|Wisdom on Knowledge
Wisdom is an advanced knowledge
transmission tool that is developed and applied by Arthur Andersen to deal with three weak
points in current KBS-technology.
Weak points in Knowledge-based Systems
Business processes in trade and
industry such as diagnosis, monitoring, engineering and
planning, increasingly depend on rapid and easy processing of large amounts of complex
knowledge. Knowledge-Based Systems technology is considered an essential enabling
technology that will shape the pattern of computer applications in the emerging
There is, however, a widespread
consensus in the KBS-community that a successful deployment of KBS-applications requires a
more rigorous development process. Initiatives for improvements include the KADS-(II) and
VITAL ESPRIT funded projects.
1. Lack of a theory of the nature
Though considerable progress has been
made in tools and methodologies, the process of transferring knowledge to computer systems
for knowledge management purposes is still susceptible to major improvements. This can
partly be attributed to the phenomenon that current method ologies and tools lack an
underlying theory that defines the nature of knowledge. A theory of the nature of
knowledge can significantly advance the process of knowledge engineering. To pave the way
for such a theory, Newell introduced the knowledge level (1981): a separate computer
systems level that focuses purely on knowledge. At this level, user interface and
implementation aspects are irrelevant.
What did not happen yet, however, is
using the knowledge level to develop a full-blown
theory of the nature of knowledge, that, by explaining the fundamental nature of
knowledge, provides guidelines in the knowledge modelling process and that can form the
basis for the development of new and powerful methodologies, tools and languages.
1. Lack of an adequate
A second weak point in current
KBS-technology can be spotted at the level of knowledge representation languages. Business
practice reveals that industrial clients impose high demands on knowledge representation
formalisms. They require a representation formalism that:
- is easy understandable for laymen,
computer personnel as well as computers so that
knowledge represented can be discussed and validated on completeness, consistency and
correctness in comprehensive ways
- has a high expressiveness so that
many forms of knowledge can be expressed
- is very effective and efficient so
that knowledge can be represented in a computer system with a minimal effort
- is based on mathematical logic so
that knowledge can be represented in non-ambiguous ways
- is easy to adapt so that the
maintenance knowledge bases is facilitated
- stimulates the structuring process
- provides simulation facilities so
that the behaviour of the computer possessing knowledge can be directly shown
Analysis reveals that the
representation languages of current KBS-Technology do not yet cope with all these
requirements. The languages made available by the KADS(-related) research, e.g. the CML
(informal language) and (ML)2 (formal language). are not easy to learn, understand and
manipulate for laymen.
3. Lack of adequate tools
A third weak point is the lack of tools
that provide advanced management facilities centred around a representation formalism that
can cope with all the requirements above. Important here is that these facilities should
be based on new graphical intuitive facilities and that they enable platform-independent
and network-independent access to knowledge.
Wisdom: A New Functional Direction in
To deal with the first weak point, the
development of Wisdom is based on the functional theory of the nature of knowledge. In
theories on the nature of knowledge concepts play an important role as classificatory and
storage mechanisms. A concept has an intension and an extension. The intension of a
concept is a set of constraints (or conditions) that should be satisfied by an object to
belong to the class covered by the concept. The intension refers to an object-type. The
extension of a concept consists of the set of objects complying with the object-type.
Knowledge-based systems perform a
matching task: they all apply constraints (the object-type) to a set of real-world
referents (the objects) to obtain a match. An object matches if it can be classified as an
object-type. As we can see from Figure 1, the description of the object-type and the
objects contains abstraction mechanisms such as generalisation, specialisation and
Figure 1: Matching Object-types and
There are several basic views on how to
reconstruct the object-type of a concept, one of which is the functional view (Hendriks
1986; Lucardie, 1994; Reitsma; 1990; Van Der Smagt & Lucardie, 1991). Two important
characteristics of the functional approach are the explicit account for goals and
The goal-oriented selection principle
The reconstruction of an object-type
normally leaves us with a sheer enumeration of conditions that appears intractable. A
selection principle is needed. The functional approach operationalizes a
selection-principle by assuming a goal or function of classification (Stepp &
Michalski, 1986). An important aspect of this approach is the emphasis placed on the role
of functions that should be performed by objects. These functions direct the select ion of
conditions of an object-type and consequently determine the relevant attributes of
Central in the functional view is the
notion of functional equivalence: the phenomenon that objects are identical, fall in the
same concept or are similar if they possess, even quite different, attributes to perform
the same function. The functional view emphasises that an object-type of a concept is
established by a goal-oriented reconstruction process wherein possibly different objects
are functionally similar. Functional equivalence can be traced back to three basic
mechanisms: (1) conditional relevance, (2) conceptual interaction and (3) variation
limited to goal-constructed categories.
Ad 1. Conditional Relevance
The first mechanism of functional
equivalence refers to the phenomenon that, under specific conditions, other attributes
(descriptors) may become important for determining class membership. Thus, conditional
relevance denotes the phenomenon that conditions of an object-type are not always
relevant. On the contrary, their relevance is conditional upon circumstances, which need
to reconstructed and incorporated in the object-type either.
Ad 2. Conceptual Interaction
Categorisations of attributes of
objects may influence each other. This phenomenon is called conceptual interaction. It
manifests itself in the mutual influence of the categorisations of the attributes .
Ad 3. Variation limited to
The third phenomenon contributing to
functional equivalence refers to the situation that objects may have different attribute
values, but that this variation is limited to, or falls within, goal-constructed
To tackle the second weak point, Wisdom
offers an integral knowledge representation language that consists of decision table
systems, frame systems, Wislogic and Prolog in order to obtain the desired knowledge
representation language that is also compatible with the functional theory
Decision table systems
Decision tables form a goal-oriented
language. Variables and the relationships between variables are define d from the
viewpoint of a particular goal. The goal in the car selling example (Figure 2) is to
define a suitable car for a particular buyer. It appears in practice that goal-oriented
modelling of knowledge most of the times leads to occurrence of conditional relevance,
conceptual interaction and functional equivalence. This is shown in the example below. The
table suitable car is defined from the goal suitable car.
Figure 2: The decision table suitable
Decision tables can represent
conditional relevance. If, for example, a car dealer is selling a car to a customer and
the customer requires a sun roof if the cost of the car exceeds $20,000, the variable sun
roof is a conditional relevant variable. The relevance of this variable depends on the
value of the variable cost. If the cost of the car exceeds $20,000 an additional variable
sun roof becomes relevant.
Decision tables can also reflect
conceptual interaction. If, for example, a car dealer is selling a car to a customer and
the customer requires a cost that doesn't exceed $20,000, but he is willing to raise his
cost to $25,000 if the car has at least 4 doors and even to raise his cost to $30,000 if
the car has 5 doors, there is conceptual interaction between on one hand the variable cost
and on the other hand the variable doors. It's impossible to define here what the customer
requirement for cost is without taking into account the value of the variable doors.
A decision table clearly displays
conditional relevance and conceptual interaction. But it also displays another important
mechanism: variation limited to goal-constructed categories.
The decision table Suitable car
implies that a car of $22,000 with a sun-roof and 4 doors is equivalent to a car of
$24,000 with a sun roof and 5 doors. Both are suitable cars. This doesn't mean that the
cars are equivalent in all aspects, but that they are equivalent from the viewpoint of
suitability for a particular buyer.
In the majority of applications, we
need a decision table system: a set of at least two decision tables in which each decision
table is linked to another table belonging to the same system. We can distinguish two
types of links. The first type of link is established by the phenomenon that a condition
subject with at least one of its alternatives of one decision table occurs as an action
subject with corresponding action alternatives in another decision table. The first table
is then called a head table, the second table is called a condition subtable.
The second type of link is established by the phenomenon that an action subject and one of
its alternatives of one decision table occurs in the same form in another table. The first
table again is called a head table, the second table is now called an action subtable.
Usually, the action subject in question, in this action subtable, is further
specified by means of other action subjects.
Frames can be used to store a
structured representation of objects that should match certain object-types. Frames form a
powerful language to store structured representations of complex objects. The variable
values stored in frames can be used in a goal-oriented knowledge representation language
to derive goal-oriented conclusions. In this sense decision tables and frames are
languages complementary to Prolog. Their mutual power is indispensable in the development
of effective and high quality knowledge-based systems.
Wislogic is an additional logic
language that can be used in decision table systems and frame systems as a complementary
language. Wislogic offers possibilities to retrieve values present in the knowledge base
and to define operations over sets.
In Wisdom can be used as a conceptual
modelling language for the representation of knowledge in addition to decision
tables systems, frames-systems and Wislogic. By this, a number of weaknesses of Prolog are
avoided: Prolog does not provide guidelines for knowledge modelling. Prolog is just a
language specifically suited to the specification of knowledge, the gap between the
declarative and the procedural semantics of Prolog, it is difficult to assess the
knowledge import of Prolog code.
To show how we tackled the third weak
point, lack of and adequate tools, we shortly discuss three characteristics of Wisdom
Specification Facilities: The
Wisdom is equipped with a graphical
toolbox that provides a multitude of convenient graphical facilities for drawing
decision tables and frames on the screen. This toolbox has a carefully designed
multi-window, menu-driven mouse-oriented interface for optimal communication with users.
It allows users to quickly reconstruct and modify complete decision table system and frame
systems. The toolbox shows 'intelligent behaviour': it knows what a correct table or frame
looks like and applies this knowledge to support the user in the drawing process. The
editor also knows how to obtain a decision table or frame that occupies a minimal amount
of space. This knowledge is used, when necessary, to calculate the measures of a minimal
table and frame after a user's action.
Furthermore, Wisdom offers facilities
to specify knowledge models through Wislogic and Prolog. This is necessary to insert
knowledge that cannot easily be represented through decision tables or frames. These
statements can be used while simulating specified knowledge.
Simulation facilities: the
The inference machine of Wisdom is a
flexible goal-oriented engine. It departs from a list of goal variables (variables are
conditions or actions in a decision table or attributes of an object represented in a
frame or other variables occurring in a field of application). The inference machine
attempts to trace these goals. The inference machine of Wisdom uses a backward chaining
strategy, though it is possible to influence the inference machine. The inference machine
is also able to make connections with other information systems to retrieve values.
Wisdom knowledge bases can be consulted
under several operating systems such as PowerPC, Windows and Unix. A Wisdom knowledge base
is also accessible through internet.
Wisdom now is entirely developed in
Prolog. We are conducting research to use Java for the user interface to increase the
platform independency of Wisdom. The number of projects in various sectors of the economy
in which Wisdom is being applied by Arthur Andersen is rapidly growing.
Hendriks, P. H. J. (1986). De
Relationele Definitie van Begrippen. Een Relationeel Realistische Visie op het
Operationaliseren en Representeren van Begrippen. Dissertation, Nijmegen.
Lucardie, G. L. (1994). Functional
Object-types as a Foundation of Complex Knowledge-based Systems. Dissertation Computer
Science, Technical University Eindhoven.
Newell, A. (1981). The Knowledge Level.
AI Magazine, 1, 1-20.
Reitsma, R. F. (1990). Functional
Classification of Space: Aspects of Site Suitability Assesment in a Decision Support
Environment. Dissertation, University of Nijmegen, International Institute for Applied
Systems Analysis, Laxenburg, Austria/Faculty of Policy Sciences.
Stepp, R. E., & Michalski, R. S.
(1986). Conceptual Clustering of Structured Objects: A Goal-Oriented Approach. Artificial
Intelligence, 28, 43-69.
Van Der Smagt, A. G. M., &
Lucardie, G. L. (1991). Decision Making under not Well-Defined Conditions: From Data
Processing to Logical Modelling. Journal of Economic and Social Geography, 82(4), 295-304.
Arthur Anderson & Co
3001 AX ROTTERDAM
Tel: 010 242 1400
Fax: 010 242 1616