andrew@dtg.nsc.com (Lord Snooty @ The Giant Poisoned Electric Head ) (03/22/90)
An article in "EE Times", March 19th, discusses projects on Artificial Life. Perhaps the most radical proposal (by Cariani of Eaton-Peabody Labs, Boston) is for a system which "evolves its own special sensors" in response to the dynamic environment. This piques my curiosity because it sounds impossible. Given (and it's a big given) that sensor assembly technology were available to an automaton in any case, how the hell would it know what to sense, given the "requirement"? I take it that the "requirement" would be some kind of detection of nonseparability in classification of environmental variables... which is probably possible (but could also be highly nontrivial!). Despite my incredulity at the proposal, automatic sensor (and indeed effector) generation is part of what evolution is all about; the wing, the eye, etc. So the existence proof is to be found in Nature. But how to actually engineer this? I'm all ears :-) -- ........................................................................... Andrew Palfreyman andrew@dtg.nsc.com Albania before April!
kp@uts.amdahl.com (Ken Presting) (03/22/90)
In article <782@berlioz.nsc.com> andrew@dtg.nsc.com (Lord Snooty @ The Giant Poisoned Electric Head ) writes: >An article in "EE Times", March 19th, discusses projects on Artificial Life. >Perhaps the most radical proposal (by Cariani of Eaton-Peabody Labs, Boston) >is for a system which "evolves its own special sensors" in response to >the dynamic environment. > >This piques my curiosity because it sounds impossible. Given (and it's a >big given) that sensor assembly technology were available to an automaton >in any case, how the hell would it know what to sense, given the >"requirement"? I take it that the "requirement" would be some kind of detection >of nonseparability in classification of environmental variables... which is >probably possible (but could also be highly nontrivial!). > >Despite my incredulity at the proposal, automatic sensor (and indeed >effector) generation is part of what evolution is all about; the wing, the >eye, etc. So the existence proof is to be found in Nature. > >But how to actually engineer this? (I hope you're not put off by wild speculation...) A look at small-scale selection systems may be helpful. When I think about artificial life, I always think of the early, chemical stages of evolution. Ribosomes are made of the same material on which they operate, so they are nicely analogous to programs in a machine with integrated data and program storage. A ribosome is sort of an enzyme made of RNA. An enzyme does discriminate between objects in its environment, which is one important part of "perception". Of course, there is no intentionality or deliberation involved. And there are only two categories - "(my) substrate" and "not (my) substrate". It may be a problem that enzyme recognition of a substrate is unconscious, but on the other hand, there is no confusing *homunculus* suggested by enzyme behavior. Of course, no particular ribosome would "design" a new sensor for itself, but that is the whole point of imbedding them in a selection system. The immune system generates a huge variety of cells using regions in the genome which are very specifically "optimised" to be *inaccurately* expressed. These cells then make antibodies, which randomly match non-self molecules, and are cloned when one of them scores a hit. This is a case of a complex system (the body) using a generate-and-test strategy to create new detection devices. Also unconscious, but at least this system does not have to be imbedded in a population to show something simliar to selection. (this is a lousy description of a beautiful phenomenon, but any immunology text would probably give an accessible account) Finally, natural scientists are always building new machines to detect phenomena of new types. Whether selection occurs among scientific theories is perhaps a matter of metaphor rather than fact, but it is an intriguing idea. My bet is on the ribosomes. You gotta hand it to those little buggers. The first digitally controlled self-assembling machine. Ken Presting
andrew@dtg.nsc.com (Lord Snooty @ The Giant Poisoned Electric Head ) (03/22/90)
In article <8bn=02lt94=p01@amdahl.uts.amdahl.com>, kp@uts.amdahl.com (Ken Presting) writes: [eloquently, In Praise Of Ribosomes...]
I'll buy the ribosome model (or is that "paradigm" these days?).
When I first understood how the "little buggers" work, I did indeed
"hand it to them". It's much better (no flames please, this is a qualitative
assessment) than a Turing machine, because there exists this "soup" ("pool"
for softies) where everything is accessible, if you wait long enough.
This sort of reminds me (using *my* concept soup) of an Asimov vignettte
where he outlined various galaxy-wide search strategies. One involved travel
to a randomly-picked planet, standing still in a randomly-picked spot....
and waiting for the object+person to turn up. Eventually it did, of course.
Scuze the digressive analogy.
This is the way of the Nature Computer (read ribosome). It's fully parallel
with indeterminate but estimable (given solute densities) delays and also
it's associative.
My point is this: it is a very beautiful architecture, but no-one builds it;
not even Neural Nets are *that* parallel.
--
...........................................................................
Andrew Palfreyman andrew@dtg.nsc.com Albania before April!cliff@unix386.Convergent.COM (Cliff Neighbors ext 3216) (03/22/90)
In article <782@berlioz.nsc.com>, andrew@dtg.nsc.com (Lord Snooty @ The Giant Poisoned Electric Head ) writes: > Perhaps the most radical proposal (by Cariani of Eaton-Peabody Labs, Boston) > is for a system which "evolves its own special sensors" in response to > the dynamic environment. ... > Despite my incredulity at the proposal, automatic sensor (and indeed > effector) generation is part of what evolution is all about; the wing, the > eye, etc. So the existence proof is to be found in Nature. > it seems backwards to me to think of the natural organism (system) to "evolve its own special sensors", etc. as if the process is something that it performs or controls. I thought evolution was something that happens to the system, where the environment beats the shit out of it ("selection") and the only way it survives is by replication, mutation and luck -- not by "responding" to the conditions. It seems to be a process of destruction more than responsiveness, the zero degree of learning. -cn- ---- cliff neighbors: uunet!pyramid!ctnews!zardoz!cliff, cliff@zardoz.Convergent.COM ----
smoliar@vaxa.isi.edu (Stephen Smoliar) (03/23/90)
In article <792@berlioz.nsc.com> andrew@dtg.nsc.com (Lord Snooty @ The Giant Poisoned Electric Head ) writes: > >I'll buy the ribosome model (or is that "paradigm" these days?). >When I first understood how the "little buggers" work, I did indeed >"hand it to them". It's much better (no flames please, this is a qualitative >assessment) than a Turing machine, because there exists this "soup" ("pool" >for softies) where everything is accessible, if you wait long enough. > >My point is this: it is a very beautiful architecture, but no-one builds it; >not even Neural Nets are *that* parallel. Part of the problem is that we still really do not have to resources to build such an architecture. Neural nets still cannot capture the magnitude of most nervous systems which we study. If you want to talk about ribosome soup, you are dealing with a system of even larger scope . . . perhaps so large as to be beyond the ability of our imagination to design it. ========================================================================= USPS: Stephen Smoliar USC Information Sciences Institute 4676 Admiralty Way Suite 1001 Marina del Rey, California 90292-6695 Internet: smoliar@vaxa.isi.edu "Only a schoolteacher innocent of how literature is made could have written such a line."--Gore Vidal
smoliar@vaxa.isi.edu (Stephen Smoliar) (03/23/90)
In article <946@unix386.Convergent.COM> cliff@unix386.Convergent.COM (Cliff Neighbors ext 3216) writes: > I thought evolution was something that >happens to the system, where the environment beats the shit out of it >("selection") and the only way it survives is by replication, mutation >and luck -- not by "responding" to the conditions. It seems to be a process >of destruction more than responsiveness, the zero degree of learning. > What Edelman and other immune system researchers have discovered is that selection may occur WITHIN an organism, as opposed to "happening to" it. In other words, the organism starts off with far more variety than it needs and builds its own architecture through selective processes. This explains such observations as that fact that no two organism are "wired up" exactly the same way. ========================================================================= USPS: Stephen Smoliar USC Information Sciences Institute 4676 Admiralty Way Suite 1001 Marina del Rey, California 90292-6695 Internet: smoliar@vaxa.isi.edu "Only a schoolteacher innocent of how literature is made could have written such a line."--Gore Vidal
zarnuk@caen.engin.umich.edu (Paul Steven Mccarthy) (03/23/90)
> andrew@dtg.nsc.com (Lord Snooty) writes: > [... How can a machine know when/how to build its own > sensors/actuators? ...] Mankind has been building its own sensors for a couple of centuries now. The development of radar, radio, telescopes, microscopes (...ad nauseum...) represent sensors which humans have developed to extend the capabilities of the ones that were "built-in". How/why did humans develop these instruments? 1) Recognizing "interesting" characteristics of existing devices and extrapolating, experimenting. 2) Recognizing logical implications of existing "world-models" which imply the existence of some property/characteristic beyond current sensor- capabilities and applying the results of (1) above to investigate the predictions of that "world-model". 3) Serendipity. 4) Market-forces. I am not really trying to assign any "intentional" force to mankind as a whole. However, I think it is fair to assign "intention" to most of the individuals who have been primarily responsible for the development of these "sensors". I think the motivations/methods implied by (1) and (2) are machine-implementable. The program "AM" by Doug Linnet (then at Stanford) was supposed to have developed its own "tools" for investigating set-theory. It was given an initial set of characteristics which were "interesting" and a small number of simple axioms from set-theory. Not only did the program go on to expand its small set of tools (--this is the program that "discovered" prime numbers--), but it also went on to refine its meaning of "interesting". ---Paul...
kp@uts.amdahl.com (Ken Presting) (03/23/90)
In article <12533@venera.UUCP> smoliar@vaxa.isi.edu.UUCP (Stephen Smoliar) writes: >In article <792@berlioz.nsc.com> andrew@dtg.nsc.com (Lord Snooty @ The Giant >Poisoned Electric Head ) writes: >> >> . . . because there exists this "soup" ("pool" >>for softies) where everything is accessible, if you wait long enough. >> >>My point is this: it is a very beautiful architecture, but no-one builds it; >>not even Neural Nets are *that* parallel. > >Part of the problem is that we still really do not have to resources to build >such an architecture. Neural nets still cannot capture the magnitude of most >nervous systems which we study. If you want to talk about ribosome soup, you >are dealing with a system of even larger scope . . . perhaps so large as to be >beyond the ability of our imagination to design it. Blackboard-style algorithms (eg HEARSAY II) use the "wait for a substrate" approach. Newell's "production" systems use it too. My fave example, though, is JES or HASP - the spooling systems in IBM mainframes. These spoolers have "initiators", which select jobs from an input queue for execution. An initiator sits in an address space, scanning the queue for a job with matching identifiers, until it finds one. Then the initiator pounces on the string of JCL cards, linking the data sets identified on "DD" cards, and loading the programs identified on "EXEC" cards, until it reaches the end of the jobstream. There is even a facility wich is roughly the opposite of excising an "intron" - the JCL processor can insert new JCL from a "procedure library", whereafter it will merrily continue processing the expanded jobstream. I won't even get started on CICS automatic transaction initiation... Ken Presting ("System programmers do it with more imagination")