the usual Personal Report; Steven Rooke, July 2, 1998
On Thursday June 25, C. Frojen helped me hang four 3x4 foot framed IRIS prints of my genetic / evolutionary artwork at the Alife Art Show in the Wight Gallery in a building about a mile from the conference location on the UCLA campus.
-- Friday, June 26, 1998: Workshops --
Hierarchical Selection
I had planned to attend Bruce Damer's Nerve Garden workshop in the afternoon, but ended up at Hierarchical Selection all day because it was highly relevant to my topic at Biota II in Cambridge in September. This workshop was the first time I had ever heard in an Alife conference discussion favoring multi-level selection. The common wisdom is that Ernst Mayr, Richard Dawkins, and others had thoroughly debunked any notion of group or species-level selection: the only entity upon which natural selection can act is the "individual", and some carry that argument right down to an ensemble of selfish genes.
John Damuth, UCSB (a paleobiologist):
The general impression in current literature, particularly popular books on evolution, is that hierarchical selection is discredited. Wrong. Many researchers now recognize that non-individual selection is important. The key thing is context, multi-level. There is a broad consensus about hierarchical selection but most people don't know about this. The popular accounts are highly idiosyncratic and ignore the last 30 years of evolutionary theory. A residual controversy remains because different traditions have different models. Kin selection, group selection, etc., are all special cases of the general structure of Hierarchical Selection theory.
HS must be explicitly built into Alife. There is no single place to which you can go in the literature. [It is the case that many Alife programmers, myself included, get essentially all our information about biological evolution from popular books, rather than scientific journals. One simply does not have time to get a doctorate in multiple disciplines and still hope to get something actually done. Although it may appear in this report that I'm doing a certain amount of Dawkins-bashing, that really is not the case. I own and have read thoroughly all of Richard Dawkins' books, and I find his explanations brilliantly lucid. My own artwork owes its lineage back through Karl Sims to Richard Dawkins' The Blind Watchmaker, hence I see constantly the incredible power of the bottom-up genotype -> phenotype -> selection, or Neo-Darwinian paradigm. The only question I have is whether we actually know that a pure bottom-up selfish gene model is all it takes to explain the whole thing.]
Define selection as all those factors that contribute to the survival of the offspring of an individual. A bottom-up hierarchy:
population studies / group & group effects / changing focal levels
Damuth showed a succession of illustrations beginning with randomly spaced colored dots. Next, adjacent dots acquired a translucent "halo" to denote that they had just engaged in some form of interaction. No spatial boundary is drawn around those individuals purporting to partake in group interaction, just a graphic tag that the interaction had taken place.
"Groups" are defined and identified in their effects upon fitnesses of their members -- NOT in terms of a rigid a priori hierarchy of candidate entities.
Groups arise out of interactions between and among individuals. Think of it as the individuals retaining a halo as they disperse. Next time haloed individuals meet, they repeat the interaction they had before. (The "halo" is carried as the memory of the interaction; individuals encountering each other in the environment are likely to entrain and reinforce those interactions that are to their benefit). This means the obviousness of what a group is, or its persistence, says nothing about how important it may be in generating group effects upon the fitnesses of its constituent individuals.
No new higher level of reproduction or inheritance is required for the process to explain the evolution of individual characters.
In a real population the individuals tend to be spatially located. Anything that might cause individuals to associate with each other, once started, enhances coagulation of groups. You now add an outer halo around a cluster of already once-haloed individuals, creating a group character at the next higher hierarchical level of organization. Over time the forces of selection will become more tightly integrated, and the lower-level individuals in the group will tend to move together spatially. Now selection processes operate on all individuals in the group.
Read John Maynard Smith and Eors Szathmary, 1995: "The Major Transitions in Evolution" on transitions to higher levels of organization. Previously independent individuals at lower levels combine into functional larger wholes. Division of labor. Reproduction of lower levels becomes subordinated to higher levels.
The notion of evolutionary transitions encourages the observer to frameshift upward one focal level. Interactions of individuals determine fitness of group. For a long time we didn't, but we now have statistical tools; it's hard, but it can be done.
Heritability at multiple levels not well addressed yet, and do not appear here in this talk. Heritable group effects on individual fitness arise by:
You as the observer can make the choice to frameshift upward. The Evolutionary Transition view combined with that of Hierarchical Selection is what is needed. Standard biologists only recognize highly organized social insect colonies as a higher order "individual", where it's really a continuum.
Group effects are modelled as adjustments to individual fitness of all members equally. This creates a definition of groups as having effects on individual fitnesses.
higher level groups -> Phenotype <- groups of protocells, local communities ("Patch", "Trait-Group")
Individuals may belong to large number of groups at a variety of hierarchical levels. Inheritance and reproduction are part of "selection" at lower level. Track these through time, and __you do not get a genealogy (not a binary genealogical tree). Ecological entities are defined differently.
[I asked the first question: many of us, programmers in the Alife community, derive our understanding of selection from Dawkins, but you can't model Lynn Margulis' symbiogenesis readily with any selfish gene theory; How do you think Alifers should modify their systems to incorporate hierarchical selection?] John Stuart[?sp] in a recent Alife journal proposed adding "transition managers"; Damuth would rather see the managers emerge than imposing them by force.
[Skipping forward in time to lunch: I ate lunch with Joao Munoz, the workshop organizer, and John Damuth, to explore specific strategies for incorporating hierarchical selection in Alife.] No managers; read Stuart's article, but it's not going to be designed managers. Higher level "individuals" have to be able to emerge.
The environment carries its own, entirely extra-genetic, heritability. Memory in nature: genetics + physical law; phenotype-genotype interactions carry through time as their own kind of heritability with natural selection. Sounds like it's heading right back to morphic resonance. Something is missing to attract formerly independent individuals at lower levels of organization to become composite newly-defined individuals at a higher level. Just where is it carried outside genetics? (in nature. It's easy to add things in software to track and remember past interactions of individuals and use that information when determining individual fitness).
I said to Damuth "Then you're giving me permission to take Context into account when I assign fitness to an individual?" -- "No: You cannot ignore Context when looking at the fitness of an individual." It is necessary to take group effects into account in order to know the fitness of an individual. Ernst Mayr and Dawkins are wrong when they claim that all group-level effects are, in fact, only the sum of individual effects, in spite of Mayr's 1979 proof. See "Alternative Formulations of Multilevel Selection", John Damuth and I. Lorraine Heisler, Biology & Philosophy 3 (1988), 407-430. [1988 or 1998?]
This sets up the possibility of engineering an Alife system with a contextual analyzer: you could run it switched on and observe the emergence of complexity -- then switch it off, and see what happens (probably the equivalent of weeds sweeping through an ecosystem - the kudzu vine). Not only might the system enable the emergence of symbiogenesis, but you could use it to run experiments to test the validity of hierarchical selection.
At lunch we also talked more specifically about what you would need to put into a system to enable its equivalent of the Eukaryotic Transition to emerge. Say one type of organism produces oxygen, which is toxic to a second (anaerobic bacteria), but desirable to a third (ancestors of mitochondria). My fitness is higher in the presence of oxygen when there are mitochondria in close proximity to metabolize it. Putting the mitochondria inside my membrane has to enhance their fitness by protection or something. Easy to put it in by design, but harder to set up a rich enough system that it can emerge spontaneously.
James Valentine, phanerozoic paleontologist, UC Berkeley (back before lunch):
There are two kinds of evolution of genes:
If it looks like a duck, walks like a duck, etc. ... Hierarchical structure written all over the map of biological evolution. Ediacaran from 580-543 million years ago (m.y.); 1 mm wide tracks of unknown animal forms in mud. Before the Ediacaran are only fossil embryos. At 543 m.y., start of Cambrian, all of a sudden you see these 1 cm tracks. The 13 million year period from the base of the Cambrian to 530 m.y. was the really big explosion. Then after the 10 million years from 530 to 520, you had all the modern phyla.
Valentine showed a slide of the opening page of Darwin's On the Origin of Species. Look at the hierarchy here: letters -> words -> syntax -> semantics. Words are linear in time, but to make sense of them, need hierarchical organization.
15,000 - 20,000 structural genes specify the body. About the same as the number of words in the average person's vocabulary. Shakespeare used 40,000 words, better. Humans have 60,000 - 80,000 translated genes, about four times more complex [in a linear sense] than inferred at the start of the Cambrian.
The gene expression cascade that in Drosophila is responsible for dorsal-ventral separation during morphogenesis, is the same cascade that in vertebrates is responsible for the immune system. The lesson is that the code, the gene-switching, could be anything. In both fruit fly and vertebrate the same enhancer molecules are synthesized at the end of that cascade, but they perform entirely different functions. Microevolution is fine-tuning. The real stuff isn't Mendelian, but rather in the patterning of gene expression.
50% of genes are homologous across all phyla. Yeast and human share 98% of the same genes [this is not the same as the 98% similarity between human and chimpanzee; in the latter, it's not only the genes, but the sequence that's 98% the same]. So it takes only a slight repatterning of yeast genes to get a lawyer.
Evolution inside genes is not what's going on. At this point someone asked Yes, but isn't the patterning itself controlled by the microgenetics? Valentine: Well, partly, but there's a whole lot extra. [This is a very thorny question, and cuts right to the debate between those who feel lower levels fully determine emergence at higher levels, and those who feel higher-level effects can reach back down to affect lower levels, leading to emergence being causally determined by more than just the lowest levels, e.g. the microgenetics].
Valentine put up some slides I didn't have time to copy fully, setting out the characteristics of four types of hierarchy [come back to this; it needs elaboration and seems important]:
=> Trees, branched positional structures, are not hierarchical. A genealogical tree looks like a hierarchy, but is not. Think of the family tree of cells in your body. You can trace the origin of each cell in the body back to the zygote; the mapping of the history of cell division is a tree, but the tissues in the body represent a hierarchy. Hierarchies have nested structure, where some parts relate to things above and below.
Genes are linear strings of nucleotides. The developmental stage is trees, not hierarchical. But bodies are hierarchical - organs, tissues, cells. Modules are hierarchical. If going to evolve modules, need hierarchy.
[From a programmer's perspective, hearing a paleontologist describe evolution this way was like learning about data structures: linear sequential GA's, tree structures, directed graph networks. The evolution of data structures through geological time!]
-- lunch with Joao Munoz & John Damuth, followed by: --
Steen Rasmussen, Los Alamos National Laboratories & Santa Fe Institute
Western philosophy: objects. Eastern: dynamics of objects.
Emergent properties do not reside at specific objects. "One of the Holy Grails in dynamical systems research is how to define a mathematical system from which objects can emerge at the next hierarchical level up".
"Repellon - bondon model in simulated water solution with hydrophobic / hydrophylic polymers: makes you think about quarks and gluons, fields, hard not to, makes you wonder about higher order systems". [That's a pseudo-quote. Steen, a respected physicist, was courageously veering into metaphysical territory. What he was suggesting is morphic fields, morphic resonance, without using those words.]
Combinatorics, even a 5 or 6 cell distance, approaches the number of particles in the Universe. Showed results of simulations with atoms at the bottommost level, faithful physics, polymers in water dynamically constructing micelle-like structures [prebiotic cells - nonliving membrane bounded spheroidal structures]. Scheduling is a deep issue in polymer simulation, from a lattice gas to polymer on a 2D triangular grid. Gridlock is a big problem. Our description collapses when you have dense structures.
Level 0: point particles. Level 1: short polymers with hydrophilic head and hydrophobic tail. "Dynamics for `individuals' are different at different levels". Level 2: micelles. Level 3 or 4: cells, structures that can self-replicate, which lower order structures cannot. Reviewed Per Luisi's self-replicating micelles from Alife III in 1992. Picture at first order a 3-mer approaching another 3-mer. In order to get higher-level interaction, where you could say the first individual was interacting with the second individual, at the lower level a single constituent particle in each 3-mer has to begin the process by interacting with a single constituent particle in the other 3-mer. Steen says in order to get 2nd order behavior, the process has to start off with 1st-order interaction, with the results of the interaction of a single particle propagating quickly to its adjacent particles in its own 3-mer. Only then can it be said the 2nd-order individuals are interacting with each other.
Steen: "We only assume that complex dynamical structures arise only from simple rules and states. This is unexamined dogma." -- heady stuff from a LANL dynamical physicist!
Gail Kennedy, UCLA anthropologist, on the possibility of hierarchical selection in hominid evolution:
"Ernst Mayr proved in 1979 that hierarchical selection is impossible". Maynard Smith lamented the lack of actual data that would help resolve the question. Kennedy studies hominid evolution in the Plio-Pleistocene, from 4 m.y. to about 1 m.y. When your view of hominid evolution is linear Darwinism, there's no need to think about hierarchical selection.
She then reviewed the history of fossil hominid discoveries, beginning with the discovery by Leakey in 1964 at Olduvai showing the coexistence of an Australopithecine with Homo habilis. Homo occupied a different kind of ecological niche which was defined as culture-bearing. Then at 2.5 m.y. came a radical climate change. Now Homo erectus known to have been contemporaneous with H.habilis and an Australopithecine.
Between 2 and 1 m.y. we have multiple Australopithecines and multiple Homo species (H.habilis, ergaster, erectus; A.robustus, afarensis). Absolute minimum of 4 species of Homo and Australopithecine coexisting. Homo rudolfensis at 2 - 1.75 m.y., high cranium. Her own research pushes Homo erectus all the way back to 2.0 m.y. (general consensus still starts erectus at 1.8 m.y.). Homo erectus has prominently retreating forehead. H.erectus at Turkana at 1.8 m.y. is identical to H.erectus at Peking at 0.5 m.y. Clearly you had multiple hominid species competing.
So: what factors allowed multiple species to coexist for such a long time in a deteriorating environment? (from 2 to 1 m.y. H.erectus coexisted with lots of others; then at 1 m.y. all the others were wiped out). None of the Australopithecines or Homos other than erectus made it out of Africa until []. Australopithecines were efficient grass eaters; Homo had a wider resource mix. [The talk had a lot of problems with A/V glitches -- which were to plague nearly every talk in every room all weekend -- and Gail never got to explore how hierarchical selection might provide a better explanation of hominid evolution. All she had time to say was that probably the group-level interactions were more important, ultimately, in determining which species of Homo survived.]
Closing Panel on Hierarchical Selection:
Damuth doesn't think emergent properties themselves have an effect on selection, even hierarchical. He adds that they don't tell us anything about causal structures.
Steen: engineer a system with eating, simple interactions, see if cooperation can emerge. Thinks doable. Says the computational intransigence of his own polymer-micelle simulations is due to starting at the level of atoms and being true to physics. [To get from atoms to water molecules and polymers to micelles required massive amounts of supercomputer time, weeks or months.] A hierarchical selection version of Tierra is not necessarily impossible. Someone: "It's all in the level you're focusing upon".
The adaptive fitness landscape for an individual would get more rugged as a higher order group evolves. If you're outside the group, you're dead. Steen: David Wolpert says it's all information, and hierarchical model gives you greater computational success [meaning greater computational success for real living organisms in their environment].
[The wrap-up in Hierarchical Selection kept getting back to whether Emergence was really something new, extra, in-its-own-right. Steen Rasmussen even ventured to question whether there existed some new field at the hierarchical organization level (a highish one) of `emergent phenomena'. Everyone in the room was pro-HS until the last questioner said he'd been keeping quiet all day, he was a confirmed selfish-genes-are-all-you-need-to-explain-everything man, and could Steen Rasmussen tell him differently? Steen: That's a tough one, but I don't like the heated debate between it's-all-selfish-genes vs. it's-all-holistic-higher-order. There has to be a lot in between, No?]
This was the first appearance of my prints in an exhibit with other works, so that was fun. Charles Taylor and Chris Langton came over and said they'd like to use some images in the Alife journal.
Rodney Barry from Sydney had a really cool interactive exhibit in its own room, with cube-shaped critters moving around a plane, emitting short musical pieces; their chromosomes coded for short MIDI sequences, and they decided whether to mate based on a distribution of consonance vs. dissonance (too similar, don't mate; too different, don't mate). An encounter would leave a little cube that would get bigger if it passed over a food pyramid. The audio was quadraphonic, and you sat in the driver's seat steering around the landscape with the mouse, hearing the sounds emitted by the critters nearest you, so it was sensory-directional. Rodney was originally from Tasmania, and he said that accounted for his adding inbreeding to the genome.
Met Jeffrey Ventrella for the first time, after hearing about him for quite a while. Nice guy, does interesting stuff. Later in a talk he had a 2D simulation running that exhibited Sims-like creature behavior, but on a PC, not a CM-5. The simplifications necessary to do that on a PC were collapsing from 3D to 2D, cutting in at a higher level of organization, and using a much-simplified model for interaction with the environment (Karl says the physical simulation part of his worlds took enormous computational resources). What Jeffrey has done turns out to be relevant to the theme of Digital Biota II in Cambridge next September.
Marcos Novak had four prints that looked to me like snapshots out of a 4d space. Later he came up and explained that, in fact, they were. He does "liquid architecture" in a very complex and interesting system which includes musical genes evolving and determining the shape of the 4d physical space.
There were lots of interesting and worthwhile exhibits. Several people made the comment that this was better than any Siggraph art show: if this isn't just hometown pride, then Alifers do computer graphics better than computer graphicists do. (Or at least better than those who run the gauntlet to get their stuff into the Siggraph show, which is fraught with logistical difficulties).
-- Barbecue on the grass --
Hooked up with Linda Emme, Ken Musgrave, Jeffrey Ventrella, Peter Hughes, Charles Ostman, and others at a dinner table. Programming war stories. How hard will it be when I have to port everything from IRIX to some Windows-NT box [Silicon Graphics is going down in flames; all the UNIX workstation manufacturers are having to switch over to Intel chips and the general feeling is you have to switch to the closest thing to UNIX, Windows-NT, awful as it is]. "Expect pain, man." "Better have a lot of Valium." How many lines of code? Me: about 100,000. Expect lots of pain. Ken: took me a half day just to get Hello World. Jeffrey: Use Code Warrior, start with template sample code. Ken: yeah, I just started from the library reference. Me: OK, 100k lines, 14 hours a day -- how long? Not six months, but painful. SGI has a two year transition plan, but you're just prolonging the inevitable; might as well just bite it and plunge into NT. We got onto resolution and the visual cortex. I passed around a copy of Through Caverns Measureless; Peter Hughes: universal patterns in visual cortex. How many pixels - no, how many rods & cones in the retina? No one knew, but the assumption is that it isn't billions, rather that there is very fast front-end processing on the `pixels' in the foveal region before compressing complex information down into the optic nerve, then a lot of back-brain processing filling in the details around the edges so it seems like super-high resolution, much more than IMAX / Showscan, but it might actually be of the same order or even less, just with much better software.
Gerald Joyce, UCSD. What's Evolving in Wet Alife? (invited talk)
Wet Alife [done in test tubes] is mostly funded by NASA, for the origins of life on earth and elsewhere in the solar system. Unlike DNA, RNA has secondary and tertiary shape structures that enable it to act as an enzyme catalyzing reactions. Genetic in bits, but phenotypic in function. A wet Alife evolution system couples a loop cycle mutating double-stranded DNA, coupled with an amplification loop cycle where RNA reads the mutated DNA bit string, then is subjected to selection for activity, producing complementary single-stranded DNA, which then closes the loop back to the double-stranded DNA. Mutation is accomplished by "dirty PCR" and other means.
Why RNA? We know it can be both chicken and egg. DNA as a double helix is locked in, cannot go out and play and be functional. Joyce showed results of several-day runs searching through astronomically large combinatorial spaces, with mutation producing millions of variants then selected down by comparable factors, resulting in a functional RNA molecule. The "business end" of that molecule contained a 15-nucleotide loop. With four types of nucleotides (A,C,T,G) that is 4^15 possible configurations (> 1 billion). Experimentally, Joyce found that swapping just the 7th nucleotide lowered the molecule's activity by a factor of 100. Swap any of the other nucleotides and the function is gone. This is a DNA enzyme.
Introduce predator-prey thing. Massive kill-off; RNA enzyme undergoing rapid continuous evolution. "We intentionally did with PCR what the L.A. crime lab in the O.J. Simpson case did unintentionally". You get what you select for, but you don't always know exactly what it is you're selecting for.
A questioner asked how could something like this have started on the prebiotic earth. Joyce: "For me, the question is How do you start Darwinian evolution without already having Darwinian evolution?"
Linglan Edwards and Yun Peng, U.Maryland, Computational Models for the Formation of Protocell Structures
[This talk is relevant to current interests in that it provides a contrast to Steen Rasmussen's totally bottom-up physics simulations of micelles that took weeks of supercomputer time; here, they began the simulation at a higher level of organization, using forces appropriate to the general notion of polymers in solution rather than actual physical forces down at the atom level.]
In earlier work, Drouffe et al. got membrane-like structures to form in 1991 for 1962 particles after 14 days of straight computation on a Sparc-1 workstation. In 1993 Heller et al. ran an atomic-level sim with 27,000 atoms beginning with a hand-constructed initial state of a rectangular patch of bilayer lipids in water. It took 14,640 hours to reach thermal equilibrium on a 60-node MIMD parallel supercomputer, approximately equivalent to 20 months of Cray-2 time.
In this work the authors chose to enter the game at a higher level of organization, focusing on the different hydrophobicity of the lipid's head and tail at their bottommost level. They used 200 lipid-like structures on a 900x900 grid with the particles initially randomly distributed. Two simulations produced micelle-like structures in 260,000 iterations. The first was on an SGI R4400, taking 76 hours; the second on an SGI R8000 at 57 hours. [Lesson: to get interesting life-like emergent behavior anywhere in the near future - like, 10-20 years - you have to cut in at an even higher level of organization; you simply can't be mucking about down at the atomic or even generalized lipid level.]
[I spent the next couple of talks trying to work out whether one could use actual C++ objects in an evolutionary system, where mutation & sexual crossover would modify source code subject to compiler syntax restraints, then request compilation - RNA reads code, produces functional proteins as executable code, which then do stuff with the environment and with the parent source code].
By Saturday afternoon I was getting the feeling that the Alife talks had lost their quirky creative character from 1992 and 1994, and were more like mainstream scientific talks. If the talk wasn't full of incomprehensible equations, it wouldn't have gotten in. From now on I spent as much time out in the halls talking to people as I did in the lecture rooms.
Anil Seth, U. Sussex, The Evolution of Complexity and the Value of Variability
Externalism: be able to say things about a system by looking outside it. Variation of `environment' in iterated Prisoner's Dilemma was accomplished by adding random noise in each individual's interpreting its memory of past interactions. Without random noise, complexity did not emerge, while with it, it did. In another experiment, the evolved neural networks of simulated robots learning to `home' to where their batteries got recharged, the conclusion was "Noise is good. Better than crystalline stasis."
Jeffrey Ventrella, jeffrey@ventrella.com, Attractiveness vs. Efficiency (How Mate Preference Affects Locomotion in the Evolution of Artificial Swimming Organisms)
Jeffrey's talk was of interest because it exhibited Sims-like behavior of evolved locomotion and emergent behavior modeled on PC's, not Connection Machines. The difference is that he reduced the dimensionality of the world from three to two, employed a much-simplified physical interaction with the environment, and constrained the physical structure of his "Swimbots" to a rectangular parts connected in a tree-like topology. The genetics was a linear GA, not Karl's nested directed graphs. With such simplifications, Jeffrey was still able to observe some emergent behavior, though to my eye it was sometimes so coarse-grained it was hard to get a good feel for it. The point is, there exist interesting avenues that can be explored on affordable computers, and I guess exploring the envelope surrounding the practical is the theme of Digital Biota 2.
-- MIT Press reception after the Saturday talks --
Bruce Damer grabbed me to say that Hiroaki Kitano, with whom he was talking, had wanted to come to Digital Burgess last year but couldn't... however, Kitano had just suggested to Bruce that we should hold Digital Biota 3 in 1999 at Shark Bay, Australia, in honor of the living stromatolites and to focus on the Eukaryotic Transition. This was of course a favorite topic of mine, and we made great plans for 1999 at the reception, then later at dinner in Westwood Village. More on this elsewhere than in this report.
Leonard Adleman, USC (invited talk)
Adleman is the `A' in RSA encryption, and one of the first to recognize and study computer viruses, to notice the organic nature of computation, and to come up with computational models of the organic.
DNA computation: 4 letter alphabet, quartenary. DNA encodes RNA, which builds proteins, which do everything. 100,000 proteins. Some proteins - DNA polymerase - act on DNA to make Watson-Crick complement (reproduction). Others split, cut, and patch the DNA that constructed them.
1 gram of DNA contains as much information as 1 trillion CD's. 10^18 operations / second. For one joule of energy you get 2 x 10^19 operations. He showed a Sierpinski's Gasket produced using block cellular automata with DNA computation. Can have 4-stranded DNA, 4 nanometers high, 12.6 nm long: planar DNA.
Biology meets Computer Science. Life / Computation / Self-Assembly. Tools: polymerase, ligase, ... so go build stuff with these tools that's cool.
Tom Ray, ATR, Kyoto: Evolution of Differented Multi-threaded Digital Organisms
Tierra-1 exhibited an intial rich increase in complexity, then reached a stasis period. Read John Maynard Smith & Eors Szathmary, "The Major Transitions in Evolution". I'm not interested in the Precambrian / Cambrian transition, rather in the steep increase in complexity right after the transition. The interesting stuff happened in the steep part of the curve right after the start of the Cambrian. [By `not interested', I think Tom means as a practical matter he's not trying to induce the transition itself - too hard, too computationally intensive to wait for it to emerge on its own; have to start with something you have a hope of seeing in your lifetime].
Now starting with two cell types, an already differentiated state. We should test whether our model works before we study its spontaneous emergence. The new Tierra organisms:
The model is:
A mature organism has two reproductive cells and eight sensory cells. The sensory cells gather data about machines on the net. Genome contains fecundity, virtual machine cpu speed, number of cells, average age, soup size, ... how long does it take for a ping to return, uptime, ... The complex speed / time landscape gives scope for evolution.
The original loss of sensory cells Tom reported at Digital Burgess was due to a bug. Now both sensory and reproductive cells evolve together.
At the end of the talk, the first questioner was Chris Adami, the conference organizer, and author of Avida, Chris' two-dimensional descendant of Tierra. Chris kept telling Tom he should put something or other into the new Tierra. Tom respectfully let Chris know he could do whatever he wanted, but Tom was going to do his own research his own way. [Personal note: Some of us had been getting increasingly unhappy with the hard-science reductionist direction in which Adami had taken the conference, and I thought Adami's comments to professional Tom Ray were out of line and rather arrogant.]
-- after evening banquet and NASA extrasolar life search speech --
A nucleus of Biotians and Alife artists formed out on the plaza. Most of us agreed that the Alife conferences had peaked at AL3 in Santa Fe, then begun a gradual decline into boring hard science accelerating at AL5 in Nara, Japan 2 years ago [which I missed], into the present sorry state. Where were the lively philosophical debates at AL3? Where was the creative spark? Well, they're alive and well in the Biota group and among the Alife artists, thank you. Chris Adami had rejected Steve G's Creatures paper because it was "too commercial". That's shooting himself (and the Alife field) in the foot.
Now it's all stuffy equations. If the talk is comprehensible, it can't be science. Use of metaphor and analogy to illuminate complex subjects is frowned upon. Word is that the Santa Fe Institute is likewise experiencing hardening of the arteries, ossification. Chris Langton is out after a bitter break-off. `Reductionist-sucking-pig-ism' seems to have triumphed for the moment...
Well, too bad. Now it's the Biotians and artists carrying the torch (with still a fraction of new and old Alifers). If we can really pull off Digital Biota 3 at Shark Bay in 1999, pull the Lynn Margulis / symbiogenesis faction [and dare I say it, Lovelock / Gaia / bio-geo-metabolism faction?] into the action, I think the movement will gain even greater creative momentum.
Christos Papadimitriou and Martha Sideri, UC Berkeley, Computational Complexity in the Life Sciences -- invited talk
[On the face of it, this could have been a boring mathematical talk, but Papadimitriou was better than that, and raised some deep questions.]
Alifers believe algorithms emerge through selection. Papadimitriou doesn't. An NP-complete problem [that takes 2^N operations to solve] is bad, like the traveling salesman problem. Very many hard problems can be reduced, mapped, to the traveling salesman problem. Some think mathematical selection leads to algorithms, like the protein-folding problem: linear DNA -- sequential protein construction -> complexly folded 3D geometry -> biological function.
But what is being optimized here, between the linear DNA sequence and the complex 3D geometry? "It is remarkable that protein shape is consistently unique". The conventional wisdom since the 1960's is that this is an energy minimum. But Levinthal's paradox: how can a 3000-atom protein molecule fold perfectly in far, far less than one second? Picture an undulating energy landscape. If it rains marbles, how come they all end up just here, in the lowest-energy configuration, and not spread out in adjacent, but not quite-so-deep wells? [I.e. why isn't there a distribution of different folding patterns of the same protein, representing those with similarly low, nearby energy minima.]
The current work simplifies the still impossibly-hard protein folding model by making a 2D model with instead of a 20-letter amino acid alphabet, only a 2-letter hydrophobic-hydrophilic alphabet. Maximize p-p adjacency on a 2D grid from the collapsed protein string of hpphhhpphp..., and you get a remarkably successful model of actual 3D protein structure. There is now a mathematical proof that "Protein folding in the 2D h-p model is NP-complete". Papadimitriou says it is a 50 page proof, must be the hardest NP-complete proof known.
So, is Nature solving an NP-complete problem? CP: I think not. Nature is solving extremely easy instances of an NP-complete problem. The 2D hp model says the landscape itself may be flatter, so naturally all the marbles fall into the same place. But then how do easy instances of hard problems evolve? The hp model hopes to find an hp sequence with __unique optimum folding, but this is ongoing work.
Moving on to the iterated Prisoner's Dilemma, why does Tit-for-Tat (TFT) evolve? Why don't we always write bad checks? Because we have to shop tomorrow. The fact that we know the interaction must be repeated guides the current interaction. But in formal Game Theory, if we decide to play 100 games, we defect on the 100th play... but then, why not defect on the 99th play? In formal Game Theory, there is no easy answer [excluding, obviously, any notion of higher-order transcendent altruism not explicitly encoded in the selfish genes running the whole show down below, or maybe just insufficient knowledge of how many plays you're going to make].
CP: TFT evolved because its competition was not too formidable. If you play the Prisoner's Dilemma with very simple players, then you get cooperation, which is exactly what evolution produced. With two possible states (cooperate or defect), from 64 possible strategies there are only four rational ones:
A trait X would not evolve if X too complex.
It is thought that the prions responsible for Kreuzfeld-Jacob's Syndrome [like "mad cow disease"] are because there are two possible ways for the particular protein to fold. Get the wrong one, your brain metabolism is cooked. [Note to RS: something to consider - multiple stable chreodes?]
In Q&A, the guy next to me said that in statistical mechanics and kinetics, only a small class of proteins has a unique folded solution. [Another note to RS - above some complexity threshold, is the Universe just too young for stable configurations to have settled out?]
C. Taylor: I wonder if diseases themselves are easy-to-solve genetic problems?
Chris Langton, Closing Remarks
Chris drew a graph with Inquiry on the vertical axis, and "Reality" on the horizontal. At the base of the vertical axis was "Artifacts", while at the top was "Natural". At the left of the horizontal axis was "The Possible - abstract theory", and at the right was "Actual, verifiable".
He drew a small circle in the upper left quadrant titled AL1, the first Artificial Life conference in 1987. Then came AL2 in 1990, enclosing AL1, but extending about halfway down and a bit further to the right. AL3 in 1992 continued the trend, still fully including 1 & 2 but extending now more toward the lower right. Then came AL4 at MIT in 1994, and now it became bimodal, with the MIT robotics people extending a large pseudopod into the lower right, while there remained a pod still in the upper left. AL5 in 1996 continued the trend, with now only a residual bit still up in the what's-possible-in-the-natural-world corner.
Chris applauded the maturing of the discipline and its increasingly verifiable nature, but was concerned that there was too much of a shift away from the Possible. CL: "The difference between Natural and Artificial is in the nature of the word "made". Nature makes things like itself, while we reached outside ourselves to make other artifacts.
"But, man was made by Nature, therefore artifacts are natural. Drop the distinction between Los Angeles and a termite mound. Our artifacts will not replace us, we'll be a part of it. It's all Biology!"
[And with that resounding theme, the conference ended. Chris' last statement resonates with those of Lynn Margulis and Dorion Sagan in recent books. (When the first oceanic life forms reached outside themselves and started constructing hard shells out of calcium carbonate, was that viewed as alien technology by the rest of the biota?)]