Dynamic tactical position evaluation functions are procedures that use static and dynamic information about the world to make tactical decisions at run-time. By having NPCs use a procedural description of the solution to a tactical decision, the solution can vary depending on the inputs. Using static information about the world as input makes sure the tactics are applicable at any place on the map. Using dynamic inputs makes sure they are applicable in multiple situations.

1. Killzone’s AI:
Dynamic Procedural Tactics
Arjen Beij & Remco Straatman
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 1

2. Contents
 Dynamic procedural tactics
 Position picking
 Position Evaluation functions
 Tactical path-finding
 Suppression Fire
 World Representation
 Conclusions
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 2

3. Dynamic Procedural Tactics
Dynamic procedural tactics =
 Having the rules and concepts to
dynamically compute
solutions to tactical problems
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 3

4. AI Tactics
Common approach
 Level designer placed hints
 Triggered scripted behavior
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 4

5. Dynamic Procedural Tactics
Game-play benefits
 Works for any number of players
 Fights anywhere
Production benefits
 Reuse of AI behavior
 Quick level prototyping
 Roll out improvements game wide
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 5

6. Position Picking Outside
Wall
danger zone
Hugging
Nearby
 Picking a good position Partial
Cover
Cover
is half the battle
in ‘fire & maneuver’ combat Line of Fire Proximity
Preferred
Range
 Account for dozens of
(dynamic) factors
 How to pick that position?
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 6

7. Position Evaluation Functions
Determining the most appropriate position
 Given n potential destinations
 To compare positions, need a single value
Use position evaluation function
 Weighted sum
 Combines basic factors
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 7

8. Example: Picking an Attack
Position 1
Example: Picking an attack position
 Multiple threats
 Attack primary threat
 Cover from other threats preferred
 Partial cover from primary
 Proximity
 Preferred fighting range
Combine into a single value per position
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 8

9. Example: Picking an Attack
Position 2
Initial situation: three threats, some walls and a waypoint graph
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 9

10. Example: Picking an Attack
Position 3
Selected nearby waypoints, annotated with proximity (weight 20)
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 10

11. Example: Picking an Attack
Position 4
Annotations for LoF to primary threat (40 if partial cover, 20 otherwise)
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 11

12. Example: Picking an Attack
Position 5
Annotations for cover from secondary threats (weight 20)
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 12

13. Example: Picking an Attack
Position 6
Annotations positions inside preferred fighting range (weight 10)
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 13

14. Example: Picking an Attack
Position 7
+
Adding up all the annotations yields the most promising attack position
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 14

15. Example: Picking an Attack
Position 8
Single threat Multiple threats
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 15

16. Configurable and Dynamic
Personality:
Configurable Action:
What factors are
What kind of position?
important?
Position evaluation function
Dynamic Current situation: Squad:
Where are the threats? Additional constraints
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 16

17. Tactical Path-finding
Shortest paths often aren’t tactically sound
1 2 3
For example:
1. Paths traversing friendly lines-of-fire
2. Paths ignoring cover from hostile fire
3. Paths leaving the AI exposed unnecessarily
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 17

18. Tactical Path-finding
Position evaluations functions in path finding
 In A*, add to the costs of traveling a link:
 Traversing a friendly lines-of-fire
 Being under fire from one or more threats
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 18

19. Example: Tactical Path-finding
1
Costs to travel each waypoint link, and the resulting shortest path.
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 19

20. Example: Tactical Path-finding
2
Link costs now reflect being under fire, and the corresponding tactical path.
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 20

21. Example: Tactical Path-finding
3 B B
T T
shortest path versus tactical path
A A
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 21

22. Suppression Fire
Suppression fire:
 Fire near hidden threats to pin them down
Not:
 Firing into the wall behind which the threat is
hiding
 Firing into locations the threat cannot reach
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 22

23. Example: Suppression Fire 1
What we want
 Deny the threat use of good attack positions
 By shooting bursts at these positions
How we implement it using position evaluation
 Evaluate threat’s attack positions
from his perspective
 Select those attack positions we can fire into
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 23

24. Example: Suppression Fire 2
A hidden threat (left), attacker (right) and the intended suppression fire.
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 24

25. Example: Suppression Fire 3
Selecting the waypoints in close proximity to the presumed threat position
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 25

26. Example: Suppression Fire 4
Annotations (20, 30) for offering the threat a LoF to the attacker (right).
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 26

27. Example: Suppression Fire 5
Annotation of 20 for positions offering the threat nearby cover from the attacker.
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 27

28. Example: Suppression Fire 6
Selecting those positions with score 40 yields the suppress targets.
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 28

29. Example: Suppression Fire 7
Merging suppression targets that overlap in yaw and pitch.
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 29

30. Example: Suppression Fire 8
The attacker’s view of suppression targets near the hidden threat.
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 30

31. Example: Suppression Fire 9
attacker PoV target PoV
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 31

32. World Representation
Pre-conditions for dynamic procedural tactics
1. Navigation info, with fine enough resolution
to represent cover locations
 Waypoints, nav meshes, cells: any will do
2. Fast Line-of-Sight/Line-of-fire checks
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 32

33. Lines-of-Fire Representation 1
Tactical decisions
 Involve hundreds of LoF checks
 Ray casts typically expensive
Use a small pre-computed look-up table
 Multiple stances, minor movement
 Pessimistic about cover
 Polar representation
Killzone’s LoF table for 4000 waypoints: 64KB
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 33

34. Lines-of-Fire Representation 2
Compression:
For every waypoint, per radial
sector, record the largest
distance from where an attacker
within that sector can fire at the
waypoint.
Inaccurate, but consistent.
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 34

35. Conclusions
Killzone was released for PS2 late 2004
 Tactical position picking
 Tactical path-finding
 Suppression by infantry and mounted MGs
 Indirect fire for hand grenades and tank shells
 Single and multiplayer AI
 Up to 14 AI characters fighting simultaneously
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 35

36. Conclusions - Killzone’s AI
Pros
 Share/reuse behavior across games
 Quick level prototyping
 Can be combined with scripting
 Fights anywhere, anytime, SP or MP
Cons
 Harder to control in detail
 Harder to test
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 36

37. Conclusions – Game AI
Position evaluation functions are powerful
 Robust decisions with many (dynamic) inputs
 Different behavior through configuration
 Many tactics can be implemented through
position evaluation
Dynamic procedural tactics
 Responsive AI behavior
 Within reach of many games
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 37

38. Further Info
Off-line
 Our proceedings paper provides more details, notably on the terrain
representation and grenade handling
 Lars Lidén, Strategic and Tactical Reasoning with Waypoints in AI
Game Programming Wisdom, Charles River Media, 2002
 Paul Tozour, Using a Spatial Database for Runtime Spatial Analysis
in AI Game Programming Wisdom 2, Charles River Media, 2003
 William van der Sterren, Tactical Path-Finding with A* in
Game Programming Gems 3, Charles River Media, 2002
On-line
 William van der Sterren’s collection of links to on-line papers at:
www.cgf-ai.com/links.html
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 38

39. Questions?
August 31, 2005 Killzone’s AI:
Game Developer Conference Europe Dynamic Procedural Tactics Guerrilla Games - 39