Artificial Intelligence in Games Week 7

1
Artificial Intelligence in GamesWeek 7

• Steve Rabin
• steve.rabin_at_gmail.com
• www.aiwisdom.com/uw

2
Pathfinding Integration Project

• How did it go?
• Any problems?
• How much time did it take?

3
Euphoria Demo
4
Tactics and Terrain Analysis
5
Terrain Analysis

• Algorithmic analysis of terrain
• Automated with no human intervention
• Alternative to hard-coding regions of interest
• In general
• Are there spots on the terrain that are
  advantageous to me?
• How can these be determined algorithmically?
• Given the current terrain, how can I tease out
  this info?

6
Terrain AnalysisReal-Time Strategy Games

• What are static points of interest on an RTS map?
• What are dynamic points of interest during an RTS
  battle?

7
Terrain AnalysisReal-Time Strategy Games

• Chokepoints
• Barriers force units through narrow path
• Barriers can be cliffs, walls, water, etc.
• Ambush sites
• High terrain overlooking low terrain chokepoint
• Wall building sites
• Low cost for building walls
• Natural barriers so not much wall required
• Cliffs
• Water
• Edge of map
• Identify front of the battle
• Identify flanking routes
• Identify inferior forces

8
Terrain AnalysisFirst-Person Shooter Games

• What are static points of interest in an FPS map?
• What are dynamic points of interest during an FPS
  fight?

9
Terrain AnalysisFirst-Person Shooter Games

• Cover points
• Shielding
• Darkness
• Sniper points
• Good cover point (shielding, darkness)
• Potential high visibility to a large area
• Rear coverage
• Height advantage
• Escape route
• Safe paths to move from area to area
• Low casualties in path
• Low visibility to other areas
• Good cover along path
• Line-of-sight vulnerability
• Stay out of other's line-of-sight

10
Tactical Reasoning in a First-Person Shooter

• Visibility is a good tactical indicator
• Does spot provide cover?
• Front cover and rear cover?
• How safe is it?
• Does spot provide accessibility?
• Is a given spot good to shoot from?
• Can we escape easily?
• How dangerous is it?

11
Tactical Reasoning with WaypointsHalf-Life 2
(first-person shooter)

• Implementation strategy
• Preprocess terrain, storing visibility info
• Can area A see area B?
• Store visibility info in waypoint nodes
• Node As ability to see every other node
• How would you store it?

12
Tactical Reasoning with WaypointsStoring
Visibility Information

• Store visibility info with bit strings
• Can perform operations on groups
• and
• or
• not

13
Tactical Reasoning with WaypointsSafe and
Dangerous Nodes

• Which locations are
• Potentially safe?
• Potentially dangerous?
• Safe pathfinding
• Penalize dangerous nodes
• These are node visibilities
• What about the actual enemies?

14
Tactical Reasoning with WaypointsAttack
Positioning

• Visible nodes
• Show where you could attack from
• But no cover
• Better strategy
• Find a node to attack enemy X from
• Which is also shielded from other enemies

15
Tactical Reasoning with WaypointsAttack
Positioning

• 1. Get nodes visible to selected enemy
• 2. Get nodes visible to other enemies
• 3. Get safe nodes to attack from 3 1 !2

16
Tactical Reasoning with WaypointsAttack
Positioning

• Other strategies
• Take cover from everyone
• Find a safe node that is close to a completely
  safe node (in case NPC needs to reload)
• Flanking
• Eliminate attack points that the enemy is facing
• For sneaking up on enemies

17
Tactical Reasoning with WaypointsStatic
Waypoint Analysis

• Determine danger of each node in a static
  environment
• Weight each node based on number of nodes that
  can see it
• If pathfinding takes into consideration
• Paths are less likely to be visible to enemies

18
Tactical Reasoning with WaypointsStatic
Waypoint Analysis

• Nodes marked for visibility count

• Sniper nodes have high exposure next to nodes
  with low exposure

19
Tactical Reasoning with WaypointsPinch Points
and Squad Tactics

• Identify pinch points as ambush spots
• 1. Find node (N) with only two neighbors
• 2. One neighbor node must be in confined area
• 3. Nodes connected to other neighbor nodes are
  ambush spots

20
Tactical Reasoning with WaypointsLimitations

• Must use hierarchy of networks to reduce size of
  bit strings
• Ducking action can make some nodes both good
  cover and good attack locations
• Annotate these nodes
• Unidirectional connections makes it a little more
  complicated
• Density of waypoints impacts quality of reasoning

21
Influence Mapping
22
Influence Mapping

• Abstracts distribution of power
• 2D grid superimposed on terrain
• Unit strength is summed into grid cells
• Influence declines exponentially with distance
• Helps determine
• Weaknesses
• Neutral routes to flank enemy
• The front of the battle

23
Influence MappingBasic Influence Map
24
Influence MappingEngagement
25
Influence MappingAnalysis of Options
26
Influence MapsOther Uses

• SimCity
• Non-violent use
• Influence of crime, pollution, police, fire
  stations, etc

27
Spatial Database for Spatial Analysis
28
Spatial Database for Spatial Analysis

• Generalize terrain reasoning with waypoints
• Use a grid instead
• Expand to capture dynamic features
• Capture other interesting aspects

29
Spatial Database for Spatial AnalysisMany
Layers of Grids

• Visibility
• How many squares are visible
• Openness layer
• Proximity to static obstacles
• Cover layer
• Locations where AI can take cover from gunfire
• Area occupancy layer
• Number of units in an area

• Line-of-fire layer
• Direction that weapons are being aimed
• Light level layer
• Brightness or darkness of any spot
• Area search layer
• Spots that have been searched

30
Spatial Database for Spatial AnalysisOpenness
Layer
31
Spatial Database for Spatial AnalysisArea
Occupancy Layer
32
Spatial Database for Spatial AnalysisLine-of-Fir
e Layer
33
Spatial Database for Spatial AnalysisCombine
Layers

• Combat desirability layer
• Represents desirability of moving into any given
  area in combat
• Combines openness, static cover, area occupancy
  layers

34
Project 3Implement Terrain Analysis

• Due next week (May 19th)
• (100) Create at least 3 simple layers of grids
  and 1 combined layer that can be toggled through
  (with the "L" key)
• Possible layers Openness Layer, Occupancy Layer,
  Line-of-Fire Layer, Cover Area Layer, Area Search
  Layer, and Combat Desirability Layer
• Extra Credit (10) Have the agents use the
  information to do something intelligent

35
Squad Tactics
36
Squad TacticsDecentralized vs Centralized
37
Squad TacticsDecentralized Squad
38
Squad TacticsDecentralized Squad

• Key points
• Extension of individual AI
• Easy to add to existing AI
• Robustly handles many situations
• Deals well with a variety of capabilities within
  the team
• Can be easily combined with scripted squad member
  actions
• Cooperation and group behavior emerges from
  individual behaviors
• Technique used in original Half-Life

39
Squad TacticsDecentralized Squad

• void SoloAIThink()
• CombatSitiation sit
• GetSituation()
• sit-gtObserve()
• for each available action
• if(action.IsApplicable(sit))
• determine value of action
• in sit
• if better than best action,
• make best action
• execute best action()

• void SoloAIThink()
• CombatSitiation sit
• GetSituation()
• sit-gtObserve()
• sit-gtProcessTeamIntentions()
• sit-gtProcessTeamObservations()
• for each available action
• if(action.IsApplicable(sit))
• determine value of action
• in sit
• if better than best action,
• make best action
• nearbySqdMembers-gtAnnounceAction()
• execute best action()

40
Squad TacticsSolo AI vs Squad Member

• Fire-and-maneuver behavior
• If too many static squad members, move up
• If too many engaging threat, won't stop to engage
  same threat
• Robust against squad member casualties
• Interactions easily translated to audio for
  effect/realism

41
Squad TacticsSquad Assault

• Tactical behavior (on top of fire/move)
• Stay close to cover
• Don't block other's line-of-fire
• Take weapon ability into account
• Maintain team cohesion / line-of-sight
• Spread out to avoid being a single target
• Keep a line-of-sight on enemy

42
Squad Tactics Squad Assault Squad Members
Mental Picture

• For each squad member
• Current pos and activity
• Claimed destination
• Line-of-fire
• For each opponent
• Last known position and state
• Estimated position
• Squad members engaging this opponent
• Line-of-fire
• For other threats (grenades, fires)
• Known/estimated position
• Damage radius

43
Squad Tactics Squad Assault with Four-Man Patrol
in Hostile Situation
44
Squad TacticsLeader evaluating positions and
picking one (center, top)
45
Squad TacticsExecuting L-Shaped Ambush

• Wait until enemy enters kill zone, aggressively
  attack, pull back to predefined rally point
• Problems
• Decentralized squad has trouble waiting until
  right moment to attack. Attack moment requires
  some kind of coordination
• Squad lacks autonomy
• Can't create L-shape itself, can't abort if enemy
  doesn't show up

46
Squad TacticsDecentralized Squad (reference)
47
Squad TacticsRelation Between Squad and Squad
Member
48
Squad TacticsCentralized Approach

• Advantages
• Separates concerns reduces number of problems
  each AI has to deal with
• Saves CPU do computations once
• AIs can be specialized independently from each
  other

49
Squad TacticsCentralized Approach

• Authoritarian command style
• No "I" in "team"
• Squad can sacrifice a member to achieve a goal
• Squad member's view might conflict with squad's
  objective
• Told to defend a front, but sniper is behind
• Coaching command style
• Each squad member given tasks
• Can choose to ignore and tell squad he can't
  comply
• Deals with sniper case
• But squad members get distracted easily

50
Squad TacticsSituation and Extracted Features

• Force ratio 53, line-of-fire ratio 21

51
Squad TacticsFSM for Valid Maneuvers

• Avoids oscillations
• FSM prevents attacking before completing pullback

52
Squad TacticsBounding Overwatch Pullback

• Tail Squad Members Behavior
• Teammates lay down suppression fire

53
Squad TacticsBounding Overwatch Pullback

• Pullback path, Cover fire locations, and two
  bounding moves in progress

54
Squad TacticsStages of Performing a Maneuver
55
Team Member AIAvoiding Friendly Fire

• If player is shooting, how can buddy AI avoid
  friendly fire?
• If player is moving around, how can buddy AI
  avoid friendly fire?

56
Team Member AIAvoiding Friendly Fire

• Two options

57
Planning
58
Planning

• Find series of actions for AI that changes the
  current configuration of the game world into a
  target configuration
• Planning allows innovative solutions when the
  game world is sufficiently rich
• Requires
• Planning algorithm
• World representation
• Choosing set of actions
• Action
• Preconditions
• Effects

59
Planning
60
Planning

• Config1 Target not dead
• Action Attack
• Config2 Target is dead, Weapon must be loaded
• Action Load Weapon
• Config3 Target is dead, Weapon loaded, Weapon
  must be armed
• Acton Draw Weapon
• Config4 Target is dead, Weapon is loaded, Weapon
  is armed