KNOWLEDGE-BASED MODELLING
Andreas M. Kleinhans
Betriebswirtschaftliches Institut
University of Stuttgart
West Germany
Abstract
Knowledge-based modelling consists of qualitative models which are implemented in a
hybrid, rule- and frame-oriented programming language. Qualitative models achieve
flexible and detailed description of both the simulation entities and relationships. This
Paper presents a simulation rt system which is based on a multi-level representation
scheme of causal diagrams. It offers a qualitative "presimulation", that is conclusions about
the sensitivity of the elements, the restrictions the possible behaviour of the model. It
enables explanation of the various implicit structural and dynamic relationships and the
user to be guided to efficient quantitative simulation.
1, Introduction
In traditional simulation the kind and quality of effects are usually perceived only by
several appropriate simulation runs /1/. oth the model structure and namic behaviour
require the interpretation of the (human) model user. Complex models claim high model
knowledge both on the model builder and user and there is need for detailed descriptions
and explanations on every level. For that reason the acceptance of traditional si ition.
systems is still very low /2/.
Many authers express the urgent demand to simplifiy the model use and evaluation, to
support the validation process and to guide the user to efficient simulation runs /3/. The
user should be able to concentrate on the model and to ask the system. uestion about
the structure and dynamic of the model. This requires a user friendly di: ¢ module,
ape procedures and explanation utilities during every modelling amd simulation stages
The advantages of Artificial Intelligence (AI) techniques for simulation have been widely
discussed /516/ and several useful combinations have been presented /7/8/9/. This paper
shows how knowledge-based, rule and frame oriented techniques support jitative struc-
tural descriptions of models and how they enrich traditional modelling by representation
of application oriented knowledge.
528
2. BAMBOO II
As a basis for the discussion the hybrid prototype system BAMBOO II is introduced. It has
been developed at our institute as a general simulation expert system demonstrating the
usefulness of object oriented techniques for simulation /9/. We got most experience with
models representing flexible assembling systems, however the aspects of qualitative
modelling are fundamental.
BAMBOO II offers useful and detailed model knowledge to the user in order so puice him
to more efficient simulation. That is the system behaves like an experienced model expert.
Its knowledge consists of diverse structural and Gyaamic peculiarities such as statements
about critical paths, the sensitivity of elements, relations and loops, effects of restrictions
and the possible behaviours resulting from different initial states.
The main objects are:
1, Representation and processing of contextual knowledge of the simulation entities
2. Qualitative "presimulation" of relationships and restrictions
3. Disclosure of application oriented backgrounds
4, Representation of generic structures
5. Handling diverse models simultaneously
In order to improve the transparency of models the system must be able to answer all
estions about the structure and ic interactions. That requires a capability of
letailed element, causal path and loop analysis including the determination of the sensiti-
yy of all entitities. So-called "presimulation” is based on the model structure and provides
inference about the structural pecularities and the possible behaviour of the model /10/.
Direct decision support is offered Dy processing qualitative, application oriented informa-
tion and knowledge, giving semantical descriptions and explanations and offering feasable
polices.
Often there are several versions of the same model, i.e. the same models with different
constant values and initial states. Or, in a more complex situation, there are different, but
close models. In order to handle diverse models appropriately, the system must build gene-
ric models, which keep the knowledge of what they share in common.
3. Datastructures of the model entities
Like in System Dynamics the basis for the model structure are causal diagrams. Though in
System mics the model. structure is appropriated by transforming it to the flow
diagram, in BAMBOO II the causal diagram itself is processed and enriched by contextual
knowledge. Furthermore the model may be viewed in a hierarchy of different levels. They
either consist of a structural order of element-, path and loop objects or of application
oriented strategy and policy objects. Every objects holds its own particular attributes.
Frames are very useful datastructures for representing objects. In general they consist of a
frame name, references to superclasses and slots (attributes) which store the properties of
the object. In theory there are several statements and functions ("flavours") for every slot,
like the specification of value-sets, default values, constraints and "demons", which are
automatically activated in reading, changing or removing the slot-values /11/12/.
The most important structural frames are the element, relation, path and loop frames (see
fig. 1). Elements represent the real objects, relations their causal connections, paths
529
provide chains of elements and relations and loops represent circular paths. Every frame
owns different attributes, which however are often based on each other.
RELATION
type: (system/processibit) type: (matiintfinit)
focus: (target/source/viewed/hidden) pred:
value: suce:
value-min: dir: (plus/minus)
value~max: delay:
dimension: force:
rel-force:
constraints:
pai enessteninnsne| class: path
dir: (positive/negative) type: (core, wide)
delay: stability: (stable/unstable)
force: dir:
delay:
force
Fig. 1 Basic frames of the causal diagram
The element type indicates the nature of the element: it may be a system, a process or a
bit (of information). They roughly correspond to the System Dynamics concepts of level,
rate and auxiliary respectively. The nature of a aystem is usually (but not necessarily)
material, and that of a bit is always immaterial. The focus attribute determines how
elements are used respecting the users interest. Target elements are those which are after
all important for the user, but which he cannot manipulate directly. Source elements are
those which can be changed and which act like operators, i.e adjustable parameters. The
attribute value "viewed" makes an element transparent to the user during the session, i.e.
only viewed elements are mentioned in a system dialogue, others are omitted. Targets and
sources are always viewed.
Arelation of the type information indicates an information flow, that of the type material,
a material flow. All types are analogical to the elementtypes. Usually an information flow
leads to an information bit, a material flow to a system and an activation starts a process.
But different combinations are allowed and there are as many interpretation rules as
combinations possible. The pres/succ attributes store the predecessor and successor of a
relation. These statements hold the basic structural information of the model. The dir
(direction) property indicates whether the relation influence is positive (i.e. in same direc-
tion) or negative (i.e. in opposite direction). The delay attribute stores the time lag, the
force attribute its strength and the rel-force attribute, its influence relatively to other
relations with the same successor. The force may be strengthening (+), steady (=)
weakening (-) or a qualitative change (#). Finally constraints store the description of all
restrictions which limit the relation.
530
"Qualitative change" signifies a non-linear, possibly discontinuous and only vague
formalizable relationship.
Note, that type attributes of the both the elements and the relations refer to corresponding
miperciassee In these classes general information is stored, e.g. the fact that a information-
relation always succeeds without any delay, that the strength of a material-relation must
not be strengthening or that the minimum value of a system must be always equal to or
greater than zero.
Paths and Foops are based on elements and relations. In general they describe the
property sum of their components. They especially refer to the attributes "dir", "delay" und
"force", however the interpretations are different. While a "negative" direction of a path
indicates an opposite effect between the first and last element, a "negative" direction of a
loop shows the loop to be goal-seeking. Furthermore it proved to be useful to distinguish
between relevant ("core") and less relevant ("wide") loops. Finally they consist of an attri-
bute which describes their stability.
The main advantage of the object-oriented representation is the easily modification and
expansion of the model entities. The modification can take place at every time even dyna-
mically during the simulation run. Note that the same attributes in different objects may
et a different interpretation, too. Furthermore the knowledge based simulation procedure
as several advantages. The relations need not be implemented by numerical equations
but can also consist of qualitative rules. Moreover the simulation process may be triggered
by forward rules which are implemented as "demons" and are activated automatically.
4, Structural Analyses
According to the model objects there exist several structural analysis procedures of the
causal diagram. They consist both of elementary anaylses of elements and relations and of
more complex evaluations of paths, path sets, loops and the total model. The main
features are:
-- determination of the peculiarities of the elements and relations
-- description of the model dynamic
- generation and evaluation of any path/loop
-- determination of the most/least sensitive element/relation in the path/loop
respecting dir/delay/force/rel-force .
-- determination of the most/least sensitive path/loop of the model respecting
dir/delay/force
-- evaluation of critical restrictions in paths/loops
-- description of the elements/relations of nested paths/loops
Element and relation analyses primarily deal with evaluations of predessors and succes-
sors.’Path and loop analyses offer an evaluation of every detail in the chain including the
disclosure of all restrictions. Fig. 2 shows example protocols of a given path. The first
example demonstrates an ex-ante-analysis ("presimulation") of the path, the other example
an evaluation at a certain time during the actual simulation. For instance the causal step
between production and supply is indicated as in the same direction (+) with a delay
between 4 and 6 periods and a steady force. At time 22 the-force counts 60% relatively to
other influences on supply.
531
“presimulation" mode:
dir delay force
production
+ 4/6 = — supply
+ Q = aiff
+ 1 a marketing
+ 2 é demand
+
7/9 «= total
actual path analyis at time 22:
dir delay force force
‘production
+ 6 = 60 supply
+ oO = 40 diff
+ 1 # 100 marketing
+ 2 # 15 demand
Fig. 2 Path analyses examples
Loop analyses are comparable to path analyses. They consist of some additional analysis
procedures which deal with the different ‘ypes of loops. E.g. there are local adjustment
loops, which only act as a delay and which often may be simply replaced by a constant. On
the other side the importance Kforce) of a loop relative to others can generally be only
determined by appropriate simulation runs.
Because of their complex combination possibilites path and loop analyses proved to be
very unwieldy. In particular the compact and user friendly representation and explanation
of the results are very difficult. However, dealing with high levelled applications, like stra-
tegic management problems, the most efficient information for the user was given by
disclosing only elementary relationships and inferences. That is because strategic problems
consist rather of little but contextually compact data.
5. Application oriented objects
Causal diagrams are a very convenient for modelling real structures and dynamic interrela-
tions. Unfortunately elements and their relations are often too complex to be appropria-
tely represented by simple causal objects. Especially when the entities of material and
information flows are not homogeneous, or when different attributes of the flow objects
are addressed, a well-suited representation is essential. In general a causal diagram like in
fig. 3 is intuitively well understood, however the relations are too inexact and looking
closer may lead to some severe misunderstandings. While the supply-sale and demand-sale
relation addresses the same object, that is the product in total, the sale-profit relation only
points to one aspect, that is the product price. The connection sale-satisfaction is even
more complex, it refers to the product quality, which consists of all product attributes
itself.
532
supply demand
wees ft
i
profit tee sale —e-> satisfaction
product
function: product quality|
services: function:
delivery period: |) | services:
aa delivery period:
price:
price:
Fig. 3 Product and quality objects
While traditional simulations have to use diverse tricks to represent such problems, the
object-oriented techniques allow an easy and application-oriented way by selecting the
relevant attributes directly from the object frames. By this means all model entities are
appropriately representable.
6. Application oriented policies
Policies are formed during an analysis of the entire model and are based on the elemen-
tary analyses. The questions are global and require distinguished knowledge about all
model pecularities:
- Which external parameters are how sensitive respecting which target variables?
- Which restrictions must be changed in which way to pass which bottle-necks?
According to the overall scanning of the diagram structural analyses usually offer a
huge number of policies. However not the possibilities of the model on principle but
rather the realizable strategies and their appropriate formalization are decisive for the
user.
Like the differentiation of the flow objects, a suitable refinement of policies and strategical
peace is necessary. E.g. usually the goal hierarchy of a company consists of multiple
levels and strategic meanings. The formalization into frames can be directly performed.
Fig. 4 shows an example hierarchy of an existing company which is operating in three diffe-
rent markets.
533
Total
Market 1:
Market 2:
Market 3:
Main Targets
Startegic | deivery gered
noe |e — Ge
{monopoi)
machine capacity: --
Operators personal capacity:
production strategy: +
number of variants: -~
number of set-up:
cost-price-rate:
price:
Fig. 4 Example goal hierarchy
The main goal naturally focuses on the profits which usually cannot be influenced directly.
Nor can the subgoals which are expressed by the strategic success factors for each market.
The strategy frame finally focuses on the operators which can be adjusted in users seality
Every attribute of the strategy frame points to the individual operator frame which holds
the description of initial conditions, restrictions, operation and sensibility.. This semantical
net supports the user in searching efficient and realizable strategies.
Besides the necessity for representing goal hierarchies there is a need for representing
different goal-oriented model versions, that is representing knowledge about the incorpo-
rated strategies of a model. Models act differently depending on which strategies are
already implemented. In fact, to compare diverse policies they must rely on the same
(base) model. This kind of knowledge is stored in a generic model whereby different
instances of the model represent model versions with different instances.
7. Restriction handling
Restrictions are of different natures, too. They appear in rule conditions, as range limitors
or simply as an evaluation constant. On principle all parameters in a simulation program
are changeable, so technically there is no differentiation between the restrictions either.
However in real systems not only the knowledge about the existence of restrictions is
534
important but also information about their status. Are they physical, technical, organisa-
tional or even political in nature? Restrictions or constraints might be passed by some
actions not mentioned yet. Therefore it is useful when the system may question the restric-
tions found during the structural and strategical analyses.
In BAMBOO II parameters, values and rules which are not definite for the user can be
marked as “situational”. Also a differentiation in diverse classes of restrictions is projected.
Furthermore there is the Possibility of storing knowledge about the consequences of the
hyphothetical violation of restrictions and to keep it on hand for explanations. These
features allow an adequate restriction management required by real problems.
8. Summary
It has been shown how object-oriented Al-techniques offer advantages for the differentia-
tion and analysis of the model structure. Several hierarchies of object classes have been
presented. Especially application-oriented refinements allow the representation of
contextual knowledge which offers a distinct support in searching efficient and realizable
strategies.
9. References
1/ Shannon, R.E. Mayer, R. und Adelsberger, H.H. (1985). ert Systems and
m Simulation. In: simulation 44:6, 275-284. e (88) Ep ”
/2/ Richmond, Barry (1985), STELLA: Software for Bringing System Dynamics to the
other 98%. In: Proceedings of the 1985 Internation: Conference ‘of the System
Dynamics Society, Keystone, Colorado, 706-718.
/3/ Young, David F. (1985). KBSIM - A Knowledge based Tool and its Use in Model
Preprocessing. In: Proceedings of the International Conference of the System
Dynamics Society, Keystone, Colorado, 1070-1080.
/4/ Kerckhoffs, E.J.H. und Vansteenkiste, G.C. (1986). The Impact of advanced
Information Processing on Simulation - An illustrative Review. In: Simulation 46:1,
/5/ Gren, Tuncer I. und Zeigler, Bernard P. (1987). Artificial intelligence in Modelling
and Simulation. In: Simulation 48:4, 131-134.
/6/ Reddy, Y.N.R. (1987). Epistemology of Knowledge Based Simulation. In:
Simulation 48:4, 162-166.
/1/ O’Kneefe, Robert (1986). Simulation and Expert Sytems - a Taxonomy and some
Examples. In: pert (986). 46:1, 10-16.
/8/ Moser, Jorge (1986). Integration of Artificial Intelligence and Simulation in a
Comprehensive Decision-Support System. In: Simulation 47:6, 223-279.
‘9/ Kleinhans, Andreas (1986). A Behavioral Analysis rt System for System
f Dynamics Models. need et al. (Hrsg., 1986), 1039 - ioe
/10/ Kleinhans, Andreas (1989). Wissensverarbeitung im Management. Frankfurt: Lang.
/11/ Harmon, P. and King, D. (1985). Expert Systems - Artificial Intelligence in
Business. John Wiley & Sons.
/12/ Puppe, Frank (1988). Einfiihrung in Expertensysteme. Berlin: Springer.
