This is an interesting and valuable analysis. But while applying selection and mutation to generate software might be a fun problem to tackle, it seems like a terrible way to produce useful software.
As a model, the evolution of human intelligence is not something we know remotely enough about to reproduce artificially. Our ancestors evolved in environments that were constantly changing. They had social relationships with other creatures of their kind. They competed with other organisms. They were predators and prey. The path from the first multicellular organisms to the humans of today was not a straight one.
How would you go about judging the evolutionary suitability for reproduction of a particular individual instance of the software? Not to mention that sex surely plays a role in our development. How do you model sexual reproduction and the recombinatorial process of mixing genes from the parents? How are the parents to judge the suitability of each other for mating? The feedback loop of mating behavior, sexual selection, and human intelligence is not a minor thing.
What about development? Humans are created through a process. Our genes alone do not specify our composition. There is feedback from the environment, from our nutrition, from the symbiotic organisms that live in our guts and on our skin, from our social relationships, from our parenting, from what skills we learn, and what sports we play, and which books we read, all of which manifest physical changes on our bodies that are not specified in our genes.
I cannot envision a situation where coming up the algorithm that could successfully produce an intelligent piece of software would not be far more complicated and impossible to create than the resulting intelligent software.
But if it were possible to create this software by this method, how inefficient it would be. Ask any evolutionary biologist how efficient natural selection is at coming up with ideal solutions to problems. There are plenty of ingenious inventions generated by natural selection, but very few if any are remotely "efficient". And the process of generating those systems via thousands and millions of generations of self-replicating organisms is about the least efficient imaginable system possible, as demonstrated by this paper.
There is a well-established part of evolutionary computing in which there is a developmental process modeled as well. The direct translation from genotype to phenotype without an intermediate level, might namely make the search process harder! The most often used ones are artificial gene regulatory networks.
Also, Eiben who worked with Winfield in the Symbrion project, had one of the first papers that I know on multiple genders. "A multi-sexual genetic algorithm for multiobjective optimization"
Regarding "efficiency", this is precisely the angle of attack of Alan here, and he makes this argument better than in general terms of "efficiency", namely by trying to formulate the "computational energy costs". That evolutionary search uses more resources than another type of search is however just you saying so. Some arguments would be great. :-)
When you claim that to evolve an intelligent agent, the evolutionary algorithm itself needs to be more intelligent, you are essentially arguing for intelligent design. In the light of modern science, this is clearly wrong. The world is full of emergent complex phenomena (such as intelligent humans) that are result of several simpler "rules".
You're confusing things that are good when explaining with things that are good when optimizing.
Simplicity is a huge explanatory advantage, because simple things generally have exponentially higher prior probabilities. But simplicity is only a minor advantage, or even a disadvantage, when you want to make something efficient. It's too constraining.
The dumb algorithm still works... it just takes longer. Nature took nonillions of cells and trillions of generations to make something intelligent; we'd prefer something a bit less brute force.
Well, remember that intelligence (in the form that we would apparently like to develop) evolved as a by product of natural selection. For one thing, our level of intelligence only really makes sense once a whole set of other niches have been filled out.
Applying a better selection/objective function to an otherwise 'dumb' process would likely drastically reduce the cost to reach a 'solution'.
IIRC the thermodynamic lower bound is ln 2 kT per bit of final output; for the human genome size that's well under 10^-10 J. Reversible computing may be far off, but I wouldn't assume it's farther off than evolving our successors. (The paper doesn't mention any of this.)
As a model, the evolution of human intelligence is not something we know remotely enough about to reproduce artificially. Our ancestors evolved in environments that were constantly changing. They had social relationships with other creatures of their kind. They competed with other organisms. They were predators and prey. The path from the first multicellular organisms to the humans of today was not a straight one.
How would you go about judging the evolutionary suitability for reproduction of a particular individual instance of the software? Not to mention that sex surely plays a role in our development. How do you model sexual reproduction and the recombinatorial process of mixing genes from the parents? How are the parents to judge the suitability of each other for mating? The feedback loop of mating behavior, sexual selection, and human intelligence is not a minor thing.
What about development? Humans are created through a process. Our genes alone do not specify our composition. There is feedback from the environment, from our nutrition, from the symbiotic organisms that live in our guts and on our skin, from our social relationships, from our parenting, from what skills we learn, and what sports we play, and which books we read, all of which manifest physical changes on our bodies that are not specified in our genes.
I cannot envision a situation where coming up the algorithm that could successfully produce an intelligent piece of software would not be far more complicated and impossible to create than the resulting intelligent software.
But if it were possible to create this software by this method, how inefficient it would be. Ask any evolutionary biologist how efficient natural selection is at coming up with ideal solutions to problems. There are plenty of ingenious inventions generated by natural selection, but very few if any are remotely "efficient". And the process of generating those systems via thousands and millions of generations of self-replicating organisms is about the least efficient imaginable system possible, as demonstrated by this paper.