1

Models, Simulations, and Experiments

Francesco Guala
Centre for Philosophy of the Social Sciences, University of Exeter, UK, [email protected]
Abstract: I discuss the difference between models, simulations, and experiments from an
epistemological and an ontological perspective. I first distinguish between
static models (like a map) and dynamic models endowed with the capacity
to generate processes. Only the latter can be used to simulate. I then criticise
the view according to which the difference between models/simulations and
experiments is fundamentally epistemic in character. Following Herbert
Simon, I argue that simulations merely require the existence of an abstract
correspondence between the simulating and the simulated system. In
experiments, in contrast, the causal relations governing the experimental and
the target systems are grounded in the same material. Simulations can produce
new knowledge just as experiments do. But the prior knowledge needed to run
a good simulation is not the same as that needed to run a good experiment. I
conclude by discussing hybrid cases of experimental simulations or
simulating experiments.
1. INTRODUCTION
Empiricist philosophies of science tend to draw a sharp distinction
between descriptive or representational devices (scientific theories) and what
is described or represented (the natural or social world). Models and
simulations are customarily placed among the representational tools,
whereas experiments are considered as parts of the natural or social world
that have been carefully designed in order to answer some specific question.
This framework is certainly useful and captures many aspects of scientific
practice. There are, however, bits of science that do not fit neatly, and for
which a different scheme of classification may be more appropriate. In this
paper I shall try to show that it is sometimes useful to think of models,
experiments and simulations as tokens of the same kind, somehow located
between our statements about the world (call them scientific laws, principles,
theories, axioms), and the world itself (see also Guala, 1998). Borrowing
from Margaret Morrison and Mary Morgan (1999), we may say that such
entities mediate between theory and reality.
First, let us notice that everyday scientific talk does often treat
experiments, models and simulations as tokens of the same kind. In one of
the earliest papers in the field of experimental economics, for example, the
2


term simulation appears three times only in the first page (see Smith, 1991,
p.8), alongside other expressions such as experiment and experimental
game. Or take medicine. Experimental physiologists make extensive use of
animals in their investigations, for well known (although controversial)
ethical reasons. Most often, these activities fall under the label of animal
experimentation. But it is not uncommon to hear or read the expression
animal models, especially when the experimenters fear that the findings
will not be easily transferable from animal subjects to human beings.
Why do scientists slip from experiment talk, to model and to
simulation talk? A plausible answer is that the difference is purely
epistemic in character: experiment and theory being the pillars upon
which all proper science should stand, scientists signal their epistemic doubts
using a special terminology. An incomplete or less than certain theory
becomes a model; a dubious experiment becomes a simulation, and so on.
However, perhaps this is not the whole story, as I shall try to show in the rest
of this paper.
2. MODELS AND SIMULATIONS
Models have been at the forefront of research in the philosophy of
science during the last couple of decades. Indeed, the latest orthodox
theories of scientific theories, the so-called Semantic View of theories,
identifies theories with sets of models. The Semantic View is more a family
of doctrines than a single, unified philosophical theory, but all its versions
share a distaste for the older syntactic approach, according to which
theories are basically sets of statements or laws. In the semantic approach the
fundamental component of a theory, the model, is in contrast a structure a
set of objects with properties and relations among them and/or their parts
that satisfies the linguistic components of the theory. The latter are
secondary, in the sense that they can be formulated in various equivalent
ways, as long as they are satisfied by the set of models that make the theory.
The axioms, laws, etc., may change depending on the language and system
of axioms scientists choose, but the models wont. The models must be put
at work by means of a theoretical hypothesis, stating that they stand in a
certain relation (of similarity, isomorphism, analogy, etc., depending on
which version of the Semantic View one subscribes to) with real-world
entities or systems. Since the Semantic View is presently the received
explication of the concept and role of scientific models, I shall take it as my
point of departure here. The next question to be tackled is: what is a
simulation?
Mario Bunge (1969) defines simulation as a relation between two
3


entities, x and y, where x simulates y if (1) there exists a correspondence
relation between the parts or the properties of x and y; and (2) the analogy is
valuable to x, or to another entity (z) that controls x (p.20). The first striking
feature of this definition is its anthropocentrism. It makes no sense to say
that a natural geyser simulates a volcano, as no one controls the simulating
process and the process itself is not useful to any one in particular. I shall
assume for the sake of this paper that the second part of Bunges definition
captures some important intuitions and connotations of the term simulation.
But the first part is unsatisfactory, because it leads to include things that we
would not intuitively call simulations at all. Consider a map: if it has been
drawn adequately, there must exist some correspondence relation between its
parts and parts of the territory it is aimed at representing.
1
Since the map is
also somehow controlled by someone (its user or its creator), and is
certainly valuable to her, it does fulfil all of Bunges criteria. Yet, it would
seem odd to say that a map simulates the territory.
Now consider a map together with a set of flags and miniaturised soldiers
and tanks, of the sort you find in military head-quarters or in games such as
Risk. If the toy-flags, mini-soldiers and mini-tanks are moved on the map
according to the appropriate rules, we can properly claim that a battle or a
military campaign is being simulated. Why? Whereas the map alone is
somehow inert, the same map, plus the miniatures, plus the players or
officials moving the miniatures according to the rules, make a dynamic
system. I shall here follow Stephan Hartmann (1996) and distinguish static
from dynamic models. A static model can only represent a system at rest. A
dynamic model can also represent the time-evolution of the system (p.82).
2

A dynamic model, then, can be in different states at different times, and
usually each state will correspond to a specific combination of values taken
by the variables in the model. Such a model then will be able to be in as
many different states as all logically or physically possible permutations of
the values its variables can take. Only dynamic systems of this sort can
properly speaking simulate. A simulation imitates one process by another
process (Hartmann, 1996, p.83), where a process is a time-ordered
sequence of states a system takes in a given time period.
3

Hartmanns characterisation, which I take to be basically correct, opens
some interesting questions. Consider my previous example: in order for the

1
I am here referring standard maps on paper only. Giere (1999, pp.44-47) provides a detailed
discussion of the function of maps as models.
2
I am paraphrasing Hartmann here, for he speaks of models as if they were linguistic entities
(made of assumptions, for example), whereas in this paper I follow the Semantic
approach and take them to be objects.
3
I am here modifying slightly Hartmanns (1996) own definition of process as some object
or system whose state changes in time (p.83).
4


map-plus-miniatures to be a simulating device, the system must be capable
of taking different states (the miniatures must change their position on the
map, for instance). This means that there must be an agent prompting the
changes in the system itself. Such a role may be played for instance by the
officials in the armys head-quarters. Thus, counter-intuitively perhaps, the
officials must belong to the simulating device itself. If simulation is an
anthropomorphic or more in general agent-dependent notion, as Bunge
seems to suggest, we should not be troubled by these features. It is just
natural that what is to be included and what to be excluded in a simulating
system is partly arbitrary and/or dependent on ones interest. Simulations are
not in nature, it is us who see them and often build them according to our
purposes. Similarly, a checkerboard and some pawns cannot by themselves
simulate anything although they can represent something: for example the
state of a given battle at time t. A chequerboard, some pawns, and two
players can simulate a battle or a war by representing a sequence of states of
that battle or war. Most often, a simulating device will have some
mechanism built into it, which once triggered will make the system go
through a series of states automatically. The agents role, then, will be
merely that of setting the initial state and starting the process, which will
then keep running until it is exogenously interrupted or runs out of steam.
3. SIMULATIONS VS. EXPERIMENTS: THE
EPISTEMIC ACCOUNT
The distinction between simulations and experiments is more tricky than
the one between models and simulations. In everyday scientific talk, such a
distinction is certainly loaded with epistemic connotations: simulations are
supposed to be somehow less fertile than genuine experiments for the
production of scientific knowledge. Their results are often qualified as
mere simulations not to be mistaken for the real thing (i.e. the real-world
system whose behaviour is being simulated, or an experiment on the real-
world system). The interesting question, however, is whether the epistemic
difference is fundamental from a semantic viewpoint, or whether it is just a
by-product of some more basic difference between experiments and
simulations.
I should make clear that I am not interested in conceptual distinctions per
se. My primary aim is to make sense of certain ambiguous tools that are
widely used in science. This is not a purely philosophical concern: scientists
worry about the same issues probably even more than philosophers do.
Take for example the sort of laboratory work done by psychologists and
economists interested in behavioural decision making. The
5


psychologist Baruch Fischhoff represents practitioners worries by means of
a graphic example. In the psychology lab, choices look like this:
Choice A In this task, you will be asked to choose between a certain loss
and a gamble that exposes you to some chance of loss. Specifically, you
must choose either: Situation A. One chance in 4 to lose $200 (and 3
chances in 4 to lose nothing). OR Situation B. A certain loss of $50. Of
course, youd probably prefer not to be in either of these situations, but, if
forced to either play the gamble (A) or accept the certain loss (B), which
would you prefer to do? (Fischhoff, 1996, p.232)
But in the real world, choices look like this:
Choice B My cousins ordinarily, Im like really close with my cousins
and everything. My cousin was having this big graduation party, but my
friend she used to live here and we went to like started preschool
together, you know. And then in 7th grade her stepdad got a job in Ohio,
so she had to move there. So she was in Ohio and she invited me up for a
weekend. And Ive always had so much fun when Id go up there for a
weekend. But, it was like my cousins graduation party was then, too
like on the same weekend. And I was just like I wanted to go to like both
things so bad, you know. I think I wanted to go more to like up Ohio, you
know, to have this great time and everything, but I knew my cousin I
mean, it would be kind of rude to say, Well, my friend invited me up,
you know for the weekend. And my cousins from out of town were
coming in and everything. So I didnt know what to do. And I wanted
mom to say, Well, you have to stay home, so then I wouldnt have to
make the decision. But she said Im not going to tell you, you have to
stay home. You decide what to do. And I hate when she does that
because its just so much easier if she just tells you what you have to do.
So I decided to stay home basically because I would feel really stupid
and rude telling my cousin, well, Im not going to be there. And I did
have a really good time at her graduation party, but I was kind of
thinking I could be in Ohio right now. (Fischhoff, 1996, p. 232)
What do choices in environments like the first one tell us about behaviour
in environments like the latter? And what are environments like the former
anyway? Are they simulations of real-life situations, or are they experiments
on human decision-making? One possible answer is that experiments like
Choice A test subjects pure cognitive capacities. But this is
unsatisfactory: decision processes may be completely different in the two
circumstances, and purity is a poor consolation if it is unlike anything we
are ultimately interested in explaining and understanding. A more reasonable
answer is that in situations like Choice B there is just too much going on,
6


and simplified settings like Choice A are just intermediary steps on the
way towards the understanding of complicated real-world decision making.
Indeed, it is always useful to think of experimental work as involving (at
least) three different systems and two distinct hypotheses (see Figure 1).
Model Experimental
System
Target
System
Internal
Validity
Hypothesis
External
Validity
Hypothesis
Theoretical
Hypothesis

Figure 1. From models to target systems
As semantic theorists point out, models must be put in correspondence
with their real-world target by means of a theoretical hypothesis specifying
what sort of relation (isomorphism, analogy, similarity, etc.) is supposed to
hold between them. But the path between a model and the real world can be
split into a number of sub-steps. Most often, the relation between a model
and an appropriately designed experimental system is tested first (in the lab).
Then it is further hypothesised that the experimental system stands in some
specific relation with the target. This is typical of sciences, like medicine or
economics, which are ultimately aimed at explaining phenomena in the
unconstrained world outside the lab. Since the quality of field data is usually
too poor to answer specific questions concerning causal mechanisms and
processes, physiologists and social scientists find it useful to try and answer
them first in the laboratory. Then, they show (or argue, or assume) that the
laboratory was a correct replica of the real-world systems they were
interested in the first place.
What kind of correspondence relation is specified by the external validity
hypothesis? Given the amount of controversy concerning the nature of
theoretical hypotheses in the semantic approach, one may be pessimistic
about the prospects of finding a general answer to this question. But,
minimally, an external validity hypothesis should map elements of the
experimental system onto elements of the target system. These elements
should be entities, properties, and relations between these entities and
properties. (Causal relations should be prominent, one would think.)
7


Experimental psychologists like Fischhoff are worried because the fact that
X causes Y in circumstances A (the lab), does not imply that X is a cause of
Y also in circumstances B (the target). An external validity hypothesis or
assumption bears the weight of any inference from the social science
laboratory to the real social world.
Model and simulation talk is more common in sciences, like
economics or medicine, in which external validity worries are widespread
and taken seriously. Indeed, you do not find in physics the sort of a priori
scepticism towards experiments that you find, for instance, among
economists. But the question is: should the distinction between experiments
and other mediating devices be based on such epistemic differences? At
least two arguments can be levelled against this solution. Firstly, epistemic
degrees of confidence in a particular scientific tool or device change in time.
Thus, Galileos experiments were not immediately greeted with enthusiasm,
but it would seem odd to claim that for this reason they should have been
labelled as simulations until they were accepted by the scientific
community. What a scientific device is should not depend on whether we are
confident in the reliability of its results.
4

Secondly, simulations require an external validity hypothesis too, which
may or may not be true depending on the circumstances. If I simulate the
battle of Waterloo using little soldiers and horses, I work on the hypothetical
assumption that, for example, the speed of horse miniatures stands
approximately in the same relation with the speed of little soldiers on my
map as the speed of infantry units stood with the speed of cavalry units in
1815. Only under this hypothesis can I use the simulating model to
investigate, for instance, what would have happened if Napoleon had chosen
a different strategy. The difference must lie elsewhere. Perhaps in the fact
that psychologists and economists use human beings (students, for example)
as subjects in their experiments. The intuition is that, unlike pawns and
armies, toys and troops, human beings are the same in and out of the lab. But
how exactly?

4
Notice that I am not claiming that the difference between experiments, models and
simulations is not anthropocentric. Indeed, as I shall clarify later, the distinction depends
very much on the way we use these devices. The point here is merely that the distinction
does not depend on the degree of confidence we attach to the correspondence relation
between mediating and target system.
8


4. SIMULATIONS VS. EXPERIMENTS: THE
ONTOLOGICAL ACCOUNT
A material model of the propagation of light, according to the wave
theory, can be built with the aid of water in a ripple tank. At a general level
of analysis any kind of wave can be modelled as a perturbation in a medium
determined by two forces: the external force producing the perturbation, and
the reacting force working to restore the medium at rest. General
relationships such as Hookes law or DAlemberts equation may hold for all
kind of waves. More fundamental relationships, such as Maxwells
equations, describe the properties of the electric and the magnetic field only.
The values given by Maxwells equations can be used in DAlemberts wave
equation in order to obtain, for instance, the velocity of propagation of an
electromagnetic wave, because electricity behaves like a wave, although the
fundamental principles at work are different from those at work in case of,
e.g., water waves. The terms appearing in the equation describing the target
and the model-systems are to be interpreted differently in the two cases: the
forces are different in nature, and so are the two media in which waves
travel. The similarity between the theoretical model of light waves and the
ripple-tank model holds at a very abstract level only. The two systems are
made of different stuff: water waves are not light waves. Because of the
formal similarity, though, the behaviour of light waves can be simulated in a
ripple tank. Both light waves and water waves obey the same non-structural
law, despite their being made of different stuff. This is due to different
reasons in each case: different underlying processes produce similar
behaviour at an abstract level of analysis.
5
Similarly, human behaviour can to
a certain extent be simulated by means of computerised models, but arises
from machines made of flesh, blood, neurons, etc. rather than silicion
chips.
Herbert Simon (1969, pp.15-18) puts it as follows: simulations rely on a
process of abstraction from the fundamental principles governing the
behaviour of the simulating and the target systems. If similar organisational
properties arise at a given non-fundamental level from different substrata, it
is possible to abstract from the substrata and simulate the behaviour of a
system A by observing the behaviour of another system B which happens to

5
Of course, if one believes in the reductionist story according to which everything physical is
made of the same fundamental sub-atomic particles, then both light and water waves are
made of the same stuff. But the reductionist story is controversial (photons seem to have
different properties from other particles), and at any rate the fact that everything is made of
the same stuff does not play any relevant role in explaining why both systems display
certain non-fundamental relations.
9


(or which is purposely built so as to) display those non-fundamental
properties. Working on this idea, we can devise a criterion to demarcate
genuine experiments from mere simulations. The difference lies in the kind
of relationship existing between, on the one hand, an experimental and its
target system, and, on the other, a simulating and its target system. In the
former case, the correspondence holds at a deep, material level, whereas
in the latter the similarity is admittedly only abstract and formal. It is
tempting to claim that in a simulating device the simulated properties,
relations, or processes are generated by different (kinds of) causes
altogether. Such a claim is problematic if you endorse a formalistic view of
(type-level) causation for example a view that defines causation in terms of
purely probabilistic relations. If, in contrast, one takes causation to be a
substantive property of specific kinds of systems ( la Wesley Salmon, for
instance)
6
, the problem may disappear. In a genuine experiment the same
material causes as those in the target system are at work; in a simulation
they are not, and the correspondence relation (of similarity or analogy) is
purely formal in character.
7

5. THE METHODOLOGY OF PURE
SIMULATIONS
To sum up: according to Simon, the difference between experiments and
simulations has nothing to do with the (lack of) confidence attached to the
correspondence relation between the simulation/experiment and its target,
but rather with its very nature. The mapping in a simulation is purely formal,
whereas experiments are materially analogous to their target systems.
Because of the different nature of the correspondence relation, simulations
and experiments are appropriate research tools in different contexts.
Typically, simulations are used in one of two different ways: either (1) to
bootstrap from the fact that a given effect (which we have observed in
system A) can be produced by means of simulation B, to the fact that the
relations governing the behaviour of B also govern the behaviour of A. Or
(2) to argue that a certain effect observed by simulating with B will also be
observed in the case of A because the two are governed by similar relations.
8

Both procedures are knowledge-producing ones. The point to be stressed is

6
The latest versions of Wesley Salmons theory of causation, however, are of little help
outside the realm of physics.
7
See Mary Hesses (1963) distinction between material and formal analogies (p.63f).
8
One may be unable to experiment with A, or the equations describing A may be so
complicated that they can be solved only by means of some brute-force solution in B.
10


that in both cases the governing relationships have to be fully specified for
the simulations to be carried on. Systems of this kind are transparent boxes,
to which the old dictum applies: a simulation is no better than the
assumptions built into it (Simon, 1969, p.18).
9

Geologists working on stratigraphy, for instance, study the structure of
rock layers below the earths surface. They also investigate the process of
strata formation, but have to face very serious obstacles, such as the
impossibility of doing controlled experiments (processes of sedimentation
last for millennia, and of course the target systems are too large to be
manageable in a lab), the difficulty to gather data even about the present
geography of the strata, the strong theory ladeness of the interpreted data, the
complex interdependencies within geological systems, and so on. In order to
solve at least some of these problems, geologists have devised simulation
techniques like STRATAGEM, a computer-based modelling package used
by large companies such as Shell Oil.
10
This simulation device works on the
basis of a number of structural equations taken from the theory of sequence
stratigraphy. The equations model the systems outcome (the actual
sedimentation) as a function of a number of variables including the
hydrodynamics of sediment deposition, the subsidence patterns, the global
sea level, the amount of sediment supplied to the basin, etc. The outcome of
the simulation is dependent on the approximate validity of the theory of
sequence stratigraphy, and also on the correct specification of the initial
conditions and of the values assigned to the free parameters in the equations
(incidentally, these are all problematic assumptions to be made in the
specific case). Geologists try to simulate systems A (real-world geological
structures) by means of a computer-model B, and all the fundamental
relations in B must be known and specified in advance.
In figure 2, and in the diagrams that follow, I represent the presently
unknown features of a system by means of dotted lines.
11
The question mark
denotes an aspect of a system whose nature or functioning the scientists is
investigating in a given case.

9
Which does not mean that they are just as good as that: to run a good simulation involves
the use of approximations, computer implementation, etc. and thus requires more
knowledge and skills than simply specifying the correct basic equations.
10
I have learned about simulation techniques in geology from Francis Longworth and his
unpublished paper on the methodology of STRATAGEM.
11
In fact, a more precise representation woud involve at least two systems (say, A and B),
each one made of entities/properties and relations (say, X
A
Y
A
, X
B
Y
B
), plus some
external validity hypothesis stating that the two systems stand in a certain relation to
each other. The dotted lines in my figures are shorthands for cases in which such a
hypothesis is not known or well-confirmed.
11


Global sea level
Sedimentation
Global sea level
Sedimentation
?
?

Figure 2. Pure simulations in geology
Notice that some features of simulating systems like STRATAGEM is
exactly opposite from that envisaged in the epistemic account. In order for
these simulations to be successful, geologists must be very confident that the
(formal) correspondence relation between relations in the simulating device
and relations in the target holds to a certain degree of approximation.
Experiments, as we shalls see in the next section, do not require that much.
6. THE METHODOLOGY OF PURE
EXPERIMENTS
The crucial presumption behind experiments is that relevant components
of the laboratory system are made of the same stuff as those of the target
system. (We shall see in a moment what relevant means.) The experiment
should feature the same causal processes that are at work in the real world,
rather than just display some formal relation by means of a device made of
different stuff. Experiments are particularly useful when one has an
imperfect understanding of some basic causal mechanism of the system
under study. They can be used in these contexts because the laboratory
stuff is assumed to be the same as the non-laboratory stuff.
What is unknown is often (but not always) what is under test. An
experiment can give us more confidence in a theoretical model, if the theory
makes some contestable assumption about some component of the target
system, and if the experiment includes the real component (for example real
human behaviour, as in experimental psychology and experimental
economics). An experiment that merely reproduces all the assumptions of
the model, for example by paying subjects to act according to the
behavioural theory of the model, does not test anything at all (except perhaps
the incentive system). But notice that not all that is imperfectly understood
needs to be under test. For instance, one can test the efficacy of a drug
without a detailed understanding of the epidemiological characteristics of a
disease. The efficacy of the drug, rather than the process of infection, is what
12


is under test. Or you can do experiments on market behaviour even without a
proper understanding of the mechanisms of individual choice and belief
formation. Market institutions, instead of individual behaviour, are under test
in these experiments. Subjects may trade at a certain equilibrium price
because they are acting in a fully rational way, or perhaps because they are
following some rule of thumb, or even by sheer imitation. Whatever the real
causal process, we can use laboratory tests to study selected aspects of
specific real-world economies as long as we are confident that the same
(unknown) basic principles of behaviour apply in both cases.
Considerations of this kind have sometimes been proposed by
experimentalists to defend their methodology. The economist Vernon Smith
(1991), for example, argues that the laboratory becomes a place where real
people earn real money for making real decisions about abstract claims that
are just as real as a share of General Motors. For this reason, Laboratory
experience suggests that all the characteristics of real world behavior that
we consider to be of primitive importance [...] arise naturally, indeed
inevitably, in experimental settings (pp.100-1). This reasoning supports
experimenters confidence in their results. To them, the real character of
experimental markets helps to bridge the gap between a theory and its
intended target of application. Laboratory microeconomies are real live
economic systems, which are certainly richer, behaviorally, than the systems
parametrized in our theories (pp.254-5). Experimental economies are
indeed supposed to work according to the same principles as the target
systems in the intended domain of economic theory, because the relevant
components of the laboratory system are made of the same stuff.
Thus, both experiments and simulations are knowledge-producing
devices. But the knowledge needed to run a good simulation is not quite the
same as the one needed to run a good experiment. When reproducing a real-
world system in the laboratory, the relationships describing the behaviour of
both systems may not be known in advance. But one does not have to
specify the full set of structural equations governing the target system. The
trick is to make sure that the target and the experimental system are similar
in most relevant respects, so as to be able to generalise the observed results
from the laboratory to the outside world. Experimenters make sure that this
is the case by using materials that resemble as closely as possible those of
which the parts of the target system are made. They also make sure that the
components of the mimetic device are put together just like those of the
target, and that nothing else is interfering. Of course, quite a lot of
knowledge is required in order to do so, but no fundamental theory of how
the target system works is needed. Parts of the laboratory system can be put
between brackets and used as black boxes. The same processes, the same
causal principles are supposed to be at work in both cases. Experimental
13


systems are reliable if they are made of the same stuff as real world
economies. No process of abstraction from the material forces at work is
needed in order to draw the analogy from the laboratory to the outside world.
One may abstract from negligible causal factors, but not from the basic
processes at work. The similarity is not merely formal, but holds at the
material level as well.
7. HYBRIDS: SIMULATING EXPERIMENTS,
EXPERIMENTAL SIMULATIONS
The distinction between simulations and experiments borrowed from
Simon seems to be of the black-or-white sort. Either the stuff is the same
as that of the target system, or it is not. But it is more realistic to allow for
intermediate cases between the two extremes. This can be done in at least
two ways. Mary Morgan (in this volume) discusses hybrid entities that are
neither entirely simulations nor entirely experiments, but a little bit of both.
She focuses on cases in which some materiality is transferred from the
target system to the mediating entity, and reproduced therein, only to a
certain extent. These hybrids are quasi-material entities, as she puts it. The
second strategy, in contrast, which I would like to explore in this paper, is to
look at hybrids that combine purely experimental and purely simulating
components. In fact, if simulations and experiments produce novel scientific
knowledge in different ways, they must be partly complementary, and we
should be able to combine them in the same project to exploit the potential
of both.
My example comes once again from the social sciences. Experimental
psychologists and economists are often concerned with designing
experiments that reproduce in all relevant respect real-world decision
situations. Subjects, for example, are invited to trade goods in an
environment governed by the rules of some real-world institution (say, a
certain auction system). Even where realism is sought, however,
experimenters sometimes have to make use of artificial devices. Take, for
example, experiments about so-called common value goods items whose
value is the same, but unknown, to all traders. In experiments of this kind
uncertainty is customarily implemented by means of a random draw. The
subjects trade lottery tickets, in other words, which will be played out at the
end of the experiment. Here uncertainty is simulated by means of a random
draw. Uncertainty arises from the interaction of experimental subjects with a
random lottery device, rather than with a real good of unknown value (say,
a concession for an oil tract or a licence for mobile phone frequencies).
Before you use such a device, you need to be very confident that such a
14


way of modeling that particular aspect of the target system is legitimate. One
has to be reasonably sure, in other words, that that part of the theory is right.
Here more paradigmatic sciences like economics are seemingly (but
misleadingly, perhaps) better off: economists are confident to take such a
move because auction theory prescribes to model agents uncertainty as a
probability distribution of this sort. If challenged, they reply (as Robert
Sugden, for example, has done in private correspondence) that the two
phenomena (uncertainty faced by oil companies bidding for real tracts, and
uncertainty faced by subjects in a laboratory auction) are particular instances
of the same phenomenon uncertainty, full stop. To ask whether uncertainty
has been produced correctly in the laboratory would be like asking whether
in an experiment to investigate the properties of steam the boiler has been
heated by burning coil rather than by a Bunsen burner: heat is heat however
it is produced, and obeys to the laws of thermodynamics.
12
Psychologists, in
contrast, tend to be suspicious of assumptions of this sort and argue that
human behaviour (peoples reactions to the same stimuli) vary widely from
context to context. Of course there may be mechanisms general enough to
support inferences to a wide range of circumstances, but that is an empirical
matter to be settled after, not before experimentation takes place.
Uncertainty Institution
Behaviour
?

Figure 3. Experimental simulations
The structure of such experimental simulations may be represented as
in Figure 3. The hypothesis under test here is signalled by the question mark.
Other causal factors are part of the target system, but must not necessarily be
reproduced identically in the experimental design. For the experiment to
work, the effect of the environment may be simulated by a device that does
the same thing as the corresponding real-world factor. But for that to be the
case, we must know that the relation between simulating and simulated
factors is one of formal similarity. Or, if you are not confident that the
simulated features really closely resemble those of the target system, you

12
I owe this example to Bob Sugden.
?
15


must make sure that they falsify reality in the right way. If you want to test
the cognitive capacities of human beings, for example, you may want to
create an environment that is much less ambiguous and indeed tidier than the
real world. The reasoning being that if subjects cannot read the situation
correctly there, surely they will not be able to do it in messier circumstances.
Sometimes hyper-realism is better than no realism or even tentative
realism.
8. CONCLUSION
Experiments, simulations, and models belong to the same category of
scientific tools. They all are somehow in between what we say about the
world and the world itself. When we use them, we must be careful that they
stand in the right relation both to what we say, and to the part of the world
we are interested in. This is quite different from the standard account, where
models and simulations stand on the theorys side at one end of the
spectrum, and experiments stand on the worlds side at the other end. It is
also different from the standard view of testing, according to which scientists
ask very general questions about the world, design very tight and controlled
experiments to answer such questions, and the data will hit the theory no
matter what. For the data to hit the target, the experiment, simulation, or
simulating experiment must mirror the target in a number of relevant
aspects. Whether the mirroring should be purely formal or material in
character, depends on the kind of question we are asking, and on the amount
and quality of background knowledge we have accumulated regarding the
target system itself. Take Fischhoffs Choice A: is it an experiment or a
simulation? A priori there is no answer to such a question. It depends what
the scientists were aiming at in the first place, what sort of target systems
they had in mind, which aspects of the target they were investigating, and
crucially what they already knew about the relationship between their targets
and experiments (or simulations) like Choice A.
ACKNOWLEDGEMENTS
Many helpful criticisms and suggestions were provided by Dan
Hausman, Mary Morgan, Bob Sugden, John Worrall, and members of the
audience at MBR01. All the remaining mistakes are mine.
16


REFERENCES:
Bunge, M., 1969, Analogy, simulation, representation, Revue internationale de philosophie
87:16-33.
Fischhoff, B., 1996, The real world: what good is it?, Organizational Behavior and Human
Decision Processes 65:232-248.
Giere, R., 1999, Using models to represent reality, in Model-Based Reasoning in Scientific
Discovery, L. Magnani, N.J. Nersessian, and P. Thagard, eds., Kluwer, New York, pp.41-
57.
Guala, F., 1998, Experiments as mediators in the non-laboratory sciences, Philosophica 62:
901-918.
Hartmann, S., 1996, The world as a process: simulations in the natural and social sciences, in
Modelling and Simulation in the Social Sciences from the Philosophy of Science Point of
View, R. Hegselmann, U. Mueller, and K. Troitzsch, eds., Kluwer, Dordrecht, pp.77-100.
Hesse, M.B., 1963, Models and Analogies in Science, Sheed & Ward, London.
Longworth, F., 1998, Computer modelling in the earth sciences, Imperial College of Science,
Technology and Medicine working paper.
Morgan, M.S., 2001, Model experiments and models in experiments, this volume.
Morrison, M. and Morgan, M.S., 1999, Models as mediating instruments, in Models as
Mediators, M.S. Morgan, and M. Morrison, eds., Cambridge University Press, Cambridge,
pp.10-37.
Simon, H.A., 1969, The Sciences of the Artificial, MIT Press, Boston.
Smith, V.L., 1991, Papers in Experimental Economics, Cambridge University Press,
Cambridge.