the halo 2 ai in 10 minutes or less damián isla bungie studios
TRANSCRIPT
The Halo 2 AI In 10 Minutes or Less
Damián IslaBungie Studios
Halo
• First person shooter for XBox• Interesting sci-fi world• Lots to do (combat, exploration, vehicles, weapons)• Lots of decisions to make• Slower-paced More strategic in nature?
The AI
Allies• Not terribly effective• “Kid brothers”• ~10 seconds or ~30 minutes
Enemies• Large variety• Clear behavioral patterns for the player
to learn and exploit• 4 - 30 second lifespans
Single-player campaign only (no multiplayer)
Character-based
The “I” in “AI”• Perception and world-modeling
– Each AI maintains mental world model
– Internal mental object-state updated according to perceivability via the “senses” (e.g., line of sight, sound events, etc.)
– Perception sharing (via explicit communication)
• Behavior– Take cover when taking damage
– Uncover hidden enemies
– Use vehicles when appropriate
– Joint behaviors
• Skills– Use weapons appropriately
– Drive vehicles
– Navigate the environment
Strongly agent-based approach
“Strategy” controlled through scripting
Sense
Plan
Act
Behave
BehaviorDecision-making: Behavior DAG
(HFSM)
Behavior: a program that takes temporary control of the actor in order to get something done.
• Good– Intuitive– Modular– Scalable (more IS better)
• Bad– Memory is hard– Debugging is hard– Never really that simple
Root Self-preservation
Engage Search
Charge
Fight
Guard
Grenade
Cover
Presearch
Uncover
Guard
Grenade
Vehicle
Investigate
Suppressing fire
Grenade
Postcombat Shoot corpse
Check corpse
Idle Guard
Moving, aiming, shooting, animation
Perception and World Modeling
Object states, events
Orders
An ethological interlude…
Root Self-preservation
Engage Search
Charge
Fight
Guard
Grenade
Cover
Presearch
Uncover
Guard
Grenade
Vehicle
Investigate
Suppressing fire
Grenade
Postcombat Shoot corpse
Check corpse
Idle Guard
Baerends, The Functional Organization of Behavior, 1976
What about Reasoning and Planning?
• Spatial Reasoning– Firing point selection
• Path-planning– A*– Dynamic obstacle interaction
… but no concept-reasoning or action-planning
So in the end…
• Second-to-second decision-making• Reactive in nature• Closer to A-Life than classical AI• Closer to animal than to human intelligence