Winston Ewert — With pro-ID grad students like this, Darwinian profs don’t stand a chance

Thanks for pointing this out. I have to say that I’ve read many posts here about the Avida program, and as a person who is not well informed on these matters, I was at a loss as to what is going on with the program. Mr. Ewert explained the basics very well, such that now I have a basic (though limited) understanding of Avida. He also did not resort to any motive mongering towards the Avida developers, which I found added to his integrity. Avida is apparently many fold more sophisticated than earlier programs, such as the Weasel program, yet it seems that the same intelligent input is necessary for these programs to work as desired. In short, they demonstrate intelligent design more than they do Darwinian evolution.

Mr Yankee,

Avida is apparently many fold more sophisticated than earlier programs, such as the Weasel program, yet it seems that the same intelligent input is necessary for these programs to work as desired.

I think that ‘work as desired’ has been shown to be ‘do x with some inputs and y with some other inputs’. If the paper shows that evolution works when partial rewards are present, and works less well or not at all when partial rewards are not present, have we really learned anything new?

The relevant question to ask at this point is whether the real world rewards gradual changes. If yes, the paper has strengthened the evidence for evolution being relevant to the real world. Since real world problems are a heavily biased sample of all possible problems, we shouldn’t be surprised that a search procedure does exist that works better than randomly for a significant group of these problems.

For those interested in Avida, you should also check out my recent paper on Irreducible Complexity, which includes a section on Avida (section 3.4). In fact, my paper points out (correctly) how to tell which parts of an Avida program were explicitly designed!

It is interesting how many levels of design exist in Avida. Ewert’s paper showed how the stair-step selection process was a mark of design. Other parts of the design include the matching-up of the problem domain (logic functions) with the language domain (assembly language). The third is the incorporation of a designed loop to accomplish reproduction.

A more interesting evolutionary scenario would be one in which the language used to evolve did not match so closely the problem domain they were trying to solve. In fact, the problem domain they were trying to solve is so limited, there are only 68 possible functions the Avida system can detect for! (see the documentation for the REACTION command)

What’s really interesting is the general place of evolutionary algorithms in computer science – they are almost always classified as optimization algorithms, which, I think, reflects the fact that they always are used on problems which are at least roughly parameterized ahead of time.

This podcast looks very interesting:

Darwin as the Pinball Wizard: Talking Probability with Robert Marks
http://intelligentdesign.podom.....8_02-08_00

This video argues how self organization happens because of simple rules and randomness, see: http://www.youtube.com/watch?v.....E&NR=1

I am not buying it though I would be interested in seeing an article on uncommondescent where it is refuted.

Mofi
I don’t see much to refute in that video as chess programs also use random search engines that can play better than the average person. Chess uses searches though millions of combination of move to find the most fittest. When it comes to tactics it pretty much flawless yet can be beat by a good long term strategy.

Nakashima,

“The relevant question to ask at this point is whether the real world rewards gradual changes. If yes, the paper has strengthened the evidence for evolution being relevant to the real world.”

I fail to see this as relevant at all. The question of whether the real world rewards gradual changes is part of the concern. You can’t use that as a correlary without begging the question, because that is part of what you are trying to prove. That’s at least how I see it. By rewards, I’m assuming you mean increased complexity, and other Darwinian ‘benefits.’ What other processes can account for complexity? If you haven’t addressed the alternative, you haven’t addressed your own premise either. The programs (Weasel and Avida) inadvertently address both processes (Darwinian and ID), and ID apparently wins.

Dr Dembski,

If the point were to show that partial rewards can lead to success where their absence doesn’t, then you don’t need a program as sophisticated as Avida — Dawkins’s WEASEL does the job quite nicely, thank you.

Agreed. Those who use Avida regularly do seem to have higher aspirations. But if partial reward functions are found in the real world, and some researchers use partial reward functions in Avida to model certain aspects of evolution as it might happen in the real world, what is the sharpness of the criticism? All you’re saying is that Avida is a tool that is fit for purpose.

Simply pointing out that the program let’s you set up a certain class of models isn’t a telling criticism unless you feel those models are inappropriate or invalid. I realize that research proceeds by small steps, and that it is sometimes necessary to publish many small intermediate results before getting the heart of your research program into print. I hope that this result will be followed up by further work that shows why partial reward models are not valid models of the real world (at some level presumed to be subject to evolution) and therefore it is “bait and switch” to publish results based on such models and call them relevant to the real world.

BTW, I think it is much more relevant that Avida biases its initial populations. Such bias could be justified in uses of Avida which are effectively just optimization runs. There’s nothing wrong in optimization in engineering applications with bringing as much domain knowledge and tool knowledge to the problem as possible. However, if you are trying to study the process of evolution, or the structure of populations, then the choice of initial population structure should definitely get mentioned in the Materials and Methods section.

“To have integrity, computer simulations of evolutionary search like Avida should make explicit [such sources of intelligence].”

That is an interesting gloss of the actual content of the paper.

To have integrity, computer simulations of evolutionary search like Avida should make explicit
(1) a measure or assessment of the difficulty, I?, of the problem being solved,
(2) the prior knowledge that gives rise to the active information in the search algorithm, and
(3) a measure or assessment of the active information, e.g. either I+ or I?, introduced by the prior knowledge.

Relative to 1, I don’t know if it is possible to quanitfy the expectation that a problem can be solved more easily by a given algorithm than by a random search. As was mentioned recently on another thread, the class of problems we are interested in heavily biased towards small problems. One source of active information would seem to be just being small.

On 2, is prior knowledge of the researcher the only possible source of active information? The conclusion of Ewert, et. al. seems to leave open the possibility of other sources.

Mutation, fitness, and choosing the fittest of a number of mutated offspring [5] are additional sources of active information in Avida we have not explored in this paper.

Glossing all of the above as intelligence seems to be assuming what you are trying to prove.

Mr Yankee,

I hope my reply to Dr Dembski which follows yours will address some of your points.

Avida allows different kinds of models to be set up and run. Ewert, et al prove this by setting up and running different kinds of models. Agida within the ID community isn’t directed at Avida per se, it is directed at specific kinds of models created within Avida, and authors publishing the results of those models as relevant to the real world. If I ran a model in Avida with very powerful instructions, a finely crafted initial population, and two kinds of time dimension, and it showed evolution worked exponentially quickly, I could publish that result and the ID community would collectively yawn, because it was obvious from the model structure that it was not in any analogous to the real world. People would say, “Nice example of ID.”

So why does evolving EQU bother people? Because of the claim that model structure is analogous to the real world. And one important aspect of that model structure criticized by Ewert, et al is partial reward.

Partial reward in this circumstance means the fitness surface is not flat, except for a single point. Evolutionary algorithms can’t guess passwords because there is no partial reward. Evolutionary algorithms can design antennas better than humans because there is partial reward.

Nakashima, you state “I don’t know if it is possible to quanitfy the expectation that a problem can be solved more easily by a given algorithm than by a random search.”

It is possible. In the paper we showed via monte carlo simulation that producing an EQU by random search has a probability of about 1 in 10 ^ 14. There is a variety of algorithms discussed in the paper, all of which have a considerably higher probability of success then that random search.

“One source of active information would seem to be just being small.”

No, actually. A small problem will have a small level of endogenous information. The result of that is that it will require little to no active information in order to produce a successful search. As a result, small problems are “uninteresting”, because they are easily solved by just about any algorithm.

“is prior knowledge of the researcher the only possible source of active information?”

The No Free Lunch Theorem and other Conservation of Information results show that apart from making valid assumptions about the search space no algorithm will perform better than any other. The consequence of this is that the only way to produce a better search algorithm is to exploit information known about the particular search problem. That is, we need prior knowledge in order to know how to construct an algorithm with better than average performance.

The other sources referenced as you quoted are also based on prior knowledge. Consider, for example, mutations. Avida’s mutations replace, insert, or delete instructions. They do not have, for example, anything equivalent to a frameshift mutation. In the context of a pseudo-assembler language a frameshift would almost certainly break the avidian program. The point is that the style of mutation has been chosen in a way that makes sense for the problem at hand. Prior knowledge about the genome’s structure was used to determine the mutations available.

The problem with partial rewards is identifying the partial solutions to reward. Avida has nicely demonstrated that it can be done if you already know what the partial solution looks like. A naturalistic evolutionary process does not have that luxury.