dkomo

Evolution produces organisms which show the appearance of design, but
are not actually designed. Evolution uses random variation and
selection to do this.

In the computer world, genetic algorithms and genetic programming
closely mimic evolution. Genetic algorithms find good approximate
solutions to difficult optimization problems using random variation and
selection for fitness. Genetic programming generates computer programs
which are good solutions to specific problems, again using random
variation and selection to accomplish this.

In the general world of human design, is there anything more than random
variation and selection? Yes, an evaluative memory. As a design
proceeds, it is *guided* using a combination of what has worked in the
past with foresight into what may work in the future. A large and
highly complex neural network continuously monitors the progress of the
design, providing highly discriminatory feedback. This type of guidance
is what is missing in evolution and the computer science examples.


--dkomo@crs.com

r norman

Your suggestion of "evaluative memory" requires mechanisms not present
at all in evolution. You can't argue that evolution is "deficient" in
failing to account for this phenomenon -- evolution simply "is" and
works with the natural laws that nature provides. Computer science
might be another story, but then the technique you suggest would no
longer be "genetic programming" that mimics to some extent evolution.

Your suggestion also may fall into the trap of discovering only local
optima of function. Limiting the equivalent of mutation or potential
variation to "what has worked in the past" eliminates the adventitious
discovery of entirely new solutions. Similarly, who is to provide the
"foresight" to predict what may work in the future? What you suggest
seems to be exactly what human designers attempt to do in their daily
lives. The genetic programming techniques of computer science were
developed to augment or supplement human problem solvers in tasks
where the traditional "evaluative memory" strategies don't seem to be
sufficiently productive.

perplexed in peoria

I'm not happy with that word "foresight" since it is evocative of
fortune-telling or perhaps of "inspiration". I would prefer to say
that what is involved are "models" that provide guidance as to what
will happen if we try something that has never before been tried.

Well, back before genetic algorithms became popular, the field of
"classical AI" sought to capture that "guidance" in an algorithm.
A good example might be the "General Problem Solver" of Newell, Shaw,
and Simon. I think that this work captured a good deal of the truth
about human (intelligent) design and problem solving.

But at the end of the day, no good human designer actually trusts the
designs that he has come up with. So he builds a prototype and tests
it. And, if the prototype fails, it is "back to the drawing board".
So ultimately, there is not that big of a difference between us and
Nature/NS after all. We both ultimately depend upon empiricism. If
it works, we keep it. It is all trial-and-error.

I think that you are claiming that human designers can accomplish a
design using fewer trials and fewer errors. Very true, but there is
a cost. As compared to what nature comes up with, human designs are
simple and pedestrian. Nature, with a much larger "empirical testing
budget", and less constrained by models and past experience, sometimes
comes up with designs that exploit whole new patterns of interaction -
whole new physical phenomena that have never been utilized before.

Also, I think that we have to admit that Nature is working in a much
more challenging design space than anything humans have yet attempted.
Her designs have to feed themselves, fend off predators, repair themselves
and grow to adulthood from a single egg (all the while feeding, fending,
and repairing).

Your goal in your posting was to highlight the differences between
"intelligent design" and "design by natural selection". My opinion is
that it is also worthwhile to highlight the similarities. To my
mind, the similarities are strong enough so that I like to say that
design by natural selection IS intelligent design. Nature IS an
intelligent Agent. Or better, there are a whole pantheon of intelligent
Agents (one per species) busily designing and redesigning the biosphere.

tim tyler

What has been missing in the past from evolution.

These days we have genetic engineering.

We can apply the full range of design strategies available to us to
organisms' genomes - and in some cases we are doing exactly that.
--
__________
|im |yler http://timtyler.org/ tim@tt1lock.org Remove lock to reply.

dkomo

I'm not sure why you use the word "suggestion". I'm not suggesting that
this is the way human design should be carried out. I'm attempting to
identify the main characteristics of the way such intelligent design
*is* carried out. And getting trapped inside local optima is a
universal problem not just of human design, but of genetic algorithms
and biological evolution as well.


Which can be remedied by brainstorming and applying lateral rather than
vertical thinking during the design.

Similarly, who is to provide the


It seems to be a characteristic of the human mind to be able to
"envision" the future. Such visions often provide a target for the
design process. That's part of what I meant by foresight. Also,
foresight can be plain ordinary prediction -- an extrapolation from what
worked in the past to what probably will work in the future.
What you suggest


--dkomo@cris.com

dkomo

Part of foresight is ordinary prediction. Part of it is specifying what
the design is to be in advance. The spec then serves as a target for
the design process. Finally, foresight could be a fairly complete
conceptual model which can be simulated or iterated in the mind's eye
before starting the design proper.

I would prefer to say


Yep, building models, whether physical or conceptual, and running those
through simulations would be part of it.

Since when does evolution build prototypes of organisms before actually
"trying them out". Evolution completely lacks foresight and simulation ability.

Because nature can use massive parallism to explore an immense search
space. It can literally "experiment" on tens of millions of species
simultaneously. And it has tens of millions of years in which to do it.
These significant advantages over what humans have to work with.

But we can produce a complex design directly with very little trial and
error. Given a design spec for a digital filter, for example, we merely
feed the parameters into a computer, and presto voila there's the
filter. Nature has absolutely no capability to do this.

And what do you suppose humans could come up with if they had hundreds
of millions of years of continuous designing? As it were, we went from
the discovery of electricity to today's computer chips in about 200
years. That's many orders of magnitude faster than nature could it (if
she could do it at all). Actually, you could say nature did evolve the
silcon chip indirectly by evolving man, but it took her 3.5 billion
years starting with some basic biochemical ingredients.


It's more important to point out that human design is an emergent
behavior of evolution a number of levels above the base level of
biological variation and selection.

It could be said that nature is herself intelligent because is composed
in part of billions, if not trillions, of intelligent organisms. But
this could quickly get us into a philosophical argument. Is the
universe itself a Great Mind? There have been many philosophers of the
past who have thought so, and these were not theists.


--dkomo@cris.com

dkomo

True, nature has evolved genetic engineering indirectly by first
evolving a highly intelligent primate who can practice it. That's
progress! :-> From now on evolution will be relying a little less on
trial and error.

Who ever said evolution is directionless? Well, it may have been
directionless in the past, but now...


--dkomo@cris.com

tim tyler

Here you are using terminology which divorces man from nature.

I know there's a long tradition of that - going back to Darwin and
beyond. However I wish people would stop using such terminology.

One of the Darwinain revelations is that man is part of and a
product of nature - not something apart from or divorced from it.


That's more like it ;-)

A useful perpective. The evolutionary process has become intelligent.

No longer can it be argued that evolution can't perform fitness
evaluations under simulation - or that it lacks foresight, aims
and goals. Those things may have been missing back in the dark
ages, true - but these days the process includes intelligent agents
that have these attributes and abilities - and their actions can
influence the path evolution takes.
--
__________
|im |yler http://timtyler.org/ tim@tt1lock.org Remove lock to reply.

perplexed in peoria

Ah! This evokes some thoughts. One area of human designing that I
have some familiarity with is computer software and hardware design.
Such designs are typically very complex and heirarchical. IMO, they
are the only examples of human designs that even approach the complexity
of the designs produced by NS.

There are currently three competing models for the best design practice
to be used in computer software system design. These are "top-down
design", "bottom-up design", and "incremental design".

In "top-down design", you proceed somewhat as you have suggested.
You start with a specification of what the system is supposed to
accomplish. You then break the task down into subsystems, carefully
specify what each of those subsystems is supposed to accomplish, and
carefully specify the interfaces between the subsystems. This completes
design at that level, but the task of designing each subsystem remains.
You keep designing lower and lower level subsystems until you reach
"bedrock". Then you are done. Clearly NS does not use a process remotely
similar to "top down design". Top down design has the property that
you cannot even begin testing your design until the whole thing is
complete.

In "bottom-up design", you start with a much vaguer specification of
what the overall system is supposed to accomplish. You proceed to do
a rough division into subsystems, but you only vaguely specify what
each of these subsystems is supposed to do. You barely specify the
interfaces between subsystems at all.

Next, you choose the most challenging or interesting of those subsystems,
and design that. Continuing in this way, you eventually reach "bedrock"
on one tiny sub-sub-system. You have implicitly specified the behavior
of that subsystem and the interfaces of that subsystem with the rest of
the system. Stepping back up a level, you revisit the design at that
next higher level. The completed specification of one subsystem and its
interfaces now acts as a constrait allowing you to revise the division
into subsystems. Now choose the most interesting of the remaining
undesigned subsystems, and iterate on it.

Moving up and down the tree in this way, you eventually complete the
design. As compared to the top-down approach, this bottom-up approach
does result in the early completion of some subsystems, and it does
permit empirical testing of those subsystems before the entire design
is complete. The downside is that it doesn't allow much design
parallelism. Furthermore, you frequently get some unpleasant surprises
when you get around to designing the "uninteresting" subsystems which
had been put off until last. You discover that what should have been
a simple piece of the system has become an impossible piece, since the
frozen interfaces of the already designed subsystems with which it
interacts just don't provide what the final "easy" subsystem needs.

I might note that the problem-solving algorithm of the General Problem
Solver (Newell, Shaw, and Simon) is a tree-walking process similar to
the "bottom-up" design process that I have described. I will also
note that "bottom-up" design doesn't match very well with what Nature
does either. But one frequently sees objections to NS in which the
objector seems to assume that Nature's design process must be
"bottom-up". These people seem to assume that subsystem designs must
be complete before the overall system can work. How, they ask, could
nature have produced feathers before there were wings to use them, and
how could wings have arisen before there were feathers to make them
useful?

As compared to "top down design", "bottom up design" permits less
design parallelism. If the interfaces and specs of each subsystem
are specified before design of the subsystem starts, then there
is no reason why multiple subsystems cannot be designed in parallel.
But you get into trouble if you attempt bottom up design in parallel.
The interface specifications flowing from below will not match each
other.

Finally, we have the new ideology of "iterative design". This is a
fairly new idea arising in association with something called "Extreme
Programming". The idea here is that you get something up and running
as quickly as possible. It doesn't necessarily do everything that
the ultimate system will do, but at least it does SOMETHING. Then
proceed to make improvements to this design. Each stage of the process
produces a system that does a little more, or does it more efficiently,
or is restructured so as to make further progress possible. At each
stage, we are designing some new subsystems, redesigning existing ones,
and occasionally completely revising the interfaces between subsystems.
Subsystems may even become obsolete.

It is interesting that in this ideology, it is not only the design and
the system that evolves. The target specification evolves as well.
Continual interaction with the "customer" or marketing (a customer
surrogate) is considered essential to the success of this approach.

Of the three approaches, this one permits testing earliest, and it
permits a greater amount of design parallelism than does the
"bottom-up" approach. It does involve a lot of building of components
that you are later going to revise and discard - for this reason it
is probably better suited to software designs than to the design
of space probes.

It can be seen that this "iterative design" approach has a lot in common
with the way that Nature does Her designs. About the only thing that
I can see that is different is in the mutability of interfaces between
subsystems. "Iterative design" requires that interfaces are occasionally
completely revised - sometimes without much change to the resulting
functionality. But Nature seems to find it much more difficult to
revise her interfaces. In this regard, Nature's approach seems to have
more in common with "bottom-up design".

guy hoelzer

This is a side comment directed at the Subject of this thread, rather than
any of the previously posted comments. IMHO it is both arrogant and without
any scientific basis (empirical or theoretical) to distinguish between human
design and natural design. This false distinction is based on the notion
that humans somehow stand outside of nature.

Guy