer Vision",
McGraw-Hill, 1975, pp. 211-277.

[10] P. HARMON, D. KING, "Expert Systems. Artificial Intelli-

gence in Business.", John Wiley & Sons, Inc., New York,
1985.

[11] T. O'SHEA, M. EISENSTADT, "Artificial Intelligence.

Tools, Techniques, and Applications.", Harper & Row,
Publishers, New York, 1984.