Re: Modeling vs. Simulation [from "fundamental problems in physics"]

[Tim Gwinn <***>]

JohnK,
See interposed comments.
Regards,
Tim


> -----Original Message-----
> From: ROSEN Forum [mailto:*** Behalf Of John
> Kineman
> Sent: Friday, December 17, 2004 2:44 PM
> To: ***
> Subject: Re: Modeling vs. Simulation [from "fundamental problems in
> physics"]
>
>
> Tim, Howard,
>
> Sorry to raise this, but it seems to me that the definition of
> simulation is too broad. Correspondance of behavior of a formal system
> (model, simulation, description, etc.) with behavior of the natural
> system, as indicated by observables, describes all of modeling in my
> opinion, i.e., the modeling relationship.
> We cannot "see" the underlying
> causality structure, so it must be inferred from behahvior and the
> modeler's insight(?).


TG: Clearly both models and simulations should create a correspondence of
their behavior with the behavior of the system under study. But this is not
yet a commuting modelling relation. Models have a further stipulation that
there is an establishment of congruence of entailment structures between the
model and the system under study.

Actually, to be a little more specific, I believe Rosen uses "simulacra" to
refer the general class of entities which only mimic behavior [EL 40-42],
and "simulation" refers specifically to those simulacra which are
machine-based (whether as a physical device or as a formal device) [LI 185].
Boris nicely summarized "simulation" in his recent post: "let us understand
that, through simulation, inferential entailments, which play the role of
efficient causes, become material causes. And that makes a clear distinction
between simulation and the modeling relation, because the modeling relation
"respects" inferential structures".

Finally, one of the purposes of a model is to allow us to probe a natural
system by asking "why?". With a model, we can ask "why?" and get answers
that are meaningful because of the congruence of entailment structures. If
we ask "why?" of a simulation, we generally cannot learn anything about the
natural system because the entailment structure of the simulation is foreign
to the natural system. In practice, we may typically know the model on which
a simulaiton is based, and therefore we can mentally jump between the two,
so we have the feeling that the entailment structure si somehow there, but
if we were only given a simulation, for example made out of a mechanical
clockwork computer, we could not ask "why?" questions of it which would tell
us anything of the system which it is simulating.



  The only difference between what present day
> scientists would call a simulation and a model would seem to be that we
> call it a model if we believe its sub-models represent real processes in
> nature. We evaluate various model structures based on beliefs about how
> nature is structured projected from these sub-model experiences; but the
> final test is through correspondence between formal and natural behavior
> - there is nothing else that can be tested.



TG: If we use different definitions for 'model' and 'simulation', we may get
different results. (I am also not sure what a "sub-model" is.) I was
speaking specifically of Rosennean definitions.



> Plant dispersal models, for example, can be based on process dynamics,
> i.e., structuring the model based on prior knowledge about water and
> nutrient uptake rates, growth rates corresponding to these processes,
> climate effects on the rates, etc. Simulations of general behavior can
> be done statistically, based on generalized statistical correlations
> with environmental variables. For example one can produce a statistical
> estimate of net annual primary production based on average rainfall,
> nutrients, temperature, etc.  Or one can attempt to model exactly what
> each species is doing physiologically in terms of time-dependent
> equations and wire all that together to get a production estimate. The
> first approach can be iterated on varioius time steps to re-introduce
> the dynamics an that produces a "simulation." The rate-based approach is
> generally not described as a simulation, because it employs sub-models
> of what everyone thinks are "actual" causes or rate-based controls on
> productivity. The word "simulation" is used when one hasn't attempted to
> emply all that is known about underlying causes. It is thus considered
> only to be characteristic of what can happen, but not "accurate" enough
> to give detailed predictions. Hence, simulation seems to be most related
> to characterizing the type of behavior, vs. attempting to predict as
> much detail as we can about it.  Simulations are good for producing
> "scenarios" of possible future system behavior - i.e., a
> characterization of what could be. Whereas process models are sought
> after to make more accurate predictions of future states, such as
> quantities, flows, storage amounts, etc. taking into account all the
> significant exchanges in an ecosystem.
>
> Another example of simulation might be fractal landscapes. These would
> be a simulated landscape because they "look" like topography and one can
> match the general fractal dimension (wiggliness) to an actual landscape
> to make it look characteristic of some real landscape, but the details
> will be all wrong - hills in different places, etc. but looking like
> real hills. The fractal simulation does not correspond to observations
> of the actual landscape at a detailed level, but it does represent one
> aspect of that landscape, its general wiggliness.
>
> Given this vernacular (I'm sure it can be better described than i just
> did), the Rosen view enters the picture in a very odd way, I think.
> Essentially, it is saying that the dyanmical model is not really any
> more causal that the simulation in the case of modeling organisms.


TG: I disagree with this. A dynamical model of some aspect of an organism
can be entirely valid. Rosen's argument is that a finite number of dynamical
models, or any other kinds of predicative models, are inadequate to fully
model organisms.


> That
> is is no better a model, and in fact because it tries to hard it may
> even be worse. Experiment bears this out, as currently the statistical
> models of ecosystem patterns actually seem to out-perform the supposedly
> dynamical/causal ones. In other words, given that we really don't know
> enough about the causalities to model them, it is better to model some
> more general aspect of the system, using fewer assumptions and fewer
> parameters subject to error. There is a growing principle of appropriate
> modeling, developing in some circles, like appropriate technology, where
> one does not use more than is necessary for a given problem. That's
> where most ecosystem modeling seems to be these days, with regard to the
> biotic interactions. With regard to the biogeochemical flows and rates,
> however, the dynamical models do best, except for the uncertainty of the
> biotic component's influence; for example how the biota may compensate
> in unpredictable ways over time to changed physical conditions.
>
> Now if Rosen's view is taken seriously, it explains why the supposed
> causal models of the biota don't work very well -- namely that they
> don't understand the true causalities and may be comming at them all
> wrong (i.e., from a mechanical mind set, because that mindset does work
> best for the physical parts).


TG: I would say that the models may have an incomplete (rather than true/not
true) representation of their causal counterparts. Perhaps incomplete in
terms of the relevant causal influences modeled and/or perhaps incomplete in
terms of the relations (e.g., impredicativities) modeled between the causal
influences.


> Most causes arethus  imagined to be
> physical in nature and law-like. Closed-loop entailments can only be
> accounted for by incorporating iterative loops in the law-like process
> outputs, or the statistical estimates. It can be applied to both kinds
> of model, statistical or process/dynamical; but it can only iterate
> law-like equations, and there is no control of accumulation of error. So
> the result of the iteration is generally fanciful, but nevertheless
> informative because it supposedly represents one possibility for future
> development. Then what some people do is run multiple simulations,
> varying the parameters reasonably and filling out the random number
> spaces, to try to map out the full range of future possibilities. Then
> managers can at least say, here's the best and worst case scenarios.
>
> But even the above approach fails, because there are non-linearities
> that can't be predicted from prior behavior. The entire system may
> "flip" into a different entailment structure, where the same
> purtabations in parameters or the range of random variables, corresponds
> to entirely new behavior because of the entirely new organization of the
> system. In essence, an ecosystem can reach a limit of its ability to
> retain its identity, and then it simply degrades and re-organizes into a
> new identity. Organisms do the same thing, but their identities are so
> obvious and stable in one mode, that we call any reorganization "death."
> That is the kind of causality - the causalities involved in  stability
> and reorganization, that I think Rosen was most concerned about, and to
> my knowledge there are no successful models that deal with it.


TG: I think these kinds of reorganizations between stabilities which appear
to not be entailed by the prior stable organization, dynamics, etc. would be
cases of "emergence" as Rosen uses it:
        "One of the most puzzling aspects of biodynamical phenomena, such as
development or evolution, is the sudden appearanece of apparently entirely
new modes of organization and behavior, which seem quite unpredictable from
anything which preceded them. Such apparently discontinuous changes in the
structure and function of biological ysstems are characterized collectively
by the term emergence, and the new structures or organizations are terms
emrgent novelties." [FM 90-91]


> The
> non-linear mathematical approach itself seems to be no more than a
> simulation, i.e., an instrumentalism (matching exercise between equation
> and natural behavior with no underlying causal theory).
>
> My sense is that because of the existence and importance of this deep
> level of organizational causality (pardon the invention of new terms),
> Rosen is then saying that everything model we currently have is really a
> simulation,


TG: I do not see Rosen at all asserting this conclusion. We do have models.
For example, it is always possible to have valid predicative models of
impredicative structures -- the issue is whether those predicative models
alone are adequate to answer the particular question or concern regarding
the system under study. If not, then the addition of impredicative models
may be required.


> even though many scientists humored themselves into thinking
> they were dealing with causalities. My own view is that Rosen's point is
> a good one, but a bit too harsh for many honest scientists who know
> their limits. The physical causes are important and constraining, but
> the problem is just that they are not a complete enough description of
> the system.


TG: Agreed.


>Where his "irritating" statement is valid and appropriate,
> however, is in countering the dogmatist who does not know his limits and
> claims that the physical process model is all there should ever be, and
> that it is capable of eventually explaining even the non-linear
> behavior. I think a higher-level theory will be needed at the system
> information/organization level to deal with the organizational problem.
>
> Sorry for the length and rambling, but I hope this paints a useful
> picture of ecosystem modeling.
>
> John Kineman