The resulting data file of some 1,800 bits appears to contain complex, specified, and novel information.

Is it in fact CSI? If so, was it "smuggled in" from elsewhere, and how? Some analysis of a relatively simple and concrete example such as this might help clarify these concepts.

quote:

---

"Thompson’s genetic algorithm created 50 random bit strings of artificial dna, each 1,800 bits long, the number of configuration bits needed to describe fully the wiring in those 100 cells. These bit strings constituted the initial 50 individuals to be run through the evolution process. "For each of the individuals in turn," Thompson explains, "I take the bit string and download it onto the chip. Now that individual is physically in silicon, and I pump in five bursts of 1 kilohertz and five bursts of 10 kilohertz in random order." He then tests these individuals to see how well they do at producing an output that’s different for the different inputs, considering that unless he is extraordinarily lucky, all the bit strings in the first few generations "will be as bad as possible at doing any task." For this reason, it’s not sufficient to say one individual works and another doesn’t. Thompson had to develop a fitness test that allows him to say one individual may perform slightly less abysmally than the next, which he did by looking for cases in which the average of the outputs during a 1 kilohertz burst was as different as possible from the average during a 10 kilohertz burst. After testing and scoring all 50, the genetic algorithm randomly chooses parents for the next generation, with a built-in preference for those that scored best on the fitness test. The single best individual on the test is also copied over unchanged to the next generation, a useful addition that computer scientists call "elitism." These chosen parents are mated, their bit streams commingled, with a pinch of mutations thrown in ("You don’t want to screw things up too much," he says) to make 50 new offspring. Then the process begins again. After 5,000 generations and two weeks of computer time, the computer was distinguishing between the two tones."

---

Seeing that Brainstorms is in the early stages of existence, we are going to take opportunities like this to set the context and direction that this board takes.

Brainstorm posters should always frame initial topics in the form of an intuition, argument or hypothesis to help fellow Brainstorm readers better understand and respond to the context of your data. For example, (using Mark's topic) you could have started the post by indicating that you have an intuition/prediction that we will find that genetic algorithms can readily produce large levels of CSI. You could then have cited the intriguing results of Adrian Thompson's work (1800 bits of CSI) and possibly others to back up your hypothesis.

[ 26 February 2002: Message edited by: Moderator ]

My intuition here is that GAs are capable of producing some seemingly substantial information, and therefore within the ID framework it becomes necessary to determine if this infomation is CSI, and if it is de novo or merely imported CSI.

A related question is, Can small amounts of CSI be generated de novo by GAs? Because of the probability factor which seems to underpin ID theory, the answer would be yes. Yes, you can have small amounts of each quantity: Complexity and/or Specificity, but only small amounts, as a function of the probability resources available. Allowing this concession, if valid, removes the need to defend the fallacy of "zero de novo CSI". That in turn sharpens the focus on the question of practical probabilistic bounds (which I understand is already happening).

Further, I see value in closely watching the progress and limitations of Evolutionary Computing, as an empirical "sanity check" and hopefully confirmation of theoretical ID work.

For example, the following subjective observation of this field tends to back up ID intuition:

quote:

---

Synthetically reproduced protolife and artificial evolution in computers have already unearthed a growing body of nontrivial surprises. Yet artificial life suffers from the same malaise that afflicts its cousin, artificial intelligence. No artificial intelligence that I am aware of-be it autonomous robot, learning machine, or massive cognition program-has run more than 24 hours in succession. After a day, artificial intelligence stalls. Likewise, artificial life. Most runs of computational life fizzle out of novelty quickly. While the programs sometimes keep running, churning out minor variation, they ascend to no new levels of complexity or surprise after the first spurt (and that includes Tom Ray's world of Tierra). Perhaps given more time to run, they would. Yet, for whatever reason, computational life based on unadorned natural selection has not seen the miracle of open-ended evolution that its creators, and I, would love to see.

http://www.well.com/user/kk/OutOfControl/ch19-b.html

---

[ 26 February 2002: Message edited by: Mark Elkington ]

quote:

---

The single best individual on the test is also copied over unchanged to the next generation, a useful addition that computer scientists call "elitism." These chosen parents are mated, their bit streams commingled, with a pinch of mutations thrown in ("You don’t want to screw things up too much," he says) to make 50 new offspring.

---

I think we need to make clear exactly what "elitism" means. It is a widely used heuristic used in Genetic Algorithms to speed up the process of reaching a solution. What it means is that the best performing individual in the current population is preserved unchanged to the next generation. When the quote says at the end that 50 new offspring are created via mutation and crossover, it is important to realise that this 50 contains an exact copy of the best one from the previous generation. Therefore, by definition, the fitness of the best individual from generation to generation cannot decrease, it can only increase if one of the offspring happens to be better than it.

In my experience in the past with playing with GA's it was found to be almost essential to use elitism. Not only did it vastly improve the performance; in most cases it made it work, where it simply made no progress at all without elitism. The reason for this is that the best individual otherwise has to compete for resources probabilistically - selection probability being apportioned out corresponding to fitness. In a particularly large population, the fittest individual may only have a tiny fitness advantage over the rest, and may therefore only have a 1% higher probability of being selected. So it quite often occurs that beneficial mutations are just lost through bad luck. The problem is less severe for really small populations, because of the "big fish in a small pool" syndrome.

What is clear from the above is that once elitism is used, one simply can't make any claims that it's evolution via the Darwinian process of Natural Selection, because the best one is effectively immortal until someone better comes along. If the researcher had been able to demonstrate success in the algorithm without elitism, then this would be much more interesting.

As far as the other question about smuggling in CSI, there are two avenues whereby this could have been achieved:

(1) Via the "design" of the fitness function.
(2) Via the decoding strategy from the genome coding to the actual circuit in the silicon. Successful design of a GA involves careful consideration of the strategy you use.

[ 27 February 2002: Message edited by: Iain Strachan ]

I think this is interesting, but the more I thought about it, the less impressive it became. As I recall, output tone is produced in first generation. Then the programmable hardware acts as high pass filter. The capability (constraint of search space) was accidentally produced by intelligent designer when the hardware was manufactured. (It’s like figuring out that a pop bottle can produce sound when you blow into it just right.) And then Fitness Function guides the "organism" through this limited space. But here’s the key sentence for me:

quote:

---

Because Thompson seems to be a pragmatist at heart, he has temporarily given up on evolving chips that could do anything fancier than distinguish between two sounds (in a follow-up experiment, he evolved a chip that could tell apart two spoken words, stop and go).

---

A few months back I was rereading Altenberg’s “The Schema Theorem and Price’s Theorem,” which proffers the “canonical,” “standard,” or “generic” dynamical form of the evolutionary algorithm. I found myself wondering if the algorithm is itself an instance of CSI?

I think Dr. Dembski would insist that CSI is a reliable indicator of design. So, if the canonical evolutionary algorithm is an example of CSI do we have a false positive or have we simply inferred that evolution is designed? A bit of a dilemma?

Many questions follow: The algorithm is definitely “Darwinian,” but does it accurately describe the process of biological evolution? The algorithm also certainly resembles those evolutionary design strategies employed by engineers. But does it accurately represent an intelligent design strategy?

If the answer to both the immediately preceding questions is positive then the dilemma remains, particularly for the CSI criterion.

If the answer to the first is no and the second is yes, then Darwin has detected design!

If the answer to the first is yes and the second is no, then Dembski has detected Darwin! (LOL)

If the answer to both questions is positive then biological evolution is indistinguishable from intelligent design, and this represents a problem for the CSI criterion as a discriminatory test.

If the answer to both is negative, then the key is answering the first question first: Is the standard evolutionary algorithm an example of CSI?

I think the answer to this question has to be very, very carefully considered.

I’ve admitted that I don’t fully understand Dembski’s CSI. But I have some basic understanding of two of its terms, information and specification. So, what if I considered Altenberg’s equation from the perspective of “specification” as it used by engineers, rather than Dr. Dembski’s more formal approach?

It appears that the theoretical object of evolutionary computation research is to somehow originate a system that exceeds, not merely the expectations of the designer, but the specifications of the designer. I would submit that the object of the research is therefore ultimately and paradoxically subversive of the theory that ostensibly supports it. In every instance where the algorithm is said to do something remarkable, all that it has really done is exceed the very vaguely defined expectations of the researcher. It’s an interesting case of how a researcher’s theoretical expectations of experimental results condition the design and conduct of the experiment, and the interpretation of results. No doubt, this is always true in science, but it’s usually the case that the experimenter’s expectations can be made concrete, presumably quantifiable, and clearly limited. But in this case the expectations cannot possibly be made concrete, quantifiable, and are definitively unlimited! If we could say exactly how the algorithm is to exceed its specifications then those would become specifications! If we could say exactly how the algorithm is to exceed our expectations then those would become our expectations! So its difficult to say exactly what this line of research is intended to achieve or how we are to understand exactly what it has achieved.

(So much of the evolutionary research program is “quasi-experimental,” so it’s difficult to know exactly what is being researched, what the expectations are, and what the results really mean. I suspect that if the algorithm ever actually met or exceeded the expectations it would be the most stunning discovery in the history of science and technology. We would have successfully harnessed the virtually limitless power and potential of evolution. Obviously, that hasn’t happened. Otherwise we would have all heard about it. Honorariums awarded, press conferences held, grant money would flow like the springtime torrents in my beloved Rockies.)

Now the requirement that the algorithm somehow exceed its specifications may be a more realistic test than the researcher’s expectations. The requirement is also a bit of a conundrum: What we are requiring is that the designer specify that his specifications be exceeded. Usually these little paradoxes and contradictions are ignored. That the algorithm can be rewritten to exceed its specifications is achievable in some narrow sense. (Since it probably is more a function of limiting expectations than exceeding specifications.) But in its (ultimate) achievement we have effectively falsified the theory that motivates the research! Even if we eliminate from consideration, what seems ineliminable, the role of the designer, his specifications, and expectations, what is really required of the algorithm is that it emulate exactly the role of the designer! This would be something of a Phyrric victory. What is really required of such an algorithm is that it emulate or implement some form of intelligence.

This is plainly seen in that body of research focused on transcending the inherent limits of the algorithm by specifying some “self-adapting” regression operator--by making it an observer/controller over its own evolutionary convergence. This is nothing more or less than an implementation of the most elementary form of machine learning. Is the solution to the basic problem in evolutionary computation to be found in artificial intelligence or machine learning?! But this approach is not obviously consistent with a Darwinian perspective on biological evolution. From the Darwinian perspective evolution is not something that life “does.” It is something that “happens” to it. Life does not directly observe and control its own evolution, except where some form of intelligence is involved. Evolution is epiphenomenal upon naturally given contingencies imposed, as it were, upon life and is not an emergent property of some sort of genomic intelligence.

But the "failure" of the “Darwin algorithm” must be placed in some perspective. The algorithm hasn’t so much failed as it has done everything that it could reasonably be expected to do. Its ironic success/failure is due to the theoretical expectations behind it and has nothing to do with its actual implementation or the results it yields.

I suspect that the idea that evolution is ultimately function optimization needs to be closely examined. What evolutionary computation has most successfully demonstrated are the limits of optimization. Limits already well understood in engineering. Innovation is required. Accordingly, the problem is most appropriately recast as one of artificial intelligence, or machine learning, of “self-adaptation,” rather than the fortuitous adaptation of Darwin.

Innovation is widely regarded as a definitive characteristic of what we call intelligence. If what we expect is that the algorithm exceed its specifications then isn’t that exactly what we are expecting of it—that it innovate?

One might argue that “innovation” is not well defined in biology. (And doesn’t even seem to be required by Darwin, although I think its fair to say that there’s a widely held belief, however ill-defined “innovation” may be, that something very much like it is required in the course of evolution, biological or computational.) But “innovation” can be defined negatively as anything that optimization cannot do. E.g., one cannot optimize something into existence. Optimization requires prior specification. (I think this fact is emphasized by Dr. Dembski.) Indeed, can anything be said to be “optimized” that in any way exceeds specifications? If in the process of optimizing we exceed specifications, we can be said to have left off optimizing and begun innovating. This partly accounts for the fact that the algorithm doesn’t meet or exceed (realistic) expectations. It can’t specify itself or revise its specifications. Specification requires innovation. Optimization does not.

Innovation is exactly what the designer/researcher does. Presumably he does so by the application of algorithms--algorithms that are not well understood by science. Because “innovation” isn’t well understood every instance of innovation definitively exceeds expectation/specification.

Now if it is indeed innovation that is required and not merely optimization than the “Darwin algorithm” serves well as a null hypothesis. Problems that it solves are not solutions to the greater problem we are confronted with. It therefore does serve a very important function in research, because knowing what the solution is not is half the job. (Frustrating as usual, the Darwin-driven research program offers us half a solution, or solutions to problems that aren’t the real problem at hand.)

So the problem as I see it is not that Dembski has detected Darwin. But that maybe Darwin has detected design, or that at least in exploring the (computational) limits of Darwin’s theory we have been led to the point where we are forced to admit that the “design of life” has far exceeded our expectations and that we cannot specify it without recognizing some “higher order” of design (innovation) is present there. Its not that Dembski’s CSI criterion fails, its that as any “mechanical” or “natural” evolutionary process converges upon CSI, it increasingly resembles what we would in any case unhesitatingly call “intelligent.” (Isn’t that what Dr. Dembski is saying?)

The algorithm plainly does not accurately represent the process of biological evolution, but that process as it really is may be indistinguishable from some (non-Darwinian) evolutionary process or evolutionary design strategy. I suspect, however, that the question remains…