A Genetic Algorithm Approach To Interactive Narrative Generation

1
A Genetic Algorithm Approach ToInteractive
Narrative Generation

• TeongJoo Ong and John Leggett
• Texas AM University

2
Introduction

• Stories are used to convey information, cultural
  values, and experiences
• New technologies have constantly provided
  increasingly sophisticated means to tell stories
• Recent trend is the convergence of technology,
  entertainment, and art in the computer

3
Background

• Interactive storytelling is a major research area
• Many overlapping approaches have been used
• AI community (Narrative intelligence)
• Immersive storytelling
• Emergent storytelling
• Plot-based systems
• Interactive authoring of stories
• Character-based systems

4
Background (cont.)

• Hypertext community
• Hypertext narratives
• Adaptive hypermedia
• Sculptural hypertext
• Partial listing of related work from these areas
• CHOROS (N. M. Sgouros)
• Façade (M. Mataes, A. Stern)
• Card Shark and Storyspace (M. Bernstein)
• Metalinear Cinematic Narrative (K.M. Brooks)
• StoryBeads (B. Barry)

5
Motivation

• The HEFTI storytelling engine attempts to merge
  results and ideas from both communities
• Recombination, mutation, selection of authored
  story elements with Genetic Algorithm (GA)
• Generate, remove and traverse links in the story
  elements as the story unfolds using the authors
  predefined rules
• Story elements and rules are encoded in XML
• Provides a small drag-and-drop tool and tree-view
  tools for authors to create their stories

6
Definitions

• Story elements Smallest units in HEFTIs story
  search space
• Story template Describes combinations of
  various story elements at a particular stage in
  the story
• Template sequences Subdivision of story
  template to depict time-based relationships among
  story elements
• Story component Combination of story elements,
  agent scripts and variables generated by HEFTI
  based on the given set of story templates and
  elements

7
Generating a Story

• A story is divided into multiple time steps
• Chromosomes represent a story component
• Genes represent a collection of story elements
  pertaining to a template sequence within a story
  template
• Genes are encoded as floating point numbers for
  convenience of manipulation (as shown later)

8
Generating a Story (cont.)

• A gene is generated by constructing valid story
  element sets based on the current state of the
  story and various conditions and rules imposed by
  the author (Encoding)
• The encoding process is sequential due to
  dependencies and ordering of story elements
• The encoding process is repeated several times to
  generate individuals in the GA population

9
Generating a Story (cont.)

• After the evolution process, the decoding process
  steps through each of the genes, decoding each
  gene based on its story context
• The end result is a story component that
  describes the next story sequence
• The process is repeated to generate subsequent
  story sequences

10
Generating a Story (cont.)
11
Operations on Story Components

• The original genetic operators are modified to
  handle authored ordering of story elements
• Chromosome level operators
• Single point crossover
• Multi point crossover
• Gene level operators
• Crossover operators
• Mutation

12
Chromosome Operators
Chosen offsets for multipoint crossover
Chromosome A




Chromosome B
Chromosome A




Chromosome B

• The offsets are chosen to share similar starting
  conditions and consequences
• The operator generates permutations of the
  template sequences

13
Chromosome Operators (cont.)

• An example from The Three Little Pigs
• Given two chromosomes (A and B) with
• Starting conditions
• Wolf shows up in front of Angela Pigs house
• Consequences
• Wolf eats Angela Pig
• If the state of the story is the same for both,
  the intermediary stages can be swapped between A
  and B

14
Gene Operators

• Alter selection of story elements
• Crossover operators are the same as at the
  chromosome level except they crossover story
  elements instead of genes
• Mutation operator changes the offset into a list
  of possible story elements

Gene A (Before mutation)
Gene A (After mutation)


1
1
Chosen offset for mutation
3
1
6
6
4
4


15
Story Threads

• Authors create potential story threads
• Story thread is an indicator for
• Length of story
• Choice of story elements from a story template
• Sets of story templates to be used at certain
  story time steps
• Rules that change the flow of the story
• Evaluation criteria for chromosome fitness

16
Fitness Evaluation

• HEFTI is free to generate stories that adhere
  only so closely (within a threshold) to a story
  thread
• Fitness evaluation takes place on the chromosomes
  of a story thread
• Authors can indicate preference, indifference, or
  dislike towards certain story elements by
  assigning positive or negative values to the
  story elements
• The existence of certain story elements in story
  components are influenced accordingly through the
  selection mechanisms of GA

17
Story Thread with Fitness Evaluation

• lttimestep order"4" name"wolf's plan"
• loop"isPigStillAvailable.value true
• AND isWolfDead.value false"gt
• ltset type"wolf's plan" gt
• ltelement name"act_13" addfitness"0.8"/gt
• lt/setgt
• lt/timestepgt

18
Scenarios of Use

• Interactive fiction A murder mystery with
  multiple dynamically generated story threads
• Sub-module of a computer game engine that
  dynamically combines various story and agent
  scripts from existing story elements
• An educational environment that can present new
  concepts in various forms while preserving the
  goals of the story

19
Conclusions

• HEFTI needs a large enough set of story elements
  and templates for it to create the story
  components
• Authoring a story from scratch can be tedious
• Story templates allow authors to create story
  elements at different granularities
• Story elements can be reused in similar story
  settings

20
Future Work

• Genetic programming might be used to generate
  agent scripts given the starting and ending agent
  actions
• Evaluation of fitness data on all of the story
  variants that HEFTI is capable of generating
• Better support for user interaction and control
  of the story