[info] is evolution slow?

[Eugen Leitl]

http://www.greythumb.org/blog/index.php?/archives/195-Is-evolution-slow.html

Thursday, March 22. 2007 Is evolution slow?  Evolution This post is a
followup to my previous post on why I've chosen to work on evolutionary
rather than neuroscience-based approaches to intelligence. A reader of my
previous post asks:

    You make a good point. On the other hand, evolution is an exceeding long
process of trial and error. How can you be certain that your algorithm that
generates algorithms (that generates algorithms that...) will generate any
interesting algorithm at all in our life time? In the end, perhaps the
neuroscience approach is a shortcut through eons of evolutionary process to
true machine intelligence.


That evolution is slow is conventional wisdom, even in biology. Here's a
collection of popular statements to this effect.

I question this conventional wisdom a bit.

It certainly seems to us that cultural evolution is very rapid while
evolutionary change is slow. This, I think, is one of the reasons why most
people are working on studying the brain as a route to machine intelligence.
If cultural evolution is so blazingly fast, then it seems like it's the
process that we should want to duplicate.

What I wonder though is whether this is an illusion. I lean toward the view
that cultural learning and modern evolutionary learning are roughly
equivalent in speed.

By "modern evolutionary learning" I am referring to evolution with all of its
evolvability adaptations: virally-mediated gene transfer, sexual
recombination, variable-mutation-rate genome architectures, stress-mediated
mutation rate variation, crossover, genomic redundancy,
evolvability-optimzied protein and nucleotide codings, etc. Simple
straightforward differential selection is probably slower than cultural
evolution, but real biological evolution is not so simple. This in turn is
why I'm working on open-ended autoconstructive artificial life systems as
evolutionary computation systems rather than straightforward classical
genetic algorithms. I want systems where evolvability can evolve and that can
"learn how to learn." But I digress...

To get a more accurate idea, it is useful to compare the artifacts produced
by the two processes.

First, consider some of the most complex artifacts that human culture has
produced: microprocessors and power plants for example. These artifacts are
certainly complex, but they are few and far between. The majority of what
human cultural evolution has produced is far less complex; most human
inventions consist merely of a few parts pulled from the local natural and
cultural environment recombined in a new way. Artifacts like modern
microchips represent a kind of pinnacle that is not representative of the
whole.

Second, consider natural biological evolution. To get a sense of the tiniest
fraction of biological complexity, take a look at some diagrams from the
visual complexity site. Those diagrams seem to describe biological systems at
least similar in complexity to a microprocessor, and they represent only tiny
samplings of the biological world. Look at a tropical rain forest: billions
of species, each one made up of trillions of cells containing virtual cities
of complexity. How many microprocessor designs has human culture produced? A
hundred? A thousand? Certainly nowhere close to the number of species in a
square mile of rain forest.

Subjectively, I find the artifacts of biological evolution just as suggestive
of speed and efficiency as the artifacts of human culture despite the
billions of years they took to produce. We greatly underestimate the
complexity, subtilety, and sheer volume of biological innovation. As someone
who has studied biology, I can tell you that the almost unimaginable
complexity and genius of biological systems is just as awe-inspiring as the
breadth of the universe revealed by astronomy. How long would it take human
engineers to design the equivalent of the global biosphere? I'm betting on at
least a billion years at the drafting table.

We also underestimate the difficulty of the task being performed by
biological evolution. Another factor that must be considered is the
combinatorial space in which it operates. Natural biological evolution must
build complex information-processing devices capable of self-reproducing
physically, not virtually. It must build them out of atoms and molecules with
messy chaotic properties in a heterogeneous environment full of toxic
impurities and ionizing radiation. This is not a simple task. Consider that
after fifty years of research human culture has yet to produce a fuel cell
capable of reliably metabolizing anything dirtier or more complex than
hydrogen or methane, while even the earliest replicators of the "primordial
soup" must have been capable of subsisting on a gigantic variety of complex
and perhaps even dangerous molecules.

So when people say that "evolution is slow," my question is "compared to
what?"

... finally, a few words about this as it relates to evolutionary computation
...

In evolutionary computation of the sort I'm working with, we are designing an
environment to support universal computation, self-reproduction, and
efficient evolutionary change "out of the box." This means that we are
instantiating the evolutionary process in an ideal, rarefied environment in
which all of the elements of advanced information processing and ecological
interaction are already available as features of the "physics" of this
virtual world. We are thus saving a tremendous amount of time. Evolution does
not have to design mechanisms for computation out of atoms and molecules.
Instead, its building blocks are already computer instructions in a
Turing-complete computer.

Furthermore, we are artificially driving the evolving system toward evolving
the kind of information processing that we want. As far as we know, no
outside force drove the evolution of intelligence on Earth... it had to
happen almost accidentally as a result of the increasing complexity of the
ecosystem and the unpredictable variation of the environment. I wonder how
quickly intelligence would have appeared had there been some alien force
deliberately sprinkling "sugar" on anything on Earth that displayed
intelligent behavior?

If biological evolutionary processes can design the biosphere in a billion
years out of atoms and molecules, I think that a similar process can probably
design useful computer algorithms out of ready-made instructions in a
rarefied virtual environment on human-friendly time scales. The fact that
classical genetic algorithms using straightforward differential selection
have already produced patentable electronic circuits suggests that I'm not
nuts.

In a followup to this post here in the next week or two, I'll take a look at
a concrete example: programming enterprise software in Java vs. what the
immune system does when it encounters a novel pathogen. The immune system
works by evolutionary principles, while a human programmer works by cognitive
principles. How do these tasks compare in terms of complexity?

Epilogue:

So am I saying that evolution "thinks" then? Well, sort of. I don't think it
thinks in any way resembling the way we do, but I do think that it represents
a type of intelligence that is equal to ours in many ways. I take a broader
view of intelligence than the AI field usually does. To me, anything that
adapts, invents, or solves problems is intelligent. I sometimes use the
phrase "organic intelligence" to refer to this broad concept of intelligence,
and "human intelligence" or "cognition" to refer to our peculiar brand.

What the human brain seems to be is a fabulously efficient containerized
version of intelligence that has evolved to permit our behaviors to adapt.
There are other examples of "containerized" intelligence. The immune system
for example is an evolutionary learning machine that uses variation and
selection to develop defenses against invasive organisms. My point is that
what the brain has inside is also found outside distributed throughout
nature. All of biology is intelligent.

Edit: (or another Epilogue...)

We often hear about the "wild west" of physics: parallel universes,
n-dimensional geometries, etc. If you want to know what the "wild west" of
evolutionary theory looks like, read John Stewart. While I think he treads a
little further than science might on occasion I think his general thinking on
evolvability is right on. More and more empirical evidence for the evolution
of many evolvability adaptations has been appearing in recent years.