chemart / arms

hub:
  repo_type: generator
  builtin: arms
  tags:
  - banzhaf-yamamoto
chemistries:
- id: arms
  name: ARMS (Abstract Rewriting System on Multisets)
  intuition: 'The simplest possible rewriting chemistry: a bag of symbols and a list of rules like ''aaa
    -> c'' or ''a -> abba''. Everything interesting comes from how the rules are applied. Rules that enlarge
    a molecule are called heating and rules that shrink it cooling, and the order they are tried in decides
    the outcome: one order drives the same starting bag to a quiet fixed point, the reverse order makes
    it grow without bound. A minimal demonstration that in an artificial chemistry the reactor algorithm
    is part of the model, not an implementation detail.

    '
  aliases:
  - chemical ARMS
  - abstract rewriting system on multi-sets
  - symbolic chemical system
  origin: Suzuki & Tanaka, 1997
  book: 9.4; ecology application 8.2.3; p53 model 18.3.2
  refs:
  - '[830]'
  - '[831]'
  - '[832]'
  - '[828]'
  - '[829]'
  family: rewriting
  kind: generator
  constructive: false
  S:
    repr: symbols from a finite alphabet A; the reaction vessel is a multiset of symbols
    definition: explicit
  R:
    definition: explicit
    arity: variable
    scheme: 'Gamma = (A, R) with R a set of indexed multiset rewriting rules r_1..r_n, written in the
      reaction syntax "3 a -> c", "a -> 2 a + 2 b" (book eq. 9.15: aaa -> c, b -> d, c -> e, d -> ff,
      a -> abba, f -> h). A rule is heating when it enlarges the multiset (|rhs| > |lhs|) and cooling
      when it shrinks it; input symbols are reactions ∅ -> s ([831] def. 6).

      '
  A:
    reactor:
    - maximally-parallel
    - well-stirred-multiset
    dilution: none; an optional maximal multiset size blocks inputs and rewrites that would exceed it
    notes: 'Book 9.4: a rule order fixes the sequence of rule application; kinetics is modelled as differing
      rule-application frequencies rather than rate constants. [831] fig. 2: at each iteration input symbols
      are injected, one rule is selected at random and applied if its left side is present and the maximal
      size is respected; only rewrites count as steps. [831] fig. 14: a heating rule is selected with
      probability p, a cooling rule with 1 - p.

      '
  params:
  - name: system
    type: enum
    default: book-example
    choices:
    - book-example
    - ru1
    - brusselator
    - two-symbol
    - custom
    role: structural
    meaning: 'rule set: book-example (book eqs. 9.13-9.16, rules r1-r6 from M0 = {a,b,a,a}), ru1 ([831]
      p. 132, aaa -> b, b -> a, b -> c, a -> bb, input a, maximal size 4, M0 = {a,a,a,a}), brusselator
      ([831] fig. 8, inputs A and B, maximal size 5000, empty M0), two-symbol ([831] ''rule set'': all
      rules with 1-5 symbols from {a,b} on each side, maximal size 10, random M0 of 1-10 symbols), custom
      (rules, initial, inputs)'
  - name: selection
    type: enum
    default: published
    choices:
    - published
    - order
    - reverse-order
    - random
    - heating-probability
    role: structural
    meaning: 'how rules are applied in the recorded run: published (order for book-example, heating-probability
      for two-symbol, random otherwise), order / reverse-order (book 9.4: one maximally parallel pass
      per step with rules by ascending / descending index), random ([831] fig. 2: one uniformly chosen
      rule per iteration), heating-probability ([831] fig. 14: a heating rule with probability p, a cooling
      rule with 1 - p)'
  - name: p
    type: float
    default: 0.5
    min: 0.0
    max: 1.0
    range: '[831] sweeps 0 to 1 in steps of 0.01'
    role: stochastic
    meaning: 'heating-probability only: probability of selecting a heating rule (a cooling rule is selected
      with 1 - p)'
  - name: input_probability
    type: float
    default: 1.0
    min: 0.0
    max: 1.0
    role: stochastic
    meaning: 'probability that each input symbol is injected at an iteration ([831]: ''frequency of inputs'';
      the published examples input at every iteration)'
  - name: steps
    type: int
    default: 1000
    min: 0
    max: 100000
    range: '[831] runs 1000 steps per initial state (two-symbol) and 10000 (brusselator)'
    role: population
    meaning: rewriting steps of the recorded run (count-step n of [831] figs. 2 and 14; one parallel pass
      in the ordered disciplines); the run stops earlier at a normal form or, for a deterministic ordered
      run, at the first repeated multiset
  - name: max_size
    type: int
    default: 0
    min: -1
    max: 1000000
    role: population
    meaning: 'maximal multiset size: 0 keeps the system''s published value (none for book-example and
      custom), -1 removes the limit, a positive value sets it'
  - name: rule_count
    type: int
    default: 0
    min: 0
    max: 380
    role: structural
    meaning: 'two-symbol only: 0 uses all 380 rules, otherwise a random subset of this many rules'
  - name: rules
    type: list
    default: []
    range: e.g. ['3 a -> c', 'b -> d', 'a -> 2 a + 2 b']
    role: structural
    meaning: 'custom only: rewriting rules r1, r2, ... in chemart reaction syntax (coefficient, space,
      symbol; terms joined by '' + ''); an empty left side is an input rule'
  - name: initial
    type: dict
    default: {}
    range: 'e.g. {a: 3, b: 1}'
    role: population
    meaning: 'custom only: initial multiset M0, symbol -> copies'
  - name: inputs
    type: list
    default: []
    range: e.g. [a]
    role: structural
    meaning: 'custom only: symbols input at every iteration (subject to input_probability and the maximal
      size)'
  provides:
  - topology
  - stoichiometry
  - catalysts
  - initial-state
  fidelity: reconstructed
  sources:
  - 'Banzhaf, W. & Yamamoto, L. (2015). Artificial Chemistries, section 9.4 (Gamma = (A, R), eqs. 9.13-9.18:
    the six-rule example, forward order halting in {e,h,h}, reverse order expanding; the lambda_e order
    parameter with cycles at intermediate values), 8.2.3 and 18.3.2 (ecology and p53 applications, no
    rules given).'
  - 'Suzuki, Y. & Tanaka, H. (1998). Order parameter for a symbolic chemical system. In C. Adami, R. Belew,
    H. Kitano & C. Taylor (eds.), Artificial Life VI, pp. 130-139, MIT Press (book [831]). Definitions
    3-7, the rewriting algorithm (figs. 2 and 14), the Ru1 example (fig. 3), the Brusselator rules (fig.
    8), the lambda_e parameter (eq. 3), the two-symbol rule set and the termination/cycle results (figs.
    15-17, table 2). Read from the open proceedings PDF, equations and figures checked on rendered page
    images: https://mitp-content-server.mit.edu/books/content/sectbyfn/books_pres_0/1431/9780262510998_ALIFE_VI.pdf?dl=1'
  decisions:
  - 'Papers not accessible: [830] (Artif. Life Robotics 1:211-219, 1997), [828] (LNCS 2235, 2001), [829]
    (Artif. Life Robotics 6:129-132, 2002) and [832] (Applications of Membrane Computing, 2006) are closed
    access with no open copy found (Springer, OpenAlex: closed). The ecology (plant/herbivore/carnivore
    with herbivore-induced volatiles) and p53 rule sets are therefore not offered; the book gives no rules
    for them and inventing them is not allowed. The membrane (P system) structure, membrane growth and
    division of the 2001 ARMS are not implemented, so ''compartments'' was removed from provides.'
  - 'Ordered discipline: the book says rules are ''executed according to index'' without defining a step.
    Chemart reads one step as one maximally parallel pass: rules in index order, each applied as often
    as its left side fits in what remains of the current multiset, products available only at the next
    step. This reproduces eq. 9.17 ({a,b,a,a} -> {c,d} -> {e,f,f} -> {e,h,h}) and eq. 9.18 ({a,b,a,a}
    -> 6 a, 6 b, 1 d) exactly.'
  - 'Random discipline ([831] fig. 2): inputs happen before the rule is tried, at every iteration, and
    only successful rewrites count as steps; this reproduces the fig. 3 walk-through. [831] figs. 4 and
    5 (Ru1 with printed rule orders) could not be reproduced by any of 16 readings of the algorithm (input
    before/after the rule, order advancing per iteration or per rewrite, states recorded per iteration
    or per rewrite); they are not used as tests.'
  - 'Ru1 initial state: the text prints {a,a,f,a} with a symbol f that no rule uses, but its successors
    {b,a}, {c,a,a}, {c,a,a,a} only follow from {a,a,a,a}; ru1 starts from {a,a,a,a}. Figs. 4 and 5 start
    from {f} and keep f throughout; they are not reproduced (previous decision).'
  - 'Normal form ([831] def. 7): no selectable rule can rewrite the multiset within the maximal size and
    no symbol can be input. Under heating-probability, heating rules are selectable only if p > 0 and
    cooling rules only if p < 1; rules that keep the size (|lhs| = |rhs|) are never selected, because
    the published protocol gives heating p and cooling 1 - p and leaves no probability for them. An empty
    selected class is a failed iteration. A run is cut at 1000 + 50 * steps * |R| iterations (never reached
    in the published systems).'
  - lambda_e is computed as printed in [831] eq. 3, sum of heating-rule applications / (1 + (sum of cooling-rule
    applications - 1)), which equals the heating/cooling ratio; it is null when no cooling rule was applied
    (the paper says only that it is greater than 1 then). The book calls it 'the ratio between the number
    of heating rules to the number of cooling rules'; [831] counts rules used, which Chemart reads as
    applications.
  - 'Two-symbol rule set: [831] builds both sides by sampling a and b with replacement, lengths 1-5, ''as
    the overall permutation of both sides'', and states eq. 4, {sum_{k=1..m} C(n+k-1, k)}^2, and 30976
    rules. For n = 2, m = 5 eq. 4 gives 400, not 30976 (= 176^2), and the prose matches neither exactly.
    Chemart uses the multiset reading of eq. 4: all 400 pairs of multisets minus the 20 identities, 380
    rules, with rule_count to sample subsets. Initial states have 1-10 symbols a/b drawn at random as
    in [831].'
  - 'Trajectory class in extras.analysis: ''halting'' (normal form reached), ''periodic'' (a deterministic
    ordered run repeats a multiset; its exact period is recorded and the run stops), ''recurrent'' (a
    stochastic run revisits multisets without halting), ''non-recurrent'' (no normal form and no revisit
    within the steps, e.g. the expanding reverse order of eq. 9.18). The paper''s measures of order are
    recorded as revisits and kinds_of_periods (distinct return times to a multiset), with heating_uses,
    cooling_uses and lambda_e. These map to [831] table 2: halting with lambda_e near 0 or >> 1 (strong
    termination, Wolfram class I/II), many kinds of periods near lambda_e = 1 (weak termination, class
    III).'
  - 'Brusselator ([831] fig. 8): the paper gives no numeric rule frequencies (it only says a higher rate
    means a higher application frequency), so rules carry no rate and random selection is uniform. The
    products D and E are never removed in the published rules, so with inputs A and B the multiset fills
    to its maximal size of 5000; the oscillations of [831] figs. 9-11 depend on the unpublished frequencies
    and are not claimed.'
  - Sequential, free and parallel rewriting (book 9.4, from [830]) are not defined in the accessible sources
    beyond their names; Chemart offers the ordered parallel pass of the book and the random single-rule
    rewriting of [831]. The rule order of [831] fig. 3 (a fixed cyclic sequence of rules) is available
    as run(..., sequence=...) in the module, not as a parameter.
  - 'v1 parameters: alphabet (matrix) is implied by the rules, initial multiset and inputs; rules (callable)
    became the system choice plus the custom rules text; rule_order (matrix) became the selection enum
    (order / reverse-order); M0 became initial; lambda_e is an output (analysis) controlled by p, as in
    [831].'
  phenomena:
  - 'book eq. 9.17: with rules applied in index order, {a,b,a,a} -> {c,d} -> {e,f,f} -> {e,h,h}, a normal
    form (the system halts)'
  - 'book eq. 9.18: with the reverse rule order the multiset expands without bound ({a,b,a,a} -> 6 a,
    6 b, 1 d -> ...)'
  - '[831] figs. 15-17: with heating probability p near 0 or 1 most calculations terminate (lambda_e =
    0 or no cooling); for intermediate p calculations do not terminate and many cycles with many kinds
    of periods appear, the most kinds near p = 0.5 (lambda_e near 1)'
  - '[831] fig. 3: Ru1 with maximal size 4 from {a,a,a,a} and the rule order r4, r1, r3, r2 gives {a,a,a,a}
    -> {b,a} -> {c,a,a} -> {c,a,a,a}'
  - Brusselator-like oscillations of X and Y ([831] figs. 9-11, rule frequencies not published)
  - plant/herbivore/carnivore ecology ([828], [829]) and p53 signalling feedback loops ([832]); rules
    not accessible
  notes: 'The module exports run (both disciplines, with an optional fixed rule sequence and a state trace),
    Rule, two_symbol_rules and the published rule sets BOOK_RULES, RU1_RULES and BRUSSELATOR_RULES. ARMS
    is a P system specialised for artificial life (book 9.5); the membrane version of [828] is not implemented.

    '