Generative Story Worlds as Linear Logic Programs

Talk for the paper of the same title, available here: http://www.cs.cmu.edu/~cmartens/int7.pdf

Presented at (INT)7 (Intelligent Narrative Technologies) in Milwaukee on June 18, 2014.

Transcript

1. Generative Story Worlds
   as
   Linear Logic Programs
   Chris Martens, João Ferreira,
   Anne-Gwenn Bosser, and Marc Cavazza
   1
   Presented at Interactive Narrative Technologies
   June 18, 2014
   

   View Slide

2. Generative Story Worlds
   as
   Linear Logic Programs
   Chris Martens, João Ferreira,
   Anne-Gwenn Bosser, and Marc Cavazza
   2
   Presented at Interactive Narrative Technologies
   June 18, 2014
   

   View Slide

3. 3
   new programming languages
   for digital narratives
   Praxis & Prompter:
   

   View Slide

4. 4
   OLD programming languages
   for digital narratives(?)
   LOGIC
   

   View Slide

5. 5
   old programming languages
   for digital narratives
   LINEAR LOGIC
   

   View Slide

6. 6
   used linear logic as a programming language (Celf)
   for specifying narrative worlds;
   used proof search to generate structured stories.
   What we did:
   

   View Slide

7. THREE ACTS
   7
   I. Setup: agent simulation, linear logic, and
   logic programming
   II. Payoff: Stories as Proofs
   III. Promise and Limitations
   

   View Slide

8. View Slide

9. 9
   greet
   insult
   compliment
   gossip
   flirt
   ...
   social verbs: social state:
   knowledge
   sentiment
   emotional state
   personality
   belief
   

   View Slide

10. 10
    verbs: rules
    causal relationships: resource dependencies
    

    View Slide

11. RULES
    11
    Narrative Actions
    ~
    Logical implication
    

    View Slide

12. 12
    the problem with standard logical implication:
    buy_soda :
    has(Alice, $1) ^ wants(Alice, soda) 㱺
    has(Alice, soda)
    

    View Slide

13. LINEAR LOGIC
    13
    A -o B
    

    View Slide

14. 14
    buy_soda :
    has(Alice, $1) * wants(Alice, soda) -o
    has(Alice, soda)
    

    View Slide

15. 15
    A * B
    A -o B
    !A
    

    View Slide

16. 16
    A * B
    A -o {B}
    !A
    

    View Slide

17. Logic Programming
    17
    proof search as program execution
    

    View Slide

18. 18
    initial
    state
    Σ Δ0
    Celf
    query
    (optional)
    A
    Δ0, …, Δ
    Δ ⊢ A?
    narrative
    actions
    

    View Slide

19. Logic Programming
    19
    A -o {B}
    Δ, A → Δ, B
    Δ = arbitrary context of resources A1, …, An
    

    View Slide

20. 20
    Example in our paper:
    Shakespearean tragedy world
    

    View Slide

21. 21
    Shakespearean tragedy world
    state components:
    character location, possession,
    sentiment toward other characters,
    goals
    

    View Slide

22. 22
    Shakespearean tragedy world
    at
    has
    anger
    philia
    depressed
    

    View Slide

23. 23
    Shakespearean tragedy world
    !dead
    !killed
    

    View Slide

24. 24
    do/insult :
    at C L * at C’ L * anger C C’
    -o {at C L * at C’ L * anger C C’
    anger C’ C * depressed C’}.
    

    View Slide

25. 25
    do/compliment :
    at C L * at C’ L * philia C C’
    -o {at C L * at C’ L *
    philia C C’ * philia C’ C}.
    

    View Slide

26. 26
    do/murder :
    anger C C’ * anger C C’ * anger C C’ *
    anger C C’ * at C L * at C’ L *
    has C weapon
    -o {at C L * !dead C’ * !murdered C C’ *
    has C weapon}.
    

    View Slide

27. 27
    do/mourn :
    at C L * philia C C’ * dead C’
    -o {philia C C’ * at C L *
    depressed C * depressed C}.
    

    View Slide

28. 28
    do/becomeSuicidal :
    at C L * depressed C * depressed C *
    depressed C * depressed C
    -o {at C L * suicidal C *
    wants C weapon}.
    

    View Slide

29. 29
    do/loot
    : at C L * dead C' * has C' O * wants C O
    -o {at C L * has C O}.
    

    View Slide

30. 30
    do/comfort :
    at C L * at C' L *
    suicidal C' * philia C C' * philia C' C
    -o {at C L * at C' L *
    philia C C' * philia C' C *
    philia C' C}.
    

    View Slide

31. 31
    initial state
    

    View Slide

32. 32
    story_start :
    init -o
    { at romeo town * at montague mon_house *
    at capulet cap_house * at mercutio town *
    at nurse cap_house * at juliet town *
    at tybalt town * at apothecary town *
    has tybalt weapon * has romeo weapon *
    has apothecary weapon *
    ...
    

    View Slide

33. 33
    ... *
    anger montague capulet * anger capulet montague *
    anger tybalt romeo * anger capulet romeo *
    anger montague tybalt *
    philia mercutio romeo * philia romeo mercutio *
    philia montague romeo * philia capulet juliet *
    philia juliet nurse * philia nurse juliet *
    neutral nurse romeo * neutral mercutio juliet *
    neutral juliet mercutio *
    neutral apothecary nurse *
    neutral nurse apothecary}.
    

    View Slide

34. 34
    final state
    

    View Slide

35. 35
    ending_happy : nonfinal *
    actor C * actor C’ *
    at C L * at C’ L * married C C’ -o {final}.
    ending_vengeance : nonfinal *
    actor C1 * actor C2 * actor C3 *
    killed C1 C2 * philia C3 C2 * killed C3 C1
    -o {final}.
    

    View Slide

36. 36
    proofs as stories
    

    View Slide

37. 37
    query:
    init -o {final}
    

    View Slide

38. Nondeterminism
    38
    A -o {B}
    Δ, B
    A -o {C}
    Δ, C
    Δ, A
    

    View Slide

39. Nondeterminism
    39
    A -o {B} C -o {D}
    

    View Slide

40. Nondeterminism
    40
    A -o {B}
    Δ, B, C
    C -o {D}
    Δ, A, D
    Δ, A, C
    

    View Slide

41. Nondeterminism
    41
    A -o {B}
    Δ, B, C
    C -o {D}
    Δ, A, D
    Δ, A, C
    Δ, B, D
    

    View Slide

42. 42
    proof of
    init -o {final}:
    function (x:init) =
    let [xs] = r [ys] in …
    

    View Slide

43. 43
    ...
    let {[X73, [X74, [X75, [X76, X77]]]]}
    = do/insult/private [a-tybalt, [a-romeo, [X68, [X66, X72]]]] in
    let {[X85, [X86, X87]]}
    = do/becomeSuicidal [a-romeo, [X79, [X41, [X59, [X52, X77]]]]] in
    let {[X88, [X89, [X90, [X91, X92]]]]}
    = do/comfort [a-mercutio, [a-romeo, [X78, [X85, [X86, [X81, X83]]]]]] in
    let {[X101, [!X102, [!X103, X104]]]}
    = do/murder [a-romeo, [a-tybalt, [X58, [X40, [X76, [X51, [X94, [X96, X27]]]]]]]] in
    let {[X105, [X106, [X107, X108]]]}
    = do/compliment/private [a-nurse, [a-juliet, [X46, [X47, X30]]]] in
    let {[X109, [X110, [X111, X112]]]}
    = do/compliment/private [a-juliet, [a-nurse, [X106, [X105, X108]]]] in
    let {[X113, X114]}
    = do/loot [a-romeo, [a-tybalt, [X101, [X102, [X26, X87]]]]] in
    ...
    

    View Slide

44. Concurrent Equality
    44
    r1: A -o B r2: C -o D
    “apply r1; apply r2” ~= “apply r2; apply r1”
    if A and C are disjoint
    

    View Slide

45. 45
    graphical representation of traces
    

    View Slide

46. 46
    init
    do/insult/private
    do/formOpinion/dislike
    do/compliment/witnessed
    do/travelTo
    do/compliment/private
    do/murder
    do/marry
    ending_1
    do/steal
    do/thinkVengefully
    do/mourn
    do/becomeSuicidal
    cleanup/1
    do/eroticize
    do/suicide
    do/flirt/discrete
    

    View Slide

47. 47
    do/insult/private
    do/formOpinion/dislike
    do/compliment/witnessed
    do/travelTo
    private
    do/murder
    do/marry
    do/steal
    do/thinkVengefully
    do/mourn
    do/becomeSuicidal
    do/suicide
    ete
    

    View Slide

48. 48
    queries on sets of traces
    

    View Slide

49. 49
    > exists ending_1
    > exists do/thinkVengefully &&
    ̃link do/thinkVengefully do/murder
    

    View Slide

50. generative exploration
    informative structure
    character perspective-agnosticism
    50
    what does this accomplish?
    

    View Slide

51. 51
    how do simulation designs fail their authors?
    

    View Slide

52. 52
    orphan verbs
    vs causally connected chains
    

    View Slide

53. ACT III
    Potential & Limitations
    53
    

    View Slide

54. 54
    combining reactive/exploratory
    + goal-driven behavior:
    A -o G vs A -o {B}
    (backward & forward chaining)
    Additional potential of LL
    

    View Slide

55. 55
    higher-order rules: (A -o B) -o C
    Additional potential of LL
    

    View Slide

56. 56
    higher-order rules: (A -o B) -o C
    “serendipity:”
    Forall L.
    at romeo L * at juliet L
    -o
    {eros romeo juliet
    * at romeo L * at juliet L})
    Additional potential of LL
    

    View Slide

57. 57
    Additional potential of LL
    dynamic object generation: Exists x:A. ...
    makeRoom : has(C, magic door) * at C R -o
    {Exists r:room. !adjacent R r}
    

    View Slide

58. 58
    limitations
    

    View Slide

59. 59
    limitations
    (of the tool)
    (of the underlying formalism)
    

    View Slide

60. 60
    limitations of the tool:
    interactivity
    history tracking
    debugging facilities
    drama management
    

    View Slide

61. 61
    limitations of the formalism:
    ability to broadcast actions
    nondeterminism and control over probabilities
    

    View Slide

62. 62
    what I plan to address:
    interactivity
    debugging facilities
    ability to “broadcast” actions
    

    View Slide

63. 63
    PHASES
    

    View Slide

64. Phases (Example)
    64
    deduplication
    phase act {
    …
    } quiesced act -o {phase dedup}.
    phase dedup {
    wants C O * wants C O -o {wants C O}.
    } quiesced dedup -o {phase act}.
    

    View Slide

65. Phases
    phase p1 = {...}
    phase p2 = {...}
    quiesced p1 -o
    {some_token * phase p2}.
    quiesced p2 -o {phase p1}.
    Blocks connected by specification of quiescence behavior
    65
    

    View Slide

66. Phases (Example)
    interactivity
    phase world = {
    rule1 : do Action * … -o {…}.
    rule2 : do Action * … -o {…}.
    }
    phase player = {
    get_input : player_turn -o {read #STDIN}.
    }
    66
    

    View Slide

67. 67
    also possible to implement: broadcast
    

    View Slide

68. 68
    also ongoing:
    phase invariants
    for each phase, constraints on the context…
    - as an invariant of the phase
    - at quiescence
    programmer-specified, machine-checked!
    

    View Slide

69. FINALE
    69
    

    View Slide

70. 70
    the point of formalism
    is to be able to run
    readable specifications
    

    View Slide

71. 71
    the point of formalism
    is to be able to reason about
    readable specifications
    

    View Slide

72. View Slide

73. View Slide

74. 74
    linear logic can be used as a programming language
    for specifying narrative worlds;
    to generate narratives
    & analyze their structure;
    potentially the basis for a richer PL.
    Takeaway
    

    View Slide

75. FINALE
    75
    Celf: https://github.com/clf/celf
    Example from the paper:
    http://tinyurl.com/int2014-clf
    Me: Chris Martens
    @chrisamaphone
    cmartens@cs.cmu.edu
    Planning to graduate Summer 2015
    

    View Slide

76. MINUS WORLD
    (bonus slides)
    76
    

    View Slide

77. Cut
    77
    Δ1 ⊢ A Δ 2, A ⊢ C
    Δ 1, Δ 2 ⊢ C
    

    View Slide

78. Frame Property
    78
    If Δ1 ⊢ A
    then
    Δ1, Δ2 ⊢ A * ⨂Δ2
    

    View Slide

79. Frame Property (alt)
    79
    If Δ1 → Δ2
    then
    Δ1, Δ’ → Δ2, Δ’
    

    View Slide

80. Relationship between step and
    deductions
    80
    Δ1 → Δ2
    iff for all C
    Δ2 ⊢ C implies Δ1 ⊢ C
    

    View Slide

81. 81
    LOGICS
    are formalisms with
    composable consequences
    

    View Slide

82. 82
    LOGICS
    are formalisms with
    composable consequences
    Δ ⊢ A Δ’, A ⊢ C
    Δ’, Δ ⊢ C
    

    View Slide

83. Phases (Example)
    Broadcast
    phase act {
    do C A * in_room C R -o {broadcast C A R}.
    }
    quiesced act -o {phase broadcast}.
    phase broadcast {
    broadcast C A R * in_room C' R -o
    {notice C' C A * broadcast C A R}.
    }
    quiesced broadcast -o {phase notice}.
    83
    

    View Slide

84. Phases (Example)
    Broadcast
    84
    phase notice {
    broadcast C A R * notice C' C A -o
    {react C' C A * in_room C' R * broadcast C A R}.
    }
    quiesced notice * broadcast _ _ _ -o {phase react}.
    phase react {
    react C C' (murder ...) * philia C C' -o {...}
    }
    quiesced react -o {phase act}.
    

    View Slide

85. Phases
    delimited subsignatures connected by
    specifications of quiescence behavior.
    85
    quiesced P * State -o {phase P’ * State’}.
    arbitrarily many phases
    looping + branching
    

    View Slide