[FRIAM] The theory of everything

[thompnickson2 at gmail.com]

Eric, 

 

Ok. So I have read your post and, while I still don’t understand everything you write, you seem to understand my question precisely and answer it voluminously.  We know what natural design is and it is a property of the relation between things and their environments, which individual organisms can instantiate but not by themselves exhibit.   To be humiliatingly honest, I had no awareness of the literature that you cite when I asked the question.  Nor, given the highly technical nature of that literature do I imagine I ever will.  

 

So, I have simply to put myself at your mercy and ask, Has anybody used that concept of natural design predictively? If I understand Darwin’s theory correctly, it asserts that better designed organisms should have more offspring than those less well designed.   

 

How does such an understanding of natural design handle such phenomena as the grossly hypertrophied plumage of some birds, or the extraordinary masculinized genitalia of the female spotted hyena, etc.  Are these instances of bad design?  I am hoping that the answer is yes, because they certainly seem like bad designs to me.  Thus, the contrary, that natural selection always produces natural design, is not validated and we need additional constructs if we are to accurately predict what natural selection produces.  

 

Thanks, as always, for your taking the time. 

 

Nick

 

Nicholas Thompson

Emeritus Professor of Ethology and Psychology

Clark University

 <mailto:ThompNickSon2 at gmail.com> ThompNickSon2 at gmail.com

 <https://wordpress.clarku.edu/nthompson/> https://wordpress.clarku.edu/nthompson/

 

 

From: Friam <friam-bounces at redfish.com> On Behalf Of David Eric Smith
Sent: Sunday, July 5, 2020 5:42 PM
To: The Friday Morning Applied Complexity Coffee Group <friam at redfish.com>
Subject: Re: [FRIAM] The theory of everything

 

Not good for me to do this, so let me try to be brief (never a thing I am good at).

 

0.  To repeat myself for the 100th time, of course if one structures conversations around writing down single words (“evolve”, “emergent”, “complexity") and arguing about what they _really_ mean, one can go on forever and happily never settle.  More complete sentences and operational definitions create a much more disappointing landscape, in which one can either reach an understanding, or reach a hard problem of insight for which one lacks a new idea, and either way the conversation does tend to end.  So one reboots with the one-word questions, re-opens the exploration phase, and tries another anneal.

 

On Jul 6, 2020, at 6:03 AM, <thompnickson2 at gmail.com <mailto:thompnickson2 at gmail.com> > <thompnickson2 at gmail.com <mailto:thompnickson2 at gmail.com> > wrote:

 

But this is useful. 





To put the question baldly, wtf is natural design, anyway, as a descriptive property.  

 

One can construe the question and its answer in many valid ways, and I only mean to say a specific thing about one of them; not to assert that is is better or only, just that it can be understood.

 

Natural design, interpreted descriptively, or causally, is among other things a distributional concept.  Basic commitments are these:

 

1. There is some sensible way to make a division of objects from environments.  It is not necessary to think of _the_ most naive way one might try to do this, and then fret that there are lots of cases it would get wrong; we have many tools for such divisions, and the ability to think precisely about many distinct kinds of relations.  But suppose there is some such division.  Animals are objects.  They might live on land or they might live in water.

 

2. Because there is a notion of factorability from 1, there are ways to talk about marginal distributions over the factors, and joint distributions over the combinations.  Animals might have lots of forms, which are variable properties _of_ them.  Environments can be described in terms of many properties, and particular classes of environments might have distributions of properties that are concentrated in one region or another of the values the properties can take.  With regard to “being a place to live”, the “land” environments are concentrated over where weight-bearing is a challenge; the water environments are concentrated over where resistance to relative motion is a challenge.

 

3. Starting conditions for some joint distribution on objects and their environments can be different from later conditions for the same pairs.  So animals that have history in one set of settings can find themselves spending more time in other settings.  They are still good at what they were doing before, but not necessarily at what they mostly spend their time doing now.  Scientists can experience this phenomenon too, most especially when they are put into managerial positions.

 

4. One can put forth various null hypotheses for what a joint distribution could be, including the product of marginal distributions drawn from other conditions.  A joint distribution _could_ look like a product of a marginal distribution over tortious-like or elephant-like animals that are good at weight-bearing, with a marginal distribution over deep-water environments where weight-bearing is not a problem to be solved, but surmounting resistance to motion is a problem.  Often we imagine that the starting conditions in 3 have this character.  This is the early output of using something in a new way (the “ex” part of “exaptation”) before selective dynamics has produced the “aptation” part.

 

5. Joint distributions can change with time (evolve in the dynamical systems sense) so that at late times they are different from the products of marginals that were the null model or the starting condition.  One can speak of “information flow” from regularities in the environment (the fact that its distribution is concentrated in one part of the parameter space and not in another), “into” the marginal on the object, reducing some relative entropy for that marginal distribution.  The descendant animals in a population can have attenuated traits good for weight-bearing, relative to the traits expressed by their ancestors, and can have enhanced traits for reducing resistance to relative motion.  The information flow is the same as repeated application of a Bayesian updating rule (see Cosma Shalizi and Andrew Gelman, Electronic Journal of Statistcs, in the appendix somewhere) though I think this result has been known since antiquity).  In that direction, it is the “aptation” of adaptation, exaptation, or whatever.  Of course, events reducing a relative entropy for the environment could also flow the other way.  Phytoplankton can change viscosity, oxygenation, light-penetration, etc., of their environment, and this is what the niche-construction and ecosystem-engineering guys go on about.  There probably is a Bayesian construction for that too.

 

6. Instead of talking about information “flow” from one marginal to the other, one can instead speak of the increase in “mutual information” as the difference of some relative entropy of the joint distribution from relative entropies of the marginals, and can characterize a degree of adaptation through the increase in that mutual information.  Mutual information can increase through either adaptation or environmental modification, so merely its increase does not report on a directionality.  Chris Adami in early papers liked to emphasize the mutual information approach, because it is the correct zeroth-order answer to facile versions of “Shannon information isn’t semantic information”.  Nihat Ay and collaborators have done a lot, instantiated in robotic problems, with notions of information flows.  Probably David Wolpert up at SFI could reel off 400 citations summarizing the leading 1% of activity in the field as of the last 24 hours, if asked.  There are some nice things to say about how to cut up systems and environments, and to talk about how each affects the other, which come out of recent “stochastic thermodynamics”, and which I will get back to trying to write up after I exit this email.  It’s all more or less elementary, though.

 

 

I guess everybody in the list already knows this, so Nick’s question “wtf is natural design” is meant to take up where the above leaves off.  I guess Nick means “after the above trivialities are all acknowledged, then what does the expression _really_ mean”.  But those questions are always hard for me to understand when expressed at the level of 4 words, if one hasn’t first solved as much with the trivialities as one can, and then been more explicit about what is left unsatisfying or unclear.  Those things tend to be in the eye of the beholder, and cannot readily be guessed if there isn’t a starting platform of using the trivialities to their fullest extent.

 

Then, however, I think we encounter Jochen’s and I guess Russ’s point about “everything and nothing” (a point I also routinely make).  The statistical formulation above is enough to show constructively that there is a thing to describe, that it is not tautological and that it can vary in degree, following which we can attach a name to that construction.  But how it is achieved mechanistically can be through endlessly variable relations.  

 

This is the sense in which Darwinism per se was never much of a scientific revolution, and was and unfortunately still is mostly a social revolution of trying to wrestle thoughts about life away from the vitalists.  The scientific advancement, even in Darwin’s work, seems to have been mostly in showing that one could get enough control over the relations for particular classes of cases, to make Bayesian updating a plausibly _useful_ explanation.

 

>From there one can branch into sub-categories: how important is individuality and population structure, versus continua, etc.  How much does one get from the substrate (chemistry, code, cognitive events, social norms and institutions?), and what part of the potential for design derives from the substrate and not from the general-purpose filtering problem?

 

But, enough.  

 

Eric

 

 

 

 

 

Nicholas Thompson

Emeritus Professor of Ethology and Psychology

Clark University

 <mailto:ThompNickSon2 at gmail.com> ThompNickSon2 at gmail.com

 <https://wordpress.clarku.edu/nthompson/> https://wordpress.clarku.edu/nthompson/

 

 

From: Friam <friam-bounces at redfish.com> On Behalf Of David Eric Smith
Sent: Sunday, July 5, 2020 5:42 PM
To: The Friday Morning Applied Complexity Coffee Group <friam at redfish.com>
Subject: Re: [FRIAM] The theory of everything

 

Not good for me to do this, so let me try to be brief (never a thing I am good at).

 

0.  To repeat myself for the 100th time, of course if one structures conversations around writing down single words (“evolve”, “emergent”, “complexity") and arguing about what they _really_ mean, one can go on forever and happily never settle.  More complete sentences and operational definitions create a much more disappointing landscape, in which one can either reach an understanding, or reach a hard problem of insight for which one lacks a new idea, and either way the conversation does tend to end.  So one reboots with the one-word questions, re-opens the exploration phase, and tries another anneal.

 

On Jul 6, 2020, at 6:03 AM, <thompnickson2 at gmail.com <mailto:thompnickson2 at gmail.com> > <thompnickson2 at gmail.com <mailto:thompnickson2 at gmail.com> > wrote:

 

But this is useful. 





To put the question baldly, wtf is natural design, anyway, as a descriptive property.  

 

One can construe the question and its answer in many valid ways, and I only mean to say a specific thing about one of them; not to assert that is is better or only, just that it can be understood.

 

Natural design, interpreted descriptively, or causally, is among other things a distributional concept.  Basic commitments are these:

 

1. There is some sensible way to make a division of objects from environments.  It is not necessary to think of _the_ most naive way one might try to do this, and then fret that there are lots of cases it would get wrong; we have many tools for such divisions, and the ability to think precisely about many distinct kinds of relations.  But suppose there is some such division.  Animals are objects.  They might live on land or they might live in water.

 

2. Because there is a notion of factorability from 1, there are ways to talk about marginal distributions over the factors, and joint distributions over the combinations.  Animals might have lots of forms, which are variable properties _of_ them.  Environments can be described in terms of many properties, and particular classes of environments might have distributions of properties that are concentrated in one region or another of the values the properties can take.  With regard to “being a place to live”, the “land” environments are concentrated over where weight-bearing is a challenge; the water environments are concentrated over where resistance to relative motion is a challenge.

 

3. Starting conditions for some joint distribution on objects and their environments can be different from later conditions for the same pairs.  So animals that have history in one set of settings can find themselves spending more time in other settings.  They are still good at what they were doing before, but not necessarily at what they mostly spend their time doing now.  Scientists can experience this phenomenon too, most especially when they are put into managerial positions.

 

4. One can put forth various null hypotheses for what a joint distribution could be, including the product of marginal distributions drawn from other conditions.  A joint distribution _could_ look like a product of a marginal distribution over tortious-like or elephant-like animals that are good at weight-bearing, with a marginal distribution over deep-water environments where weight-bearing is not a problem to be solved, but surmounting resistance to motion is a problem.  Often we imagine that the starting conditions in 3 have this character.  This is the early output of using something in a new way (the “ex” part of “exaptation”) before selective dynamics has produced the “aptation” part.

 

5. Joint distributions can change with time (evolve in the dynamical systems sense) so that at late times they are different from the products of marginals that were the null model or the starting condition.  One can speak of “information flow” from regularities in the environment (the fact that its distribution is concentrated in one part of the parameter space and not in another), “into” the marginal on the object, reducing some relative entropy for that marginal distribution.  The descendant animals in a population can have attenuated traits good for weight-bearing, relative to the traits expressed by their ancestors, and can have enhanced traits for reducing resistance to relative motion.  The information flow is the same as repeated application of a Bayesian updating rule (see Cosma Shalizi and Andrew Gelman, Electronic Journal of Statistcs, in the appendix somewhere) though I think this result has been known since antiquity).  In that direction, it is the “aptation” of adaptation, exaptation, or whatever.  Of course, events reducing a relative entropy for the environment could also flow the other way.  Phytoplankton can change viscosity, oxygenation, light-penetration, etc., of their environment, and this is what the niche-construction and ecosystem-engineering guys go on about.  There probably is a Bayesian construction for that too.

 

6. Instead of talking about information “flow” from one marginal to the other, one can instead speak of the increase in “mutual information” as the difference of some relative entropy of the joint distribution from relative entropies of the marginals, and can characterize a degree of adaptation through the increase in that mutual information.  Mutual information can increase through either adaptation or environmental modification, so merely its increase does not report on a directionality.  Chris Adami in early papers liked to emphasize the mutual information approach, because it is the correct zeroth-order answer to facile versions of “Shannon information isn’t semantic information”.  Nihat Ay and collaborators have done a lot, instantiated in robotic problems, with notions of information flows.  Probably David Wolpert up at SFI could reel off 400 citations summarizing the leading 1% of activity in the field as of the last 24 hours, if asked.  There are some nice things to say about how to cut up systems and environments, and to talk about how each affects the other, which come out of recent “stochastic thermodynamics”, and which I will get back to trying to write up after I exit this email.  It’s all more or less elementary, though.

 

 

I guess everybody in the list already knows this, so Nick’s question “wtf is natural design” is meant to take up where the above leaves off.  I guess Nick means “after the above trivialities are all acknowledged, then what does the expression _really_ mean”.  But those questions are always hard for me to understand when expressed at the level of 4 words, if one hasn’t first solved as much with the trivialities as one can, and then been more explicit about what is left unsatisfying or unclear.  Those things tend to be in the eye of the beholder, and cannot readily be guessed if there isn’t a starting platform of using the trivialities to their fullest extent.

 

Then, however, I think we encounter Jochen’s and I guess Russ’s point about “everything and nothing” (a point I also routinely make).  The statistical formulation above is enough to show constructively that there is a thing to describe, that it is not tautological and that it can vary in degree, following which we can attach a name to that construction.  But how it is achieved mechanistically can be through endlessly variable relations.  

 

This is the sense in which Darwinism per se was never much of a scientific revolution, and was and unfortunately still is mostly a social revolution of trying to wrestle thoughts about life away from the vitalists.  The scientific advancement, even in Darwin’s work, seems to have been mostly in showing that one could get enough control over the relations for particular classes of cases, to make Bayesian updating a plausibly _useful_ explanation.

 

>From there one can branch into sub-categories: how important is individuality and population structure, versus continua, etc.  How much does one get from the substrate (chemistry, code, cognitive events, social norms and institutions?), and what part of the potential for design derives from the substrate and not from the general-purpose filtering problem?

 

But, enough.  

 

Eric

 

 

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://redfish.com/pipermail/friam_redfish.com/attachments/20200705/06cf42c5/attachment.html>