Newsgroups: comp.ai.alife,comp.ai.genetic,comp.theory.self-org-sys,comp.theory.cell-automata,comp.parallel,comp.ai.neural-nets,comp.ai
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!news.mathworks.com!uunet!newsflash.concordia.ca!nstn.ns.ca!dragon.acadiau.ca!bounce-back
From: Eugen Leitl <ui22204@sunmail.lrz-muenchen.de>
Subject: new computation paradigms
In-Reply-To: <Pine.SUN.3.91.950213135124.6506A-100000@sun1>
X-Sender: ui22204@sun1
Status: RO
Content-Type: TEXT/PLAIN; charset=US-ASCII
To: raja@santafe.edu
Message-ID: <Pine.SUN.3.91.950214153241.4541A-100000-100000@sun1>
Originator: rmuise@dragon.acadiau.ca
Lines: 194
Sender: news@relay.acadiau.ca
Nntp-Posting-Host: sun1.lrz-muenchen.de
Organization: Leibniz-Rechenzentrum, Muenchen (Germany)
References: <Pine.SUN.3.91.950213135124.6506A-100000@sun1>
Mime-Version: 1.0
Date: Fri, 24 Feb 1995 17:59:33 GMT
Approved: rmuise@dragon.acadiau.ca
Lines: 193
Xref: glinda.oz.cs.cmu.edu comp.ai.alife:2579 comp.ai.genetic:5049 comp.theory.self-org-sys:805 comp.theory.cell-automata:3345 comp.parallel:11854 comp.ai.neural-nets:22264 comp.ai:27756

 (This is a cc of my email to the authors of the recently
 advertised "Evolving Globally Synchronized Cellular Automata"
 paper. Since matters mentioned might be of general interest, I
 thought I'd crosspost it to some appropriate comp.newsgroups.)
 
 Dear Rajarshi/James/Melanie/James,
 
 thank you for your applaudable effort to bring a computational
 paradigm, commonly regarded as cryptic and irrelevant to matters
 of practice, to the attention of the general public.
 
 The paper (advertised in some recent newsgroup broadcast
 posting) is excellent and I read it with joy.
 
 (In this context I think I should mention a German book
 "Algorithmen in Zellularautomaten" by R. Vollmar, B. G. Teubner
 Stuttgart, 1979. Synchronization problems (FSSP etc.) and
 patterns transformation/recognition in CAs are the major issues
 of his readable book, which also contains an extensive
 bibliography. Doubtless you know all about this matters, but CA
 references being hard to find :( one can quite never be
 certain...)
 
 Since I did a superficial (vgrep) random search in 1d/2d
 low-state CA rule space in my early youth (about 1986/1987) I
 easily recognized the CAs in your paper (their archetypal
 phenotypes, that is ;). I also found both your results and your
 methods of approach to GA CA/CA-rule space search quite
 exciting.
 
 Becoming increasingly interested in molecular electronics (ME)
 about the same time I realized that CA computation paradigm is
 highly ME-suitable. (That's the reason why I wound up studying
 bio/chemistry instead of a Real Science in the first place ;).
 
 Some facts:
 
 1) Regular low-defect 3d lattices of high-complexity cells
    are not unheard of in real life (virus crystals).
 
 2) In vitro as in vivo viral crystals autoassembly on the macroscopic
    scale (crystal) as well on the nanoscopic (viral component) scale.
 
 3) Low-state/complexity 2d/3d automatons can emulate wires, signals and
    logical gates thus being equivalent/compatible to the computation
    paradigm, currently en vogue, plus offering the additional bonus of
    "soft circuitry" (SC).
 
 4) The problems of translating an algorithm into a finite state
    register machine till down to the level of individual logic gates
    had been solved and routinized to the point of automagicability
    (DACAPO III silicon compilers, Abelson & Sussman's Scheme chip
    (wizard book) approach).
 
 5) Recent advances in photonically/electronically switchable
    molecules have shown, that CA ALUs, while still quite beyond
    the state of the art of organic synthesis, are, at least
    theoretically, possible.
 
 6) Estimation of ME-CAM computational performance/integration
    density yielded, er,     impressive	   figures.
 
 7) Antibodies of high recognition specifity and/or catalytic
    activity are now being routinely (Pl"uckthun et al.)
    prepared with protein engineering methods, which
    is good news for synthesis/autoassembly.
 
 I think most (alas, not all) physical obstacles, riddling all ME
 concepts (flames to infamous Drexler et al.), are not very
 relevant for ME-CAs (construction, energy, I/O: autoassembly,
 laser irradiation/chemical energy sources, optical
 pattern-encoded I/O). There is one (logical) problem I still
 regard as quite formidable: that of the initial, or bootstrap
 virtual machine (BVM) configuration/state pattern.
 
 In theory, a small initial configuration can be written using
 the arrived raster tunnel microcopy (RTM) technique. It has then
 to reproduce itself, spreading through crystal bulk,
 initializing some parts differently (thus creating Meta-CA cell
 individuality/addressability/distinguishability; "switching
 genes in morphogenesis process" metaphor), thus preparing the
 necessary circuitry for the download of the optically encoded
 final/transient virtual machine (VM). The problems of designing
 the BVM and its different initialization in different parts of
 the crystal are formidable, yet (fragmentally) solved already.
 
 The hard problem is to discover the simplest possible CA (number
 of states, transition rule complexity/error tolerance) as well
 as the smallest possible BVM, since these two form the boundary
 conditions for the ME-CA/RTM complexity, a hard limit on scarce
 resources, being posed by current restrictions of the art of
 organic synthesis and RTM write time. I've thought about using
 GA search upon embed particles/glider-featuring CAs to discover
 new methods of self-replication, since I don't find von
 Neumann's read/write head approach very practicable. (Maybe a
 pair of movable heads, read and write, communicating by
 signals-in-extruded-wire or a particle stream? gliders of
 programmable life time? interacting glider streams, yielding the
 desired pattern?)
 
 Low-interaction-embed-particle ("neutrino glider")
 emitters/receptors can also be used to increase hypergrid
 communication bandwidth (broader usable Meta-CA cross-section)
 as well as packing density (less space for bus "wires"
 necessary) of Meta-CAs (higher-layer CAs, defined in terms of
 underlying CA layer(s).
 
 The metaphor of a shadow/ghost/mirror space, reserved solely for
 information transfer (temporaspatially-binary coded messages as
 sequences of gliders/particles set conformations) does also seem
 especially attractive, though probably ME-hard (number of
 states/transition complexity) to implement. Though one can use
 the above mentioned communicating r/w head pairs in shadowspace
 to do any computation, one should regard them as (inferior)
 additions to the signal/wire/logic gate approach, since they are
 nothing but a fancy Turing (ergo sequential) machine, constantly
 running a "read/endcode/transmit/decode/write" loop, operating
 on a 3d tape with state/write locality condition suspended.
 
 The ability of a MetaCA to achieve cross-space "strange loop"
 r/w to the shadow space with special/reserved switchable state
 configs allows it to steer a r/w head pair, empowering it to
 construct/copy other MetaCAs (but beware of runaway head
 pairs..). This facility should be used sparingly, however, since
 neutrino gliders _do_ interact in the shadow space (heavy
 choreography issues, that is, unless you'd want to implement a
 hierarchy of shadow spaces? ;)
 
 
 These are some of CA things I find currently quite interesting
 (I'm probably not alone in this):
 
  - high-entropy CAs in fast secure cryptography
    (a great source of one-time pad pseudo-rnd stream
    XOR-type cryptosystems)
 
  - CA trapdoor codes and public-key cryptosystems
    (Are they really possible? If yes: how does one derive
    public/private key-rule pairs? Would it stand a serious
    cryptoattack/Is the achievable cryptosecurity level
    theoretically provable?)
 
  - CA in fluid/particle dynamics simulations (direct Navier-Stokes
    type-differential-equation-to-rule-mapping to reduce the amount of
    computation; saw some very encouraging results already)
 
  - CAs as a toy model for non-linearity discipline
 
  - implications of CA upon neural nets (NN) science and
    vice versa (since they are both special cases of each other)
 
  - mapping message-passing hypergrid (high N) topology
    architectures (and implementing NNs in terms of them)
    onto 2d and 3d CA and routing therein (_very_ interesting)
 
  - fast CA algorithms (lookup-table code primary-cache optimized,
    the half-height-inverse-light-cone-hashing approach to
    shorten computation time ("blitz'n!"), flexible-resolution
    grids for increased computation space/time efficiency
    (array of (hookable)-mono-method objects to allow
    space/time-varying neighbourhoods/rules), etc.)
 
  - fast CAM (hardware) implementations (semiconductor "smart RAM"
    and ME, heterogenous (von Neumann/CA mix) and pure CA machines)
 
  - CAs world model in physics (Is Reality a CA? Analogies
    in thermodynamics (irreversibility, the arrow of time) spacetime
    quantization (Planck length and time), relativistic (light cone)
    and particle (gliders) physics)
 
  - CA/GA in A-Life (efficient population map dynamics modeling)
 
  - modeling excitable media with CAs (chemical (auto)catalysis
    system dynamics; exhaust cat, Belousov-Zhabotinsky/Dictyostelium
    discoideum, etc.)
 
  - biological morphogenesis with CAs (Turing's late works (with
    diffeqs))
 
  - CAs in information- and computability theory (???)
 
 CU,
 
   Eugene "Daedalus" Leitl.
 
 
 P.S. Have you heard of/done any research in this directions?
      It would be highly interesting to learn more about it.
 
 2P.S. BTW, do you have any sources of CA-relevant stuff on the net?
 
 3P.S. Was I talking sense?
 
