A New, Improved, Completely Whacky Theory of Evolution
This blog posts presents some really weird, speculative science, that I take with multiple proverbial salt-grains ... but, well, wouldn't it be funky if it were true?
The idea came to mind in the context of a conversation with my old friend Allan Combs, with whom I co-edit the online journal Dynamical Psychology.
It basically concerns the potential synergy between two apparently radically different lines of thinking:
Morphic Fields
The basic idea of a morphic field is that, in this universe, patterns tend to continue -- even when there's not any obvious causal mechanism for it. So that, for instance, if you teach thousands of rats worldwide a certain trick, then afterwards it will be easier for additional rats to learn that trick, even though the additional rats have not communicated with the prior one.
Sheldrake and others have gathered a bunch of evidence in favor of this claim. Some say that it's fraudulent or somehow subtly methodologically flawed. It might be. But after my recent foray into studying Ed May's work on precognition, and other references from Damien Broderick's heartily-recommended book Outside the Gates of Science (see my previous blog posts on psi), I'm becoming even more willing than usual to listen to data even when it goes against prevailing ideas.
Regarding morphic fields on the whole, as with psi, I'm still undecided, but interested. The morphic field idea certainly fits naturally with my philosophy that "the domain of pattern is primary, not the domain of spacetime"
Estimation of Distribution Algorithms
EDA's, on the other hand, are a nifty computer science idea aimed at accelerating artificial evolution (that occurs within software processes)
Evolutionary algorithms are a technique in computer science in which, if you want to find/create a certain object satisfying a certain criterion, you interpret the criterion as a "fitness function" and then simulate an "artificial evolution process" to try to evolve objects better and better satisfying the criterion. A population of candidate objects is generated at random, and then, progressively, evolving objects are crossed-over and mutated with each other. The fittest are chosen for further survival, crossover and mutation; the rest are discarded.
Google "genetic algorithms" and "genetic programming" if this is novel to you.
This approach has been used to do a lot of practical stuff -- in my own work, for example, I've evolved classification rules predicting who has cancer or who doesn't based on their genetic data (see Biomind); evolved little programs controlling virtual agents in virtual worlds to carry out particular tasks (see Novamente); etc. (though in both of those cases, we have recently moved beyond standard evolutionary algorithms to use EDA's ... see below...)
EDA's mix evolutionary algorithms with probabilistic modeling. If you want to find/create an object satisfying a certain criterion, you generate a bunch of candidates -- and then, instead of letting them cross over and mutate, you do some probability theory and figure out the patterns distinguishing the fit ones from the unfit ones. Then you generate new babies, new candidates, from this probability distribution -- throw them into the evolving population; lather, rinse, repeat.
It's as if, instead of all this sexual mating bullcrap, the Federal gov't made an index of all our DNA, then did a statistical study of which combinations of genes tended to lead to "fit" individuals, then created new individuals based on this statistical information. Then these new individuals, as they grow up and live, give more statistical data to throw into the probability distribution, etc. (I'd argue that this kind of eugenics is actually a plausible future, if I didn't think that other technological/scientific developments were so likely to render it irrelevant.)
Martin Pelikan's recent book presents the idea quite well, for a technical computer science audience.
Moshe Looks' PhD thesis presents some ideas I co-developed regarding applying EDA's to automated program learning.
There is by now a lot of mathematical/computational evidence that EDA's can solve optimization problems that are "deceptive" (hence very difficult to solve) for pure evolutionary learning. To put it in simple terms, there are many broad classes of fitness functions for which pure neo-Darwinist evolution seems prone to run into dead ends, but for which EDA style evolution can jump out of the dead ends.
Morphic Fields + EDA's = ??
Anyway -- now how do these two ideas fit together?
What occurred to Allan Combs and myself in an email exchange (originating from Allan reading about EDA's in my book The Hidden Pattern) is:
If you assume the morphic field hypothesis is true, then the idea that the morphic field can serve as the "probability distribution" for an EDA (allowing EDA-like accelerated evolution) follows almost immediately...
How might this work?
One argument goes as follows.
Many aspects of evolving systems are underdetermined by their underlying genetics, and arise via self-organization (coupled to the environment and initiated via genetics). A great example is the fetal and early-infancy brain, as analyzed in detail by Edelman (in Neural Darwinism and other writings) and others. Let's take this example as a "paradigm case" for discussion.
If there is a morphic field, then it would store the patterns that occurred most often in brain-moments. The brains that survived longest would get to imprint their long-lasting patterns most heavily on the morphic field. So, the morphic field would contain a pattern P, with a probability proportional to the occurrence of P in recently living brains ... meaning that occurrence of P in the morphogenetic field would correspond roughly to the fitness of organisms containing P.
Then, when young brains were self-organizing, they would be most likely to get imprinted with the morphic-field patterns corresponding to the most-fit recent brains....
So, if one assumes a probabilistically-weighted morphic field (with the weight of a pattern proportional to the number of times it's presented) then one arrives at the conclusion that evolution uses an EDA ...
Interesting to think that the mathematical power of EDA's might underly some of the power of biological evolution!
The Role of Symbiosis?
In computer science there are other approaches than EDAs for jumping out of evolutionary-programming dead ends, though -- one is symbiosis and its potential to explore spaces of forms more efficiently than pure evolution. See e.g. Richard Watson's book from a couple year back --
Compositional Evolution: The Impact of Sex, Symbiosis, and Modularity
on the Gradualist Framework of Evolution
and, also, Google "symbiogenesis." (Marginally relevantly, I wrote a bit about Schwemmler's ideas on symbiogenesis and cancer , a while back.)
But of course, symbiosis and morphic fields are not contradictory notions.
Hypothetically, morphic fields could play a role in helping organisms to find the right symbiotic combinations...
But How Could It Be True?
How the morphic fields would work in terms of physics is a whole other question. I don't know. No one does.
As I emphasized in my posts on psi earlier this year, it's important not to reject data just because one lacks a good theory to explain it.
I do have some interesting speculations to propound, though (I bet you suspected as much ;-). I'll put these off till another blog post ... but if you want a clue of my direction of thinking, mull a bit on
http://www.physics.gatech.edu/schatz/clocks.html
The idea came to mind in the context of a conversation with my old friend Allan Combs, with whom I co-edit the online journal Dynamical Psychology.
It basically concerns the potential synergy between two apparently radically different lines of thinking:
- Rupert Sheldrake's idea of "morphic fields"
- The notion of EDA's (Estimation of Distribution Algorithms) in computer science
Morphic Fields
The basic idea of a morphic field is that, in this universe, patterns tend to continue -- even when there's not any obvious causal mechanism for it. So that, for instance, if you teach thousands of rats worldwide a certain trick, then afterwards it will be easier for additional rats to learn that trick, even though the additional rats have not communicated with the prior one.
Sheldrake and others have gathered a bunch of evidence in favor of this claim. Some say that it's fraudulent or somehow subtly methodologically flawed. It might be. But after my recent foray into studying Ed May's work on precognition, and other references from Damien Broderick's heartily-recommended book Outside the Gates of Science (see my previous blog posts on psi), I'm becoming even more willing than usual to listen to data even when it goes against prevailing ideas.
Regarding morphic fields on the whole, as with psi, I'm still undecided, but interested. The morphic field idea certainly fits naturally with my philosophy that "the domain of pattern is primary, not the domain of spacetime"
Estimation of Distribution Algorithms
EDA's, on the other hand, are a nifty computer science idea aimed at accelerating artificial evolution (that occurs within software processes)
Evolutionary algorithms are a technique in computer science in which, if you want to find/create a certain object satisfying a certain criterion, you interpret the criterion as a "fitness function" and then simulate an "artificial evolution process" to try to evolve objects better and better satisfying the criterion. A population of candidate objects is generated at random, and then, progressively, evolving objects are crossed-over and mutated with each other. The fittest are chosen for further survival, crossover and mutation; the rest are discarded.
Google "genetic algorithms" and "genetic programming" if this is novel to you.
This approach has been used to do a lot of practical stuff -- in my own work, for example, I've evolved classification rules predicting who has cancer or who doesn't based on their genetic data (see Biomind); evolved little programs controlling virtual agents in virtual worlds to carry out particular tasks (see Novamente); etc. (though in both of those cases, we have recently moved beyond standard evolutionary algorithms to use EDA's ... see below...)
EDA's mix evolutionary algorithms with probabilistic modeling. If you want to find/create an object satisfying a certain criterion, you generate a bunch of candidates -- and then, instead of letting them cross over and mutate, you do some probability theory and figure out the patterns distinguishing the fit ones from the unfit ones. Then you generate new babies, new candidates, from this probability distribution -- throw them into the evolving population; lather, rinse, repeat.
It's as if, instead of all this sexual mating bullcrap, the Federal gov't made an index of all our DNA, then did a statistical study of which combinations of genes tended to lead to "fit" individuals, then created new individuals based on this statistical information. Then these new individuals, as they grow up and live, give more statistical data to throw into the probability distribution, etc. (I'd argue that this kind of eugenics is actually a plausible future, if I didn't think that other technological/scientific developments were so likely to render it irrelevant.)
Martin Pelikan's recent book presents the idea quite well, for a technical computer science audience.
Moshe Looks' PhD thesis presents some ideas I co-developed regarding applying EDA's to automated program learning.
There is by now a lot of mathematical/computational evidence that EDA's can solve optimization problems that are "deceptive" (hence very difficult to solve) for pure evolutionary learning. To put it in simple terms, there are many broad classes of fitness functions for which pure neo-Darwinist evolution seems prone to run into dead ends, but for which EDA style evolution can jump out of the dead ends.
Morphic Fields + EDA's = ??
Anyway -- now how do these two ideas fit together?
What occurred to Allan Combs and myself in an email exchange (originating from Allan reading about EDA's in my book The Hidden Pattern) is:
If you assume the morphic field hypothesis is true, then the idea that the morphic field can serve as the "probability distribution" for an EDA (allowing EDA-like accelerated evolution) follows almost immediately...
How might this work?
One argument goes as follows.
Many aspects of evolving systems are underdetermined by their underlying genetics, and arise via self-organization (coupled to the environment and initiated via genetics). A great example is the fetal and early-infancy brain, as analyzed in detail by Edelman (in Neural Darwinism and other writings) and others. Let's take this example as a "paradigm case" for discussion.
If there is a morphic field, then it would store the patterns that occurred most often in brain-moments. The brains that survived longest would get to imprint their long-lasting patterns most heavily on the morphic field. So, the morphic field would contain a pattern P, with a probability proportional to the occurrence of P in recently living brains ... meaning that occurrence of P in the morphogenetic field would correspond roughly to the fitness of organisms containing P.
Then, when young brains were self-organizing, they would be most likely to get imprinted with the morphic-field patterns corresponding to the most-fit recent brains....
So, if one assumes a probabilistically-weighted morphic field (with the weight of a pattern proportional to the number of times it's presented) then one arrives at the conclusion that evolution uses an EDA ...
Interesting to think that the mathematical power of EDA's might underly some of the power of biological evolution!
The Role of Symbiosis?
In computer science there are other approaches than EDAs for jumping out of evolutionary-programming dead ends, though -- one is symbiosis and its potential to explore spaces of forms more efficiently than pure evolution. See e.g. Richard Watson's book from a couple year back --
Compositional Evolution: The Impact of Sex, Symbiosis, and Modularity
on the Gradualist Framework of Evolution
and, also, Google "symbiogenesis." (Marginally relevantly, I wrote a bit about Schwemmler's ideas on symbiogenesis and cancer , a while back.)
But of course, symbiosis and morphic fields are not contradictory notions.
Hypothetically, morphic fields could play a role in helping organisms to find the right symbiotic combinations...
But How Could It Be True?
How the morphic fields would work in terms of physics is a whole other question. I don't know. No one does.
As I emphasized in my posts on psi earlier this year, it's important not to reject data just because one lacks a good theory to explain it.
I do have some interesting speculations to propound, though (I bet you suspected as much ;-). I'll put these off till another blog post ... but if you want a clue of my direction of thinking, mull a bit on
http://www.physics.gatech.edu/schatz/clocks.html
posted by Ben at 9:56 AM
7 comments
