chemart / reflexive-ac

hub:
  repo_type: generator
  builtin: reflexive-ac
  tags:
  - banzhaf-yamamoto
chemistries:
- id: reflexive-ac
  name: Reflexive artificial chemistry
  intuition: 'Machines that compose with machines, including themselves. A molecule is a finite state
    machine drawn as a graph with a current state; a reaction feeds one machine''s output into the other
    and keeps the reachable graph of paired states as the product. Composition multiplies state spaces,
    so products grow quickly and are self-similar. The reflexive case is the interesting one: a machine
    composed with itself is a machine reading its own structure, which is the sense in which it can act
    on itself.

    '
  aliases:
  - graph-based reflexive artificial chemistry
  - graph-constructing graphs
  - reflexive composition of state machines
  origin: Salzberg, 2007
  book: '9.8'
  refs:
  - '[737]'
  - doi:10.1016/j.biosystems.2005.12.008
  - doi:10.7551/mitpress/1429.003.0084
  - arXiv:2505.07186
  family: rewriting
  kind: generator
  constructive: true
  S:
    repr: 'finite state machines (Mealy machines) drawn as labelled directed graphs: every node has at
      most one link per input symbol, labelled input/output, and one node is the current state (the pointer).
      Input data are graphs of the same kind (missing links allowed). Species id: canonical text with
      nodes numbered breadth-first from the pointer, node j listing destination/output per input, e.g.
      the adding machine M45 at S0 is 0/0,1/0;1/1,0/1

      '
    definition: implicit
    notes: only the part reachable from the pointer is kept, so pointed isomorphic graphs are one species;
      two-state two-symbol machines may be named by number (M45, M61@S1)
  R:
    definition: implicit
    arity: 2
    scheme: 's1 + s2 -> s1 + s2 + s1 o s2: the reachable composite state graph of s1 (sender) and s2 (receiver)
      under reflexive composition; elastic if it has more than max_states nodes'
    notes: 'Reflexive composition (Salzberg & Sayama 2025, eqs. 1-2): on input i the composite outputs
      phi_r(phi_s(i)) and moves to (delta_r(phi_s(i)), delta_s(i)), the two machines swapping the sender
      and receiver roles at every step, so neither is only processor or only data. A graph composed with
      itself reads its own structure (s + s -> 2 s + s o s). Repeated composition is the n-machine lattice
      of eqs. 3-4, and state spaces grow exponentially.

      '
  A:
    reactor:
    - graph-rewrite
    - well-stirred-multiset
    dilution: 'closure: none; soup: each reactive collision adds the product and removes a random molecule
      (population M constant)'
    notes: The source describes the reaction and its products, not a reactor; the flow-reactor soup is
      the book's standard AC reactor, a Chemart choice.
  params:
  - name: method
    type: enum
    default: closure
    choices:
    - closure
    - soup
    role: structural
    meaning: 'closure: every composition reachable from the seed machines (chemart.expand over ordered
      pairs), complete or cut off by max_species; soup: a stochastic flow reactor of M molecules, observed
      reactions with firing counts'
  - name: machines
    type: list
    default:
    - M45
    range: e.g. ['M45', 'M61'], ['M54'], ['0/0,1/1;1/1,0/1@1']; the arXiv:2505.07186 machines of interest
      are M7, M44, M45, M54, M60, M61
    role: structural
    meaning: 'seed machines: elementary numbers M<n> (optionally @S0/@S1 for the pointer) or machine texts
      ''dest/out,...;...'' with ''-'' for a missing link and optional ''@j'' pointer; all must share the
      alphabet. Empty: M random machines'
  - name: M
    type: int
    default: 50
    min: 2
    max: 100000
    role: population
    meaning: 'soup: reactor size (given machines are split equally); without machines, the number of random
      machines drawn (closure seeds or initial soup population)'
  - name: states
    type: int
    default: 2
    min: 1
    max: 1000
    role: population
    meaning: 'random machines: number of states before trimming to the part reachable from state 0; links
      are uniform over states and outputs'
  - name: symbols
    type: int
    default: 2
    min: 1
    max: 16
    range: the sources use the binary alphabet {0, 1}
    role: structural
    meaning: 'random machines: input/output alphabet size k'
  - name: max_states
    type: int
    default: 64
    min: 1
    max: 100000
    role: structural
    meaning: 'largest product accepted: a composition with more reachable nodes is elastic'
  - name: max_species
    type: int
    default: 100
    min: 1
    max: 100000
    role: structural
    meaning: 'closure only: species budget; reactions that would add more are dropped and status becomes
      truncated'
  - name: collisions
    type: int
    default: 2000
    min: 0
    max: 10000000
    role: population
    meaning: 'soup only: number of collisions (elastic ones included)'
  provides:
  - topology
  - stoichiometry
  - catalysts
  - flow
  - initial-state
  - sequence-structure-function
  fidelity: reconstructed
  sources:
  - 'Banzhaf, W. & Yamamoto, L. (2015). Artificial Chemistries, section 9.8 (one sentence: finite state
    machines and their input data deconstructed into a graph that rewrites a graph, and in principle itself).'
  - 'Salzberg, C. (2007). A graph-based reflexive artificial chemistry. BioSystems 87(1):1-12 (book [737]);
    abstract only, the full text was not accessible: FSM machine graph and input string dissolved into
    one space of directed graphs, graphs interact with their own topology to generate a product, self-similar
    topologies, novel reaction pathways and open-ended diversity. https://pubmed.ncbi.nlm.nih.gov/16733079/'
  - 'Salzberg, C. & Sayama, H. (2025). Reflexive composition of elementary state machines, with an application
    to the reversal of cellular automata rule 90. arXiv:2505.07186: the formulation of the ''composite
    state machine'' introduced in Salzberg 2006a, 2006b and 2007 (Mealy machine G = (Q, A, delta, phi);
    reflexive composition eqs. 1-4 with sender/receiver alternation; composition generating exponentially
    larger state-space graphs), the naming scheme of Table 1, Table 2, figs. 2-17. https://arxiv.org/abs/2505.07186'
  - 'Salzberg, C., Sayama, H. & Ikegami, T. (2004). A tangled hierarchy of graph-constructing graphs.
    Artificial Life IX, pp. 495-500, MIT Press (open access): Mealy machines with a shared input/output
    alphabet, tapes as chains of links, pointers to the current states, symmetric engagement of machine
    and tape graphs. https://doi.org/10.7551/mitpress/1429.003.0084'
  decisions:
  - The primary source [737] and its companions (Salzberg 2006a, ALife X; 2006b, Artificial Life 12:487-512)
    were not accessible (paywalled; no preprint, thesis or archived copy found). The reaction is reconstructed
    from the 2025 restatement of the formulation by the same author, which says the earlier studies composed
    distinct sender and receiver machines repeatedly into exponentially larger state-space graphs. The
    exact product construction, elastic conditions and reactor of the 2007 paper are therefore Chemart's
    reading, not a transcription.
  - 'Molecule: a deterministic Mealy machine over the alphabet {0..k-1} with a pointer (the current state,
    the dark node of the 2025 figures), trimmed to the nodes reachable from the pointer. Missing links
    are allowed so that input data (e.g. the tape chains of the 2004 paper, with a symbol playing xi)
    are graphs of the same kind; a composite link exists only if both links it passes through exist.'
  - 'Product s1 o s2: the composite of eqs. 1-2 with s1 as sender, whose node (x, y) holds the sender-position
    and receiver-position states; after each input the machines swap positions (the 2025 ''folding'').
    For distinct species the node also records which machine sends, so the product has at most 2|s1||s2|
    nodes; for s + s the machines are identical and it has at most |s|^2 (the uniform lattice of the 2025
    paper). Its pointer is the pair of reactant pointers and only reachable nodes are kept.'
  - 'Reaction scheme: both reactants are catalysts, s1 + s2 -> s1 + s2 + s1 o s2, as in the book''s constructive
    chemistries (AlChemy, eq. 9.4); the 2007 abstract says only that interacting graphs ''generate a product''.
    Ordered pairs: s1 o s2 and s2 o s1 differ.'
  - 'Not modelled: the 2004 graph-constructing-graph machinery (xi-flows chosen by a decision tree over
    N-loops, translation of the output string by codons into construction-head instructions that grow
    and fold the graph). The 2025 restatement does not use it, and the 2004 paper leaves conflict resolution
    open.'
  - 'Elastic boundary: products with more than max_states reachable nodes (Chemart choice, since state
    counts multiply at every composition). No rates are published, so reactions carry rate None.'
  - 'Canonical id: breadth-first numbering from the pointer with links taken in input order; for accessible
    deterministic graphs this is a complete isomorphism invariant, so no general graph canonisation is
    needed.'
  - 'Soup: chemart.soup with constant dilution draws an ordered pair (sender, receiver), adds the product
    and removes a random molecule; initial population: the given machines in equal copies, or M random
    machines (duplicates allowed). Random machines have uniform links over states and outputs, trimmed
    from state 0. These are Chemart choices; the defaults (closure from M45, M = 50, 2000 collisions,
    max_states 64, max_species 100) are small, not published values.'
  - 'The v1 entry had no parameters; all parameters are new. v1 said ''dilution: none'', which holds for
    the closure; the soup is an added reactor.'
  - 'Published numbers checked by computation: Table 1 encoding (M61 = 00111101, M45 = 00101101), 76 machines
    unique under complement and mirror, the Table 2 equivalents, transposition M45 <-> M54 and M60 ->
    M60 (fig. 17, footnote 3), M45 equal to rule 90 on even steps with the fig. 7 boundary inputs, M54
    even steps as rule-90 preimages with the eqs. 6-7 inputs, M44 inverted by flipping the lattice. The
    fig. 12 caption''s input sequence (0 up to t = 5, then 1, 1) was not reproduced: with the constraint
    algorithm as described, the run from two centred 1 cells needs input 1 at t = 0; the figure''s step
    indexing cannot be recovered from the text. Likewise fig. 13''s statement that M60 with the same initial
    states and boundary inputs has the same even rows as M54 did not hold when M60 was fed M54''s inputs,
    so only M54''s reversal is tested.'
  phenomena:
  - 'composition multiplies state spaces: repeated composition yields exponentially larger, self-similar
    state-space graphs (M45 gives 2, 4, 16, 32, 64 states)'
  - 'a graph composed with itself reads its own structure (reflexivity): s + s -> 2 s + s o s'
  - the n-fold composite of a machine is the lattice of reflexive composition (arXiv:2505.07186 eqs. 3-4)
  - M45 (adding machine modulo 2) reproduces elementary CA rule 90 and the Sierpinski triangle on even
    steps (figs. 5-7)
  - 'M54 and M60 run rule 90 backwards: each even step is a rule-90 preimage of the previous one (figs.
    10-15); M45 and M54 are transposes of each other (fig. 17)'
  - M44 shows reversible billiard-ball dynamics, inverted by flipping the lattice (fig. 8)
  - of the 256 two-state two-symbol machines, 76 are unique under complement and mirror (Table 2)
  - 'Salzberg 2007 (abstract): continuous emergence of complex self-similar topologies, novel reaction
    pathways, seemingly open-ended diversity'
  notes: 'Chemart observation, not a published result: random soups of two-state machines usually collapse
    to one species within a few thousand collisions, a machine of tens of states that reproduces itself
    under self-composition (s o s = s). The module exports canonical, parse, compose, elementary and number
    for working with the machines directly.

    '