Compositional Creativity (ICFP 2017)

Transcript

1. ICFP 2017, Oxford, UK Compositional Creativity Chris Martens Assistant Professor Computer Science Department North Carolina State University
2. About Me Assistant Professor (UK equiv.: Lecturer) at NCSU Postdoc UC Santa Cruz, Expressive Intelligence Studio PhD CMU 2015, POP group 2
3. About Me Conferences I used to go to had words like: functional, types, logic, deduction, languages 3 Conferences I go to now have words like: artificial intelligence, narrative, interactive storytelling, games, play
4. My Lab Conferences I used to go to had words like: functional, types, logic, deduction, languages 4 Conferences I go to now have words like: artificial intelligence, narrative, interactive storytelling, games, play
5. Research Mission Use principles from logic and formal methods to enable more expressive human-computer collaboration 5
6. Human-Computer Collaboration ❖ Interactive storytelling ❖ Intelligent tutors ❖ Conversational agents ❖ Mixed-initiative creative tools 6
7. “Artificial Intelligence”? ❖ Interactive storytelling ❖ Intelligent tutors ❖ Conversational agents ❖ Mixed-initiative creative tools 7
8. Unifying themes: logic ❖ Logic as an intermediate language between human and computational thought ❖ Reasoning about synthesized behavior 8
9. Unifying themes: social reasoning ❖ Simulating social agents ❖ Designing languages and other human-computer interfaces through a metaphor of conversation 9
10. This talk From classical AI to modern FP Plots as plans as proofs Intentional reasoning Epistemic reasoning 10
11. From classical AI to modern FP From classical AI to modern FP Plots as plans as proofs Intentional reasoning Epistemic reasoning 11
12. A common origin of FP and AI, 1956 12
13. Influence on AI First system to codify 
 reasoning as search Presented at McCarthy’s AI-defining Dartmouth Research Project, Summer 1956 AI historian Pamela McCormick: “proof positive that a machine could perform tasks heretofore considered intelligent, creative and uniquely human” 13
14. Influence on FP To build LTM, they implemented a new language called IPL Influenced McCarthy’s design of LISP Use of lists rather than arrays Hierarchical subroutines and recursion 14
15. “Current Developments in Complex Information Processing” Allen Newell and Herbert Simon, 1959 15
16. Reasoning as Search: 3 examples * Theorem proving * Program synthesis * AI planning 16
17. Reasoning as Search: 3 examples * Theorem proving * Program synthesis * AI planning 17
18. Theorem Proving Γ ⊢A 18 Initial Assumptions Goal Defined by atomic inference rules
19. Theorem Proving Γ ⊢A 19 Γ ⊢A Composition of inferences
20. Reasoning as Search: 3 examples * Theorem proving * Program synthesis * AI planning 20
21. Reasoning as Search: 3 examples * Theorem proving * Program synthesis * AI planning 21 Conor McBride (since at least 2013): Type inference is so last century! Let’s write more detailed types, do program inference!
22. (Type-based) Program Synthesis Γ ⊢A 22 Library functions Specification Typing rules of the language
23. (Type-based) Program Synthesis Γ ⊢A :A let rec reverse l = case l of nil => … | cons x::xs => … Osera and Pierce - “Type-and-example-directed program synthesis,” POPL 2015; Polikarpova et al., “Program synthesis from polymorphic refinement types,” POPL 2016 23 Composition of program constructs
24. Reasoning as Search: 3 examples * Theorem proving * Program synthesis * AI planning 24
25. AI Behavior Synthesis (Planning) Γ ⊢A 25 Initial State Goal state Atomic operators that can change the state
26. AI Behavior Synthesis (Planning) Γ ⊢A move a1 a2 pickup block1 move a2 b2 place block1 : { at(bot, a1), at(block1, a2)} —-> {at(bot, b2), at(block1, b2)} 26 Plan (sequence of operations)
27. Creative machines? * Theorem proving * Program synthesis * AI planning 27
28. Plots as plans as proofs From classical AI to modern FP Plots as plans as proofs Intentional reasoning Epistemic reasoning 28
29. AI Planning Domain - sets of operators (parameterized actions) Problem - initial and goal configuration —PLANNER—> plan, i.e.
 sequence of operators applied to terms to reach goal from initial 29
30. AI Planning Domains PDDL/STRIPS - domain specific langs. for planners Actions defined in terms of of preconditions and effects Preconditions, effects, configurations, are collections of logical predicates 30
31. Story Generation Planning for Plot Generation
 (Young 1999; Cavazza and Charles 2001-2) Linear Logic Programming for Plot Generation
 (Martens, Ferreira, Bosser, Cavazza 2013) 31
32. Story Configurations initial: { dead(jon_arryn, hand),
 role(robert_baratheon, king),
 heir(robert_baratheon, joffrey),
 allies(robert_baratheon, eddard_stark),
 role(eddard_stark, warden_of_north) } sets of atomic propositions describing the world (Game-of-Thrones-inspired example) 32 ! " ❄ (
33. Story Operators narrative events, e.g. ways characters can bid for power name_hand(Ruler, NewHand) :
 preconditions: need_filled(hand),
 role(Ruler, ruler_of_westeros),
 allies(Ruler, NewHand) effects: offered_role(NewHand, hand)
 delete need_filled(hand) 33
34. Story Operators As a linear logic formula: name_hand(Ruler, NewHand) : need_filled(hand) *
 has_role(Ruler, ruler_of_westeros) *
 allies(Ruler, NewHand) -o offered_role(NewHand, hand) *
 has_role(Ruler, ruler_of_westeros) *
 allies(Ruler, NewHand) 34
35. Compositional notation for plots 35 " ! ❄ " ! ( "
36. 36 " names ! " ! ❄ " ! ( " " " ! ! Compositional notation for plots
37. ! observes ( 37 " names ! " ! ❄ " ! " " ! ( " ! … Compositional notation for plots
38. ! observes ( 38 " names ! " ! ❄ " ! " " ! ( " ! … … … Compositional notation for plots
39. SEASON 1
 EPISODE 2 39 " ! ❄ " ! ( " … … … Compositional notation for plots
40. SEASON 1
 EPISODE 2 40 " ! ❄ " ! ( " … … … Compositional notation for plots c.f. symmetric monoidal categories
41. Story Generation goal: 
 has_role(!, ) 41 Searching for plots with specified endings
42. Example: Story Generation Output: computer-generated fan fiction 42
43. Narrative variation? goal: 
 Exists C. 
 has_role(C, ruler_of_westeros) 43 Searching for plots with varied endings
44. Planners & theorem provers find the most efficient solution 
 
 : Exists C. has_role(C, ruler_of_westeros) 44 Narrative variation?
45. Planners & theorem provers find the most efficient solution has_role(rb, ruler_of_westeros) |- 
 
 : Exists C. has_role(C, ruler_of_westeros) 45 Narrative variation?
46. Planners & theorem provers find the most efficient solution has_role(rb, ruler_of_westeros) |- 
 
 “The end.” : Exists C. has_role(C, ruler_of_westeros) 46 Narrative variation?
47. Planners & theorem provers find the most efficient solution has_role(rb, ruler_of_westeros) |- 
 
 “The end.” : Exists C. has_role(C, ruler_of_westeros) Stories are anti-efficient! 47 Narrative variation?
48. A Partial Solution: Nondeterministic Proof Construction Linear logic for story generation: Martens et al. 2013 (LPNMR) Use forward chaining nondeterminism to leave goals unspecified,
 randomly explore the search space 48 Sacrifice completeness of search for variation in outcome
49. Another Limitation Stories generate interest e.g. through characters with conflicting goals Subterfuge, betrayal, suspense, etc. all depend on characters acting on goals with limited knowledge 49
50. Meanwhile, in program synthesis… 50
51. Adding constraints Program synthesis:
 Refining types with input-output examples (test cases)
 
 51 f : int -> nat
 | f 1 => 1
 f 2 => 4
 f 5 => 25
52. Adding constraints Story generation: Enforcing intentional behavior in characters 52
53. Intentional reasoning From classical AI to modern FP Plots as plans as proofs Intentional reasoning Epistemic reasoning 53
54. Game AI: from opponents to cooperators Computer opponents (e.g. Chess, Go):
 moves need not be intelligible to human Computer cooperators:
 need to move in a legible, intentional-seeming way 54
55. Grice's Maxims of Conversation ❖ Quantity - say neither too much nor too little ❖ Quality - tell the truth ❖ Relation - what is said should be relevant to the topic/stated intentions ❖ Manner - avoid obscurity and ambiguity 55 (Paul Grice, “Logic and Conversation,” 1975)
56. Hanabi Your hand is visible only to others Your shared goal is to build sequentially- ordered stacks of like- color cards 56
57. Hanabi One turn: Play, Discard, or Give hint (spends a token) 57
58. Hanabi: hints 58
59. Computational cooperation in Hanabi ❖ Model the player’s knowledge, including their model of our knowledge ❖ Give hints to human players that will be recognized as intentional: point out cards when unambiguous whether to play or discard
 
 59
60. Computational cooperation in Hanabi ❖ Model the player’s knowledge, including their model of our knowledge ❖ Give hints to human players that will be recognized as intentional: point out cards when unambiguous whether to play or discard
 
 60 }“Intentional”
 model
61. Computational cooperation in Hanabi ❖ Model the player’s knowledge, including their model of our knowledge ❖ Give hints to human players that will be recognized as intentional: point out cards when unambiguous whether to play or discard ❖ Interpret hints from human players as carrying the same intent; play or discard appropriately 61 }“Intentional”
 model { “Full”
 model
62. Computational Intelligence in Games (CIG) 2017 3 implementations (base, intentional, full) + experimental evaluation 1. Simulation of computer:computer 2. Human subjects study, human:computer 62
63. Computational Intelligence in Games (CIG) 2017 63
64. Computational Intelligence in Games (CIG) 2017 64 Players rated more intentional agents both more enjoyable and more intentional-seeming
65. Grice's Maxims of Conversation Quantity - say neither too much nor too little ❖ Enforced by game rules 65
66. Grice's Maxims of Conversation Quality - tell the truth ❖ Players have shared goals 66
67. Grice's Maxims of Conversation Relation - be relevant ❖ Only give hints that can be interpreted as play/discard instructions 67
68. Grice's Maxims of Conversation Manner - avoid ambiguity ❖ Only hint sets of cards for which the actionable subset is unambiguous 68
69. Limitation Only one step of look-ahead How to apply intentional reasoning to search? 69
70. Intentionality depends on Belief Transition from ad-hoc encoding of knowledge to formal Describe how beliefs influence actions and vice versa 70
71. Epistemic reasoning From classical AI to modern FP Plots as plans as proofs Intentional reasoning Epistemic reasoning 71
72. Theory of Mind Forming mental models of other people’s mental models 72
73. Theory of Mind Forming mental models of other people’s mental models A key form of social reasoning in humans 73
74. 3 Logicians Walk into a Pub 74 Will everyone have a beer?
75. 3 Logicians Walk into a Pub 75 Will everyone have a beer? I don’t know
76. 3 Logicians Walk into a Pub 76 Will everyone have a beer? I don’t knowI don’t know
77. 3 Logicians Walk into a Pub 77 Will everyone have a beer? I don’t knowI don’t know Yes!
78. Agent a believes A
 Epistemic Logic ☐a A
 78
79. Dynamic Epistemic Logic 79 ☐a A
 [α]A Agent a believes A
 A holds under action α
80. Agent a believes A
 A holds under action α Dynamic Epistemic Logic ☐a A
 [α]A 80 Actions may change the 
 actual world or the
 set of worlds an agent considers possible
81. Epistemic Actions (Baltag 2002) α, β ::= flip p 
 | ?A 
 | α + β 
 | α ; β 
 Change truth value
 Precondition
 Nondeterministic choice
 Sequence
 81
82. Epistemic Actions (Baltag 2002) α, β ::= flip p 
 | ?A 
 | α + β 
 | α ; β 
 | αa 
 | α* Change truth value
 Precondition
 Nondeterministic choice
 Sequence
 Appearance to a
 Public action 82
83. Semantics: Possible Worlds Epistemic states: sets of world each agent considers possible 83
84. Semantics: Possible Worlds An actual world 84 Set of world an agent considers
 possible
85. Semantics: Possible Worlds These sets can expand and contract based on actions 85
86. Semantics: Possible Worlds 86 α Apply α
87. α α α α Semantics: Possible Worlds 87 α Apply α
88. Semantics: Possible Worlds 88 Apply α + β
89. Semantics: Possible Worlds 89 α Apply α + β β or α α α α β β β β
90. Semantics: Possible Worlds 90 Apply ?( v ); α
91. Semantics: Possible Worlds 91 α Apply ?( v ); α α α
92. –Petyr Baelish “Fight every battle everywhere, always, in your mind. Everyone is your enemy, everyone is your friend. Every possible series of events is happening all at once. Live that way and nothing will surprise you. Everything that happens will be something that you’ve seen before.” 92
93. A tiny example Alice flips a coin in secret;
 Bob and Carol observe 93 coin(heads)Alice + coin(tails)Alice; 
 (coin(heads) + coin(tails)){Bob, Carol}
94. A practical action language? 94 Aggregate sums/products for expressing more than Boolean changes
 Term size blowup
95. Ostari - Eger & Martens 2017 Parameterized actions
 Parameters can be secret to specific sets of agents 95 ⟨action⟩ ::= ⟨identifier⟩(⟨params⟩) <cmd>
96. Ostari - Eger & Martens 2017 <cmd> ::=
 <property> := <value>
 | learn <agents> <fact>
 | public <cmd>
 | <cmd> ; <cmd> 96 ⟨action⟩ ::= ⟨identifier⟩(⟨params⟩) <cmd>
97. Ostari - Eger & Martens 2017 <cmd> ::=
 <property> := <value>
 | learn <agents> <fact>
 | public <cmd>
 | <cmd> ; <cmd> 97 ⟨action⟩ ::= ⟨identifier⟩(⟨params⟩) <cmd>
98. Ostari - Eger & Martens 2017 <cmd> ::=
 <property> := <value>
 | learn <agents> <fact>
 | public <cmd>
 | <cmd> ; <cmd> <fact> ::=
 exists x.<cond>
 | all x.<cond>
 | each x.<cond>
 | which x.<cond> 98 ⟨action⟩ ::= ⟨identifier⟩(⟨params⟩) <cmd>
99. Ostari - Eger & Martens 2017 <cmd> ::=
 <property> := <value>
 | learn <agents> <fact>
 | public <cmd>
 | <cmd> ; <cmd> <fact> ::=
 exists x.<cond>
 | all x.<cond>
 | each x.<cond>
 | which x.<cond> 99 ⟨action⟩ ::= ⟨identifier⟩(⟨params⟩) <cmd>
100. Epistemic Quantifiers 100 learn (p): exists c in Cards: at(deck, 1) == c. Useless knowledge
101. Epistemic Quantifiers 101 learn (p): which c in Cards: at(deck, 1) == c. Useful knowledge
102. Example 102 hintcolor(p: Players, c: Colors) learn (p): Each i in HandIndex: color(at(p, i)) == c Hinting a color in Hanabi
103. Implementation in Haskell
 (laziness useful for possible worlds representation!) Sequences of actions and queries Goals that generate action sequences Ostari - Eger & Martens 2017 103
104. Epistemic Narrative Generation WorldA:
 at(Sherlock) = BakerStreet
 looks like (Sherlock): WorldA
 looks like (Moriarty): WorldA, WorldB WorldB:
 at(Sherlock) = ScotlandYard
 looks like (Sherlock): WorldB
 looks like (Moriarty): WorldA, WorldB 104
105. Epistemic Narrative Generation goal: 
 murdered(Moriarty, Victim) ^
 ☐(Moriarty): ☐(Sherlock): exists M. 
 M != Moriarty ^ murdered(M, Victim) 105
106. Epistemic Narrative Generation Generates a story in which Moriarty frames someone for murder goal: 
 murdered(Moriarty, Victim) ^
 ☐(Moriarty): ☐(Sherlock): exists M. 
 M != Moriarty ^ murdered(M, Victim) 106
107. Summary 3 investigations into using logic for creativity: 1. Generating narrative with planning/linear logic 2. Intentional reasoning in the context of cooperation 3. Codifying intentionality with Epistemic Logic 107
108. From modern FP to future AI From classical AI to modern FP Plots as plans as proofs Epistemic reasoning Intentional reasoning From modern FP to future AI 108
109. Compositionality FP’s secret weapon [A ∘ B] == [A] [∘] [B] 109
110. Compositionality 110 x f g z y y Function composition (meaning [-] is interface)
111. f ∘ g Compositionality Function composition (meaning [-] is interface) 111 x z
112. Compositionality Proof composition (use of lemmas) 112 C A ⊢ Δ, ⊢ A
113. Compositionality 113 , C Δ ⊢ Proof composition (use of lemmas)
114. Compositionality In planning: Frame Principle (“horizontal” composition) 114 Δ Δ’ π ’ ρ
115. Compositionality In planning: Frame Principle (“horizontal” composition) 115 Δ Δ’ π
 ||
 ρ ’
116. Compositionality In planning linear logic: Frame Principle (“horizontal” composition) 116 Δ Δ’ ’ π
 ||
 ρ
117. Research Challenges Dynamic logics with functional action languages Constructive/lambda-calculable DEL Epistemic session types 117
118. Compositional Creativity? asking computers to generate 
 insights, behaviors, narrative
 based on logical specifications 118
119. Why Games and Stories? Logic is a great intermediate language between humans and machines, but insufficient Humans have been communicating through play and storytelling for even longer than they have been doing logic The best abstractions are discovered in diverse disciplines! 119
120. Thanks! Chris Martens http://go.ncsu.edu/martens martens@csc.ncsu.edu @chrisamaphone github.com/chrisamaphone 120