Re: Behaviorism vs. evolutionary psychology

On Jul 13, 6:44 am, c...@xxxxxxxx (Curt Welch) wrote:
JGCASEY <jgkjca...@xxxxxxxxxxxx> wrote:

JC:
And how is that different from trial and error which
quickly falls victim to the combinational explosion?

CW:
Trial and error with memory of success doesn't fall victim
to combinational explosions as long as there is path which
can be taken one step at a time.

JC:
Yes there must be a path and possible mechanism (memory of
success as you put it) to follow that path in an efficient
and fast way. For example your hill climber might have a
simple mutation that causes jumps to prevent it being caught
in a smaller hill than is really available. So it is more
than just a path it is also an effective path follower.

With a simple version of your net to sort the input according
to the gaps the problem of setting the values in the nodes
quickly became victim to the combinational explosion although
knowing how it worked the programmer could easily insert the
correct values.

======================

I am suggesting that not just anything can evolve even if
it could form complex structures (say by human design) simply
because the parts are not capable of building themselves up
in working stages by chance alone in a sufficiently short time
span or the even if they form a configuration that they might
not have what it takes to hold that configuration.


CW:
That might be true.

And the whole area of studying how different substrates effect
their ability to evolve I think would be an area someone may
like to study. But I don't see it as the great mystery you
seem to see it as.

JC:
Not so much a mystery as an unknown. Use the right algorithm
and you can quickly solve the rubiks cube. Until then the
ability for someone to do this is a "mystery". Often a problem
has a simple solution once you know what that solution is :)

CW:
There's on a few very trivial requirements for a substrate
to evolve.

JC:
? I assume you meant "is" not "on". What are those
very trivial requirements?

=========================

JC:
Take reinforcement learning and the combinational explosion.
In theory trial and error should, given enough time, bring
about any pattern to be selected. But in fact the Universe
will not be around long enough for that to happen. Somehow
there are constraints in these systems that mean that they
can, within a reasonable time limit, produce the right
patterns to actually be selected.


CW:
It's not so complex John. All you need to solve the combinational
explosion problem is memory - aka the substrate must be able to
form a structure and hold the structure. It's called descent with
modification. It's a well known and well understood part of the
puzzle of evolution. This is what allows evolution to take one
step at a time slowly changing the structure. For complex
structures to grow in a reasonable time, there must also be a path ...

JC:
Yes there must be a path. But what is required by the basic
units for there to be such a path? Clearly rocks as basic
units form no such path if shaken, magnetic buttons have no
such path if shaken, molecules that make up crystals have no
such path but the molecules of life do have such a path and
whatever they have is the right stuff I am talking about and
that is part of what the explanation must encompass.

I might add that your current networks also have no path
or if you like no mechanism to find that path effectively :)


CW
... evolution can take from simple structures to complex
structures where each advance creates a stronger structure
(stronger in terms of whatever fitness function the system
is using).

But for evolution, survival of the structure in the universe
is the only fitness function. And if you have a system of
building blocks that can create a near infinite number of
structures, is it really so hard to believe that at all states
of evolution, there will always be slight modifications possible
to the current structure to make it's odds of survival a bit
stronger?

JC:
If you want to "evolve" the sequence 123456789 there may be
something that can generate every possible 9 place number and
test it for success. But as that number gets bigger the
chances become exponentially unlikely of success.

Evolution requires the more correct digits in each place the
better the chances of survival.

That such a path exists in the form of molecules in an earth like
environment is a fact of observation but not deducible from what
we understand so far about molecules. I do not consider it so
much a mystery as an unknown as indicated earlier in this thread.

That there is such a path flags, for me, that there are constraints
which limit the number of possibilities to a manageable number
for the given number of trials possible. An eye that can see
various light levels is better than no eye at all. An eye that
can see change in light levels is better than an eye that cannot
detect changes and so on .. each stage adding something better.

But if no such path existed, that is, if it was all or nothing,
it would never happen. If you want to search a list of words
that are constrained say in alphabetical order you can do a
binary search. Constraints in the way things are, both in the
structure of living things and in the environment, can be taken
advantage of to eliminate the combinational explosion.

I don't think millions of years or billions of trials is an
explanation because it doesn't take too many variables to
explode well beyond the number of combinations that can be
tested in the lifetime of the known Universe. A pathway of
working stages has to be inherently possible to start with.

I think computers are a fun way to explore these things.

We can simulate anything that can be precisely defined and
is not just hand waving.

An example I gave before was "evolving" by pure chance a full
adder. This would take a long time. However if there is a
pathway in the form of working stages, evolve a module to get
the sum, and separately another to get the carry, and combine
those modules to make a half adder ...

We end up with an unlikely to happen in one go to likely to
happen, a full adder, in working self contained modular stages.



[...]


Take Dawkin's biomorphs.

GS:
I've never heard of them.

JC:
I have mentioned them before.

http://en.wikipedia.org/wiki/Weasel_program

http://www.panspermia.org/computrs.htm

Of course there are many creationists that call it a rubbish.

Richard Dawkins does explain its limitations as an evolutionary
model as it really only explains the requirement for small changes
to accumulate in working stages.

You might get something out of reading Richard Dawkins book,
"Climbing Mount Improbable".

CW:
But we can look at how the universe works and come rather quickly
to a basic understand of what's needed to make a good substrate
for evolution.


< JC: snip stuff I essentially agree with >

CW:
So it seems to me, that we need a substrate that can create long
term stability, but at the same time, the universe as a whole
must never be allowed to stabilize.

JC:
That is more or less what I wrote. If it is too stable it will not
evolve, if it is too unstable it will not evolve although stable
and unstable parts may be part of the mix.


CW:
I think if you were to play with this ideas, it would not be hard
to create many simple (game of life like simple) substrates that
show the power to constantly evolve into increasingly stronger
structures.

JC:
I have played with these ideas. It _is_ hard that is why there
has been limited success although I have had some limited success
with some of my ideas as others have had with theirs. One of my
interests is the same as yours but I like to try things out as
things that sound good often don't pan out when you actually test
them in an actual program. But the failures or limitations give
you insight as to where to look next.

CW:
I really don't see it as being the great mystery you seem to see
it as.

JC:
Not so much a mystery as an unknown.

A bit like the right kind of topology and the right kind of node
and the right way to reinforce it is, at this stage, an unknown to
you with regards to your RL networks. In other words all the hard
parts.



.

Quantcast