Re: the Temple of Aggregated Optimizations Re: The First Church of Robotics

c...@xxxxxxxx (Curt Welch) wrote:
Yevgen Barsukov <evgen...@xxxxxxxxx> wrote:
c...@xxxxxxxx (Curt Welch) wrote:
Yevgen Barsukov <evgen...@xxxxxxxxx> wrote:
casey <jgkjca...@xxxxxxxxxxxx> wrote:
Evolution is a simple process and so is all intelligence.
There is no magic complexity in any of it, it's just
different forms of reinforcement learning systems - which
at their core, are all dead simple.

And you are dead wrong. Yes RL and natural selection are
simple mechanisms but the environment to bring about all
these complexities is not itself simple.

I would like to give it yet anther spin. Both of the involved
entities:
1) mechanisms of the evolution (selection, memory, replication), and
2) the environment laws that shape evolution (second law of
thermodynamics, catalysis
accelerating entropy increase and use of released energy
to counteract own complexity)
...are simple.
But the RESULT of evolution that continued for several billions of
years
under continuously changing, while still simple, environment, is NOT
simple,
just because the process (1) is ACCUMULATIVE. It is accumulating and
reconsiling
the empirically found answers posed by (2) over extremely long
periods of time.
The resulting aggregated knowledge can be arbitrarily complex.

To give you an example, imagine that you are entering a labyrinth
with a notepad in your hand. You have a simple rule - you record a
right and left
turn that you are making, and crossing out all the "bad" turns. After
some
long time you emerge on the exit of the labyrinth with your notepad
containing
the optimal walk-through map.
Now, both the method you used to record your turns,
and the goal you had (coming through) are very simple.
The labyrinth structure could also be very simple, in fact it can be
completely
random so it would not have any structure.
But the complexity of the map you got as the result is going to be
determined
only by the size of the labyrinth!

The same thing is with complexity of Life and of our "cross organism"
type of
intelligence. Life has accumulated its "map", now stored in the
genetic code, over
2 billions of years. The complexity of the map is immense and
immeasurable.

It's not so big. =A0It's only a few hundred MB of map data in the DNA
(if=
 I
remember correctly) and some hard to measure (but not so huge) amount
in the structure of the cell that interprets it. =A0It's far from
immense in=
 my
book seeing the cheap computer in front of me has a few orders of
magnitu=
de
more complexity to it.

It is important to differentiate complexity of the result, and
"computing resources"
or in case where physical experiments are needed, generally "energy
resources and time
resources". For example for typical example of NP problem (selecting
optimal
permutions between 100 students in a dorm) the result might be just a
100 bytes but it
takes million years worth of computations to obtain it.

Sure, because the complexity in question we are interested in there is the
size of the search space, not the size of the input, or output, of the
process.



The same is
true for example
for factoring numbers created by multiplying large primes. In both
cases, single
steps to get the information are well understood, however the immense
amount of
these operations is what makes it unattainable and result very
precious.
For obtaining information by interaction with physical world it is
even more costly.
For example to find that sun raises on the east, you need to actually
wait for 1 day.
To find that some particular plan was toxic, someone had to die.

To summarize, what makes _results_ of evolution unique is not the size
of resulting information,
but billions of years of effort (as well as unique combination of
physical environment factors)
needed to obtain this information. Which makes this information
irreplaceable, despite of knowledge
of the simple steps involved in getting it.

Sure, the search space was huge and we are just lucky enough to be here to
see the result of the work to date.



Its
loss would be completely unrecoverable (except by going through the
labirith
again!) therefore we should not get careless or disrespectful to it
even though
we know basic principles behinds its aquisition.
If there anything that we would call "magic" (in the sense of great
respect
and admiration), that map of optimization accumulated over billions
of years
would definitely deserve such designation.

Sure.

"Cross organism" intelligence that our society possess and which is
due
to existence of conceptual language, is a much more recent
phenomenon, and its "map" has been accumulating only for
approximately 100 000 years.
But even that is nothing to sneer at, considering much accelerated
rate of accumulation.
It is is sufficient time to accumulate amazing complexity and to
force us to (mentally)
recline our heads in respect.

This ACCUMULATED COMPLEXITY is the temple to the legitimate God of
the new enlightened religion of information age. It is well worthy of
our admiration.

Well, those few hundred MB of DNA data that took 4 billion years to
accumulate in us is nothing compared to the information accumulated
already
by man in his head (social knowledge), and in books, and on the
internet and the additional information that is being created daily by
the actions of this global brain.

Again, I would like to differentiate between "amount" of information,
and computational
cost of it.
Closely related to "resource cost" of information, we could also
evaluate "usefulness" of
information (just like 8 letter password to a bank is much more useful
than 8 letter expletive).
From this point of view, information in DNA that accumulated over
billions of years of optimizatinos
defines an extremely robust algorithm of Life operation, which we know
has survived billions of years
over highly variable conditions. Those it is extremely useful (and is
proven so,

Well, the concept of "useful" is another whole point here.  Usefulness is
relative, it's not absolute.  There is no universal measure of value.  

This is where the heresy comes out! Bevare of offending the one true
God!
Until you understand that Life is a very specific physical process,
distinct from
all other physical processes by its purpose - maximally accelerating
the increase of entropy
due to their high complexity, while maintaining their complexity by
using part of the released
energy, of cause it will look to you that everything is relative and
there is no "measure of value".

However, if you are standing on the solid foundation of understanding
what Life is,
than THERE IS a very clear measure of value. It is "the integral of
entropy increase
caused by Life from time 0 to infinity". Clearly to maximize this
integral, you need to have
both
1) high catalytic activity
2)long (or preferably, indefinite) survival

Lets compare the DNA programming success with cultural evolution
success on these two criteria:

1) catalytic activity: can be measured in terawatts of energy of
catalyzed processes.
Humanity, year 2008 - 15 terawatt
http://en.wikipedia.org/wiki/World_energy_resources_and_consumption
Phytoplankton: 63 terawatt
http://news.mongabay.com/2006/1013-fsu.html

Clearly, humanity did not come even close to even to just one of more
successful sub-classes
of biological life forms, not to mention all other Life combined, in
therms of catalytic activity.

2) long survival
Here, there is really no contest. Humanity only boasts 100 000 years
(an more modern
type of humanity that includes written language, more like 20 000
years), while
DNA (and even more recent subclass, the multi-cell animal DNA) is over
500 million
years.

In summary, sorry but there is a clear criteria for Life's program
success,
and according to this criteria, so far humanity cultural program is
lagging behind DNA program,
although it might get even in the future yet.


We can likewise dream up some random search problem and let a computer
search for a million years to produce the answer, but just because the
answer was expensive to create, doesn't mean it has value.  It just means
it was expensive to create.

That is why I specifically separated "usefulness" and "cost". They are
related,
because _typically_ more useful information is also more costly to
get. But
there are of cause exceptions, so it does make senses to treat these
concepts
individually.

Human culture is created by a process which is simply far more efficient
than the way our DNA based design was created.  Human culture is in effect,
"running" on a machine which is many orders of magnitude better at
searching the design space than DNA life.  DNA life evolves though a
process of trial and error that is spreading changes though the population
very slowly (improved genes might take 10,000 years to spread though a
majority of the population).  An improved brain design (good idea created
by culture) can spread through the majority of population in years now.  So
you can't compare years of DNA evolution to years of culture evolution and
assume the two are accurate measures of the value of the result.

I quite agree that cultural evolution is more efficient in finding
solutions,
that is why in only 100 000 years we came close (1/5 is quite
impressive for a contender!)
to power generation of the entire oceans planktons. But lets not
forget two things:

- humanity still derives part of its program from DNA, and so part of
its success
is still due to great flexible and robust bodies and agile brains of
monkeys into which the cultural
program is loaded. If we had only robots and no monkeys (powered by
DNA) to load cultural program into,
that would be a clean comparisons of "culture vs DNA".
But in this case both the efficiency and survivability of these hybrid
will be very questionable.
In fact it is not yet technically feasible to completely load all
aspect of cultural program into DNA-free
carriers, and once it will be, it might still be completely wiped out
after only 10 years if
left alone. So, a lot of humility should be still on our minds and the
word "robustness"
has to wake up at night everybody who is dreaming of completely
eliminating reliance on
DNA.

- we still have to prove that cultural evolution (even one using DNA-
based carriers)
has enough internal consistence, self-correcting and self-preserving
properties to
prevent self-destruction. 100 000 years is just not enough to prove
this point. What we
do know that technical means for the self-destruction has already been
created, while
control mechanisms for them remain highly questionable.
We have to keep proving it every day and of cause hope for the best.
Well, after all
we can always fall back to cockroaches DNA program if the grand
experiment of
cultural evolution fails :-)

Regards,
Yevgen

--
Tune in to "Strange Drawing of the Day" buzz:
http://www.google.com/profiles/100679771837661030957#buzz
.

Quantcast