Jason Hutchens

1
(No Transcript)
2
Jason Hutchens
AI Team Lead Team Bondi
3
Interactive Drama
Dynamic Dialogue and Sandbox Worlds
4
ABOUT ME

• Research in Grammatical Inference and Data
  Compression.
• Worked at Lionhead Studios on Black White.
• Former Chief Scientist of Artificial
  Intelligence Inc. in Israel.
• Recently joined Team Bondi as AI Team Lead.

5
BLACK WHITE
6
SCOPE

• No State Machines, Scripting, A or Flocking.
• Use AI to create a dramatic experience for the
  player.
• What can be achieved on Sony's next-gen console?

7
ARTIFICIAL INTELLIGENCE

• A contentious term. Slippery to define
• Turing un-asked the question what is
  intelligence via his Test.
• We're specifically interested in behaviour that
  appears intelligent.

8
INFORMATION THEORY

• Intelligent behaviour is unpredictable yet
  relevant.
• Eliminate randomness and repetition maintain
  reasonable surprise.
• Behaviourism predicts the slot machine
  phenomenon.

9
EXAMPLE

• Stochastic language models generate unpredictable
  sentences
• MegaHAL WITHIN MY PENGUIN LIES A TORRID STORY OF
  HATE AND LOVE.
• Keyword-based generation helps to make them
  relevant.
• User In 1793 the french king was executed.
• MegaHAL HA HA! CORRECT. ALTHOUGH, EXECUTED HAS
  MULTIPLE MEANINGS.
• Users talk to this thing for hours on end, and
  respond emotionally.

10
SANDBOX WORLDS

• Graphics got fast and detailed, so players wanted
  more interaction.
• Physics improved, and is now practically
  mandatory.
• It's fun to muck around within these sandbox
  worlds.
• AI required to populate the world with believable
  characters.

11
INTERACTIVE DRAMA

• All the world's a stage add a compelling story
  to the mix.
• Façade, by Mateas and Stern, is an example
  (InteractiveStory.net).
• Dialogue is a necessary component.
• AI required to present the story, and process the
  dialogue.

12
AUDIENCE OR PROTAGONIST?

• Cut-scenes are non-interactive and quickly become
  repetitive.
• Sandbox worlds lack dramatic tension.
• Rather than being the protagonist, the player
  may act in their role.
• AI required to interpret player action with
  respect to its context.

13
GAME AI

• Typically, only 5 of resources spent on AI.
• We'll soon have the power of a desktop PC of the
  early 1990's.
• Create a compelling experience on such
  technology.
• We need to look beyond the search-and-optimise of
  GOFAI.

14
SUMMARY

• AI is required to populate the world with
  believable characters.
• AI is required to present the story.
• AI is required to process the dialogue.
• AI is required to interpret player action.

15
THE CHALLENGE
16
ARCHITECTURE
17
PLAYER MODEL
18
PLAYER MODEL

• Requirement Interpret player action.
• Goal Balance player freedom with narrative
  consistency.
• Goal Enable cinematic presentation of gameplay
  sequences.
• Abstracted commands context provided by story,
  world and subject.

19
DIALOGUE ENGINE
20
DIALOGUE ENGINE

• Requirement Process the dialogue.
• Goal Ubiquitous, believable conversations.
• Voice acting, localisation and input devices
  limit the possibilities.
• Move beyond menu-based conversation and taunts.

21
DIALOGUE ENGINE

• Adopt Mann's Dialogue Macrogame Theory.
• A conversation is a sequence of short dialogue
  games.
• Players take turns speaking, and negotiate game
  entry and exit.
• Dialogue Engine moderates the games, and issues
  potential actions.

22
DRAMA MANAGER
23
DRAMA MANAGER

• Requirement Present the story.
• Goal Select a story thread and direct other
  components.
• Provide potential dialogue games to the Dialogue
  Engine.
• Constrain the behaviour of the Player Model and
  the Cast.
• Inform the Camera, etc.

24
STORY ENGINE
25
STORY ENGINE

• Requirement Present the story.
• Goal Provide potential story threads to the
  Drama Manager.
• Story threads specify cast, dialogue, locations,
  props, etc.
• Story Engine selects relevant story threads
  opportunistically.
• Each thread is a network of potential dramatic
  moments, or beats.

26
CAST MANAGER
27
CAST MANAGER

• Requirement Populate the world with believable
  characters.
• Goal A world of 1.5 million inhabitants.
• Goal Embodied, autonomous, reactive and
  conversational NPCs.
• Extras provide background, and ad lib when
  necessary.
• Heroes participate in the overarching story.

28
CAST MANAGER

• NPCs have personality, emotion and a hierarchy of
  relationships.
• Extras are spawned at random, according to
  demographics.
• Extras behave reactively, foraging the
  environment and exiting ASAP.
• Heroes are hand-crafted, and inhabit the world
  persistently.
• Heroes behave deliberatively, as specified by
  story and routine.

29
WORLD MODEL
30
WORLD MODEL

• Provide world knowledge to Story Engine and Cast
  Manager.
• Manage non-story events that occur throughout the
  world.
• Provide appropriate dialogue games for talking
  about them.

31
WORLD MODEL

• Adopt Milgram's Six Degrees of Separation.
• A grapevine extends from the NPC to an item of
  world knowledge.
• Probability of existence may be estimated.
• Each NPC appears to have their own knowledge of
  the world.

32
CONCLUSION

• Suspension of disbelief requires believable
  worlds.
• Technology only worth using if it enhances the
  player's experience.
• We have the resources to build an AI team to
  achieve them.
• And the opportunity to create a groundbreaking,
  immersive game.
• JasonHutchens_at_gmail.com

33
(No Transcript)