1. AI Game Programming Wisdom 2 Chapter 3.3
Team Member AI in a FPS
AI Game Programming Wisdom 2
Chapter 3.3

2. Goals Teammates that back the player up
Teammates that follow instructions
Teammates that survive until end of game
Don’t want: teammates that get in the player’s way or wander off

3. Correct Positioning Player’s F.O.V. should not be obstructed
Balance between various aspects:
Visible teammates add to feeling of the game
Consider accuracy of the player’s weapon
Should always leave area between player and targets clear

4. Balancing NPC Movement
NPC movement while in player’s F.O.V. should be smooth so as to not detract from game play
Should move to clear player’s F.O.V., but should move as little as possible
Another consideration: stance, could the teammates crouch or crawl past the player? (Again, take into account accuracy of player’s weapon, and lines of fire)

5. Correct Movement
Multiple targets means a greater chance that at least one line of sight is obstructed
Moving out of the way of one target could put a teammate in front of another target
Solutions:
Consider all targets & find the smaller angle that would leave a clear view for player
Consider the player’s heading and move away from it

6. Correct Movement E E T P E E

7. Correct Movement E E T P E E
May result in a longer path for teammate, but ensures the teammate will not cross player’s line of sight

8. New Considerations Target Threat:
Proximity
Level of awareness
Readiness
AI must be able to act to keep the target with the highest threat clear
Ex. Teammate might spot a patrol 500m away while there is an enemy just 10m away
Full Spectrum Warrior!

9. Guidelines
Maximum straying distance: don’t want the player to have to chase teammates
AI should not take the fun from the player (e.g. Should not be the first around corners, through doors)
If ordered to take the initiative, teammates should behave strategically
Should move at same speed as player but lag behind slightly

10. Guidelines
Remove?
Context: teammates should always respect the context and be able to intelligently change between contexts.
Ex. In a stealth game teammates would avoid firing on targets unless they were instructed to do so, or unless they were engaged by the target
Can set context for a individual teammates or entire team

11. Correct Behavior
NPC behavior can be defined in rule based terms, focusing on strategic aspects of individual and the entire team
3 behaviors common to FPSs
Use of cover
Selective firing
Reloading

12. Correct Behavior – Use of Cover
Remove?
Use of cover keeps the teammates alive as long as the player Should favor use of cover in path finding algorithms
When entering a new area, teammates should map out possible points of cover
Must take into consideration maximum straying distance

13. Correct Behavior – Selective Firing
Remove?
Should only fire when target is in range and in view
Realistic games: fire in short bursts to improve accuracy
Balance accuracy with response time
Low response time could result in AI that appears unaware and less intelligent
Response time should be >= enemy response time to ensure the team survives more often

14. Correct Behavior - Reloading
Remove?
Teammates should count the number of rounds used & find safe cover to reload
Staggering firing of teammates ensure only one will need to reload at a given time
While one teammate is reloading, another should be able to keep up the attack

15. Supporting the Player
At least one teammate should cover player’s flanks
Report enemy sightings so player can take action and to help guide player through game
If enemy sighted and team too busy, teammate should engage autonomously
Should concentrate on targets other than the player’s target
Should take into consideration threat level
Might chose target that is farthest from player’s line of sight

16. Table of Actions
A table of actions can help the team function more effectively and give a greater illusion of intelligence
Action
Immediate action?
Whole team required?
Cover Me No
I need backup
Yes
Attack my target
Close in
Spread out
Move to waypoint
Hold

17. The Player is Most Important
Player should be allowed to play the game, supported by the team – team should never hinder game play
NPCs should chose weapon according to player’s weapon
NPCs should defer objects to player
If an NPC does get an object, most needy one should have first choice
NPCs should have different reaction times for distant and near enemies, to allow the player to respond first

18. Rainbow Six Demo

19. Implementation Team Manager Environment Threat Player Layered approach
Four layers
Lower layers take higher priority
Caters to player to improve game play
Team Manager
Environment
Threat
Player

20. Implementation – Player Awareness
Primary concern is to avoid player’s line of sight
Player’s position, orientation, current target, and possible targets should all be taken into consideration
Secondary concern is with avoiding other NPCs’ lines of sight
A hierarchical ordering of NPCs solves this

21. Implementation – Threat Awareness
NPCs should be looking for targets using line of sight algorithms
Enemies sighted should be reported
Movement gets deferred to player awareness level when not searching or engaging enemies

22. Implementation - Environment Awareness
When moving into a new area, teamates should identify cover and areas that are impermissible due to danger or exposure
Environment can be used to reduce NPC straying – boundary restrictions

23. Implementation - Team Manager
Shared by all teammates
Contains all the team’s data
Helps coordinate the whole team
NPCs query manager to find out where and how to move (sets context)
Tell NPCs if they can pick up objects
Consults environ. level to get stray dist.
Consults threat level to control rate of fire
Never consider the player (unpredictable)

24. Teammate Availability
N = # enemies in covering area
O = # enemies in range
P = # enemies threatening team
Q = supporting another team member (1/0)
(Multiply by infinity to make sure busy teammate is never selected)
Availability = (1+N) (1+O) (1+P) + (Q*Inf.)
Lower numbers indicate greater availability.
Set thresholds to determine if a teammate is busy

25. Fire Staggering Should happen naturally
Teammates should not target and fire at the same enemy
Team manager should monitor ammo levels
Could widen cover area of one NPC to reduce another NPC’s cover area, allowing them to conserve ammo
When one NPC runs out the other can then cover the whole area while the former reloads

26. AI Game Programming Wisdom 3.4
Applying Goal-Oriented Action Planning to Games Autonomous decisions to activate behaviors intended to accomplish most relative goals
AI Game Programming Wisdom
3.4

27. Goal Oriented Action Planning (GOAP)
FEAR!
A decision making architecture that allows NPCs to decide what to do next, and how to do it
Produces less repetitive, less predictable behavior
Can be adapted to fit the current situation
Facilitates authoring, maintaining, and reuse of behaviors
Active goal determines NPC’s behavior through a hard-coded sequence of state transitions

28. GOAP
An agent uses a planner to formulate a sequence of actions that will satisfy some goal
GOAP does not replace the need for FSMs, but greatly simplifies those required
Plan becomes a sequence of actions where each action represents a state transition
Separating the state transition logic from states greatly simplifies the underlying FSM

29. GOAP - Goals A goal is any condition that an agent wants to satisfy
Agent may have any number of goals
Only one goal is active at any given moment
A goal knows how to calculate its relevance
A goal knows when it has been satisfied
A goal does not include a plan, but merely conditions that must be met

30. GOAP - Plan
A plan satisfying a goal is any valid sequence of actions that move the agent from some starting state to some a state satisfying the goal
Could be multiple plans that satisfy some goal
Only need to find one
The planner can be given hints
Ex. Associate costs with actions, look for lowest cost to reach goal

31. GOAP - Action
An action is a single, atomic step that makes an agent do something
Ex. GoToPoint, ActivateObject
Could be short or infinitely long
Ex. Reload, Attack
An action knows its preconditions and effects
Allows for easy chaining of actions
Ex. Reload and Attack

32. GOAP – Plan Formulation
An agent provides a goal to a handler called a planner
The planner searches the space of actions for a sequence that will take the agent from its starting state to a goal state (formulating the plan)
If planner is successful, the agent follows the plan until completion or until another goal becomes more relevant
If another goal activates or plan becomes invalid, agent aborts and formulates another plan

33. GOAP – Plan Formulation
Kill Enemy Goal
Action Space
Sleep
Start
Idle
Draw
Weapon
Dodge
Reload
Weapon
Go to Cover
Attack
Activate
Object
Go to Point
Goal State

34. GOAP – Runtime Benefits
An agent can custom fit actions to current surroundings
Agent can dynamically find alternate solutions to problems
Ex. Soldier out of ammo, but laser near by

35. GOAP – Development Benefits
Handling every possible situation for every agent can quickly become difficult to manage
Adding design requirements in GOAP is simply a matter of adding actions and defining their preconditions and effects
Ex. Adding TurnLightsOn action, precondition to GoTo action
GOAP helps prevent invalid plans to be formed, where human error could easily contribute to invalid plans in hand-written code

36. GOAP – Variety Benefits
GOAP can be used to create a variety of characters simply by giving the planner different subsets of actions
One character’s actions can easily be replaced with functionally similar, but aesthetically different actions for another character
Ex. OpenDoor for regular soldier and brute
Extension: add preconditions to certain actions for mood, energy level, etc.

37. GOAP – Implementation Considerations
Challenges:
Determining best method for search the space of actions
Planner must be able to represent the game world in a clear, compact manner

38. GOAP – Planner Search
Similar to path finding, so the same algorithm can be used: A*
The planner need only implement its own classes for A*’s nodes, map, and goals
A* involves the calculation of the cost of a node and the heuristic distance from a node to a goal
Can search forward or backwards
Forward: exhaustive brute force search
Backwards: more efficient and intuitive

39. GOAP – Planner Search
Node Cost Calculation: sum of costs of actions that take the world to the state represented by the node
A lower cost is preferable
Hueristic Cost Calculation: sum of the number of unsatisfied properties of the goal state

40. GOAP – World Representation
Planner must be able to apply preconditions and effects of action to realize when it has reached the goal state
Only need to represent minimal number of properties of the world state that are relevant to the goal the planner is trying to satisfy
As planner adds actions, goal state grows with the preconditions of actions added

41. GOAP – World Representation
Can construct a list of world property structures for each action taken
Enumerated attribute key
A value
A handle to a subject (the character)
Pass desired properties to the goal state

42. GOAP – World Representation
Remove?
struct SWorldProperty {
GAME_OBJECT_ID hSubjectID;
WORLD_PROP_KEY eKey;
union value {
bool bValue;
float fValue;
int nValue; }
Ex. KillEnemy goal
SWorldProperty Prop;
Prop.hSubjectID = hShooterID;
Prop.eKey = kTargetIsDead;
Prop.bValue = true;

43. GOAP – Final Thought
Context preconditions – something that needs to be true for an action to take place, but the planner will never try to satisfy it
Ex. Attack action, must be in range and F.O.V.
Planner must run two functions, one to make a plan and check all preconditions and another to verify all context preconditions are satisfied for the plan

44. GOAP - Optimization Possible optimizations for GOAP:
Optimizing the search algorithm (which takes the general form of A* but is game specific)
Cache results from previous searches
Distributed plan formulation over several updates
Use of context preconditions to prune search trees.