Next-Generation Game AI

1
Next-Generation Game AI

• or, Weve figured out shooting. Now what do we
  do?

CS196-2 Innovating Game Development
Ryan Houlette Lead AI Software
Engineer Stottler Henke Associates houlette_at_stott
lerhenke.com 617-616-1293
2
Outline

• Who am I?
• State of the art
• Future directions
• Questions
• Tools

3
Who am I?

• Lead AI Software Engineer at Stottler Henke
  Associates since 1998
• AI consulting firm
• commercial government clients (NASA DoD)
• Architect of SimBionic game AI middleware
• Author section editor in AI Game Programming
  Wisdom series
• Gamer since about 1984

4
State of the Art

• In a word disappointing.
• Only a fraction of todays games have good AI
• Even then AI rarely matches the quality of the
  graphics and sound (and increasingly, physics)
• Many games have merely passable or sub par AI
• unrealistic behavior
• not challenging or unfairly difficult
• boring, predictable, or unresponsive
• just plain stupid
• Why is this?

5
A Game AI Programmers Life is Hard

• Developing game AI is hard
• one of the most complex parts of gameplay
• demands extensive tuning and balancing
• requires a mostly-working game world to test
• developers often start from scratch
• Tension between character design and character
  autonomy
• Currently game AI is written in code or scripting
  language
• game designers are often non-programmers
• separates the design of the AI from the
  implementation
• translation barrier
• long design-code-test cycle

6
and AI is Hard

• Many human-level capabilities cannot currently be
  easily replicated by computers
• vision
• natural language
• movement
• common sense
• learning
• AI algorithms tend to be computationally
  expensive
• so are nice graphics and physics
• AI ends up getting a small slice of the CPU
  (about 10)

7
Current Technology

• Code
• C, Lua, Python, UnrealScript, etc.
• Finite state machines
• A for pathfinding
• Annotated worlds (smart environments)
• A few specialized algorithms
• influence maps for RTS

8
There is Hope!

• A few ambitious developers
• e.g. F.E.A.R. (Monolith), Black White
  (Lionhead)
• More contact between academia industry
• AIIDE conference
• lots of game-development curricula
• Dawning recognition of AI as next big thing
  after graphics, physics
• Faster machines offloading of graphics, physics
  chores on specialized hardware

9
Directions for Next-Gen Game AI

• 0 Eliminate glaring stupidity
• instantly breaks the suspension of disbelief and
  ruins immersion
• usual culprits abysmal pathfinding,
  unresponsiveness, insensitivity to context,
  repetitive behavior
• Much of this is due to brittleness of current
  game AI

10
Getting to Zero

• Overcoming NIH syndrome
• Standard algorithms
• AI Game Programming Wisdom series, Gamasutra,
  Game Developer Magazine, etc.
• Middleware
• Open-source packages
• Recognizing the importance of good AI
• requires dedicated developers w/specialized
  expertise
• needs to be included in pre-production and
  tool-building
• More open-ended games are forcing more open-ended
  AI

11
Direction 1

• Increase behavioral verisimilitude
• to complement increasingly realistic character
  models
• more realistic gestures, gaze, facial
  expressions, movement
• can be done manually by animators, but content
  creation is a huge and growing cost in game
  development
• AAA title cost jumping from 2-10M for Xbox/PS2
  to 10-30M for Xbox360/PS3
• subtle AI needed to imbue characters with the
  appearance of emotion, attention, intention
• caveat aim is the illusion of life, not the real
  thing

12
The Uncanny Valley

• Described by Masahiro Mori in 1970
• Humans are more sensitive to subtle behavioral
  miscues in human-like characters
• As characters look more realistic, they run the
  risk of becoming increasingly unconvincing
• Subtle AI needs to model and capture these
  nuances to avoid the valley

GFDL Karl MacDorman, Wikipedia
13
Example Half-Life 2

• Valve pushed the state of the art in facial
  animation, gestures
• Excellent series of articles on gaze in Game
  Developer Magazine, Aug-Sep 05
• Based on study of real human behavior
• Still lots of scripting behind the scenes

14
Half-Life 2
15
Direction 2

• Engage the players mind
• games have their eyes, ears, and hands already
• goal is not to build super-AIs that always win
• relatively easy
• not much fun for the player
• better to build AI that puts up an epic fight and
  then loses excitingly
• want the player to treat characters as thinking
  adversaries (or allies!)
• encourages emotional engagement
• allows the player to reason about and outsmart
  NPCs

16
Tools of Engagement (1)

• Goal-based planning
• gives AI explicit goals that the player can
  deduce
• enables AI to choose among multiple means of
  achieving the same end
• adds variety, reduces brittleness
• used in F.E.A.R.
• must be carefully constrained to avoid search
  combinatorics

17
Tools of Engagement (2)

• Team/social intelligence
• characters are largely unaware of one another
• ability to work as a team to achieve shared goals
  creates a challenge for the player
• both player-visible and invisible inter-NPC
  communication are useful tools
• next level characters that have a model of other
  NPCs goals and intentions
• difficulties joint planning, synchronization

18
Tools of Engagement (3)

• Emotions and personalities
• NPCs who have distinct personalities are easier
  for the player to anthropomorphize
• not just another guard
• NPCs whose behavior is affected by their
  emotional state may require more subtlety from
  the player
• Even relatively simple models can be effective
• Can be tied to team/social behavior

19
Tools of Engagement (4)

• Adaptive behavior
• characters who learn and exploit players habits,
  strengths, weaknesses
• can be used to dynamically adjust difficulty to
  keep game constantly engaging
• more sophisticated than simply scaling the number
  or accuracy of enemies
• smart advisors/teammates that help out when
  player is stuck
• in-game vs between-game learning
• learning can be costly
• for some genres (e.g. RTS), between-game learning
  is appropriate
• ranges from simple statistical approaches to more
  complex induction/pattern recognition

20
Direction 3

• Enable new kinds of gameplay
• AI is intimately tied to game design, even more
  closely than physics
• game genres have been limited by the capabilities
  of game AI
• look at all of the movie/book genres yet to be
  tapped by games!
• demands more than incremental improvements to AI

21
Exploring New Frontiers (1)

• Dialogue-capable NPCs
• basically shunned in games since Infocom
• speech recognition is tough
• parsing is tough
• understanding is really tough
• speech generation is not nearly as good as voice
  actors
• the potential rewards are huge
• non-combat-based games
• richer stories
• more natural, open-ended interactions with
  characters
• new modalities for player interaction

22
Exploring New Frontiers (2)

• Virtual stagecraft
• NPCs who know how to act
• games, like movies, already make heavy use of
  selective omission
• want to emphasize the important and interesting
  stuff
• Virtual directors who manage the pacing and
  intensity of the story development
• Virtual storytellers who build customized
  storylines for the player and adapt them
  dynamically

23
Exploring New Frontiers (3)

• Persistent worlds
• current MMOGs have barely scratched the surface
  here
• e.g. player surrogates
• take over for the player when he/she is offline
  (in an RTS-style game)
• learn and mimic the players playing style
• integrating NPCs socially into communities
• rumor, history, reputation, memory

24
Direction 4

• Develop better AI tools and standards
• content creation is the single biggest challenge
  facing the industry today
• AI tools have traditionally been home-grown (if
  they exist at all)
• better tools are needed if AI is to improve
• usable by designers
• reusable across projects (incremental
  improvement)
• more data-driven
• the analogue of a higher-level language for AI
  is desperately needed
• permit designers to work at the conceptual level
  rather than micromanaging every footstep and
  eyeblink

25
Summary

• Current state of game AI C-
• Plenty of room for innovation
• ransacking academic AI for algorithms
• exploring interplay of AI and game design
• building better tools
• AI will soon be a major selling point for games
  on a par with graphics physics
• How about an AI Game Jam?

26
Questions?
27
Tools for Class Projects

• SimBionic
• game AI middleware that allows visual authoring
  of NPC behaviors
• integrates with Java/C game engine
• SimVentive
• entire visual platform for developing games and
  simulations

28
SimBionic Overview
29
SimBionic Demo
30
SimVentive Overview
Sim Authoring Tool
Sim Execution Engine
Player
Author
Sim Specification
Standardized GUI
Scenario Library

• Authoring
• Create or modify
• Test and debug

• Playing
• Standard Windows PC
• New games can be easily distributed

31
SimVentive Demo