[Date Prev][Date Next][Thread Prev][Thread Next]   [Date Index] [Thread Index] [Author Index

Re: Rosennean Complex Systems as Classifiers



Hi Tim,
 
Interesting post. Sorry it's taken me so long to respond.
 
I think it would definitely be worth writing up and posting on your website. I have been finding more and more, lately, that this issue (measurement) tends to come up  all the time. It's worth a few general comments in addition to the issues you raised in your post...
 
The concept of a meter, which any form of measurement or assessment represents, and the effect of applying any meter to a system-- which is then part of the resulting observation (measurement)-- is precisely what Einstein was on about.  If we sacrifice the relation, and with it all the information contained by it, then our result will be anomalous. There is a second problem now, in science: I suspect that the meters/modes of measurement we have created (in physics) are specifically designed to give the "right" answer in spite of this missing information. In other words, people take a set of values produced by measurements and create "transformation" equations which make that set of measurements agree with a known end sum. Then, the transformations are applied to all measurements, under the assumption that we have accurately figured out how to deal with this "measurement problem"...... But we're missing critical information. The meters don't register it because they were not designed to and the transformations are simply modes of making two sets of numbers agree with each other, as pure syntax.
 
In a sense, I think that this whole phenomenon is also what frustrated Einstein in his search for a Unified Field Theory. He found that you can make certain specific measurements obtained via Relativity Theory "agree" with measurements taken via Quantum Theory, but the solution (the set of transformations) won't apply in all cases of measurement comparisons. So, each meter requires different transformations, to make measurements obtained via the two disciplines agree with each other.

In any case, I think your post spotlights one of the biggest sources of trouble in science right now, namely; complexity in systems on one side and tools conceived around simplicity on the other side. Even if one doesn't presume simplicity, the tools and techniques (including the "scientific method") do. The current scientific notions of measurements-as-numbers/increments are based on presumptions of simplicity. As it happens, such numbers may be very useful in local, transient, specific situations, but the type of information that is specific, in nature, is usually different from the kind that is systemic, and vice versa. Increments and numbers tend to obscure that fact.
 
We see this happening in the field of medicine all the time. It's common, in medical studies, to extrapolate back up to larger systems from smaller ones, using results obtained via reductive means. But we've lost information by those reductive means, which is missing also in the extrapolation.  For example, drug trials done on men will yield results which are not necessarily applicable to female physiology because the measurements all include hidden information specific to men. Drug trial results on adults are not necessarily applicable to infants or children, for the same reason. There are times when drug trials conducted on laboratory animals may not be applicable to human physiology... Specificity and genericity...
 
I have to wonder: What if we have rejected the "magic bullet" cancer drug (or something like it), because it killed the laboratory rats when it was administered to them? We automatically assume that something which is toxic to rats will be toxic to us, as well. But this is not always the case: If we had tested chocolate on dogs before we had ever eaten it, we would consider it poisonous and it would be banned by the FDA. Their physiology can't handle it-- it's lethal to them, but not to us.
 
I see a lot of medical studies which cite tons of numbers, often with .001 level of exactitude.... I would laugh if it wasn't deadly serious. But I don't understand how they can take it seriously: Exact numbers are almost never what they seem, even with simple systems. Even if it were possible to get exact numbers in measurements, there's a tendency in humans, when reading a number registering on some meter, to round up or down, to the nearest increment. ("Is that line closer to the six or the seven?" Squint...) Then there's the fact that our various measuring equipment has a tendency to malfunction and/or to be somewhat idiosyncratic in its calibration. And we have a tendency to not notice. On top of these considerations, there is the fact that choices have to be made about what mode of measurement to employ in every measurement situation and that involves foresight and judgment in the chooser-- which may or may not be reliable for the task (especially because schools generally teach from the reductionist mindset).  
 
Clearly, all of the above adds to the Einsteinean findings, proving that the act of measurement is going to add information to any measurement and-- it can also be subtracting information as well.
 
Judith
----- Original Message -----
From: Tim Gwinn
To: ***
Sent: Saturday, June 25, 2005 3:52 PM
Subject: [ROSEN] Rosennean Complex Systems as Classifiers

The following is something I've been thinking about for awhile. Maybe I'll write it up more completely and add it to my website....
 
One of the central concepts in Rosen's Fundamentals of Measurement is that of the meter. A meter does at least two things: 1) it interacts in particular  modes with other systems, and 2) it can discriminate and classify those interactions. Those classifications can be described in formal terms as partitioning into equivalence classes the entire set of the range of possible induced responses to interactions. Each of those equivalence classes is then a possible value of a meter reading. When a meter generates a value, these values are then imputed back to observables of the system which induced the response by the meter.
 
In FM, Rosen utilizes the concept of meters along with his formalism to explore the idea of measurement and the interactions of dynamical systems generally. This idea was also explored in (at least) two earlier papers by Rosen, "Autonomous State Classification by Dynamical Systems" (BMB 1972, 14:151-167) and "Further Comments on Autonomous State Classifiers and an Application to Genetics" (BMB 1972, 14:305-310).
 
In the first of these papers, Rosen discusses the general requirements and principles involved necessary for some system, like a meter, to act as a classifier of its own states based on its intrinsic autonomous dynamics (hence the name, autonomous state classifier). Thus, a meter is mainly a discriminator of its own states ;  that those states are induced by another system (i.e., the system being measured) is what allows the imputation of the classified value back to the observable of the system being measured. As Rosen notes, this concept of classifiers applies to many areas, including measurement, pattern recognition, psychophysical discrimination, and self-organization (pattern recognition therein).
 
In this treatment, Rosen utilizes a formalism whereby dynamics of the classifier system are thought of in terms of trajectories and attractors. That is, as dynamics are induced on a classifier system, the result will be dynamics in the classifier which will follow some trajectory toward some attractor. Each of these attractors is thus an equivalence class. (Intuitively, one can think of a trajectory toward an attractor as an analog meter needle moving toward some particular value and then remaining fixed there, and not jumping around randomly. All the possible induced dynamics which result in the meter needle pointing to that same value are thus members of that same equivalence class.)
 
In the second treatment, Rosen discusses a modified approach in which an interaction is not tied to corresponding fixed trajectories, but rather is tied to rates of change in the classifier. Thus, a classifier system of this kind may operate such that its initial state is not specifically relevant to how it classifies, but its initial dynamics is.
 
I suspect this could approach be iterated, such that classification occurred based on rates of rates of change and so on. Or, more generally, that there are a myriad of possible similar and unique approaches.  The key idea that I come away with is that there are many ways in which a system may implement the role of classifier.
 
If we consider classifiers as members of the general class of systems, then it occurs to me is that there is no requirement that the classifier rely specifically on states (or their changes). Put another way, if state-based models do not exhaust the description of a Rosennean complex system, then state-based treatment of classifiers does not theoretically exhaust the set of classifiers possible in a Rosennean complex system.
 
It seems that such classifiers that could be left untreated in state-based approaches would include those classifiers whose equivalence classes are partitions of a set of organizations which admit a relational description, but not necessarily a correspondings dynamical description. That is, an interaction with another system would not result in a classification based on the dynamics of the classifier; but rather, based upon the kind of organization that resulted. For example, perhaps such a classifier could be one which generates different active sites. Or, psychophysically, classifiers might operate by generating particular select organizations of neuronal activity.
 
What is important in either the state-based approach or the relational approach, is that the system can reliably classify what is induced by the interaction. Regardless of how it classifies, it is necessary that the classifier be able to partition its responses into the same equivalence classes repeatedly. A classifier is competent if it can do so. (Intuitively, a voltmeter which reads 5 volts at one time and then 3 volts at another time, when the induced value is actually the same, is an incompetent meter - we would consider it defective.)
 
More generally, though, if neither state-based or relational-based models alone suffice to fully describe a Rosennean complex system, then it may well be that a complex system could be a perfectly competent classifier, yet seem quite incomprehensible to us. This apparent incomprehensibility would be further compounded because of the inability to apply the notion of algorithms to such a system. [see Essays p. 2-3] What this means in this arena is that if a complex classifier were classifying based upon multiple observables, it is quite possible that we would be unable - even in principle - to create a list, a set of conditions, which would specify what particular combinations of what particular values of individual observables correspond to a classification into a given equivalence class.
 
This result seems to me to be a possible reason for our inability to codify many "subjective" criteria into a list of conditions. What is beauty, what is good art, what is love, etc., may be things which we can each classify for ourselves yet be quite unable to describe to anyone precisely what are our rules for making such classifications. Natural language has remained quite immune to such codifications, despite many attempts, yet this does not impede our ability to use, comprehend, and extend natural languages.
 
It may also be that complex classifiers on a biological level are at work, perhaps partaking in the process of things such as guiding self-organization, and metabolic regulation. Just as, for example, Chomsky's transformational grammar codifies some useful portions of language structure, so too might we well be able to successfully codify portions of those aforementioned processes. But if these processes involve complex classifiers, then such successes in partial codification do not point to our being able to asymptotically achieve a complete codification.
 
This would also indicate that diagnosis would be sometimes confounding. A system which contains a complex classifier which is no longer competent may be very difficult to diagnose since the classification scheme it should have may not be possible to delineate in a list of conditions and accordingly, it could be difficult to identify any deviations from its normal classification scheme, much less correct them.
 
There are other ramifications of this line of thought, but this post is long enough already....
 
Regards,
Tim