Transcript Computational Creativity

Computational
Creativity
Richie Abraham(05005010)
Pramod Mudrakarta(05005030)
Shashank Samant(05D05011)
Sumedh Ambokar(05D05013)
Computational Creativity
•
Creativity is a process involving the generation of new ideas or concepts, or
new associations between existing ideas or concepts
-- Wikipedia.
•
•
Humans are creative. Ability to think out of the box.
Goal of Computational Creativity: To model, simulate or replicate creativity
using a computer.
“Reflections are images of tarnished aspirations”
--a quotation generated by the program RACTER, in 1984
Roadmap
• Motivation
• Formalizing the notion
• Creative Flexibility
– Meta- level
– Analogy
• Discovery Programs
• Case studies
– AARON
– BRUTUS
Why Creativity?
• To construct a program/computer capable of
human-level creativity.
• To better understand human creativity and to
formulate an algorithmic perspective on
creative behavior in humans.
• To design programs that can enhance human
creativity without necessarily being creative
themselves.
Essential Characteristics of a Creative Idea
Ideas of Newell, Shaw and Simon
• Novelty and usefulness (either for the individual
or society)
• Rejection of previous ideas
• Results from intense motivation and persistence
• Clarification of vague ideas
Margaret Boden’s view
• P-Creativity (Psychological or Individual)
• H-Creativity (Historical or Collective)
Formalizing the Philosophical Concepts
Exploration: Within a conceptual space
Transformation: Out of the box(space)
Formalization :
• Conceptual Space C a strict subset of set of all concepts U
• Axiom1:Every concept c is a distinct member of U
• Axiom2:Every conceptual space includes F (empty
concept)
Formalization (contd)
• Rules : R(Existsence in a space) and T(Transformation in a
space) R , T subsets of L
• Interpretation function: [.] and Traversal function: <.>
• Exploratory concepual space as a tuple : (U,L,R,T,E)
• Beginning of exploratory creative process : <R + T>({F})
• Evaluating concepts (E) : E(C) = value of the conceptual
space.
Formalization(contd)
• Partitioning a conceptual space into “concepts
achieved” and “concepts not achieved yet”
• Exploratory search involves experimenting with T
• Transformation search involves experimenting with R
• T is the “technique” of the individual to search
• R is the mutually agreed domain specified.
• Meta level: Rules for changing R and T
Where is the AI?
• RACTER, 1984 generates poems, stories, etc.
– Syntax directives
– Sentences too bizarre at first look
– Deeper meaning on repeated thought
• Creativity is in the reader’s mind
• Sentences become insignificant soon
• Need for more control
Need for Flexibility
• Rule-based systems are monotonous
• Example: Generating a story (TALE-SPIN)
– Each object tries to satisfy its goals
– Creativity is shown only when the plot turns an
unexpected way
– Object need not try to reach goals at every step
• Solution: Use the “meta-” approach
• Develop rules for rules
Using meta- rules
MUSCADET: Theorem prover for linear spaces
• Heuristics used in proving
• Meta- rules over heuristics
Drawbacks: Does not distinguish important and
trivial issues from a math point of view.
Example: Trying to be creative in proving “1=1”
versus trying to be creative in proving prime
factorization
Another example(problem?)
DAY-DREAMER: planner
• Operates on two interacting domains(personal,
objective)
• Each works on its own goals.
• Preprocessing: Determines situations where
personal goals are met
• In action: Tries to match the succesful plans with
the objective world situations
Drawbacks: Determining the parameters of
personal world is hard.
The meta- question; Analogies
• Meta-rule based systems not very different
• Need for meta-meta-rule based systems
• Deja vu?
• Alternate approach: Working by analogy
• Concepts from other domains applied
Analogy contd.
Mapping (electron, nucleus) to (planet, sun)
Problem: Choosing variables whichdetermine
similarity
• Planets on orbit
• Planets have moons
• Sun loses energy, emits light
Drawback: Solving the problem is hard
Discovery Programs
Shashank
Discovery Programs Overview
• Able to discover new facts on a domain
• Three major families:
– AM Family
• Domain Mathematics
• AM , Euresco , Cyrano
– BACON
• Domain experimental data
• BACON, GLAUBER, STAHL, DALTON
– GT
• Domain Graph Theory
AM Overview
• Starts with set of concepts arranged in a
specialization hierarchy
• Concept
– Definition, Examples, Domain, Range, Specializations,
Worth
– Initial concepts: Sets, List, Ordered pairs and some
operations
• Heuristics
– Fill, Check, Suggest, Interest
• Task
– Applying heuristics on set of concepts
– Output concept as a code
AM
• Discoveries
– Natural numbers, addition, primes
– Prime factorization, Goldbach’s conjecture
• Limitations
– Heuristics too theory specific
– Many theories ignored
– Interpretation of concepts ambiguous
BACON Family
• Operate on a data driven basis
• Heuristically guided process
– Mostly an ad-hoc curve fitting exercise
• BACON
– Syntactic number games to summarize data
• GLAUBER
– Generalization from specific examples
• STAHL
– Model building using three rules
• Infer, Substitute and Reduce
• DALTON
– Atomic Modelling
Graph Theorist (GT)
• Discovers and proves properties of graphs
• Graph property
– A property p represents a set of graphs P iff every
graph in P satisfies p
– Represented by a concept
– Examples: TREE, ACYCLIC, COMPLETE.
Concept
• Defined by a triple <f, S, σ>
– f : operator
• To transform a member to a new member
– S : seed set
• Minimal graphs satisfying the property
– σ : selector
• Restrictions for binding variables appearing in f
• Example
– Acyclic <Ax+AyzAz; {K1}; y in V, x, z not in V>
p-Generator
• Exhaustively generates P described by p
• Checks if particular graph is a member of P
– A* can be used
– Still quite inefficient
• Not of much interest
4 Types of Graph Theorems
• If a graph has a property p, then it has
property q
• A graph has property p if and only if it has
property q
• If a graph has property p and property q, then
it has property r
• It is not possible for a graph to have both
property p and property q
Subsumption and Merger
• Property p for class P subsumes property q for class
Q iff Q is a subset of P
• Merger of p and q is the property representing
intersection of P and Q
• The four rules rewritten as
–
–
–
–
q subsumes p
p subsumes q and q subsumes p
r subsumes merger of p and q
merger of p and q is empty
Proofs of Subsumption
• p1=<f1,S1,σ1>, p2=<f2,S2,σ2>
• p1 subsumes p2 if– f2 is a special case of f1
– Every graph in S2 has property p1.
– σ2 is more restrictive than σ1
• Example:
– GRAPH subsumes TREE.
Proof Involving Mergers
• p1=<f1,S1,σ1>, p2=<f2,S2,σ2>
• P is the merger of p1 and p2
– If p1 subsumes p2, p is p1.
– Three more complex rules.
• Example– ACYCLIC merged with CONNECTED is TREE
Construction of new concepts
• By specialization
– constrain the seed set, operator or selector
– combination of above
• By generalization
– expand the seed set, operator or selector
– combination of above
• By merger
Final word on GT
• Generates many new concepts and proves
properties.
• Power increases with increased knowledge
base
• Limitations
– does not assign worth to concepts
– only properties of “graph theory”.
Case Studies
Sumedh
AARON Overview
•
•
•
•
•
•
By Harold Cohen
Creates original artistic images
Since 1973
Initially only black and white images
Colored images since 1992
See it to believe it !
AARON Techniques
• Structure of core-figures embedded
– Body parts
– Postures
• Starts scribbling randomly
• Next step based on what is drawn so far
• Coloring after sketching
– Core-figures determine colour
Is AARON creative?
•
•
•
•
•
Can create infinite distinct images
Cannot learn imagery on its own
Output follows a noticeable formula
Real artist is Cohen
Cohen: “If it is not thinking, what exactly is it
doing?“
BRUTUS Overview
• A creative story generator
• Should have wide variability
– Plot, characters, settings, themes, imagery
• Earlier strategy
– Each variable aspect parameterized
– Wide variability not achieved
BRUTUS approach
• Designed to satisfy seven characteristics
– Capable of raw imagination
– Generate imagery
– Defines mental state and actions of characters
– Mathematize themes
– Interesting stories
• Topics like sex, money and death
– Structured stories
– Avoid mechanical prose
Conclusion
• Many philosophical issues
• Lack of universally accepted definition of
creativity
• Light at the end of tunnel
• One of the fastest growing areas of research in
AI.
Current Research
• IJWCC 2003, Acapulco, Mexico, as part of
IJCAI'2003
• IJWCC 2004, Madrid, Spain, as part of
ECCBR'2004
• IJWCC 2005, Edinburgh, UK, as part of
IJCAI'2005
• IJWCC 2006, Riva del Garda, Italy, as part of
ECAI'2006
• IJWCC 2007, London, UK, a stand-alone event
Journals
• Journal of Knowledge-Based Systems, volume
9, issue 7, November 2006
• New Generation Computing, volume 24, issue
6, 2006
• http://www.thinkartificial.org/artificialcreativity/
References
• Learning and Discovery: One System’s Search
for Mathematical Knowledge. Epstein.
Computational Intelligence 4 (1): 42-53, 1988.
• Creativity: A survey of AI approaches, J. Rowe
and D. Patridge. Artificial Intelligence Review
7, 43--70, 1993.
• Colouring Without Seeing: a Problem in
Machine Creativity. Harold Cohen, Dept. of
visual arts, UC San Diego, 2003
References
• The further exploits of AARON-painter,
Harold Cohen, 2001.
• Towards a more precise characterisation of
creativity in AI, IJWCC 2005
• www.wikipedia.org
• www.kurzweilcyberart.com