Teaching Structural Knowledge in the Control of Dynamic Systems:

Direction of Causality makes a Difference

Wolfgang Schoppek (wolfgang.schoppek@uni-bayreuth.de)
Department of Psychology, University of Bayreuth
D-95440 Bayreuth, Germany

Abstract
Recent publications about humans controlling dynamic
systems have emphasized the role of specific rules or
exemplar knowledge. Although it has been shown that
small systems can be controlled with these types of
knowledge, there is evidence that general knowledge about
the structure of a system plays an important role, too,
particularly when dealing with systems of higher
complexity. However, teaching structural knowledge has
often failed the expected positive effect. The present work
investigates details of acquisition and use of structural
knowledge. It is hypothesized that guiding subjects to focus
on dependencies rather than effects supports them in
applying structural knowledge, especially when the
application is practiced in a strategy training. An
experiment with N=95 subjects supported the hypothesis of
the usefulness of the dependency perspective, but revealed
an adverse effect of the strategy training. Differences
between subgroups studying different majors have been
found that give rise to questions about the relation between
prior knowledge and instruction. The results have
interesting implications for models of how structural
knowledge is represented as well as for methods of
teaching system control efficiently.

Humans have to deal with dynamic systems throughout

their lives. Especially in industrial environments, people
are confronted with new systems such as production lines
frequently. Therefore it is worthwhile to study how
humans learn to control dynamic systems, and how
instruction can support the learning process.
In cognitive psychology, a common paradigm for
studying the control of dynamic systems can be characterized by the following features: The systems simulate
some fictitious device or environment that most people
have no specific experience with (e.g. a tank with sea
animals in a biology lab, used by Vollmeyer, Burns, &
Holyoak, 1996). This is to ensure an equally low level of
prior knowledge. Discrete linear additive equations are
used for simulation, one equation per output variable.
There is the opportunity to assign values to the input
variables in each simulation step, which is referred to as
trial. A number of trials, e.g. six simulated hours, make

up a round. The objective for participants is to attain a

specific goal state either at the end of a round, or as soon
as possible and to maintain the state. A prominent, yet
simple example is the Sugar Factory (Berry &
Broadbent, 1984) that has been used to investigate
questions about implicit vs. explicit knowledge and about
rule vs. exemplar learning (e.g. Dienes & Fahey, 1995;
Fum & Stocco, 2003; Lebiere, Wallach, & Taatgen, 1998)
Research with this paradigm has shown that subjects
largely prefer acquiring and using exemplar knowledge
rather than structural knowledge, i.e. subjects memorize
specific actions taken in specific situations together with
their outcomes. This strategy can be successful under
certain conditions: First, when the system has a small
problem space like, for example, the Sugar Factory (144
different states); second, when the same goal state has to
be attained repeatedly (Vollmeyer et al., 1996), which
means that only a small fraction of a possibly large
problem space is relevant. Simulation studies with the
Sugar Factory have shown that it can be successfully
controlled by using either declarative representations of
specific actions (Lebiere et al., 1998), or learned
production rules that also represent specific interventions
(Fum & Stocco, 2003). In conditions, however, where
subjects have to deal with huge problem spaces (e.g.
because the system is more complex and subjects have to
attain a number of different goal states), the exemplar
strategy is no longer useful1. Instead, it is more reasonable
to use general knowledge about the causal structure of the
system to navigate through the problem space. I will refer
to this type of knowledge as structural knowledge.
In principle, complete structural knowledge is sufficient to control a system even without specific experience. Although correlations between structural knowledge
and performance have been reported (Funke, 1993),
experiments where structural knowledge was taught,
usually failed to demonstrate its superiority (Putz1

The inclination to use exemplar knowledge even when it is

inappropriate may explain why subjects generally perform at
very low levels when they are asked to control complex
dynamic systems that are new to them.

Osterloh, 1993; Schoppek, 2002). One reason for this is

that deriving specific actions from structural knowledge is
a skill that has to be practiced in addition to learning the
structure. This view is corroborated by results from
studies where the application of structural knowledge has
been practiced extensively (Preussler, 1998). A second
reason for the difficulties of applying structural
knowledge is that knowledge about causal relations is
acquired under a different perspective than it is applied
when controlling a system. This issue is elaborated in the
following paragraphs.
Verbal protocols of successful system controllers and
simulation studies (Schoppek, 2002) have helped identify
efficient strategies for acquisition and application of
structural knowledge. A good strategy for exploring the
causal structure of a system is to vary input variables one
at a time to identify the immediate effects of the input
variables and the momentum of the system, which is
produced by effects of output variables onto each other.
For example, a subject could put some lime into the
animal tank to observe the effect onto the oxygen content
of the water, then set lime input back to zero and observe
how the oxygen content changes on its own.
A common application strategy starts with (1) predicting the next state of the system under the assumption of
no interventions, continues with (2) calculating the differences between the predicted and the desired state, (3)
selecting a free input variable, (4) calculating the input
value, and ends with (5) applying the intervention. In the
course of this strategy, for each output variable all their
dependencies are considered in turn. This consideration of
dependencies is a marked difference compared with the
focus on effects that is prevalent during acquisition of
structural knowledge.
Thus we can distinguish two perspectives on causal
relations: One looking for effects of a given cause, the
other looking for possible causes of a given effect. The
first perspective is prevalent during exploration of a new
system, the second is more adaptive during system
control. In the following, I will use the word effects to
characterize constructs related to the first perspective, and
dependencies to characterize the second perspective.
The distinction of perspectives on causal relations has
a number of implications. The first has to do with the
question what given information cues the retrieval of
what other information. During exploration, when input
variables are manipulated and effects are observed,
associations from cause to effect are learned, resulting in
a structure where representations of input manipulations
act as cues for representations of changes in output
variables. When the task is to control a system and the

dependencies of output variables are considered, output

variables should be learned as cues for input variables.
A second implication concerns the mechanism of
chunking, which plays an important role in successful
problem solving (Newell, 1990, Gobet & Simon, 1996).
The effect perspective suggests chunking together single
effects of a variable (which can be an input or an output
variable), whereas the dependencies perspective suggests
chunking together all dependencies of an output variable.
Again, the second possibility seems to be more adaptive
in system control, because having all dependencies in one
chunk relieves the problem solver from extensive memory
search, a process that consumes much time, poses high
demands on working memory, and is thus error prone.
A second issue in the context of helping humans to use
structural knowledge has to do with strategy instruction.
Undoubtedly, extensive practice under supervision of
experienced operators is effective, but also very costly.
Thus it is important to find ways of getting leverage from
structural knowledge efficiently. The way followed here
was to base a training program on a strategy that has
proven successful in a computer simulated cognitive
model of controlling a system similar to the present one
(Schoppek, 2002).
To summarize, the aim of the present work is to
investigate ways of teaching structural knowledge about
dynamic systems, either indirectly by manipulating the
perspective on causal relations, or directly by practicing
the application of structural knowledge. Specifically, I
tested the hypothesis that guiding subjects to focus on
dependencies rather than effects enhances performance.
By measuring access to causal knowledge with a speeded
judgment task I investigated if the different perspectives
are also reflected in the representation of structural
knowledge. The results may show new ways of teaching
structural knowledge and extend our understanding of the
use of this type of knowledge.

Experiment
The system I used in this experiment is a simulation of the
influences of three fictitious medicines onto the levels of
three fictitious peptides in the blood. The medicines are
called MedA, MedB, and MedC; the peptides are called
Muron, Fontin, and Sugon. The effects of the substances
onto each other are simulated with the following discrete
linear equations:
(1) Muront = 0.1 Muront-1 + 2 MedAt
(2) Fontint =
Fontint-1 + 0.5 Muront-1 0.2 Sugont-1 + MedBt
(3) Sugont = 0.9 Sugont-1 + MedCt

In a neutral state with Muron = Sugon = 0 and Fontin = x,

the system is stable. Once some of the medicines are
administered, the system gains momentum. Note that the
amount of Fontin in the blood can only be reduced
through Sugon, which depends on MedC. Since Sugon
decomposes slowly, large time delays of changes in
medication have to be dealt with. Subjects interacted with
the system through an interface consisting of two tables
showing the states of the variables in all trials, and input
boxes where they could enter values for the medicines.
One round comprised six trials, introduced to the subjects
as simulated hours.
Structural knowledge was tested with a speeded causal
relation judgment task. All names of input and output
variables were shown on a screen in a spatial arrangement
that matched that of the simulation interface. This was
done to assure that variables could be identified by both,
their names and their locations. Then the name of an
output variable was highlighted on the right side of the
screen, followed by the highlighting of another variable
name on the left side with an ISI of 500 ms. The subject
was asked to respond with pressing one of two keys as
quickly and accurately as possible to indicate her
judgment if there was a causal relation between the
highlighted variables or not. All 18 possible input-output
and output-output relations were shown in one test. Eight
of these relations had to be answered with yes, 10 with
no. The procedure was arranged such that knowledge of
dependencies should result in faster judgments compared
with pure knowledge of effects. This is expected because
the variable that was highlighted first (the effect) is
assumed to act as a prime for the variable highlighted
second (the cause) only when causal relations have been
memorized under the perspective of dependencies.
Subjects and Design
N=95 subjects, studying different majors at the University
of Bayreuth, participated in the experiment. Subjects were
paid 10 for their participation.
The factor type of knowledge with the levels
knowledge of effects (Eff) and knowledge of dependencies (Dep), and the factor strategy training with the
levels no training and training were varied between
subjects. A third, quasi-experimental factor field of
study with the three levels arts/humanities,
law/economy, and science was also analyzed. In principle, subjects were randomly assigned to one of the four
conditions. A few exceptions from complete randomization were due to the objective to have approximately
equal distributions of field of study in each condition.

Procedure
The experiment began with a general instruction about the
system. All subjects went through a standardized
exploration phase guided by the experimenter. The
exploration was designed to demonstrate all causal
relations between the variables of the system. Subjects
were guided to analyze the observed effects and asked to
enter them in cards provided by the experimenter. The
procedure in this phase was different for the two
knowledge conditions: In the Dep condition, the
experimenter consistently asked for dependencies, and the
cards were sorted by the dependent variables Muron,
Fontin, and Sugon. In the Eff condition, the experimenter
consistently asked for effects, and the cards were sorted
by the independent variables MedA, MedB, and MedC.
At the end of this phase, the experimenter examined the
knowledge of the subject orally, again consistently asking
either for dependencies or for effects. Subjects had to
recall all possible relations with the respective numeric
weights before moving on to the next phase (all subjects
achieved that).
Subjects in the no strategy training condition could
then explore the system for one round (six simulated
hours) on their own. Subjects in the strategy training
condition went through a number of exercises where they
practiced a method of predicting future states of the
system. As mentioned above, this was the first part of a
strategy tested earlier in a cognitive model. Only a part of
the complete strategy was selected to keep the training
short. Nevertheless, all effects (condition Eff) or dependencies (condition Dep) were needed and rehearsed in
these exercises.
Next, all subjects were given the control problems. All
problems comprised six simulated hours and were given
with the objective that the goal states had to be reached as
soon as possible, and to be maintained. Table 1 shows the
initial states and the goal states for the four control
problems. Initially, all variables except Fontin were zero.
In order for the subjects to familiarize themselves with the
control task, they were given two rounds for Problem 1.
Table 1: The four control problems given to the subjects
Problem 1: Fontin = 50
Problem 2: Fontin = 900
Problem 3: Fontin = 2000
Problem 4: Fontin = 50

Muron = 200 , Fontin = 1000

Muron = 100
Fontin = 1000
Muron = 400 , Fontin = 900

Results
To measure control performance, the solution error was
calculated by summing the natural logs of the absolute
differences between the goal values and the actual values

Table 2: Solution error of system control in the various

conditions of the experiment

Strategy
training

Eff

Dep

yes

2.9 (1.5)
n = 24

2.4 (1.5)
n = 26

2.6 (1.5)
n = 50

no

2.4 (1.5)
n = 22

1.7 (0.9)
n = 23

2.0 (1.3)
n = 45

2.6 (1.5)
n = 46

2.1 (1.3)
n = 49

2.4 (1.4)
n = 95

4,5
4
3,5
Solution Error

for each time step of a round (Mller, 1993). A perfect

solution is indicated by a solution error of zero. Since the
results of Problem 2 were close to ceiling, they were
excluded from the analysis. I analyzed the mean solution
error of the remaining problems as dependent variable in
an ANOVA with the factors type of knowledge
(knowledge about effects, Eff vs. knowledge about
dependencies, Dep), strategy training (with vs.
without training), and the quasi-experimental factor field
of study of the participant (arts/humanities,
law/economy, science). The means aggregated across all
fields of study are listed in Table 2.
The ANOVA yielded significant main effects of all
three factors, type of knowledge (F = 3.94, df = 1,
MSE = 5.57, p = .05), strategy training (F = 5.97,
df = 1, MSE = 8.45, p < .05), and field of study
(F = 13.24, df = 2, MSE = 18.75, p < .01). As expected,
subjects who were guided to acquire knowledge of
dependencies were more successful in controlling the
system (mean solution error = 2.1, SD = 1.3) than
subjects who were guided to acquire knowledge of effects
(M = 2.6, SD = 1.5). Contrary to expectation, subjects
who underwent the strategy training performed lower
(M = 2.6, SD = 1.5) than those without strategy training
(M = 2.0, SD = 1.3). Subjects studying arts or humanities
performed worst (M = 3.2, SD = 1.4, n = 33), followed by
subjects studying law or economy (M = 2.3, SD = 1.2,
n = 30). Most successful in controlling the system were
science students (M = 1.6, SD = 1.1, n = 32).
There is a significant interaction between field of
study and type of knowledge (F = 3.29, df = 2,
MSE = 4.65, p < .05). Detailed analyses revealed that a
strong effect of type of knowledge was only present in
the group of subjects who studied arts/humanities (see
Figure 1). No other effects reached statistical significance
(all p > .05).

3
2,5

Eff
Dep

2
1,5
1
0,5
0
Arts/Hum

Econ/Law

Science

Figure 1: Means and standard errors of solution

error of controlling the system (smaller values
indicating better performance)
factors as described above and the mean response times
for hits in the first test as dependent variable. (Three
subjects with mean response times of greater than 3800
ms were excluded from the analysis. Raw values were lntransformed for the ANOVA). The expected effect of
type of knowledge was confirmed by the analysis
(F = 7.83, df = 1, p < .01), (1559 ms vs. 1237 ms, Dep
faster). However unexpectedly, there was also a main
effect of strategy training (F = 11.24, df = 1, p < .01),
(1576 ms vs. 1236 ms, with training faster). No other
effects were significant at the level of = .05. The results
of the second structural knowledge test were analogous to
the first test.
Similar analyses with the discrimination index (an
index of how well subjects can discriminate between
relations and no relations, cf. Snodgrass & Corvin, 1988)
as dependent variable yielded no significant effects.
Discrimination indices were relatively high in all
conditions (di = 0.89).

Discussion

To test the expectation that knowledge of dependencies

results in faster response times in the speeded structural
knowledge test, I calculated an ANOVA with the same

The experiment has confirmed the hypothesis that guiding

subjects to focus on dependencies of output variables
rather than on effects of input variables can enhance performance in controlling a complex dynamic system.
Although there is an effect in the complete sample, the
major contribution came from the subjects studying arts/
humanities. Presumably, this group has the least experience with abstract representations of dynamic systems
and thus learned something new when focusing on
dependencies instead of effects. If the other groups did

not benefit from the manipulation because they take the

dependencies perspective on their own, or because of
some other strategy cannot be told with the present data.
The results of the speeded causal judgment task indicate that focusing on dependencies vs. effects affects the
mental representation of causal relations. The task was
arranged to enable priming from output to input variables,
but not the other way round. Subjects in the Dep condition were significantly faster in judging the relations,
supporting the assumption that they have established
stronger associations between output to input variables
than subjects in the Eff condition.
The two findings are raising the question about their
relation. Are these stronger associations a cause for better
performance or are they just a side effect of the
experimental manipulation? If the relation was causal,
there should be a substantial (negative) correlation
between response time in the causal judgment task and
solution error in the control problems. The respective
correlation is r=.05 in the whole sample. Hence, the faster
reaction times in the Dep condition are probably a side
effect of the manipulation. This, in turn, supports the
hypothesis that the positive effect of knowledge of
dependencies on performance is based on the chunking
aspect, i.e. the integration of single effect representations
according to output variables. It is possible that especially
science students have built such chunks on their own,
even in the Eff condition. (Note that subjects in the Eff
condition were not prevented from gaining knowledge of
dependencies). Figure 2 shows a sketch of the hypothetical structure of a dependency chunk Dep01 (the
causal weights are omitted for clarity). The shaded

Figure 2: Hypothetical structure of a dependency

chunk; solid lines indicate slot-value relations, dotted
lines indicate associations.

substructure Eff01 is a chunk that represents the single

causal relation between MedC and Sripon. The structure,
whose construction in a learning process appears
straightforward, mirrors the equations defining the
behavior of the system remarkably. The solid lines
indicate slot-value relations. Dotted lines indicate the
associations between the name of the dependent variable
and names of influencing variables, which may have been
learned under the Dep condition. These associations can
explain the effects in the speeded judgment task, but are
not necessary for the usefulness of dependency chunks in
control tasks. This interpretation is in line with the
assumption of Boucher & Dienes (2003) that there are
two ways of learning associations, one resulting in
activating relations, the other resulting in chunks that
combine the associated information. Baker, Murphy and
Valle-Tourangeau (1996) suppose that these two ways
may be attributed to different modules of the mind.
Research on causal reasoning has discovered many other
cases where concept-driven symbolic processing must be
assumed in addition to pure associative learning to explain the phenomena (Waldmann, 1996).
Unexpectedly, subjects in the strategy training condition also answered the causal judgment task faster. The
above interpretation suggests that during the training,
subjects must have rehearsed relations between each
output variable y and the variables affecting y. This is
exactly what happened, even in the Eff condition, because
subjects were asked there for all variables that had an
effect on the output variable in question.
The adverse effect of the strategy training was
unexpected, too. The training had been inspired by results
from cognitive tutoring that subskills can effectively be
trained based on single production rules (Anderson,
1993), and thus, practicing only the most difficult part of
a larger strategy appeared reasonable. However, the
success of this kind of training depends on the compatibility of the practiced subskills with the subjects own
strategies. This condition was obviously hurt in the
present case. Subjects might have applied the practiced
method of predicting the next state, and after successful
completion were unclear about what to do next and how
to use the result. An alternative explanation is that the
practiced strategy has interfered with the subjects own
strategies, resulting in mixtures of incompatible strategy
fragments.
In future efforts to train the application of structural
knowledge it should be assured that subjects have at least
an idea of the whole strategy. This could be achieved by
introducing abstract labels for all subgoals and practicing
the whole strategy at least once before possibly focusing
on the most difficult part of it (Catrambone, 1998).

In general, the results of the experiment show that

variations of structural knowledge do affect performance
in the control of dynamic systems. This extends the view
that mainly exemplar knowledge or very specific rules are
used for controlling systems (Dienes & Fahey, 1995; Fum
& Stocco, 2003; Lebiere et al., 1998). It is important to
note that not knowledge about single causal relations as
measured by the discrimination index of the causal
judgment task makes the difference (there were no effects
of the experimental factors on di), but rather the way of
using it, obviously depending on prior knowledge, and the
way of chunking it into larger units.

Acknowledgments
The research reported here was supported by the
University of Bayreuth. I would like to thank Lucie
Necasova, Tereza Kvetnova, and Nha-Yong Au for
carrying out the experiment.

References
Anderson J.R. (1993). Rules of the mind. Hillsdale, NJ:
Lawrence Erlbaum Associates.
Baker, A.G., Murphy, A., & Valle-Tourangeau, F.
(1996). Associative and Normative Models of Causal
Induction: Reacting to Versus Understanding Cause. In
D. R. Shanks, K. J. Holyoak, & D. L. Medin (Eds.),
The Psychology of Learning and Motivation 34 (pp. 3 46). San Diego: Academic Press.
Berry D.C., & Broadbent D.E. (1984). On the relationship
between task performance and associated verbalisable
knowledge. The Quarterly Journal of Experimental
Psychology, 36A, 209 - 231.
Boucher, L., & Dienes, Z. (2003). Two ways of learning
associations. Cognitive Science, 27, 807-842.
Catrambone R. (1998). The subgoal learning model:
Creating better examples so that students can solve
novel problems. Journal of Experimental Psychology:
General, 127, 355 - 376.
Dienes Z., & Fahey R. (1995). Role of specific instances
in controlling a dynamic system. Journal of
Experimental Psychology: Learning, Memory, and
Cognition, 21, 848 - 862.
Fum, D., & Stocco, A. (2003). Outcome evaluation and
procedural knowledge in implicit learning. In R.
Alterman & D. Kirsh (Eds.), Proceedings of the 25th
Annual Conference of the Cognitive Science Society.
Funke, J. (1993). Microworlds based on linear equation
systems: A new approach to complex problem solving
and experimental results. In G. Strube & K.-F. Wender

(Eds.), The cognitive psychology of knowledge. (pp.

313-330). Amsterdam: Elsevier (North-Holland).
Gobet F., & Simon H.A. (1996). Recall of Random and
Distorted Chess Positions: Implications for the Theory
of Expertise. Memory & Cognition, 24, 493-503.
Lebiere C., Wallach D., & Taatgen N. (1998). Implicit
and explicit learning in ACT-R. In F.E. Ritter, & R. M.
Young (Eds.), Proceedings of the Second European
Conference on Cognitive Modelling (ECCM-98) (pp.
183 - 189). Nottingham, UK: Nottingham University
Press.
Mller, H. (1993). Komplexes Problemlsen: Reliabilitt
und Wissen. [Complex problem solving: Knowledge
and reliability]. Bonn: Holos.
Newell A. (1990). Unified Theories of Cognition: The
1987 William James Lectures. Cambridge, MA:
Harvard University Press.
Preuler W. (1998). Strukturwissen als Voraussetzung fr
die Steuerung komplexer dynamischer Systeme.
[Structural knowledge as precondition for the control of
complex systems] Zeitschrift fr Experimentelle
Psychologie, 45, 218 - 240.
Putz-Osterloh W. (1993). Strategies for knowledge
acquisition and transfer of knowledge in dynamic tasks.
In G. Strube, & K. F. Wender (Eds.), The cognitive
psychology of knowledge (pp. 331 - 350). Amsterdam:
Elsevier.
Taatgen N.A. (2001). A model of individual differences
in learning air traffic control. In E. M. Altmann, A.
Cleeremans, C. D. Schunn, & W. D. Gray (Eds.),
Fourth international conference on cognitive modeling
(pp. 211 - 216). Mahwah, NJ: Lawrence Erlbaum
Associates.
Schoppek W. (2001). The influence of causal
interpretation on memory for system states. In J. D.
Moore, & K. Stenning (Eds.), Proceedings of the 23rd
Annual Conference of the Cognitive Science Society
(pp. 904 - 909). Mahwah, NJ: Erlbaum.
Schoppek W. (2002). Examples, rules, and strategies in
the control of dynamic systems. Cognitive Science
Quarterly, 2, 63 - 92.
Snodgrass J.G., & Corvin J. (1988). Pragmatics of
measuring recognition memory: applications to
dementia and amnesia. Journal of Experimental
Psychology: General, 117, 34 - 50.
Vollmeyer R., Burns B.D., & Holyoak K.J. (1996). The
impact of goal specificity on strategy use and the
acquisition of problem structure. Cognitive Science, 20,
75 - 100.
Waldmann M.R. (1996). Knowledge-based causal
induction. In D. R. Shanks, K. J. Holyoak, & D. L.
Medin (Eds.), The Psychology of Learning and
Motivation 34 (pp. 47 - 88). San Diego: Academic
Press.