Chapter 89 15.7 Death is the best teacher

David Ackley is a researcher in the field of neural networks and genetic algorithms at Bell Communications Research Institute.I came across some of Akeley's most original ideas about evolutionary systems. Akeley was a bear-like fellow with a mouth full of wisecracks.He and his colleague Michael Littman made a hilarious video about the world of artificial life, which was shown at the 2nd Artificial Life Congress in 1990, to the laughter of the 250 serious scientists present.His "creations" are actually snippets of code, not much different from a classic genetic algorithm, but he represents these snippets with funny smiling faces, letting them swim around and bite each other, or bump into walls representing boundaries .The wise live, the stupid die.Like everyone else, Akeley discovered that his world was capable of evolving organisms that were exceptionally adapted to their environment.Successful individuals live very long lives - 25,000 "days" on the time scale of their world.These guys have figured out the system.They know how to get what they need with minimal effort and how to stay out of trouble.The "organisms" with this gene not only live a long life, but also the population composed of them thrives from generation to generation.

After doing some research on the genes of these "street fighters", Akeley found that some of their resources were not fully utilized, which made him feel like he could do something like a god: by improving their chromosomes, using these resources, making them more attuned to the environment he built for them.So, he modifies their evolved code (a move that is effectively the equivalent of early virtual genetic engineering) and puts them back in his world.As individuals, they are super capable, stand out, and adapt more than any previous predecessors. However, Akeley noticed that their populations were always lower than those from which they evolved naturally.As a group, they lack vitality.Although never extinct, they were always endangered.Akeley believes that because the number is too small, this species will not reproduce more than 300 generations.That is to say, although the artificially improved genes can best suit individuals, they are not as good as those naturally grown genes from the perspective of benefiting the entire population.Here and now, in the home-brewed world of midnight hackers, an old ecological adage is being borne out for the first time: what's best for the individual isn't necessarily for the species.

"It's hard to accept that we don't know what's best in the long run," Akeley said at the Artificial Life conference, to applause. "But I thought, hey, this is life!" The reason Bell Communications allowed Akeley to work on his mustard world was because they recognized that evolution is also a calculation.Bell Communications Research was and still is interested in better methods of computing, especially those based on a distributed model, since the telephone network is a distributed computer.If evolution is efficient distributed computing, are there other ways?If so, what improvements or changes can we make to evolutionary technology?Borrowing the library/space metaphor we often use, Akeley gushed, "The computing space is unbelievably huge, and we have only explored some very small corners of it. What I am doing now, and I want to To do further is to expand the space of computing that humans recognize."

Of all the possible types of computing, Akeley was most interested in those processes related to learning. "Strong learning" is a way of learning that requires smart teachers.The teacher tells the student what to know, and the student analyzes the information and stores it in memory.Less bright teachers teach differently.She may not know what she is trying to teach per se, but she can tell when a student has guessed the right answer—like a substitute teacher grading students on a test.If students have guessed part of the answer, the teacher can give hints of "close" or "deviation" to help students continue to explore.In this way, the less bright teacher may generate knowledge that he does not have.Ackley has been promoting the study of "weak learning," which he sees as a way of maximizing the computational space: using the least input information to get the most output information. "I've been trying to find the dumbest, most ill-informed teacher," Akeley told me, "and I think I've found it. The answer: death."

Death is the only teacher in evolution.Akeley's mission is to find out: What can one learn from death alone?We don't quite know the answer yet, but there are examples: the soaring eagle, the pigeon's navigation system, or the termite's skyscraper.Finding out will take some time.Evolution is smart, but blind and dumb at the same time. "I can't think of a dumber way to learn than natural selection," says Akeley. In the space of all possible computations and learnings, natural selection occupies a special place, a pole at which information transfer is minimized.It constitutes the lowest baseline of learning and intelligence: no learning occurs below the baseline, and smarter, more complex learning occurs above the baseline.Although we still do not fully understand the nature of natural selection in a co-evolutionary world, it remains a fundamental melting point for learning.If we can give evolution a measure (and we haven't), we can use it as a benchmark against which to judge other forms of learning.

Natural selection hides under many surfaces.Ackley was right; computer scientists today realize that there are many ways to compute—many of which are evolutionary.As anyone knows, there are potentially hundreds of ways to evolve and learn; whatever the strategy is, it's really a search of a library or space. "The shining idea—and the only idea—in traditional AI research is 'search,'" asserts Ackley.There are many ways to achieve search, and natural selection at work in natural life is just one of them. Biological life is bound to special hardware, the carbon-based DNA molecule.This particular hardware limits the search methods that natural selection can use.With the new hardware of computers, especially parallel computers, many new adaptive systems have been developed and new search strategies have been applied.For example, the chromosomes of biological DNA cannot "advertise" their code to the DNA molecules of other organisms so that they can get the information and change their code, whereas in a computer environment you can do that.

David Ackley and Michael Littman are both members of the Cognitive Science Research Group at the Bell Communications Institute.They set out to build a non-Darwinian evolutionary system on a computer.They chose the most logical option: Lamarckian evolution—that is, acquired inheritance.Lamarck's theory is very attractive.Intuitively, it has far more advantages than Darwinian evolution, because theoretically useful mutations can enter the genetic sequence more quickly.However, its computational magnitude soon taught hopeful engineers how impractical it was to construct such a system. If a blacksmith needs bulging biceps, how does his body work backwards to produce the genetically required changes?The flaw in the Lamarckian system is that any beneficial change requires going back to the genetic makeup of the developing embryo.Since any change in an organism can be caused by multiple genes, or by multiple interacting instructions during the development of the body.Intrinsic causation in any outward form is an intricate web, and untangling it requires a tracking system that is no less complex than the organism itself.Lamarckian evolution in biology is trapped by a strict mathematical law: it is extremely easy to multiply multiple prime numbers, but it is extremely difficult to decompose prime numbers.The best encryption algorithms take advantage of this asymmetric difficulty.The reason why the Lamarckian theory has not really existed in the biological world is that it requires an impossible biological decryption scheme.

However, a body is not required for calculations.In computer evolution (such as Tom Ray's electroevolution machine), computer code acts as both the gene and the body.In this way, the problem of deducing genes from appearances is easily solved. (In fact, this "same as the outside" constraint is not limited to the artificial realm, and life on Earth must have passed this stage. Perhaps any self-organized living system must start with a "same as the outside" form, just like the self-replicating molecules are as simple as that.) In the artificial world of computers, Lamarckian evolution works.Ackerly and Littman implemented the Lamarck system on a parallel computer with 16,000 processors.Each processor manages a subpopulation of 64 individuals, for a total of approximately 1 million individuals.In order to simulate the double information effect of the body and the gene, the system made a copy of the gene for each individual, and called it the "body".The code for each body is slightly different, and they all try to solve the same problem.

Scientists at the Bell Communications Institute set up two modes of operation.In the Darwinian model, the body code mutates.Some lucky guy might accidentally get a better result, so the system selects it for mating and replication.In Darwinian evolution, however, an organism must mate using an original "genetic" copy of its code—that is, the code it has inherited, rather than an acquired, improved body code.This is the way of living things.So when the blacksmith mates, he uses his "nature" code, not his "nurture" code. In the Lamarckian model, by contrast, when the lucky one with an improved body code is selected for mating, it can use the improved code it acquired as a basis for mating.It's like a blacksmith can pass on his strong arms to his descendants.

Comparing the two systems, Akeley and Littman found that Lamarck's solution was twice as good as Darwin's for the complex problems they considered.The smartest Lamarckian is much smarter than the smartest Darwinian.The hallmark of Lamarckian evolution, Ackley says, is that it crowds out the "idiots very quickly." Ackley once yelled at a room of scientists: "Lamarck is so much better than Darwin!" In a mathematical sense, Lamarckian evolution injects a bit of learning.Learning is defined as the adaptation of an individual while alive.In classic Darwinian evolution, individual learning is not important.Lamarckian evolution, on the other hand, allows information acquired while an individual is alive (how to build muscle, or how to solve equations) to be combined with the long, dull learning of evolution.Lamarckian evolution is able to produce smarter answers because it is a smarter search method.

The superiority of Lamarckian evolution surprised Akeley, because he thought nature had already done a good job: "From a computer science point of view, it would be really stupid of nature to be Darwinian rather than Lamarckian." .But nature is stuck with chemicals, and we don't." This made him think that if the objects of evolution are not limited to molecules, there may be more efficient evolution methods and search methods.