chemart / ikegami-hashimoto

hub:
  repo_type: generator
  builtin: ikegami-hashimoto
  tags:
  - banzhaf-yamamoto
chemistries:
- id: ikegami-hashimoto
  name: Machine-tape chemistry
  intuition: 'Von Neumann''s separation of constructor and description, made into a chemistry. Two populations
    coexist: machines, which act, and tapes, which are read. A machine reading a tape produces a new machine
    and a new tape, so reproduction requires both. A machine whose description is missing washes out of
    the reactor, which puts pressure on closed loops of machines that between them produce each other''s
    descriptions - including the minimal case of a machine and tape that reproduce as a pair.

    '
  aliases:
  - machine-tape interaction
  - coevolution of machines and tapes
  origin: Ikegami & Hashimoto, 1995
  book: 10.5.4
  refs:
  - '[419]'
  - '[420]'
  - doi:10.1007/3-540-59496-5_302
  family: automata
  kind: generator
  constructive: true
  S:
    repr: 'Two separate populations. Tapes: circular 7-bit strings, species T<2 hex digits> = the bits
      read clockwise from the tape''s source (the papers name a tape by its smallest rotation, extras.paper_names).
      Machines: 16-bit words, species M<4 hex digits> = T'' column, M'' column, head (4 bits), tail (4
      bits) of a transition table (tape bit, machine state) -> (tape bit'', state'').

      '
    definition: implicit
  R:
    definition: implicit
    arity: 2
    scheme: M + T -> M + T + M' + T'
    notes: 'M reads T if its head matches T at a site h (first match clockwise from the source) and its
      tail at a different site (first match clockwise after h). M rewrites the reading frame from h up
      to the tail site with its table, half of the machines starting in state 0 and half in state 1, giving
      T''. T'' is translated into M'' by reading 16 bits clockwise from the source (7 bits reused): bits
      0, 2, 4, 6 are the T'' column, 1, 3, 5, 7 the M'' column, 8, 10, 12, 14 the head and 9, 11, 13,
      15 the tail. One tape encodes different machines depending on its source, so rotations of a tape
      are different species.

      '
  A:
    reactor:
    - ode
    - well-stirred-multiset
    dilution: each population has capacity N; every generation a fraction c of each is replaced by reaction
      products (c = d = 0.6)
    notes: 'Discrete generations on integer populations: concentrations f = n/N, c N new machines and
      tapes shared among reacting pairs in proportion to f_M f_T (eq. 4), old populations reduced by the
      fraction d = c, integer part taken (eqs. 5-6). External noise flips bits in the reading frame with
      probability mu_P per bit (eq. 7).

      '
  params:
  - name: method
    type: enum
    default: dynamics
    choices:
    - dynamics
    - closure
    role: structural
    meaning: 'dynamics: the papers'' population dynamics, observed reactions with firing counts; closure:
      the noise-free reaction network reachable from the seed machines and tapes (Chemart addition)'
  - name: machines
    type: list
    default: []
    range: e.g. the minimal self-replicating loop [1002] with tapes [01]
    role: structural
    meaning: seed machines as 16-bit hex strings (T' column, M' column, head, tail); empty draws n_machines
      random machines
  - name: tapes
    type: list
    default: []
    range: e.g. [01]; the papers' T1 read from another source is 04, 10, ...
    role: structural
    meaning: seed tapes as 7-bit hex strings read from the source; empty draws n_tapes random tapes
  - name: n_machines
    type: int
    default: 10
    min: 0
    max: 65536
    range: 'papers: about 10'
    role: population
    meaning: number of random seed machines when machines is empty
  - name: n_tapes
    type: int
    default: 3
    min: 0
    max: 128
    range: 'papers: 2 or 3'
    role: population
    meaning: number of random seed tapes when tapes is empty
  - name: N
    type: int
    default: 1000
    min: 1
    max: 10000000
    range: not stated in the papers
    role: population
    meaning: 'dynamics only: capacity of each population (machines, tapes); seeds share it equally'
  - name: c
    type: float
    default: 0.6
    min: 0.0
    max: 1.0
    range: 'papers: c = d = 0.6'
    role: kinetic
    meaning: fraction of each population replaced by reaction products per generation (c = d_m = d_t);
      also the rate scale of eq. 4
  - name: noise
    type: float
    default: 0.05
    min: 0.0
    max: 1.0
    range: 'papers: 0.04 (minimal loop), 0.055 (oscillation), 0.07-0.08 (core networks), scans 0-0.1'
    role: stochastic
    meaning: 'dynamics only: external noise mu_P, bit-flip probability per bit of the reading frame'
  - name: generations
    type: int
    default: 150
    min: 0
    max: 100000
    range: 'papers: 1600-3000'
    role: population
    meaning: 'dynamics only: number of generations'
  - name: noise_off
    type: int
    default: -1
    min: -1
    max: 100000
    range: 'papers: 2000'
    role: stochastic
    meaning: 'dynamics only: generation at which external noise is turned off; -1 keeps it on'
  - name: source
    type: enum
    default: random
    choices:
    - random
    - inherit
    role: stochastic
    meaning: 'dynamics only: source of a tape whose circular pattern is new to the population: random
      site (papers) or the site it was written from'
  - name: max_species
    type: int
    default: 400
    min: 2
    max: 100000
    role: structural
    meaning: 'closure only: species budget (the noise-free closure has at most 128 tapes and 128 + seed
      machines)'
  provides:
  - topology
  - stoichiometry
  - catalysts
  - rate-constants
  - flow
  - initial-state
  - sequence-structure-function
  fidelity: reconstructed
  sources:
  - 'Ikegami, T. & Hashimoto, T. (1995). Active mutation in self-reproducing networks of machines and
    tapes. Artificial Life 2(3):305-318. Author''s manuscript dated May 31, 1996: model (sec. 2, eqs.
    2-9, fig. 1), results (figs. 2-9, sec. 4 list of self-replicating pairs). https://drive.google.com/file/d/1WwFXhB10rx8FiTPZZjUVpQIHl6BtN73D/view
    (linked from https://www.sacral.c.u-tokyo.ac.jp/publications)'
  - Ikegami, T. & Hashimoto, T. (1995). Coevolution of machines and tapes. In Moran et al. (eds.), Advances
    in Artificial Life (ECAL 95), LNAI 929, 234-245. Eq. 4 with c_ij = 0 for non-reading pairs, source
    rule, figs. 1-6. https://web.archive.org/web/20170809083155/http://sacral.c.u-tokyo.ac.jp/pdf/ikegami_alife_1995.pdf
  decisions:
  - Bit layout from paper fig. 1, not from the prose. The prose says head and tail come first and translation
    starts at the reading frame; the figure translates the rewritten tape [1110110] from its source into
    M ebbd with the table columns in bits 0-7 (T', M' alternating) and head/tail in bits 8-15 (alternating).
    That layout reproduces fig. 1 exactly, gives the machines of the published pairs (M1002 <- T1, M3006
    <- T5, M1222 <- T3), makes the five self-replicating pairs of sec. 4 self-replicate, and makes M1002
    the machine produced by the largest number of machines (sec. 5).
  - 'The rewrite runs over sites h .. tail-1 (fig. 1: head at site 3, tail at site 2, six steps), so L
    is 1-6 (the <L> axis of figs. 2-3 ends at 6). The tail is searched from h + 1 and may overlap the
    head.'
  - 'Tapes are strings read from their source, so the 7 rotations of a circular tape are different species
    (one tape encodes different machines). In the dynamics, rotations share one source (''every translational
    invariant tape has the same source''): a product whose circular pattern is present is merged into
    that rotation; a new pattern gets a random source (source: random) as in the papers. Translation of
    the product machine is from the source of the tape that was read, as in fig. 1. The closure has no
    population, so it keeps rotations apart.'
  - 'Normalisation: the Artificial Life text makes c N new machines and tapes per generation, the book
    says the population is kept constant, while ECAL eq. 4 divides by the sum over all pairs (so non-reading
    pairs waste production). Chemart follows the book and the Artificial Life text: the c N products are
    shared among reading pairs, d = c, and populations stay at N (up to integer parts). Each network reaction
    gets mass-action k = c/2 per initial state producing it (c when both states give the same products),
    and outflow constant-total: the rate equations are the continuous-time version of eqs. 4-6 without
    noise and integer parts, up to a rescaling of time by the sum of f_M f_T over reading pairs (checked
    in tests).'
  - 'Noise: the reaction terms are multiplied by 1 - eps with eps = 1 - (1 - mu_P)^L (eq. 7), and eps
    times the flux of each reaction and state becomes mutants, rounded stochastically to an integer (''at
    most eps c N mutant populations'', Monte Carlo). A mutant flips each frame bit of the rewritten tape
    with probability mu_P conditioned on at least one flip, and its machine is translated from the mutated
    tape. The papers do not say whether the error hits the read or the written bit; Chemart flips the
    written tape.'
  - 'Integer parts: populations are floored after each generation (so objects below one copy, f < 1/N,
    are removed); mutants are whole objects. Seeds share N equally; the papers do not give initial populations
    or N (default 1000).'
  - 'Observed network: every reaction that fired, with count = number of generations in which it fired.
    Error-free reactions carry the mass-action rate above and frame_length L (the noise factor (1 - mu_P)^L
    is not folded in, since noise_off changes it); reactions produced only by noise have no rate and are
    listed by index in extras.noise_induced. extras.analysis records per generation the distinct machines
    and tapes (figs. 4-6), the active mutation rate <mu_A> of eq. 9 (w/L averaged over reading pairs with
    weights m_i t_j and over both initial states) and the average frame length <L>.'
  - 'Dropped v1 params: tape_len and machine_len (the encoding is only defined for 7-bit tapes and 16-bit
    machines); compartments (the cell model of [420] is not implemented: the ALife V paper is not openly
    available).'
  - 'Paper errata: sec. 4 lists the self-replicating pair ''Mbdd1 with T37'', but bdd1''s tail 0001 cannot
    bind T37 (two zeros); Mbdd7 is the self-replicator of T37. It also writes ''T9dd3'' for M9dd3. Fig.
    1 calls the rewritten tape [1110110] Te6; its value is 76 hex (T37 as a circular tape). ECAL fig.
    1 names M1222''s tape T41, a rotation of T3.'
  - Sec. 4 says there are 5 possible self-replicating pairs (M1002/T1, M2004/T1, Mdffb/T3f, M9dd3/T1d,
    Mbdd7/T37). With this encoding all five self-replicate from initial state 1, but so do 10 more (e.g.
    M0000/T0, Mffff/T7f, M3446/Td); the papers do not say how the five were selected.
  phenomena:
  - a machine without its description tape is washed out; a closed loop of machines producing each other's
    tapes is needed
  - 'low noise: minimal self-replicating loop M1002 + T1 (Eigen-Schuster type, zero active mutation),
    metastable: it survives when noise is turned off'
  - 'higher noise (0.04-0.055): parasitic machines (M3006 with T5, M1222 with T3) invade and populations
    oscillate, with bursts of active mutation'
  - 'mid noise (0.05-0.1), after about 600-1100 generations: a core network of about 40-60 machines with
    high active mutation that keeps its diversity after noise is turned off (fixed point or oscillating);
    double autocatalytic loops (RNA-like editing) besides Eigen-Schuster loops (DNA-like replication)'
  - 'high noise: back to a minimal loop or extinction'
  - cell assemblies of core networks differentiate ([420], not implemented)
  notes: 'A von Neumann style replicator chemistry with separate description (tapes) and constructor (machines).
    Machines are only 16 bits and tapes 7 bits, so every reaction is a table lookup; the noise-free closure
    of any seed has at most 128 tapes. The papers'' core-network phenomena need long noisy runs (thousands
    of generations) with a capacity N they do not state.

    '